program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3405.2.1"}})]
{
    func main<ios18>(tensor<fp16, [1, 404, 128]> input_x, tensor<fp16, [1, 101, 1024]> pos_emb) {
            tensor<int32, [1]> input_1_axes_1 = const()[name = string("input_1_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 404, 128]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_1, x = input_x)[name = string("input_1_cast_fp16")];
            string dense_output_1_pad_type_1 = const()[name = string("dense_output_1_pad_type_1"), val = string("custom")];
            tensor<int32, [4]> dense_output_1_pad_1 = const()[name = string("dense_output_1_pad_1"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> dense_output_1_strides_1 = const()[name = string("dense_output_1_strides_1"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> dense_output_1_dilations_1 = const()[name = string("dense_output_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1_groups_1 = const()[name = string("dense_output_1_groups_1"), val = int32(1)];
            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2432))))[name = string("pre_encode_conv_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 256, 202, 64]> dense_output_1_cast_fp16 = conv(dilations = dense_output_1_dilations_1, groups = dense_output_1_groups_1, pad = dense_output_1_pad_1, pad_type = dense_output_1_pad_type_1, strides = dense_output_1_strides_1, weight = pre_encode_conv_0_dense_conv_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = string("dense_output_1_cast_fp16")];
            string sparse_output_1_pad_type_1 = const()[name = string("sparse_output_1_pad_type_1"), val = string("custom")];
            tensor<int32, [4]> sparse_output_1_pad_1 = const()[name = string("sparse_output_1_pad_1"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> sparse_output_1_strides_1 = const()[name = string("sparse_output_1_strides_1"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> sparse_output_1_dilations_1 = const()[name = string("sparse_output_1_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1_groups_1 = const()[name = string("sparse_output_1_groups_1"), val = int32(1)];
            tensor<fp16, [256, 1, 3, 3]> pre_encode_conv_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3136))), nonzero_data = tensor<fp16, [24]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3008))))[name = string("pre_encode_conv_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 256, 202, 64]> sparse_output_1_cast_fp16 = conv(dilations = sparse_output_1_dilations_1, groups = sparse_output_1_groups_1, pad = sparse_output_1_pad_1, pad_type = sparse_output_1_pad_type_1, strides = sparse_output_1_strides_1, weight = pre_encode_conv_0_sparse_conv_weight_to_fp16_sparsified, x = input_1_cast_fp16)[name = string("sparse_output_1_cast_fp16")];
            tensor<fp16, [1, 256, 202, 64]> output_1_cast_fp16 = add(x = dense_output_1_cast_fp16, y = sparse_output_1_cast_fp16)[name = string("output_1_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_85_to_fp16 = const()[name = string("op_85_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3520)))];
            tensor<fp16, [1, 256, 202, 64]> input_3_cast_fp16 = add(x = output_1_cast_fp16, y = var_85_to_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 256, 202, 64]> input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
            string dense_output_3_pad_type_1 = const()[name = string("dense_output_3_pad_type_1"), val = string("custom")];
            tensor<int32, [4]> dense_output_3_pad_1 = const()[name = string("dense_output_3_pad_1"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> dense_output_3_strides_1 = const()[name = string("dense_output_3_strides_1"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> dense_output_3_dilations_1 = const()[name = string("dense_output_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_3_groups_1 = const()[name = string("dense_output_3_groups_1"), val = int32(1)];
            tensor<fp16, [256, 256, 3, 3]> pre_encode_conv_2_dense_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [256, 256, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4096))), nonzero_data = tensor<fp16, [0]>([]))[name = string("pre_encode_conv_2_dense_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 256, 101, 32]> dense_output_3_cast_fp16 = conv(dilations = dense_output_3_dilations_1, groups = dense_output_3_groups_1, pad = dense_output_3_pad_1, pad_type = dense_output_3_pad_type_1, strides = dense_output_3_strides_1, weight = pre_encode_conv_2_dense_conv_weight_to_fp16_sparsified, x = input_5_cast_fp16)[name = string("dense_output_3_cast_fp16")];
            string sparse_output_3_pad_type_1 = const()[name = string("sparse_output_3_pad_type_1"), val = string("custom")];
            tensor<int32, [4]> sparse_output_3_pad_1 = const()[name = string("sparse_output_3_pad_1"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> sparse_output_3_strides_1 = const()[name = string("sparse_output_3_strides_1"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> sparse_output_3_dilations_1 = const()[name = string("sparse_output_3_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_3_groups_1 = const()[name = string("sparse_output_3_groups_1"), val = int32(1)];
            tensor<fp16, [256, 256, 3, 3]> pre_encode_conv_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [256, 256, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82560))), nonzero_data = tensor<fp16, [2304]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77888))))[name = string("pre_encode_conv_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 256, 101, 32]> sparse_output_3_cast_fp16 = conv(dilations = sparse_output_3_dilations_1, groups = sparse_output_3_groups_1, pad = sparse_output_3_pad_1, pad_type = sparse_output_3_pad_type_1, strides = sparse_output_3_strides_1, weight = pre_encode_conv_2_sparse_conv_weight_to_fp16_sparsified, x = input_5_cast_fp16)[name = string("sparse_output_3_cast_fp16")];
            tensor<fp16, [1, 256, 101, 32]> output_3_cast_fp16 = add(x = dense_output_3_cast_fp16, y = sparse_output_3_cast_fp16)[name = string("output_3_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_105_to_fp16 = const()[name = string("op_105_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156352)))];
            tensor<fp16, [1, 256, 101, 32]> input_7_cast_fp16 = add(x = output_3_cast_fp16, y = var_105_to_fp16)[name = string("input_7_cast_fp16")];
            string dense_output_5_pad_type_1 = const()[name = string("dense_output_5_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_5_strides_1 = const()[name = string("dense_output_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_5_pad_1 = const()[name = string("dense_output_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_5_dilations_1 = const()[name = string("dense_output_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_5_groups_1 = const()[name = string("dense_output_5_groups_1"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156928))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222528))))[name = string("pre_encode_conv_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 256, 101, 32]> dense_output_5_cast_fp16 = conv(dilations = dense_output_5_dilations_1, groups = dense_output_5_groups_1, pad = dense_output_5_pad_1, pad_type = dense_output_5_pad_type_1, strides = dense_output_5_strides_1, weight = pre_encode_conv_3_dense_conv_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = string("dense_output_5_cast_fp16")];
            string sparse_output_5_pad_type_1 = const()[name = string("sparse_output_5_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_5_strides_1 = const()[name = string("sparse_output_5_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_5_pad_1 = const()[name = string("sparse_output_5_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_5_dilations_1 = const()[name = string("sparse_output_5_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_5_groups_1 = const()[name = string("sparse_output_5_groups_1"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224512))), nonzero_data = tensor<fp16, [656]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223104))))[name = string("pre_encode_conv_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 256, 101, 32]> sparse_output_5_cast_fp16 = conv(dilations = sparse_output_5_dilations_1, groups = sparse_output_5_groups_1, pad = sparse_output_5_pad_1, pad_type = sparse_output_5_pad_type_1, strides = sparse_output_5_strides_1, weight = pre_encode_conv_3_sparse_conv_weight_to_fp16_sparsified, x = input_7_cast_fp16)[name = string("sparse_output_5_cast_fp16")];
            tensor<fp16, [1, 256, 101, 32]> output_5_cast_fp16 = add(x = dense_output_5_cast_fp16, y = sparse_output_5_cast_fp16)[name = string("output_5_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_124_to_fp16 = const()[name = string("op_124_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232768)))];
            tensor<fp16, [1, 256, 101, 32]> input_9_cast_fp16 = add(x = output_5_cast_fp16, y = var_124_to_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 256, 101, 32]> input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
            string dense_output_7_pad_type_1 = const()[name = string("dense_output_7_pad_type_1"), val = string("custom")];
            tensor<int32, [4]> dense_output_7_pad_1 = const()[name = string("dense_output_7_pad_1"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> dense_output_7_strides_1 = const()[name = string("dense_output_7_strides_1"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> dense_output_7_dilations_1 = const()[name = string("dense_output_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_7_groups_1 = const()[name = string("dense_output_7_groups_1"), val = int32(1)];
            tensor<fp16, [1, 256, 51, 16]> dense_output_7_cast_fp16 = conv(dilations = dense_output_7_dilations_1, groups = dense_output_7_groups_1, pad = dense_output_7_pad_1, pad_type = dense_output_7_pad_type_1, strides = dense_output_7_strides_1, weight = pre_encode_conv_2_dense_conv_weight_to_fp16_sparsified, x = input_11_cast_fp16)[name = string("dense_output_7_cast_fp16")];
            string sparse_output_7_pad_type_1 = const()[name = string("sparse_output_7_pad_type_1"), val = string("custom")];
            tensor<int32, [4]> sparse_output_7_pad_1 = const()[name = string("sparse_output_7_pad_1"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> sparse_output_7_strides_1 = const()[name = string("sparse_output_7_strides_1"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> sparse_output_7_dilations_1 = const()[name = string("sparse_output_7_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_7_groups_1 = const()[name = string("sparse_output_7_groups_1"), val = int32(1)];
            tensor<fp16, [256, 256, 3, 3]> pre_encode_conv_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [256, 256, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82560))), nonzero_data = tensor<fp16, [2304]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233344))))[name = string("pre_encode_conv_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 256, 51, 16]> sparse_output_7_cast_fp16 = conv(dilations = sparse_output_7_dilations_1, groups = sparse_output_7_groups_1, pad = sparse_output_7_pad_1, pad_type = sparse_output_7_pad_type_1, strides = sparse_output_7_strides_1, weight = pre_encode_conv_5_sparse_conv_weight_to_fp16_sparsified, x = input_11_cast_fp16)[name = string("sparse_output_7_cast_fp16")];
            tensor<fp16, [1, 256, 51, 16]> output_7_cast_fp16 = add(x = dense_output_7_cast_fp16, y = sparse_output_7_cast_fp16)[name = string("output_7_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_144_to_fp16 = const()[name = string("op_144_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238016)))];
            tensor<fp16, [1, 256, 51, 16]> input_13_cast_fp16 = add(x = output_7_cast_fp16, y = var_144_to_fp16)[name = string("input_13_cast_fp16")];
            string dense_output_9_pad_type_1 = const()[name = string("dense_output_9_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_9_strides_1 = const()[name = string("dense_output_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_9_pad_1 = const()[name = string("dense_output_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_9_dilations_1 = const()[name = string("dense_output_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_9_groups_1 = const()[name = string("dense_output_9_groups_1"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238592))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304192))))[name = string("pre_encode_conv_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 256, 51, 16]> dense_output_9_cast_fp16 = conv(dilations = dense_output_9_dilations_1, groups = dense_output_9_groups_1, pad = dense_output_9_pad_1, pad_type = dense_output_9_pad_type_1, strides = dense_output_9_strides_1, weight = pre_encode_conv_6_dense_conv_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = string("dense_output_9_cast_fp16")];
            string sparse_output_9_pad_type_1 = const()[name = string("sparse_output_9_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_9_strides_1 = const()[name = string("sparse_output_9_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_9_pad_1 = const()[name = string("sparse_output_9_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_9_dilations_1 = const()[name = string("sparse_output_9_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_9_groups_1 = const()[name = string("sparse_output_9_groups_1"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> pre_encode_conv_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306176))), nonzero_data = tensor<fp16, [656]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304768))))[name = string("pre_encode_conv_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 256, 51, 16]> sparse_output_9_cast_fp16 = conv(dilations = sparse_output_9_dilations_1, groups = sparse_output_9_groups_1, pad = sparse_output_9_pad_1, pad_type = sparse_output_9_pad_type_1, strides = sparse_output_9_strides_1, weight = pre_encode_conv_6_sparse_conv_weight_to_fp16_sparsified, x = input_13_cast_fp16)[name = string("sparse_output_9_cast_fp16")];
            tensor<fp16, [1, 256, 51, 16]> output_9_cast_fp16 = add(x = dense_output_9_cast_fp16, y = sparse_output_9_cast_fp16)[name = string("output_9_cast_fp16")];
            tensor<fp16, [1, 256, 1, 1]> var_163_to_fp16 = const()[name = string("op_163_to_fp16"), val = tensor<fp16, [1, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314432)))];
            tensor<fp16, [1, 256, 51, 16]> input_15_cast_fp16 = add(x = output_9_cast_fp16, y = var_163_to_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 256, 51, 16]> x_3_cast_fp16 = relu(x = input_15_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<int32, [4]> var_166_perm_1 = const()[name = string("op_166_perm_1"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_171 = const()[name = string("op_171"), val = tensor<int32, [4]>([1, 51, 1, 4096])];
            tensor<fp16, [1, 51, 256, 16]> var_166_cast_fp16 = transpose(perm = var_166_perm_1, x = x_3_cast_fp16)[name = string("transpose_987")];
            tensor<fp16, [1, 51, 1, 4096]> x_7_cast_fp16 = reshape(shape = var_171, x = var_166_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<int32, [4]> input_17_perm_1 = const()[name = string("input_17_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_11_pad_type_1 = const()[name = string("dense_output_11_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_11_strides_1 = const()[name = string("dense_output_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_11_pad_1 = const()[name = string("dense_output_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_11_dilations_1 = const()[name = string("dense_output_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_11_groups_1 = const()[name = string("dense_output_11_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> pre_encode_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315008))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4509376))))[name = string("pre_encode_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 4096, 1, 51]> input_17_cast_fp16 = transpose(perm = input_17_perm_1, x = x_7_cast_fp16)[name = string("transpose_986")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_11_cast_fp16 = conv(dilations = dense_output_11_dilations_1, groups = dense_output_11_groups_1, pad = dense_output_11_pad_1, pad_type = dense_output_11_pad_type_1, strides = dense_output_11_strides_1, weight = pre_encode_out_dense_conv_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = string("dense_output_11_cast_fp16")];
            string sparse_output_11_pad_type_1 = const()[name = string("sparse_output_11_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_11_strides_1 = const()[name = string("sparse_output_11_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_11_pad_1 = const()[name = string("sparse_output_11_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_11_dilations_1 = const()[name = string("sparse_output_11_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_11_groups_1 = const()[name = string("sparse_output_11_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> pre_encode_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4593920))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4509952))))[name = string("pre_encode_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_11_cast_fp16 = conv(dilations = sparse_output_11_dilations_1, groups = sparse_output_11_groups_1, pad = sparse_output_11_pad_1, pad_type = sparse_output_11_pad_type_1, strides = sparse_output_11_strides_1, weight = pre_encode_out_sparse_conv_weight_to_fp16_sparsified, x = input_17_cast_fp16)[name = string("sparse_output_11_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> output_cast_fp16 = add(x = dense_output_11_cast_fp16, y = sparse_output_11_cast_fp16)[name = string("output_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1]> op_191_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5118272))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5119360))))[name = string("op_191_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> x_9_cast_fp16 = add(x = output_cast_fp16, y = op_191_to_fp16_palettized)[name = string("x_9_cast_fp16")];
            tensor<int32, [4]> x_11_perm_1 = const()[name = string("x_11_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [1]> input_19_axes_1 = const()[name = string("input_19_axes_1"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 51, 1, 1024]> x_11_cast_fp16 = transpose(perm = x_11_perm_1, x = x_9_cast_fp16)[name = string("transpose_985")];
            tensor<fp16, [1, 51, 1024]> input_19_cast_fp16 = squeeze(axes = input_19_axes_1, x = x_11_cast_fp16)[name = string("input_19_cast_fp16")];
            int32 var_197 = const()[name = string("op_197"), val = int32(-1)];
            tensor<int32, [1]> x_13_axes_1 = const()[name = string("x_13_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5119936)))];
            tensor<fp16, [1024]> layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5122048)))];
            fp16 var_212_to_fp16 = const()[name = string("op_212_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_13_cast_fp16 = layer_norm(axes = x_13_axes_1, beta = layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_212_to_fp16, gamma = layers_0_norm_feed_forward1_weight_to_fp16, x = input_19_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<int32, [4]> var_231 = const()[name = string("op_231"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_15_cast_fp16 = reshape(shape = var_231, x = x_13_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<int32, [4]> input_21_perm_1 = const()[name = string("input_21_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_13_pad_type_1 = const()[name = string("dense_output_13_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_13_strides_1 = const()[name = string("dense_output_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_13_pad_1 = const()[name = string("dense_output_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_13_dilations_1 = const()[name = string("dense_output_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_13_groups_1 = const()[name = string("dense_output_13_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_0_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5124160))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9318528))))[name = string("layers_0_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_21_cast_fp16 = transpose(perm = input_21_perm_1, x = x_15_cast_fp16)[name = string("transpose_984")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_13_cast_fp16 = conv(dilations = dense_output_13_dilations_1, groups = dense_output_13_groups_1, pad = dense_output_13_pad_1, pad_type = dense_output_13_pad_type_1, strides = dense_output_13_strides_1, weight = layers_0_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = string("dense_output_13_cast_fp16")];
            string sparse_output_13_pad_type_1 = const()[name = string("sparse_output_13_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_13_strides_1 = const()[name = string("sparse_output_13_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_13_pad_1 = const()[name = string("sparse_output_13_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_13_dilations_1 = const()[name = string("sparse_output_13_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_13_groups_1 = const()[name = string("sparse_output_13_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_0_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9403072))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9319104))))[name = string("layers_0_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_13_cast_fp16 = conv(dilations = sparse_output_13_dilations_1, groups = sparse_output_13_groups_1, pad = sparse_output_13_pad_1, pad_type = sparse_output_13_pad_type_1, strides = sparse_output_13_strides_1, weight = layers_0_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_21_cast_fp16)[name = string("sparse_output_13_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_23_cast_fp16 = add(x = dense_output_13_cast_fp16, y = sparse_output_13_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_25_cast_fp16 = silu(x = input_23_cast_fp16)[name = string("input_25_cast_fp16")];
            string dense_output_15_pad_type_1 = const()[name = string("dense_output_15_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_15_strides_1 = const()[name = string("dense_output_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_15_pad_1 = const()[name = string("dense_output_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_15_dilations_1 = const()[name = string("dense_output_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_15_groups_1 = const()[name = string("dense_output_15_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_0_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9927424))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14121792))))[name = string("layers_0_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_15_cast_fp16 = conv(dilations = dense_output_15_dilations_1, groups = dense_output_15_groups_1, pad = dense_output_15_pad_1, pad_type = dense_output_15_pad_type_1, strides = dense_output_15_strides_1, weight = layers_0_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = string("dense_output_15_cast_fp16")];
            string sparse_output_15_pad_type_1 = const()[name = string("sparse_output_15_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_15_strides_1 = const()[name = string("sparse_output_15_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_15_pad_1 = const()[name = string("sparse_output_15_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_15_dilations_1 = const()[name = string("sparse_output_15_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_15_groups_1 = const()[name = string("sparse_output_15_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_0_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14206336))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14122368))))[name = string("layers_0_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_15_cast_fp16 = conv(dilations = sparse_output_15_dilations_1, groups = sparse_output_15_groups_1, pad = sparse_output_15_pad_1, pad_type = sparse_output_15_pad_type_1, strides = sparse_output_15_strides_1, weight = layers_0_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_25_cast_fp16)[name = string("sparse_output_15_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_17_cast_fp16 = add(x = dense_output_15_cast_fp16, y = sparse_output_15_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<int32, [4]> x_19_perm_1 = const()[name = string("x_19_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_266 = const()[name = string("op_266"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_19_cast_fp16 = transpose(perm = x_19_perm_1, x = x_17_cast_fp16)[name = string("transpose_983")];
            tensor<fp16, [1, 51, 1024]> var_267_cast_fp16 = reshape(shape = var_266, x = x_19_cast_fp16)[name = string("op_267_cast_fp16")];
            fp16 var_268_to_fp16 = const()[name = string("op_268_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_269_cast_fp16 = mul(x = var_267_cast_fp16, y = var_268_to_fp16)[name = string("op_269_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_27_cast_fp16 = add(x = input_19_cast_fp16, y = var_269_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<int32, [1]> q_1_axes_1 = const()[name = string("q_1_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_0_norm_self_att_weight_to_fp16 = const()[name = string("layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14730688)))];
            tensor<fp16, [1024]> layers_0_norm_self_att_bias_to_fp16 = const()[name = string("layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14732800)))];
            tensor<fp16, [1, 51, 1024]> q_1_cast_fp16 = layer_norm(axes = q_1_axes_1, beta = layers_0_norm_self_att_bias_to_fp16, epsilon = var_212_to_fp16, gamma = layers_0_norm_self_att_weight_to_fp16, x = input_27_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<int32, [3]> var_343 = const()[name = string("op_343"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_29_axes_1 = const()[name = string("input_29_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_344_cast_fp16 = transpose(perm = var_343, x = q_1_cast_fp16)[name = string("transpose_982")];
            tensor<fp16, [1, 1024, 51, 1]> input_29_cast_fp16 = expand_dims(axes = input_29_axes_1, x = var_344_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<int32, [3]> var_352 = const()[name = string("op_352"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_35_axes_1 = const()[name = string("input_35_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 101]> var_353_cast_fp16 = transpose(perm = var_352, x = pos_emb)[name = string("transpose_981")];
            tensor<fp16, [1, 1024, 101, 1]> input_35_cast_fp16 = expand_dims(axes = input_35_axes_1, x = var_353_cast_fp16)[name = string("input_35_cast_fp16")];
            string dense_output_17_pad_type_1 = const()[name = string("dense_output_17_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_17_strides_1 = const()[name = string("dense_output_17_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_17_pad_1 = const()[name = string("dense_output_17_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_17_dilations_1 = const()[name = string("dense_output_17_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_17_groups_1 = const()[name = string("dense_output_17_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14734912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14866048))))[name = string("layers_0_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_17_cast_fp16 = conv(dilations = dense_output_17_dilations_1, groups = dense_output_17_groups_1, pad = dense_output_17_pad_1, pad_type = dense_output_17_pad_type_1, strides = dense_output_17_strides_1, weight = layers_0_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_17_cast_fp16")];
            string sparse_output_17_pad_type_1 = const()[name = string("sparse_output_17_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_17_strides_1 = const()[name = string("sparse_output_17_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_17_pad_1 = const()[name = string("sparse_output_17_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_17_dilations_1 = const()[name = string("sparse_output_17_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_17_groups_1 = const()[name = string("sparse_output_17_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14869312))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14866624))))[name = string("layers_0_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_17_cast_fp16 = conv(dilations = sparse_output_17_dilations_1, groups = sparse_output_17_groups_1, pad = sparse_output_17_pad_1, pad_type = sparse_output_17_pad_type_1, strides = sparse_output_17_strides_1, weight = layers_0_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_17_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_369_cast_fp16 = add(x = dense_output_17_cast_fp16, y = sparse_output_17_cast_fp16)[name = string("op_369_cast_fp16")];
            tensor<int32, [4]> var_370 = const()[name = string("op_370"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_372 = const()[name = string("op_372"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_371_cast_fp16 = transpose(perm = var_370, x = var_369_cast_fp16)[name = string("transpose_980")];
            tensor<fp16, [1, 51, 128]> q_head_1_cast_fp16 = reshape(shape = var_372, x = var_371_cast_fp16)[name = string("q_head_1_cast_fp16")];
            string dense_output_19_pad_type_1 = const()[name = string("dense_output_19_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_19_strides_1 = const()[name = string("dense_output_19_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_19_pad_1 = const()[name = string("dense_output_19_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_19_dilations_1 = const()[name = string("dense_output_19_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_19_groups_1 = const()[name = string("dense_output_19_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14885760))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15016896))))[name = string("layers_0_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_19_cast_fp16 = conv(dilations = dense_output_19_dilations_1, groups = dense_output_19_groups_1, pad = dense_output_19_pad_1, pad_type = dense_output_19_pad_type_1, strides = dense_output_19_strides_1, weight = layers_0_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_19_cast_fp16")];
            string sparse_output_19_pad_type_1 = const()[name = string("sparse_output_19_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_19_strides_1 = const()[name = string("sparse_output_19_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_19_pad_1 = const()[name = string("sparse_output_19_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_19_dilations_1 = const()[name = string("sparse_output_19_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_19_groups_1 = const()[name = string("sparse_output_19_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15020160))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15017472))))[name = string("layers_0_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_19_cast_fp16 = conv(dilations = sparse_output_19_dilations_1, groups = sparse_output_19_groups_1, pad = sparse_output_19_pad_1, pad_type = sparse_output_19_pad_type_1, strides = sparse_output_19_strides_1, weight = layers_0_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_19_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_388_cast_fp16 = add(x = dense_output_19_cast_fp16, y = sparse_output_19_cast_fp16)[name = string("op_388_cast_fp16")];
            tensor<int32, [4]> var_389 = const()[name = string("op_389"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_391 = const()[name = string("op_391"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_390_cast_fp16 = transpose(perm = var_389, x = var_388_cast_fp16)[name = string("transpose_979")];
            tensor<fp16, [1, 51, 128]> k_head_1_cast_fp16 = reshape(shape = var_391, x = var_390_cast_fp16)[name = string("k_head_1_cast_fp16")];
            string dense_output_21_pad_type_1 = const()[name = string("dense_output_21_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_21_strides_1 = const()[name = string("dense_output_21_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_21_pad_1 = const()[name = string("dense_output_21_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_21_dilations_1 = const()[name = string("dense_output_21_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_21_groups_1 = const()[name = string("dense_output_21_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15036608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15167744))))[name = string("layers_0_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_21_cast_fp16 = conv(dilations = dense_output_21_dilations_1, groups = dense_output_21_groups_1, pad = dense_output_21_pad_1, pad_type = dense_output_21_pad_type_1, strides = dense_output_21_strides_1, weight = layers_0_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_21_cast_fp16")];
            string sparse_output_21_pad_type_1 = const()[name = string("sparse_output_21_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_21_strides_1 = const()[name = string("sparse_output_21_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_21_pad_1 = const()[name = string("sparse_output_21_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_21_dilations_1 = const()[name = string("sparse_output_21_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_21_groups_1 = const()[name = string("sparse_output_21_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15171008))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15168320))))[name = string("layers_0_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_21_cast_fp16 = conv(dilations = sparse_output_21_dilations_1, groups = sparse_output_21_groups_1, pad = sparse_output_21_pad_1, pad_type = sparse_output_21_pad_type_1, strides = sparse_output_21_strides_1, weight = layers_0_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_21_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_407_cast_fp16 = add(x = dense_output_21_cast_fp16, y = sparse_output_21_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_410 = const()[name = string("op_410"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_409_cast_fp16 = transpose(perm = var_408, x = var_407_cast_fp16)[name = string("transpose_978")];
            tensor<fp16, [1, 51, 128]> v_head_1_cast_fp16 = reshape(shape = var_410, x = var_409_cast_fp16)[name = string("v_head_1_cast_fp16")];
            string dense_output_23_pad_type_1 = const()[name = string("dense_output_23_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_23_strides_1 = const()[name = string("dense_output_23_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_23_pad_1 = const()[name = string("dense_output_23_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_23_dilations_1 = const()[name = string("dense_output_23_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_23_groups_1 = const()[name = string("dense_output_23_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15187456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15318592))))[name = string("layers_0_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_23_cast_fp16 = conv(dilations = dense_output_23_dilations_1, groups = dense_output_23_groups_1, pad = dense_output_23_pad_1, pad_type = dense_output_23_pad_type_1, strides = dense_output_23_strides_1, weight = layers_0_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_23_cast_fp16")];
            string sparse_output_23_pad_type_1 = const()[name = string("sparse_output_23_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_23_strides_1 = const()[name = string("sparse_output_23_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_23_pad_1 = const()[name = string("sparse_output_23_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_23_dilations_1 = const()[name = string("sparse_output_23_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_23_groups_1 = const()[name = string("sparse_output_23_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15321856))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15319168))))[name = string("layers_0_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_23_cast_fp16 = conv(dilations = sparse_output_23_dilations_1, groups = sparse_output_23_groups_1, pad = sparse_output_23_pad_1, pad_type = sparse_output_23_pad_type_1, strides = sparse_output_23_strides_1, weight = layers_0_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_23_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_426_cast_fp16 = add(x = dense_output_23_cast_fp16, y = sparse_output_23_cast_fp16)[name = string("op_426_cast_fp16")];
            tensor<int32, [4]> var_427 = const()[name = string("op_427"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_429 = const()[name = string("op_429"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_428_cast_fp16 = transpose(perm = var_427, x = var_426_cast_fp16)[name = string("transpose_977")];
            tensor<fp16, [1, 101, 128]> p_head_1_cast_fp16 = reshape(shape = var_429, x = var_428_cast_fp16)[name = string("p_head_1_cast_fp16")];
            tensor<fp16, [128]> var_431_to_fp16 = const()[name = string("op_431_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15338304)))];
            tensor<fp16, [1, 51, 128]> var_432_cast_fp16 = add(x = q_head_1_cast_fp16, y = var_431_to_fp16)[name = string("op_432_cast_fp16")];
            tensor<int32, [1]> q_u_1_axes_1 = const()[name = string("q_u_1_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_1_cast_fp16 = expand_dims(axes = q_u_1_axes_1, x = var_432_cast_fp16)[name = string("q_u_1_cast_fp16")];
            tensor<fp16, [128]> var_434_to_fp16 = const()[name = string("op_434_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15338624)))];
            tensor<fp16, [1, 51, 128]> var_435_cast_fp16 = add(x = q_head_1_cast_fp16, y = var_434_to_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> q_v_1_axes_1 = const()[name = string("q_v_1_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_1_cast_fp16 = expand_dims(axes = q_v_1_axes_1, x = var_435_cast_fp16)[name = string("q_v_1_cast_fp16")];
            tensor<int32, [1]> k_head_3_axes_1 = const()[name = string("k_head_3_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_3_cast_fp16 = expand_dims(axes = k_head_3_axes_1, x = k_head_1_cast_fp16)[name = string("k_head_3_cast_fp16")];
            tensor<int32, [1]> v_head_3_axes_1 = const()[name = string("v_head_3_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_3_cast_fp16 = expand_dims(axes = v_head_3_axes_1, x = v_head_1_cast_fp16)[name = string("v_head_3_cast_fp16")];
            tensor<int32, [1]> p_head_3_axes_1 = const()[name = string("p_head_3_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_3_cast_fp16 = expand_dims(axes = p_head_3_axes_1, x = p_head_1_cast_fp16)[name = string("p_head_3_cast_fp16")];
            bool var_441_transpose_x_3 = const()[name = string("op_441_transpose_x_3"), val = bool(false)];
            bool var_441_transpose_y_3 = const()[name = string("op_441_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_441_cast_fp16 = matmul(transpose_x = var_441_transpose_x_3, transpose_y = var_441_transpose_y_3, x = q_u_1_cast_fp16, y = k_head_3_cast_fp16)[name = string("op_441_cast_fp16")];
            fp16 var_442_to_fp16 = const()[name = string("op_442_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_1_cast_fp16 = mul(x = var_441_cast_fp16, y = var_442_to_fp16)[name = string("scores_content_1_cast_fp16")];
            bool x_21_transpose_x_3 = const()[name = string("x_21_transpose_x_3"), val = bool(false)];
            bool x_21_transpose_y_3 = const()[name = string("x_21_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_21_cast_fp16 = matmul(transpose_x = x_21_transpose_x_3, transpose_y = x_21_transpose_y_3, x = q_v_1_cast_fp16, y = p_head_3_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [8]> x_23_pad_1 = const()[name = string("x_23_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_23_mode_1 = const()[name = string("x_23_mode_1"), val = string("constant")];
            fp16 const_1309_to_fp16 = const()[name = string("const_1309_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_23_cast_fp16 = pad(constant_val = const_1309_to_fp16, mode = x_23_mode_1, pad = x_23_pad_1, x = x_21_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<int32, [4]> var_456 = const()[name = string("op_456"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_25_cast_fp16 = reshape(shape = var_456, x = x_23_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<int32, [4]> var_460_begin_1 = const()[name = string("op_460_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_460_end_1 = const()[name = string("op_460_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_460_end_mask_1 = const()[name = string("op_460_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_460_cast_fp16 = slice_by_index(begin = var_460_begin_1, end = var_460_end_1, end_mask = var_460_end_mask_1, x = x_25_cast_fp16)[name = string("op_460_cast_fp16")];
            tensor<int32, [4]> var_462 = const()[name = string("op_462"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_463_cast_fp16 = reshape(shape = var_462, x = var_460_cast_fp16)[name = string("op_463_cast_fp16")];
            tensor<int32, [4]> var_468_begin_1 = const()[name = string("op_468_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_468_end_1 = const()[name = string("op_468_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_468_end_mask_1 = const()[name = string("op_468_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_468_cast_fp16 = slice_by_index(begin = var_468_begin_1, end = var_468_end_1, end_mask = var_468_end_mask_1, x = var_463_cast_fp16)[name = string("op_468_cast_fp16")];
            fp16 var_469_to_fp16 = const()[name = string("op_469_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_1_cast_fp16 = mul(x = var_468_cast_fp16, y = var_469_to_fp16)[name = string("scores_pos_1_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_1_cast_fp16 = add(x = scores_content_1_cast_fp16, y = scores_pos_1_cast_fp16)[name = string("logits_1_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_472_cast_fp16 = softmax(axis = var_197, x = logits_1_cast_fp16)[name = string("op_472_cast_fp16")];
            bool var_474_transpose_x_1 = const()[name = string("op_474_transpose_x_1"), val = bool(false)];
            bool var_474_transpose_y_1 = const()[name = string("op_474_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_474_cast_fp16 = matmul(transpose_x = var_474_transpose_x_1, transpose_y = var_474_transpose_y_1, x = var_472_cast_fp16, y = v_head_3_cast_fp16)[name = string("op_474_cast_fp16")];
            tensor<int32, [1]> var_475_axes_1 = const()[name = string("op_475_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_475_cast_fp16 = squeeze(axes = var_475_axes_1, x = var_474_cast_fp16)[name = string("op_475_cast_fp16")];
            string dense_output_25_pad_type_1 = const()[name = string("dense_output_25_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_25_strides_1 = const()[name = string("dense_output_25_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_25_pad_1 = const()[name = string("dense_output_25_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_25_dilations_1 = const()[name = string("dense_output_25_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_25_groups_1 = const()[name = string("dense_output_25_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15338944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15470080))))[name = string("layers_0_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_25_cast_fp16 = conv(dilations = dense_output_25_dilations_1, groups = dense_output_25_groups_1, pad = dense_output_25_pad_1, pad_type = dense_output_25_pad_type_1, strides = dense_output_25_strides_1, weight = layers_0_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_25_cast_fp16")];
            string sparse_output_25_pad_type_1 = const()[name = string("sparse_output_25_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_25_strides_1 = const()[name = string("sparse_output_25_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_25_pad_1 = const()[name = string("sparse_output_25_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_25_dilations_1 = const()[name = string("sparse_output_25_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_25_groups_1 = const()[name = string("sparse_output_25_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15473344))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15470656))))[name = string("layers_0_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_25_cast_fp16 = conv(dilations = sparse_output_25_dilations_1, groups = sparse_output_25_groups_1, pad = sparse_output_25_pad_1, pad_type = sparse_output_25_pad_type_1, strides = sparse_output_25_strides_1, weight = layers_0_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_25_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_490_cast_fp16 = add(x = dense_output_25_cast_fp16, y = sparse_output_25_cast_fp16)[name = string("op_490_cast_fp16")];
            tensor<int32, [4]> var_491 = const()[name = string("op_491"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_493 = const()[name = string("op_493"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_492_cast_fp16 = transpose(perm = var_491, x = var_490_cast_fp16)[name = string("transpose_976")];
            tensor<fp16, [1, 51, 128]> q_head_3_cast_fp16 = reshape(shape = var_493, x = var_492_cast_fp16)[name = string("q_head_3_cast_fp16")];
            string dense_output_27_pad_type_1 = const()[name = string("dense_output_27_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_27_strides_1 = const()[name = string("dense_output_27_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_27_pad_1 = const()[name = string("dense_output_27_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_27_dilations_1 = const()[name = string("dense_output_27_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_27_groups_1 = const()[name = string("dense_output_27_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15489792))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15620928))))[name = string("layers_0_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_27_cast_fp16 = conv(dilations = dense_output_27_dilations_1, groups = dense_output_27_groups_1, pad = dense_output_27_pad_1, pad_type = dense_output_27_pad_type_1, strides = dense_output_27_strides_1, weight = layers_0_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_27_cast_fp16")];
            string sparse_output_27_pad_type_1 = const()[name = string("sparse_output_27_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_27_strides_1 = const()[name = string("sparse_output_27_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_27_pad_1 = const()[name = string("sparse_output_27_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_27_dilations_1 = const()[name = string("sparse_output_27_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_27_groups_1 = const()[name = string("sparse_output_27_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15624192))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15621504))))[name = string("layers_0_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_27_cast_fp16 = conv(dilations = sparse_output_27_dilations_1, groups = sparse_output_27_groups_1, pad = sparse_output_27_pad_1, pad_type = sparse_output_27_pad_type_1, strides = sparse_output_27_strides_1, weight = layers_0_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_27_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_509_cast_fp16 = add(x = dense_output_27_cast_fp16, y = sparse_output_27_cast_fp16)[name = string("op_509_cast_fp16")];
            tensor<int32, [4]> var_510 = const()[name = string("op_510"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_512 = const()[name = string("op_512"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_511_cast_fp16 = transpose(perm = var_510, x = var_509_cast_fp16)[name = string("transpose_975")];
            tensor<fp16, [1, 51, 128]> k_head_5_cast_fp16 = reshape(shape = var_512, x = var_511_cast_fp16)[name = string("k_head_5_cast_fp16")];
            string dense_output_29_pad_type_1 = const()[name = string("dense_output_29_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_29_strides_1 = const()[name = string("dense_output_29_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_29_pad_1 = const()[name = string("dense_output_29_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_29_dilations_1 = const()[name = string("dense_output_29_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_29_groups_1 = const()[name = string("dense_output_29_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15640640))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15771776))))[name = string("layers_0_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_29_cast_fp16 = conv(dilations = dense_output_29_dilations_1, groups = dense_output_29_groups_1, pad = dense_output_29_pad_1, pad_type = dense_output_29_pad_type_1, strides = dense_output_29_strides_1, weight = layers_0_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_29_cast_fp16")];
            string sparse_output_29_pad_type_1 = const()[name = string("sparse_output_29_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_29_strides_1 = const()[name = string("sparse_output_29_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_29_pad_1 = const()[name = string("sparse_output_29_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_29_dilations_1 = const()[name = string("sparse_output_29_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_29_groups_1 = const()[name = string("sparse_output_29_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15775040))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15772352))))[name = string("layers_0_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_29_cast_fp16 = conv(dilations = sparse_output_29_dilations_1, groups = sparse_output_29_groups_1, pad = sparse_output_29_pad_1, pad_type = sparse_output_29_pad_type_1, strides = sparse_output_29_strides_1, weight = layers_0_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_29_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_528_cast_fp16 = add(x = dense_output_29_cast_fp16, y = sparse_output_29_cast_fp16)[name = string("op_528_cast_fp16")];
            tensor<int32, [4]> var_529 = const()[name = string("op_529"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_531 = const()[name = string("op_531"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_530_cast_fp16 = transpose(perm = var_529, x = var_528_cast_fp16)[name = string("transpose_974")];
            tensor<fp16, [1, 51, 128]> v_head_5_cast_fp16 = reshape(shape = var_531, x = var_530_cast_fp16)[name = string("v_head_5_cast_fp16")];
            string dense_output_31_pad_type_1 = const()[name = string("dense_output_31_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_31_strides_1 = const()[name = string("dense_output_31_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_31_pad_1 = const()[name = string("dense_output_31_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_31_dilations_1 = const()[name = string("dense_output_31_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_31_groups_1 = const()[name = string("dense_output_31_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15791488))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15922624))))[name = string("layers_0_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_31_cast_fp16 = conv(dilations = dense_output_31_dilations_1, groups = dense_output_31_groups_1, pad = dense_output_31_pad_1, pad_type = dense_output_31_pad_type_1, strides = dense_output_31_strides_1, weight = layers_0_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_31_cast_fp16")];
            string sparse_output_31_pad_type_1 = const()[name = string("sparse_output_31_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_31_strides_1 = const()[name = string("sparse_output_31_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_31_pad_1 = const()[name = string("sparse_output_31_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_31_dilations_1 = const()[name = string("sparse_output_31_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_31_groups_1 = const()[name = string("sparse_output_31_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15925888))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15923200))))[name = string("layers_0_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_31_cast_fp16 = conv(dilations = sparse_output_31_dilations_1, groups = sparse_output_31_groups_1, pad = sparse_output_31_pad_1, pad_type = sparse_output_31_pad_type_1, strides = sparse_output_31_strides_1, weight = layers_0_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_31_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_547_cast_fp16 = add(x = dense_output_31_cast_fp16, y = sparse_output_31_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<int32, [4]> var_548 = const()[name = string("op_548"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_550 = const()[name = string("op_550"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_549_cast_fp16 = transpose(perm = var_548, x = var_547_cast_fp16)[name = string("transpose_973")];
            tensor<fp16, [1, 101, 128]> p_head_5_cast_fp16 = reshape(shape = var_550, x = var_549_cast_fp16)[name = string("p_head_5_cast_fp16")];
            tensor<fp16, [128]> var_552_to_fp16 = const()[name = string("op_552_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15942336)))];
            tensor<fp16, [1, 51, 128]> var_553_cast_fp16 = add(x = q_head_3_cast_fp16, y = var_552_to_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> q_u_3_axes_1 = const()[name = string("q_u_3_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_3_cast_fp16 = expand_dims(axes = q_u_3_axes_1, x = var_553_cast_fp16)[name = string("q_u_3_cast_fp16")];
            tensor<fp16, [128]> var_555_to_fp16 = const()[name = string("op_555_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15942656)))];
            tensor<fp16, [1, 51, 128]> var_556_cast_fp16 = add(x = q_head_3_cast_fp16, y = var_555_to_fp16)[name = string("op_556_cast_fp16")];
            tensor<int32, [1]> q_v_3_axes_1 = const()[name = string("q_v_3_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_3_cast_fp16 = expand_dims(axes = q_v_3_axes_1, x = var_556_cast_fp16)[name = string("q_v_3_cast_fp16")];
            tensor<int32, [1]> k_head_7_axes_1 = const()[name = string("k_head_7_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_7_cast_fp16 = expand_dims(axes = k_head_7_axes_1, x = k_head_5_cast_fp16)[name = string("k_head_7_cast_fp16")];
            tensor<int32, [1]> v_head_7_axes_1 = const()[name = string("v_head_7_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_7_cast_fp16 = expand_dims(axes = v_head_7_axes_1, x = v_head_5_cast_fp16)[name = string("v_head_7_cast_fp16")];
            tensor<int32, [1]> p_head_7_axes_1 = const()[name = string("p_head_7_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_7_cast_fp16 = expand_dims(axes = p_head_7_axes_1, x = p_head_5_cast_fp16)[name = string("p_head_7_cast_fp16")];
            bool var_562_transpose_x_3 = const()[name = string("op_562_transpose_x_3"), val = bool(false)];
            bool var_562_transpose_y_3 = const()[name = string("op_562_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_562_cast_fp16 = matmul(transpose_x = var_562_transpose_x_3, transpose_y = var_562_transpose_y_3, x = q_u_3_cast_fp16, y = k_head_7_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_3_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("scores_content_3_cast_fp16")];
            bool x_29_transpose_x_3 = const()[name = string("x_29_transpose_x_3"), val = bool(false)];
            bool x_29_transpose_y_3 = const()[name = string("x_29_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_29_cast_fp16 = matmul(transpose_x = x_29_transpose_x_3, transpose_y = x_29_transpose_y_3, x = q_v_3_cast_fp16, y = p_head_7_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [8]> x_31_pad_1 = const()[name = string("x_31_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_31_mode_1 = const()[name = string("x_31_mode_1"), val = string("constant")];
            fp16 const_1315_to_fp16 = const()[name = string("const_1315_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_31_cast_fp16 = pad(constant_val = const_1315_to_fp16, mode = x_31_mode_1, pad = x_31_pad_1, x = x_29_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<int32, [4]> var_577 = const()[name = string("op_577"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_33_cast_fp16 = reshape(shape = var_577, x = x_31_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [4]> var_581_begin_1 = const()[name = string("op_581_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_581_end_1 = const()[name = string("op_581_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_581_end_mask_1 = const()[name = string("op_581_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_581_cast_fp16 = slice_by_index(begin = var_581_begin_1, end = var_581_end_1, end_mask = var_581_end_mask_1, x = x_33_cast_fp16)[name = string("op_581_cast_fp16")];
            tensor<int32, [4]> var_583 = const()[name = string("op_583"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_584_cast_fp16 = reshape(shape = var_583, x = var_581_cast_fp16)[name = string("op_584_cast_fp16")];
            tensor<int32, [4]> var_589_begin_1 = const()[name = string("op_589_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_589_end_1 = const()[name = string("op_589_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_589_end_mask_1 = const()[name = string("op_589_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_589_cast_fp16 = slice_by_index(begin = var_589_begin_1, end = var_589_end_1, end_mask = var_589_end_mask_1, x = var_584_cast_fp16)[name = string("op_589_cast_fp16")];
            fp16 var_590_to_fp16 = const()[name = string("op_590_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_3_cast_fp16 = mul(x = var_589_cast_fp16, y = var_590_to_fp16)[name = string("scores_pos_3_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_3_cast_fp16 = add(x = scores_content_3_cast_fp16, y = scores_pos_3_cast_fp16)[name = string("logits_3_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_593_cast_fp16 = softmax(axis = var_197, x = logits_3_cast_fp16)[name = string("op_593_cast_fp16")];
            bool var_595_transpose_x_1 = const()[name = string("op_595_transpose_x_1"), val = bool(false)];
            bool var_595_transpose_y_1 = const()[name = string("op_595_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_595_cast_fp16 = matmul(transpose_x = var_595_transpose_x_1, transpose_y = var_595_transpose_y_1, x = var_593_cast_fp16, y = v_head_7_cast_fp16)[name = string("op_595_cast_fp16")];
            tensor<int32, [1]> var_596_axes_1 = const()[name = string("op_596_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_596_cast_fp16 = squeeze(axes = var_596_axes_1, x = var_595_cast_fp16)[name = string("op_596_cast_fp16")];
            string dense_output_33_pad_type_1 = const()[name = string("dense_output_33_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_33_strides_1 = const()[name = string("dense_output_33_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_33_pad_1 = const()[name = string("dense_output_33_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_33_dilations_1 = const()[name = string("dense_output_33_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_33_groups_1 = const()[name = string("dense_output_33_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15942976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16074112))))[name = string("layers_0_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_33_cast_fp16 = conv(dilations = dense_output_33_dilations_1, groups = dense_output_33_groups_1, pad = dense_output_33_pad_1, pad_type = dense_output_33_pad_type_1, strides = dense_output_33_strides_1, weight = layers_0_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_33_cast_fp16")];
            string sparse_output_33_pad_type_1 = const()[name = string("sparse_output_33_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_33_strides_1 = const()[name = string("sparse_output_33_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_33_pad_1 = const()[name = string("sparse_output_33_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_33_dilations_1 = const()[name = string("sparse_output_33_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_33_groups_1 = const()[name = string("sparse_output_33_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16077376))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16074688))))[name = string("layers_0_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_33_cast_fp16 = conv(dilations = sparse_output_33_dilations_1, groups = sparse_output_33_groups_1, pad = sparse_output_33_pad_1, pad_type = sparse_output_33_pad_type_1, strides = sparse_output_33_strides_1, weight = layers_0_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_33_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_611_cast_fp16 = add(x = dense_output_33_cast_fp16, y = sparse_output_33_cast_fp16)[name = string("op_611_cast_fp16")];
            tensor<int32, [4]> var_612 = const()[name = string("op_612"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_614 = const()[name = string("op_614"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_613_cast_fp16 = transpose(perm = var_612, x = var_611_cast_fp16)[name = string("transpose_972")];
            tensor<fp16, [1, 51, 128]> q_head_5_cast_fp16 = reshape(shape = var_614, x = var_613_cast_fp16)[name = string("q_head_5_cast_fp16")];
            string dense_output_35_pad_type_1 = const()[name = string("dense_output_35_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_35_strides_1 = const()[name = string("dense_output_35_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_35_pad_1 = const()[name = string("dense_output_35_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_35_dilations_1 = const()[name = string("dense_output_35_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_35_groups_1 = const()[name = string("dense_output_35_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16093824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16224960))))[name = string("layers_0_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_35_cast_fp16 = conv(dilations = dense_output_35_dilations_1, groups = dense_output_35_groups_1, pad = dense_output_35_pad_1, pad_type = dense_output_35_pad_type_1, strides = dense_output_35_strides_1, weight = layers_0_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_35_cast_fp16")];
            string sparse_output_35_pad_type_1 = const()[name = string("sparse_output_35_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_35_strides_1 = const()[name = string("sparse_output_35_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_35_pad_1 = const()[name = string("sparse_output_35_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_35_dilations_1 = const()[name = string("sparse_output_35_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_35_groups_1 = const()[name = string("sparse_output_35_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16228224))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16225536))))[name = string("layers_0_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_35_cast_fp16 = conv(dilations = sparse_output_35_dilations_1, groups = sparse_output_35_groups_1, pad = sparse_output_35_pad_1, pad_type = sparse_output_35_pad_type_1, strides = sparse_output_35_strides_1, weight = layers_0_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_35_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_630_cast_fp16 = add(x = dense_output_35_cast_fp16, y = sparse_output_35_cast_fp16)[name = string("op_630_cast_fp16")];
            tensor<int32, [4]> var_631 = const()[name = string("op_631"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_633 = const()[name = string("op_633"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_632_cast_fp16 = transpose(perm = var_631, x = var_630_cast_fp16)[name = string("transpose_971")];
            tensor<fp16, [1, 51, 128]> k_head_9_cast_fp16 = reshape(shape = var_633, x = var_632_cast_fp16)[name = string("k_head_9_cast_fp16")];
            string dense_output_37_pad_type_1 = const()[name = string("dense_output_37_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_37_strides_1 = const()[name = string("dense_output_37_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_37_pad_1 = const()[name = string("dense_output_37_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_37_dilations_1 = const()[name = string("dense_output_37_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_37_groups_1 = const()[name = string("dense_output_37_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16244672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16375808))))[name = string("layers_0_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_37_cast_fp16 = conv(dilations = dense_output_37_dilations_1, groups = dense_output_37_groups_1, pad = dense_output_37_pad_1, pad_type = dense_output_37_pad_type_1, strides = dense_output_37_strides_1, weight = layers_0_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_37_cast_fp16")];
            string sparse_output_37_pad_type_1 = const()[name = string("sparse_output_37_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_37_strides_1 = const()[name = string("sparse_output_37_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_37_pad_1 = const()[name = string("sparse_output_37_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_37_dilations_1 = const()[name = string("sparse_output_37_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_37_groups_1 = const()[name = string("sparse_output_37_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16379072))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16376384))))[name = string("layers_0_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_37_cast_fp16 = conv(dilations = sparse_output_37_dilations_1, groups = sparse_output_37_groups_1, pad = sparse_output_37_pad_1, pad_type = sparse_output_37_pad_type_1, strides = sparse_output_37_strides_1, weight = layers_0_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_37_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_649_cast_fp16 = add(x = dense_output_37_cast_fp16, y = sparse_output_37_cast_fp16)[name = string("op_649_cast_fp16")];
            tensor<int32, [4]> var_650 = const()[name = string("op_650"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_652 = const()[name = string("op_652"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_651_cast_fp16 = transpose(perm = var_650, x = var_649_cast_fp16)[name = string("transpose_970")];
            tensor<fp16, [1, 51, 128]> v_head_9_cast_fp16 = reshape(shape = var_652, x = var_651_cast_fp16)[name = string("v_head_9_cast_fp16")];
            string dense_output_39_pad_type_1 = const()[name = string("dense_output_39_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_39_strides_1 = const()[name = string("dense_output_39_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_39_pad_1 = const()[name = string("dense_output_39_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_39_dilations_1 = const()[name = string("dense_output_39_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_39_groups_1 = const()[name = string("dense_output_39_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16395520))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16526656))))[name = string("layers_0_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_39_cast_fp16 = conv(dilations = dense_output_39_dilations_1, groups = dense_output_39_groups_1, pad = dense_output_39_pad_1, pad_type = dense_output_39_pad_type_1, strides = dense_output_39_strides_1, weight = layers_0_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_39_cast_fp16")];
            string sparse_output_39_pad_type_1 = const()[name = string("sparse_output_39_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_39_strides_1 = const()[name = string("sparse_output_39_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_39_pad_1 = const()[name = string("sparse_output_39_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_39_dilations_1 = const()[name = string("sparse_output_39_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_39_groups_1 = const()[name = string("sparse_output_39_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16529920))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16527232))))[name = string("layers_0_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_39_cast_fp16 = conv(dilations = sparse_output_39_dilations_1, groups = sparse_output_39_groups_1, pad = sparse_output_39_pad_1, pad_type = sparse_output_39_pad_type_1, strides = sparse_output_39_strides_1, weight = layers_0_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_39_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_668_cast_fp16 = add(x = dense_output_39_cast_fp16, y = sparse_output_39_cast_fp16)[name = string("op_668_cast_fp16")];
            tensor<int32, [4]> var_669 = const()[name = string("op_669"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_671 = const()[name = string("op_671"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_670_cast_fp16 = transpose(perm = var_669, x = var_668_cast_fp16)[name = string("transpose_969")];
            tensor<fp16, [1, 101, 128]> p_head_9_cast_fp16 = reshape(shape = var_671, x = var_670_cast_fp16)[name = string("p_head_9_cast_fp16")];
            tensor<fp16, [128]> var_673_to_fp16 = const()[name = string("op_673_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16546368)))];
            tensor<fp16, [1, 51, 128]> var_674_cast_fp16 = add(x = q_head_5_cast_fp16, y = var_673_to_fp16)[name = string("op_674_cast_fp16")];
            tensor<int32, [1]> q_u_5_axes_1 = const()[name = string("q_u_5_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_5_cast_fp16 = expand_dims(axes = q_u_5_axes_1, x = var_674_cast_fp16)[name = string("q_u_5_cast_fp16")];
            tensor<fp16, [128]> var_676_to_fp16 = const()[name = string("op_676_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16546688)))];
            tensor<fp16, [1, 51, 128]> var_677_cast_fp16 = add(x = q_head_5_cast_fp16, y = var_676_to_fp16)[name = string("op_677_cast_fp16")];
            tensor<int32, [1]> q_v_5_axes_1 = const()[name = string("q_v_5_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_5_cast_fp16 = expand_dims(axes = q_v_5_axes_1, x = var_677_cast_fp16)[name = string("q_v_5_cast_fp16")];
            tensor<int32, [1]> k_head_11_axes_1 = const()[name = string("k_head_11_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_11_cast_fp16 = expand_dims(axes = k_head_11_axes_1, x = k_head_9_cast_fp16)[name = string("k_head_11_cast_fp16")];
            tensor<int32, [1]> v_head_11_axes_1 = const()[name = string("v_head_11_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_11_cast_fp16 = expand_dims(axes = v_head_11_axes_1, x = v_head_9_cast_fp16)[name = string("v_head_11_cast_fp16")];
            tensor<int32, [1]> p_head_11_axes_1 = const()[name = string("p_head_11_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_11_cast_fp16 = expand_dims(axes = p_head_11_axes_1, x = p_head_9_cast_fp16)[name = string("p_head_11_cast_fp16")];
            bool var_683_transpose_x_3 = const()[name = string("op_683_transpose_x_3"), val = bool(false)];
            bool var_683_transpose_y_3 = const()[name = string("op_683_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_683_cast_fp16 = matmul(transpose_x = var_683_transpose_x_3, transpose_y = var_683_transpose_y_3, x = q_u_5_cast_fp16, y = k_head_11_cast_fp16)[name = string("op_683_cast_fp16")];
            fp16 var_684_to_fp16 = const()[name = string("op_684_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_5_cast_fp16 = mul(x = var_683_cast_fp16, y = var_684_to_fp16)[name = string("scores_content_5_cast_fp16")];
            bool x_37_transpose_x_3 = const()[name = string("x_37_transpose_x_3"), val = bool(false)];
            bool x_37_transpose_y_3 = const()[name = string("x_37_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_37_cast_fp16 = matmul(transpose_x = x_37_transpose_x_3, transpose_y = x_37_transpose_y_3, x = q_v_5_cast_fp16, y = p_head_11_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<int32, [8]> x_39_pad_1 = const()[name = string("x_39_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_39_mode_1 = const()[name = string("x_39_mode_1"), val = string("constant")];
            fp16 const_1321_to_fp16 = const()[name = string("const_1321_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_39_cast_fp16 = pad(constant_val = const_1321_to_fp16, mode = x_39_mode_1, pad = x_39_pad_1, x = x_37_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<int32, [4]> var_698 = const()[name = string("op_698"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_41_cast_fp16 = reshape(shape = var_698, x = x_39_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> var_702_begin_1 = const()[name = string("op_702_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_702_end_1 = const()[name = string("op_702_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_702_end_mask_1 = const()[name = string("op_702_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_702_cast_fp16 = slice_by_index(begin = var_702_begin_1, end = var_702_end_1, end_mask = var_702_end_mask_1, x = x_41_cast_fp16)[name = string("op_702_cast_fp16")];
            tensor<int32, [4]> var_704 = const()[name = string("op_704"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_705_cast_fp16 = reshape(shape = var_704, x = var_702_cast_fp16)[name = string("op_705_cast_fp16")];
            tensor<int32, [4]> var_710_begin_1 = const()[name = string("op_710_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_710_end_1 = const()[name = string("op_710_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_710_end_mask_1 = const()[name = string("op_710_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_710_cast_fp16 = slice_by_index(begin = var_710_begin_1, end = var_710_end_1, end_mask = var_710_end_mask_1, x = var_705_cast_fp16)[name = string("op_710_cast_fp16")];
            fp16 var_711_to_fp16 = const()[name = string("op_711_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_5_cast_fp16 = mul(x = var_710_cast_fp16, y = var_711_to_fp16)[name = string("scores_pos_5_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_5_cast_fp16 = add(x = scores_content_5_cast_fp16, y = scores_pos_5_cast_fp16)[name = string("logits_5_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_714_cast_fp16 = softmax(axis = var_197, x = logits_5_cast_fp16)[name = string("op_714_cast_fp16")];
            bool var_716_transpose_x_1 = const()[name = string("op_716_transpose_x_1"), val = bool(false)];
            bool var_716_transpose_y_1 = const()[name = string("op_716_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_716_cast_fp16 = matmul(transpose_x = var_716_transpose_x_1, transpose_y = var_716_transpose_y_1, x = var_714_cast_fp16, y = v_head_11_cast_fp16)[name = string("op_716_cast_fp16")];
            tensor<int32, [1]> var_717_axes_1 = const()[name = string("op_717_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_717_cast_fp16 = squeeze(axes = var_717_axes_1, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            string dense_output_41_pad_type_1 = const()[name = string("dense_output_41_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_41_strides_1 = const()[name = string("dense_output_41_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_41_pad_1 = const()[name = string("dense_output_41_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_41_dilations_1 = const()[name = string("dense_output_41_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_41_groups_1 = const()[name = string("dense_output_41_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16547008))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16678144))))[name = string("layers_0_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_41_cast_fp16 = conv(dilations = dense_output_41_dilations_1, groups = dense_output_41_groups_1, pad = dense_output_41_pad_1, pad_type = dense_output_41_pad_type_1, strides = dense_output_41_strides_1, weight = layers_0_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_41_cast_fp16")];
            string sparse_output_41_pad_type_1 = const()[name = string("sparse_output_41_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_41_strides_1 = const()[name = string("sparse_output_41_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_41_pad_1 = const()[name = string("sparse_output_41_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_41_dilations_1 = const()[name = string("sparse_output_41_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_41_groups_1 = const()[name = string("sparse_output_41_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16681408))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16678720))))[name = string("layers_0_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_41_cast_fp16 = conv(dilations = sparse_output_41_dilations_1, groups = sparse_output_41_groups_1, pad = sparse_output_41_pad_1, pad_type = sparse_output_41_pad_type_1, strides = sparse_output_41_strides_1, weight = layers_0_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_41_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_732_cast_fp16 = add(x = dense_output_41_cast_fp16, y = sparse_output_41_cast_fp16)[name = string("op_732_cast_fp16")];
            tensor<int32, [4]> var_733 = const()[name = string("op_733"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_735 = const()[name = string("op_735"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_734_cast_fp16 = transpose(perm = var_733, x = var_732_cast_fp16)[name = string("transpose_968")];
            tensor<fp16, [1, 51, 128]> q_head_7_cast_fp16 = reshape(shape = var_735, x = var_734_cast_fp16)[name = string("q_head_7_cast_fp16")];
            string dense_output_43_pad_type_1 = const()[name = string("dense_output_43_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_43_strides_1 = const()[name = string("dense_output_43_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_43_pad_1 = const()[name = string("dense_output_43_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_43_dilations_1 = const()[name = string("dense_output_43_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_43_groups_1 = const()[name = string("dense_output_43_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16697856))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16828992))))[name = string("layers_0_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_43_cast_fp16 = conv(dilations = dense_output_43_dilations_1, groups = dense_output_43_groups_1, pad = dense_output_43_pad_1, pad_type = dense_output_43_pad_type_1, strides = dense_output_43_strides_1, weight = layers_0_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_43_cast_fp16")];
            string sparse_output_43_pad_type_1 = const()[name = string("sparse_output_43_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_43_strides_1 = const()[name = string("sparse_output_43_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_43_pad_1 = const()[name = string("sparse_output_43_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_43_dilations_1 = const()[name = string("sparse_output_43_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_43_groups_1 = const()[name = string("sparse_output_43_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16832256))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16829568))))[name = string("layers_0_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_43_cast_fp16 = conv(dilations = sparse_output_43_dilations_1, groups = sparse_output_43_groups_1, pad = sparse_output_43_pad_1, pad_type = sparse_output_43_pad_type_1, strides = sparse_output_43_strides_1, weight = layers_0_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_43_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_751_cast_fp16 = add(x = dense_output_43_cast_fp16, y = sparse_output_43_cast_fp16)[name = string("op_751_cast_fp16")];
            tensor<int32, [4]> var_752 = const()[name = string("op_752"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_754 = const()[name = string("op_754"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_753_cast_fp16 = transpose(perm = var_752, x = var_751_cast_fp16)[name = string("transpose_967")];
            tensor<fp16, [1, 51, 128]> k_head_13_cast_fp16 = reshape(shape = var_754, x = var_753_cast_fp16)[name = string("k_head_13_cast_fp16")];
            string dense_output_45_pad_type_1 = const()[name = string("dense_output_45_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_45_strides_1 = const()[name = string("dense_output_45_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_45_pad_1 = const()[name = string("dense_output_45_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_45_dilations_1 = const()[name = string("dense_output_45_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_45_groups_1 = const()[name = string("dense_output_45_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16848704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16979840))))[name = string("layers_0_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_45_cast_fp16 = conv(dilations = dense_output_45_dilations_1, groups = dense_output_45_groups_1, pad = dense_output_45_pad_1, pad_type = dense_output_45_pad_type_1, strides = dense_output_45_strides_1, weight = layers_0_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_45_cast_fp16")];
            string sparse_output_45_pad_type_1 = const()[name = string("sparse_output_45_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_45_strides_1 = const()[name = string("sparse_output_45_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_45_pad_1 = const()[name = string("sparse_output_45_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_45_dilations_1 = const()[name = string("sparse_output_45_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_45_groups_1 = const()[name = string("sparse_output_45_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16983104))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16980416))))[name = string("layers_0_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_45_cast_fp16 = conv(dilations = sparse_output_45_dilations_1, groups = sparse_output_45_groups_1, pad = sparse_output_45_pad_1, pad_type = sparse_output_45_pad_type_1, strides = sparse_output_45_strides_1, weight = layers_0_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_45_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_770_cast_fp16 = add(x = dense_output_45_cast_fp16, y = sparse_output_45_cast_fp16)[name = string("op_770_cast_fp16")];
            tensor<int32, [4]> var_771 = const()[name = string("op_771"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_773 = const()[name = string("op_773"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_772_cast_fp16 = transpose(perm = var_771, x = var_770_cast_fp16)[name = string("transpose_966")];
            tensor<fp16, [1, 51, 128]> v_head_13_cast_fp16 = reshape(shape = var_773, x = var_772_cast_fp16)[name = string("v_head_13_cast_fp16")];
            string dense_output_47_pad_type_1 = const()[name = string("dense_output_47_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_47_strides_1 = const()[name = string("dense_output_47_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_47_pad_1 = const()[name = string("dense_output_47_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_47_dilations_1 = const()[name = string("dense_output_47_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_47_groups_1 = const()[name = string("dense_output_47_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16999552))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17130688))))[name = string("layers_0_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_47_cast_fp16 = conv(dilations = dense_output_47_dilations_1, groups = dense_output_47_groups_1, pad = dense_output_47_pad_1, pad_type = dense_output_47_pad_type_1, strides = dense_output_47_strides_1, weight = layers_0_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_47_cast_fp16")];
            string sparse_output_47_pad_type_1 = const()[name = string("sparse_output_47_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_47_strides_1 = const()[name = string("sparse_output_47_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_47_pad_1 = const()[name = string("sparse_output_47_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_47_dilations_1 = const()[name = string("sparse_output_47_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_47_groups_1 = const()[name = string("sparse_output_47_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17133952))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17131264))))[name = string("layers_0_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_47_cast_fp16 = conv(dilations = sparse_output_47_dilations_1, groups = sparse_output_47_groups_1, pad = sparse_output_47_pad_1, pad_type = sparse_output_47_pad_type_1, strides = sparse_output_47_strides_1, weight = layers_0_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_47_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_789_cast_fp16 = add(x = dense_output_47_cast_fp16, y = sparse_output_47_cast_fp16)[name = string("op_789_cast_fp16")];
            tensor<int32, [4]> var_790 = const()[name = string("op_790"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_792 = const()[name = string("op_792"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_791_cast_fp16 = transpose(perm = var_790, x = var_789_cast_fp16)[name = string("transpose_965")];
            tensor<fp16, [1, 101, 128]> p_head_13_cast_fp16 = reshape(shape = var_792, x = var_791_cast_fp16)[name = string("p_head_13_cast_fp16")];
            tensor<fp16, [128]> var_794_to_fp16 = const()[name = string("op_794_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17150400)))];
            tensor<fp16, [1, 51, 128]> var_795_cast_fp16 = add(x = q_head_7_cast_fp16, y = var_794_to_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [1]> q_u_7_axes_1 = const()[name = string("q_u_7_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_7_cast_fp16 = expand_dims(axes = q_u_7_axes_1, x = var_795_cast_fp16)[name = string("q_u_7_cast_fp16")];
            tensor<fp16, [128]> var_797_to_fp16 = const()[name = string("op_797_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17150720)))];
            tensor<fp16, [1, 51, 128]> var_798_cast_fp16 = add(x = q_head_7_cast_fp16, y = var_797_to_fp16)[name = string("op_798_cast_fp16")];
            tensor<int32, [1]> q_v_7_axes_1 = const()[name = string("q_v_7_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_7_cast_fp16 = expand_dims(axes = q_v_7_axes_1, x = var_798_cast_fp16)[name = string("q_v_7_cast_fp16")];
            tensor<int32, [1]> k_head_15_axes_1 = const()[name = string("k_head_15_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_15_cast_fp16 = expand_dims(axes = k_head_15_axes_1, x = k_head_13_cast_fp16)[name = string("k_head_15_cast_fp16")];
            tensor<int32, [1]> v_head_15_axes_1 = const()[name = string("v_head_15_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_15_cast_fp16 = expand_dims(axes = v_head_15_axes_1, x = v_head_13_cast_fp16)[name = string("v_head_15_cast_fp16")];
            tensor<int32, [1]> p_head_15_axes_1 = const()[name = string("p_head_15_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_15_cast_fp16 = expand_dims(axes = p_head_15_axes_1, x = p_head_13_cast_fp16)[name = string("p_head_15_cast_fp16")];
            bool var_804_transpose_x_3 = const()[name = string("op_804_transpose_x_3"), val = bool(false)];
            bool var_804_transpose_y_3 = const()[name = string("op_804_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_804_cast_fp16 = matmul(transpose_x = var_804_transpose_x_3, transpose_y = var_804_transpose_y_3, x = q_u_7_cast_fp16, y = k_head_15_cast_fp16)[name = string("op_804_cast_fp16")];
            fp16 var_805_to_fp16 = const()[name = string("op_805_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_7_cast_fp16 = mul(x = var_804_cast_fp16, y = var_805_to_fp16)[name = string("scores_content_7_cast_fp16")];
            bool x_45_transpose_x_3 = const()[name = string("x_45_transpose_x_3"), val = bool(false)];
            bool x_45_transpose_y_3 = const()[name = string("x_45_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_45_cast_fp16 = matmul(transpose_x = x_45_transpose_x_3, transpose_y = x_45_transpose_y_3, x = q_v_7_cast_fp16, y = p_head_15_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<int32, [8]> x_47_pad_1 = const()[name = string("x_47_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_47_mode_1 = const()[name = string("x_47_mode_1"), val = string("constant")];
            fp16 const_1327_to_fp16 = const()[name = string("const_1327_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_47_cast_fp16 = pad(constant_val = const_1327_to_fp16, mode = x_47_mode_1, pad = x_47_pad_1, x = x_45_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<int32, [4]> var_819 = const()[name = string("op_819"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_49_cast_fp16 = reshape(shape = var_819, x = x_47_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [4]> var_823_begin_1 = const()[name = string("op_823_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_823_end_1 = const()[name = string("op_823_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_823_end_mask_1 = const()[name = string("op_823_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_823_cast_fp16 = slice_by_index(begin = var_823_begin_1, end = var_823_end_1, end_mask = var_823_end_mask_1, x = x_49_cast_fp16)[name = string("op_823_cast_fp16")];
            tensor<int32, [4]> var_825 = const()[name = string("op_825"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_826_cast_fp16 = reshape(shape = var_825, x = var_823_cast_fp16)[name = string("op_826_cast_fp16")];
            tensor<int32, [4]> var_831_begin_1 = const()[name = string("op_831_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_831_end_1 = const()[name = string("op_831_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_831_end_mask_1 = const()[name = string("op_831_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_831_cast_fp16 = slice_by_index(begin = var_831_begin_1, end = var_831_end_1, end_mask = var_831_end_mask_1, x = var_826_cast_fp16)[name = string("op_831_cast_fp16")];
            fp16 var_832_to_fp16 = const()[name = string("op_832_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_7_cast_fp16 = mul(x = var_831_cast_fp16, y = var_832_to_fp16)[name = string("scores_pos_7_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_7_cast_fp16 = add(x = scores_content_7_cast_fp16, y = scores_pos_7_cast_fp16)[name = string("logits_7_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_835_cast_fp16 = softmax(axis = var_197, x = logits_7_cast_fp16)[name = string("op_835_cast_fp16")];
            bool var_837_transpose_x_1 = const()[name = string("op_837_transpose_x_1"), val = bool(false)];
            bool var_837_transpose_y_1 = const()[name = string("op_837_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_837_cast_fp16 = matmul(transpose_x = var_837_transpose_x_1, transpose_y = var_837_transpose_y_1, x = var_835_cast_fp16, y = v_head_15_cast_fp16)[name = string("op_837_cast_fp16")];
            tensor<int32, [1]> var_838_axes_1 = const()[name = string("op_838_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_838_cast_fp16 = squeeze(axes = var_838_axes_1, x = var_837_cast_fp16)[name = string("op_838_cast_fp16")];
            string dense_output_49_pad_type_1 = const()[name = string("dense_output_49_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_49_strides_1 = const()[name = string("dense_output_49_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_49_pad_1 = const()[name = string("dense_output_49_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_49_dilations_1 = const()[name = string("dense_output_49_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_49_groups_1 = const()[name = string("dense_output_49_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17151040))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17282176))))[name = string("layers_0_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_49_cast_fp16 = conv(dilations = dense_output_49_dilations_1, groups = dense_output_49_groups_1, pad = dense_output_49_pad_1, pad_type = dense_output_49_pad_type_1, strides = dense_output_49_strides_1, weight = layers_0_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_49_cast_fp16")];
            string sparse_output_49_pad_type_1 = const()[name = string("sparse_output_49_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_49_strides_1 = const()[name = string("sparse_output_49_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_49_pad_1 = const()[name = string("sparse_output_49_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_49_dilations_1 = const()[name = string("sparse_output_49_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_49_groups_1 = const()[name = string("sparse_output_49_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17285440))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17282752))))[name = string("layers_0_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_49_cast_fp16 = conv(dilations = sparse_output_49_dilations_1, groups = sparse_output_49_groups_1, pad = sparse_output_49_pad_1, pad_type = sparse_output_49_pad_type_1, strides = sparse_output_49_strides_1, weight = layers_0_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_49_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_853_cast_fp16 = add(x = dense_output_49_cast_fp16, y = sparse_output_49_cast_fp16)[name = string("op_853_cast_fp16")];
            tensor<int32, [4]> var_854 = const()[name = string("op_854"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_856 = const()[name = string("op_856"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_855_cast_fp16 = transpose(perm = var_854, x = var_853_cast_fp16)[name = string("transpose_964")];
            tensor<fp16, [1, 51, 128]> q_head_9_cast_fp16 = reshape(shape = var_856, x = var_855_cast_fp16)[name = string("q_head_9_cast_fp16")];
            string dense_output_51_pad_type_1 = const()[name = string("dense_output_51_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_51_strides_1 = const()[name = string("dense_output_51_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_51_pad_1 = const()[name = string("dense_output_51_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_51_dilations_1 = const()[name = string("dense_output_51_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_51_groups_1 = const()[name = string("dense_output_51_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17301888))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17433024))))[name = string("layers_0_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_51_cast_fp16 = conv(dilations = dense_output_51_dilations_1, groups = dense_output_51_groups_1, pad = dense_output_51_pad_1, pad_type = dense_output_51_pad_type_1, strides = dense_output_51_strides_1, weight = layers_0_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_51_cast_fp16")];
            string sparse_output_51_pad_type_1 = const()[name = string("sparse_output_51_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_51_strides_1 = const()[name = string("sparse_output_51_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_51_pad_1 = const()[name = string("sparse_output_51_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_51_dilations_1 = const()[name = string("sparse_output_51_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_51_groups_1 = const()[name = string("sparse_output_51_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17436288))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17433600))))[name = string("layers_0_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_51_cast_fp16 = conv(dilations = sparse_output_51_dilations_1, groups = sparse_output_51_groups_1, pad = sparse_output_51_pad_1, pad_type = sparse_output_51_pad_type_1, strides = sparse_output_51_strides_1, weight = layers_0_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_51_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_872_cast_fp16 = add(x = dense_output_51_cast_fp16, y = sparse_output_51_cast_fp16)[name = string("op_872_cast_fp16")];
            tensor<int32, [4]> var_873 = const()[name = string("op_873"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_875 = const()[name = string("op_875"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_874_cast_fp16 = transpose(perm = var_873, x = var_872_cast_fp16)[name = string("transpose_963")];
            tensor<fp16, [1, 51, 128]> k_head_17_cast_fp16 = reshape(shape = var_875, x = var_874_cast_fp16)[name = string("k_head_17_cast_fp16")];
            string dense_output_53_pad_type_1 = const()[name = string("dense_output_53_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_53_strides_1 = const()[name = string("dense_output_53_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_53_pad_1 = const()[name = string("dense_output_53_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_53_dilations_1 = const()[name = string("dense_output_53_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_53_groups_1 = const()[name = string("dense_output_53_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17452736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17583872))))[name = string("layers_0_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_53_cast_fp16 = conv(dilations = dense_output_53_dilations_1, groups = dense_output_53_groups_1, pad = dense_output_53_pad_1, pad_type = dense_output_53_pad_type_1, strides = dense_output_53_strides_1, weight = layers_0_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_53_cast_fp16")];
            string sparse_output_53_pad_type_1 = const()[name = string("sparse_output_53_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_53_strides_1 = const()[name = string("sparse_output_53_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_53_pad_1 = const()[name = string("sparse_output_53_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_53_dilations_1 = const()[name = string("sparse_output_53_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_53_groups_1 = const()[name = string("sparse_output_53_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17587136))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17584448))))[name = string("layers_0_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_53_cast_fp16 = conv(dilations = sparse_output_53_dilations_1, groups = sparse_output_53_groups_1, pad = sparse_output_53_pad_1, pad_type = sparse_output_53_pad_type_1, strides = sparse_output_53_strides_1, weight = layers_0_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_53_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_891_cast_fp16 = add(x = dense_output_53_cast_fp16, y = sparse_output_53_cast_fp16)[name = string("op_891_cast_fp16")];
            tensor<int32, [4]> var_892 = const()[name = string("op_892"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_894 = const()[name = string("op_894"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_893_cast_fp16 = transpose(perm = var_892, x = var_891_cast_fp16)[name = string("transpose_962")];
            tensor<fp16, [1, 51, 128]> v_head_17_cast_fp16 = reshape(shape = var_894, x = var_893_cast_fp16)[name = string("v_head_17_cast_fp16")];
            string dense_output_55_pad_type_1 = const()[name = string("dense_output_55_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_55_strides_1 = const()[name = string("dense_output_55_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_55_pad_1 = const()[name = string("dense_output_55_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_55_dilations_1 = const()[name = string("dense_output_55_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_55_groups_1 = const()[name = string("dense_output_55_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17603584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17734720))))[name = string("layers_0_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_55_cast_fp16 = conv(dilations = dense_output_55_dilations_1, groups = dense_output_55_groups_1, pad = dense_output_55_pad_1, pad_type = dense_output_55_pad_type_1, strides = dense_output_55_strides_1, weight = layers_0_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_55_cast_fp16")];
            string sparse_output_55_pad_type_1 = const()[name = string("sparse_output_55_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_55_strides_1 = const()[name = string("sparse_output_55_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_55_pad_1 = const()[name = string("sparse_output_55_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_55_dilations_1 = const()[name = string("sparse_output_55_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_55_groups_1 = const()[name = string("sparse_output_55_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17737984))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17735296))))[name = string("layers_0_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_55_cast_fp16 = conv(dilations = sparse_output_55_dilations_1, groups = sparse_output_55_groups_1, pad = sparse_output_55_pad_1, pad_type = sparse_output_55_pad_type_1, strides = sparse_output_55_strides_1, weight = layers_0_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_55_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_910_cast_fp16 = add(x = dense_output_55_cast_fp16, y = sparse_output_55_cast_fp16)[name = string("op_910_cast_fp16")];
            tensor<int32, [4]> var_911 = const()[name = string("op_911"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_913 = const()[name = string("op_913"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_912_cast_fp16 = transpose(perm = var_911, x = var_910_cast_fp16)[name = string("transpose_961")];
            tensor<fp16, [1, 101, 128]> p_head_17_cast_fp16 = reshape(shape = var_913, x = var_912_cast_fp16)[name = string("p_head_17_cast_fp16")];
            tensor<fp16, [128]> var_915_to_fp16 = const()[name = string("op_915_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17754432)))];
            tensor<fp16, [1, 51, 128]> var_916_cast_fp16 = add(x = q_head_9_cast_fp16, y = var_915_to_fp16)[name = string("op_916_cast_fp16")];
            tensor<int32, [1]> q_u_9_axes_1 = const()[name = string("q_u_9_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_9_cast_fp16 = expand_dims(axes = q_u_9_axes_1, x = var_916_cast_fp16)[name = string("q_u_9_cast_fp16")];
            tensor<fp16, [128]> var_918_to_fp16 = const()[name = string("op_918_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17754752)))];
            tensor<fp16, [1, 51, 128]> var_919_cast_fp16 = add(x = q_head_9_cast_fp16, y = var_918_to_fp16)[name = string("op_919_cast_fp16")];
            tensor<int32, [1]> q_v_9_axes_1 = const()[name = string("q_v_9_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_9_cast_fp16 = expand_dims(axes = q_v_9_axes_1, x = var_919_cast_fp16)[name = string("q_v_9_cast_fp16")];
            tensor<int32, [1]> k_head_19_axes_1 = const()[name = string("k_head_19_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_19_cast_fp16 = expand_dims(axes = k_head_19_axes_1, x = k_head_17_cast_fp16)[name = string("k_head_19_cast_fp16")];
            tensor<int32, [1]> v_head_19_axes_1 = const()[name = string("v_head_19_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_19_cast_fp16 = expand_dims(axes = v_head_19_axes_1, x = v_head_17_cast_fp16)[name = string("v_head_19_cast_fp16")];
            tensor<int32, [1]> p_head_19_axes_1 = const()[name = string("p_head_19_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_19_cast_fp16 = expand_dims(axes = p_head_19_axes_1, x = p_head_17_cast_fp16)[name = string("p_head_19_cast_fp16")];
            bool var_925_transpose_x_3 = const()[name = string("op_925_transpose_x_3"), val = bool(false)];
            bool var_925_transpose_y_3 = const()[name = string("op_925_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_925_cast_fp16 = matmul(transpose_x = var_925_transpose_x_3, transpose_y = var_925_transpose_y_3, x = q_u_9_cast_fp16, y = k_head_19_cast_fp16)[name = string("op_925_cast_fp16")];
            fp16 var_926_to_fp16 = const()[name = string("op_926_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_9_cast_fp16 = mul(x = var_925_cast_fp16, y = var_926_to_fp16)[name = string("scores_content_9_cast_fp16")];
            bool x_53_transpose_x_3 = const()[name = string("x_53_transpose_x_3"), val = bool(false)];
            bool x_53_transpose_y_3 = const()[name = string("x_53_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_53_cast_fp16 = matmul(transpose_x = x_53_transpose_x_3, transpose_y = x_53_transpose_y_3, x = q_v_9_cast_fp16, y = p_head_19_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<int32, [8]> x_55_pad_1 = const()[name = string("x_55_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_55_mode_1 = const()[name = string("x_55_mode_1"), val = string("constant")];
            fp16 const_1333_to_fp16 = const()[name = string("const_1333_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_55_cast_fp16 = pad(constant_val = const_1333_to_fp16, mode = x_55_mode_1, pad = x_55_pad_1, x = x_53_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<int32, [4]> var_940 = const()[name = string("op_940"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_57_cast_fp16 = reshape(shape = var_940, x = x_55_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<int32, [4]> var_944_begin_1 = const()[name = string("op_944_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_944_end_1 = const()[name = string("op_944_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_944_end_mask_1 = const()[name = string("op_944_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_944_cast_fp16 = slice_by_index(begin = var_944_begin_1, end = var_944_end_1, end_mask = var_944_end_mask_1, x = x_57_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<int32, [4]> var_946 = const()[name = string("op_946"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_947_cast_fp16 = reshape(shape = var_946, x = var_944_cast_fp16)[name = string("op_947_cast_fp16")];
            tensor<int32, [4]> var_952_begin_1 = const()[name = string("op_952_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_952_end_1 = const()[name = string("op_952_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_952_end_mask_1 = const()[name = string("op_952_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_952_cast_fp16 = slice_by_index(begin = var_952_begin_1, end = var_952_end_1, end_mask = var_952_end_mask_1, x = var_947_cast_fp16)[name = string("op_952_cast_fp16")];
            fp16 var_953_to_fp16 = const()[name = string("op_953_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_9_cast_fp16 = mul(x = var_952_cast_fp16, y = var_953_to_fp16)[name = string("scores_pos_9_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_9_cast_fp16 = add(x = scores_content_9_cast_fp16, y = scores_pos_9_cast_fp16)[name = string("logits_9_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_956_cast_fp16 = softmax(axis = var_197, x = logits_9_cast_fp16)[name = string("op_956_cast_fp16")];
            bool var_958_transpose_x_1 = const()[name = string("op_958_transpose_x_1"), val = bool(false)];
            bool var_958_transpose_y_1 = const()[name = string("op_958_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_958_cast_fp16 = matmul(transpose_x = var_958_transpose_x_1, transpose_y = var_958_transpose_y_1, x = var_956_cast_fp16, y = v_head_19_cast_fp16)[name = string("op_958_cast_fp16")];
            tensor<int32, [1]> var_959_axes_1 = const()[name = string("op_959_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_959_cast_fp16 = squeeze(axes = var_959_axes_1, x = var_958_cast_fp16)[name = string("op_959_cast_fp16")];
            string dense_output_57_pad_type_1 = const()[name = string("dense_output_57_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_57_strides_1 = const()[name = string("dense_output_57_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_57_pad_1 = const()[name = string("dense_output_57_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_57_dilations_1 = const()[name = string("dense_output_57_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_57_groups_1 = const()[name = string("dense_output_57_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17755072))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17886208))))[name = string("layers_0_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_57_cast_fp16 = conv(dilations = dense_output_57_dilations_1, groups = dense_output_57_groups_1, pad = dense_output_57_pad_1, pad_type = dense_output_57_pad_type_1, strides = dense_output_57_strides_1, weight = layers_0_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_57_cast_fp16")];
            string sparse_output_57_pad_type_1 = const()[name = string("sparse_output_57_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_57_strides_1 = const()[name = string("sparse_output_57_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_57_pad_1 = const()[name = string("sparse_output_57_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_57_dilations_1 = const()[name = string("sparse_output_57_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_57_groups_1 = const()[name = string("sparse_output_57_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17889472))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17886784))))[name = string("layers_0_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_57_cast_fp16 = conv(dilations = sparse_output_57_dilations_1, groups = sparse_output_57_groups_1, pad = sparse_output_57_pad_1, pad_type = sparse_output_57_pad_type_1, strides = sparse_output_57_strides_1, weight = layers_0_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_57_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_974_cast_fp16 = add(x = dense_output_57_cast_fp16, y = sparse_output_57_cast_fp16)[name = string("op_974_cast_fp16")];
            tensor<int32, [4]> var_975 = const()[name = string("op_975"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_977 = const()[name = string("op_977"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_976_cast_fp16 = transpose(perm = var_975, x = var_974_cast_fp16)[name = string("transpose_960")];
            tensor<fp16, [1, 51, 128]> q_head_11_cast_fp16 = reshape(shape = var_977, x = var_976_cast_fp16)[name = string("q_head_11_cast_fp16")];
            string dense_output_59_pad_type_1 = const()[name = string("dense_output_59_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_59_strides_1 = const()[name = string("dense_output_59_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_59_pad_1 = const()[name = string("dense_output_59_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_59_dilations_1 = const()[name = string("dense_output_59_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_59_groups_1 = const()[name = string("dense_output_59_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17905920))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18037056))))[name = string("layers_0_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_59_cast_fp16 = conv(dilations = dense_output_59_dilations_1, groups = dense_output_59_groups_1, pad = dense_output_59_pad_1, pad_type = dense_output_59_pad_type_1, strides = dense_output_59_strides_1, weight = layers_0_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_59_cast_fp16")];
            string sparse_output_59_pad_type_1 = const()[name = string("sparse_output_59_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_59_strides_1 = const()[name = string("sparse_output_59_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_59_pad_1 = const()[name = string("sparse_output_59_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_59_dilations_1 = const()[name = string("sparse_output_59_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_59_groups_1 = const()[name = string("sparse_output_59_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18040320))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18037632))))[name = string("layers_0_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_59_cast_fp16 = conv(dilations = sparse_output_59_dilations_1, groups = sparse_output_59_groups_1, pad = sparse_output_59_pad_1, pad_type = sparse_output_59_pad_type_1, strides = sparse_output_59_strides_1, weight = layers_0_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_59_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_993_cast_fp16 = add(x = dense_output_59_cast_fp16, y = sparse_output_59_cast_fp16)[name = string("op_993_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_996 = const()[name = string("op_996"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_995_cast_fp16 = transpose(perm = var_994, x = var_993_cast_fp16)[name = string("transpose_959")];
            tensor<fp16, [1, 51, 128]> k_head_21_cast_fp16 = reshape(shape = var_996, x = var_995_cast_fp16)[name = string("k_head_21_cast_fp16")];
            string dense_output_61_pad_type_1 = const()[name = string("dense_output_61_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_61_strides_1 = const()[name = string("dense_output_61_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_61_pad_1 = const()[name = string("dense_output_61_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_61_dilations_1 = const()[name = string("dense_output_61_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_61_groups_1 = const()[name = string("dense_output_61_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18056768))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18187904))))[name = string("layers_0_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_61_cast_fp16 = conv(dilations = dense_output_61_dilations_1, groups = dense_output_61_groups_1, pad = dense_output_61_pad_1, pad_type = dense_output_61_pad_type_1, strides = dense_output_61_strides_1, weight = layers_0_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_61_cast_fp16")];
            string sparse_output_61_pad_type_1 = const()[name = string("sparse_output_61_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_61_strides_1 = const()[name = string("sparse_output_61_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_61_pad_1 = const()[name = string("sparse_output_61_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_61_dilations_1 = const()[name = string("sparse_output_61_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_61_groups_1 = const()[name = string("sparse_output_61_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18191168))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18188480))))[name = string("layers_0_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_61_cast_fp16 = conv(dilations = sparse_output_61_dilations_1, groups = sparse_output_61_groups_1, pad = sparse_output_61_pad_1, pad_type = sparse_output_61_pad_type_1, strides = sparse_output_61_strides_1, weight = layers_0_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_61_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1012_cast_fp16 = add(x = dense_output_61_cast_fp16, y = sparse_output_61_cast_fp16)[name = string("op_1012_cast_fp16")];
            tensor<int32, [4]> var_1013 = const()[name = string("op_1013"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1015 = const()[name = string("op_1015"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1014_cast_fp16 = transpose(perm = var_1013, x = var_1012_cast_fp16)[name = string("transpose_958")];
            tensor<fp16, [1, 51, 128]> v_head_21_cast_fp16 = reshape(shape = var_1015, x = var_1014_cast_fp16)[name = string("v_head_21_cast_fp16")];
            string dense_output_63_pad_type_1 = const()[name = string("dense_output_63_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_63_strides_1 = const()[name = string("dense_output_63_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_63_pad_1 = const()[name = string("dense_output_63_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_63_dilations_1 = const()[name = string("dense_output_63_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_63_groups_1 = const()[name = string("dense_output_63_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18207616))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18338752))))[name = string("layers_0_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_63_cast_fp16 = conv(dilations = dense_output_63_dilations_1, groups = dense_output_63_groups_1, pad = dense_output_63_pad_1, pad_type = dense_output_63_pad_type_1, strides = dense_output_63_strides_1, weight = layers_0_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_63_cast_fp16")];
            string sparse_output_63_pad_type_1 = const()[name = string("sparse_output_63_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_63_strides_1 = const()[name = string("sparse_output_63_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_63_pad_1 = const()[name = string("sparse_output_63_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_63_dilations_1 = const()[name = string("sparse_output_63_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_63_groups_1 = const()[name = string("sparse_output_63_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18342016))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18339328))))[name = string("layers_0_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_63_cast_fp16 = conv(dilations = sparse_output_63_dilations_1, groups = sparse_output_63_groups_1, pad = sparse_output_63_pad_1, pad_type = sparse_output_63_pad_type_1, strides = sparse_output_63_strides_1, weight = layers_0_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_63_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_1031_cast_fp16 = add(x = dense_output_63_cast_fp16, y = sparse_output_63_cast_fp16)[name = string("op_1031_cast_fp16")];
            tensor<int32, [4]> var_1032 = const()[name = string("op_1032"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1034 = const()[name = string("op_1034"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_1033_cast_fp16 = transpose(perm = var_1032, x = var_1031_cast_fp16)[name = string("transpose_957")];
            tensor<fp16, [1, 101, 128]> p_head_21_cast_fp16 = reshape(shape = var_1034, x = var_1033_cast_fp16)[name = string("p_head_21_cast_fp16")];
            tensor<fp16, [128]> var_1036_to_fp16 = const()[name = string("op_1036_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18358464)))];
            tensor<fp16, [1, 51, 128]> var_1037_cast_fp16 = add(x = q_head_11_cast_fp16, y = var_1036_to_fp16)[name = string("op_1037_cast_fp16")];
            tensor<int32, [1]> q_u_11_axes_1 = const()[name = string("q_u_11_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_11_cast_fp16 = expand_dims(axes = q_u_11_axes_1, x = var_1037_cast_fp16)[name = string("q_u_11_cast_fp16")];
            tensor<fp16, [128]> var_1039_to_fp16 = const()[name = string("op_1039_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18358784)))];
            tensor<fp16, [1, 51, 128]> var_1040_cast_fp16 = add(x = q_head_11_cast_fp16, y = var_1039_to_fp16)[name = string("op_1040_cast_fp16")];
            tensor<int32, [1]> q_v_11_axes_1 = const()[name = string("q_v_11_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_11_cast_fp16 = expand_dims(axes = q_v_11_axes_1, x = var_1040_cast_fp16)[name = string("q_v_11_cast_fp16")];
            tensor<int32, [1]> k_head_23_axes_1 = const()[name = string("k_head_23_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_23_cast_fp16 = expand_dims(axes = k_head_23_axes_1, x = k_head_21_cast_fp16)[name = string("k_head_23_cast_fp16")];
            tensor<int32, [1]> v_head_23_axes_1 = const()[name = string("v_head_23_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_23_cast_fp16 = expand_dims(axes = v_head_23_axes_1, x = v_head_21_cast_fp16)[name = string("v_head_23_cast_fp16")];
            tensor<int32, [1]> p_head_23_axes_1 = const()[name = string("p_head_23_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_23_cast_fp16 = expand_dims(axes = p_head_23_axes_1, x = p_head_21_cast_fp16)[name = string("p_head_23_cast_fp16")];
            bool var_1046_transpose_x_3 = const()[name = string("op_1046_transpose_x_3"), val = bool(false)];
            bool var_1046_transpose_y_3 = const()[name = string("op_1046_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_1046_cast_fp16 = matmul(transpose_x = var_1046_transpose_x_3, transpose_y = var_1046_transpose_y_3, x = q_u_11_cast_fp16, y = k_head_23_cast_fp16)[name = string("op_1046_cast_fp16")];
            fp16 var_1047_to_fp16 = const()[name = string("op_1047_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_11_cast_fp16 = mul(x = var_1046_cast_fp16, y = var_1047_to_fp16)[name = string("scores_content_11_cast_fp16")];
            bool x_61_transpose_x_3 = const()[name = string("x_61_transpose_x_3"), val = bool(false)];
            bool x_61_transpose_y_3 = const()[name = string("x_61_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_61_cast_fp16 = matmul(transpose_x = x_61_transpose_x_3, transpose_y = x_61_transpose_y_3, x = q_v_11_cast_fp16, y = p_head_23_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [8]> x_63_pad_1 = const()[name = string("x_63_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_63_mode_1 = const()[name = string("x_63_mode_1"), val = string("constant")];
            fp16 const_1339_to_fp16 = const()[name = string("const_1339_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_63_cast_fp16 = pad(constant_val = const_1339_to_fp16, mode = x_63_mode_1, pad = x_63_pad_1, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<int32, [4]> var_1061 = const()[name = string("op_1061"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_65_cast_fp16 = reshape(shape = var_1061, x = x_63_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<int32, [4]> var_1065_begin_1 = const()[name = string("op_1065_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1065_end_1 = const()[name = string("op_1065_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_1065_end_mask_1 = const()[name = string("op_1065_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_1065_cast_fp16 = slice_by_index(begin = var_1065_begin_1, end = var_1065_end_1, end_mask = var_1065_end_mask_1, x = x_65_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> var_1067 = const()[name = string("op_1067"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_1068_cast_fp16 = reshape(shape = var_1067, x = var_1065_cast_fp16)[name = string("op_1068_cast_fp16")];
            tensor<int32, [4]> var_1073_begin_1 = const()[name = string("op_1073_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1073_end_1 = const()[name = string("op_1073_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_1073_end_mask_1 = const()[name = string("op_1073_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_1073_cast_fp16 = slice_by_index(begin = var_1073_begin_1, end = var_1073_end_1, end_mask = var_1073_end_mask_1, x = var_1068_cast_fp16)[name = string("op_1073_cast_fp16")];
            fp16 var_1074_to_fp16 = const()[name = string("op_1074_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_11_cast_fp16 = mul(x = var_1073_cast_fp16, y = var_1074_to_fp16)[name = string("scores_pos_11_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_11_cast_fp16 = add(x = scores_content_11_cast_fp16, y = scores_pos_11_cast_fp16)[name = string("logits_11_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_1077_cast_fp16 = softmax(axis = var_197, x = logits_11_cast_fp16)[name = string("op_1077_cast_fp16")];
            bool var_1079_transpose_x_1 = const()[name = string("op_1079_transpose_x_1"), val = bool(false)];
            bool var_1079_transpose_y_1 = const()[name = string("op_1079_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_1079_cast_fp16 = matmul(transpose_x = var_1079_transpose_x_1, transpose_y = var_1079_transpose_y_1, x = var_1077_cast_fp16, y = v_head_23_cast_fp16)[name = string("op_1079_cast_fp16")];
            tensor<int32, [1]> var_1080_axes_1 = const()[name = string("op_1080_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_1080_cast_fp16 = squeeze(axes = var_1080_axes_1, x = var_1079_cast_fp16)[name = string("op_1080_cast_fp16")];
            string dense_output_65_pad_type_1 = const()[name = string("dense_output_65_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_65_strides_1 = const()[name = string("dense_output_65_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_65_pad_1 = const()[name = string("dense_output_65_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_65_dilations_1 = const()[name = string("dense_output_65_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_65_groups_1 = const()[name = string("dense_output_65_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18359104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18490240))))[name = string("layers_0_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_65_cast_fp16 = conv(dilations = dense_output_65_dilations_1, groups = dense_output_65_groups_1, pad = dense_output_65_pad_1, pad_type = dense_output_65_pad_type_1, strides = dense_output_65_strides_1, weight = layers_0_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_65_cast_fp16")];
            string sparse_output_65_pad_type_1 = const()[name = string("sparse_output_65_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_65_strides_1 = const()[name = string("sparse_output_65_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_65_pad_1 = const()[name = string("sparse_output_65_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_65_dilations_1 = const()[name = string("sparse_output_65_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_65_groups_1 = const()[name = string("sparse_output_65_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18493504))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18490816))))[name = string("layers_0_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_65_cast_fp16 = conv(dilations = sparse_output_65_dilations_1, groups = sparse_output_65_groups_1, pad = sparse_output_65_pad_1, pad_type = sparse_output_65_pad_type_1, strides = sparse_output_65_strides_1, weight = layers_0_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_65_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1095_cast_fp16 = add(x = dense_output_65_cast_fp16, y = sparse_output_65_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<int32, [4]> var_1096 = const()[name = string("op_1096"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1098 = const()[name = string("op_1098"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1097_cast_fp16 = transpose(perm = var_1096, x = var_1095_cast_fp16)[name = string("transpose_956")];
            tensor<fp16, [1, 51, 128]> q_head_13_cast_fp16 = reshape(shape = var_1098, x = var_1097_cast_fp16)[name = string("q_head_13_cast_fp16")];
            string dense_output_67_pad_type_1 = const()[name = string("dense_output_67_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_67_strides_1 = const()[name = string("dense_output_67_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_67_pad_1 = const()[name = string("dense_output_67_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_67_dilations_1 = const()[name = string("dense_output_67_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_67_groups_1 = const()[name = string("dense_output_67_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18509952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18641088))))[name = string("layers_0_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_67_cast_fp16 = conv(dilations = dense_output_67_dilations_1, groups = dense_output_67_groups_1, pad = dense_output_67_pad_1, pad_type = dense_output_67_pad_type_1, strides = dense_output_67_strides_1, weight = layers_0_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_67_cast_fp16")];
            string sparse_output_67_pad_type_1 = const()[name = string("sparse_output_67_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_67_strides_1 = const()[name = string("sparse_output_67_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_67_pad_1 = const()[name = string("sparse_output_67_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_67_dilations_1 = const()[name = string("sparse_output_67_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_67_groups_1 = const()[name = string("sparse_output_67_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18644352))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18641664))))[name = string("layers_0_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_67_cast_fp16 = conv(dilations = sparse_output_67_dilations_1, groups = sparse_output_67_groups_1, pad = sparse_output_67_pad_1, pad_type = sparse_output_67_pad_type_1, strides = sparse_output_67_strides_1, weight = layers_0_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_67_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1114_cast_fp16 = add(x = dense_output_67_cast_fp16, y = sparse_output_67_cast_fp16)[name = string("op_1114_cast_fp16")];
            tensor<int32, [4]> var_1115 = const()[name = string("op_1115"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1117 = const()[name = string("op_1117"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1116_cast_fp16 = transpose(perm = var_1115, x = var_1114_cast_fp16)[name = string("transpose_955")];
            tensor<fp16, [1, 51, 128]> k_head_25_cast_fp16 = reshape(shape = var_1117, x = var_1116_cast_fp16)[name = string("k_head_25_cast_fp16")];
            string dense_output_69_pad_type_1 = const()[name = string("dense_output_69_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_69_strides_1 = const()[name = string("dense_output_69_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_69_pad_1 = const()[name = string("dense_output_69_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_69_dilations_1 = const()[name = string("dense_output_69_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_69_groups_1 = const()[name = string("dense_output_69_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18660800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18791936))))[name = string("layers_0_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_69_cast_fp16 = conv(dilations = dense_output_69_dilations_1, groups = dense_output_69_groups_1, pad = dense_output_69_pad_1, pad_type = dense_output_69_pad_type_1, strides = dense_output_69_strides_1, weight = layers_0_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_69_cast_fp16")];
            string sparse_output_69_pad_type_1 = const()[name = string("sparse_output_69_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_69_strides_1 = const()[name = string("sparse_output_69_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_69_pad_1 = const()[name = string("sparse_output_69_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_69_dilations_1 = const()[name = string("sparse_output_69_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_69_groups_1 = const()[name = string("sparse_output_69_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18795200))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18792512))))[name = string("layers_0_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_69_cast_fp16 = conv(dilations = sparse_output_69_dilations_1, groups = sparse_output_69_groups_1, pad = sparse_output_69_pad_1, pad_type = sparse_output_69_pad_type_1, strides = sparse_output_69_strides_1, weight = layers_0_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_69_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1133_cast_fp16 = add(x = dense_output_69_cast_fp16, y = sparse_output_69_cast_fp16)[name = string("op_1133_cast_fp16")];
            tensor<int32, [4]> var_1134 = const()[name = string("op_1134"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1136 = const()[name = string("op_1136"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1135_cast_fp16 = transpose(perm = var_1134, x = var_1133_cast_fp16)[name = string("transpose_954")];
            tensor<fp16, [1, 51, 128]> v_head_25_cast_fp16 = reshape(shape = var_1136, x = var_1135_cast_fp16)[name = string("v_head_25_cast_fp16")];
            string dense_output_71_pad_type_1 = const()[name = string("dense_output_71_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_71_strides_1 = const()[name = string("dense_output_71_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_71_pad_1 = const()[name = string("dense_output_71_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_71_dilations_1 = const()[name = string("dense_output_71_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_71_groups_1 = const()[name = string("dense_output_71_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18811648))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18942784))))[name = string("layers_0_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_71_cast_fp16 = conv(dilations = dense_output_71_dilations_1, groups = dense_output_71_groups_1, pad = dense_output_71_pad_1, pad_type = dense_output_71_pad_type_1, strides = dense_output_71_strides_1, weight = layers_0_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_71_cast_fp16")];
            string sparse_output_71_pad_type_1 = const()[name = string("sparse_output_71_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_71_strides_1 = const()[name = string("sparse_output_71_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_71_pad_1 = const()[name = string("sparse_output_71_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_71_dilations_1 = const()[name = string("sparse_output_71_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_71_groups_1 = const()[name = string("sparse_output_71_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18946048))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18943360))))[name = string("layers_0_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_71_cast_fp16 = conv(dilations = sparse_output_71_dilations_1, groups = sparse_output_71_groups_1, pad = sparse_output_71_pad_1, pad_type = sparse_output_71_pad_type_1, strides = sparse_output_71_strides_1, weight = layers_0_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_71_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_1152_cast_fp16 = add(x = dense_output_71_cast_fp16, y = sparse_output_71_cast_fp16)[name = string("op_1152_cast_fp16")];
            tensor<int32, [4]> var_1153 = const()[name = string("op_1153"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1155 = const()[name = string("op_1155"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_1154_cast_fp16 = transpose(perm = var_1153, x = var_1152_cast_fp16)[name = string("transpose_953")];
            tensor<fp16, [1, 101, 128]> p_head_25_cast_fp16 = reshape(shape = var_1155, x = var_1154_cast_fp16)[name = string("p_head_25_cast_fp16")];
            tensor<fp16, [128]> var_1157_to_fp16 = const()[name = string("op_1157_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18962496)))];
            tensor<fp16, [1, 51, 128]> var_1158_cast_fp16 = add(x = q_head_13_cast_fp16, y = var_1157_to_fp16)[name = string("op_1158_cast_fp16")];
            tensor<int32, [1]> q_u_13_axes_1 = const()[name = string("q_u_13_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_13_cast_fp16 = expand_dims(axes = q_u_13_axes_1, x = var_1158_cast_fp16)[name = string("q_u_13_cast_fp16")];
            tensor<fp16, [128]> var_1160_to_fp16 = const()[name = string("op_1160_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18962816)))];
            tensor<fp16, [1, 51, 128]> var_1161_cast_fp16 = add(x = q_head_13_cast_fp16, y = var_1160_to_fp16)[name = string("op_1161_cast_fp16")];
            tensor<int32, [1]> q_v_13_axes_1 = const()[name = string("q_v_13_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_13_cast_fp16 = expand_dims(axes = q_v_13_axes_1, x = var_1161_cast_fp16)[name = string("q_v_13_cast_fp16")];
            tensor<int32, [1]> k_head_27_axes_1 = const()[name = string("k_head_27_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_27_cast_fp16 = expand_dims(axes = k_head_27_axes_1, x = k_head_25_cast_fp16)[name = string("k_head_27_cast_fp16")];
            tensor<int32, [1]> v_head_27_axes_1 = const()[name = string("v_head_27_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_27_cast_fp16 = expand_dims(axes = v_head_27_axes_1, x = v_head_25_cast_fp16)[name = string("v_head_27_cast_fp16")];
            tensor<int32, [1]> p_head_27_axes_1 = const()[name = string("p_head_27_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_27_cast_fp16 = expand_dims(axes = p_head_27_axes_1, x = p_head_25_cast_fp16)[name = string("p_head_27_cast_fp16")];
            bool var_1167_transpose_x_3 = const()[name = string("op_1167_transpose_x_3"), val = bool(false)];
            bool var_1167_transpose_y_3 = const()[name = string("op_1167_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_1167_cast_fp16 = matmul(transpose_x = var_1167_transpose_x_3, transpose_y = var_1167_transpose_y_3, x = q_u_13_cast_fp16, y = k_head_27_cast_fp16)[name = string("op_1167_cast_fp16")];
            fp16 var_1168_to_fp16 = const()[name = string("op_1168_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_13_cast_fp16 = mul(x = var_1167_cast_fp16, y = var_1168_to_fp16)[name = string("scores_content_13_cast_fp16")];
            bool x_69_transpose_x_3 = const()[name = string("x_69_transpose_x_3"), val = bool(false)];
            bool x_69_transpose_y_3 = const()[name = string("x_69_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_69_cast_fp16 = matmul(transpose_x = x_69_transpose_x_3, transpose_y = x_69_transpose_y_3, x = q_v_13_cast_fp16, y = p_head_27_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [8]> x_71_pad_1 = const()[name = string("x_71_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_71_mode_1 = const()[name = string("x_71_mode_1"), val = string("constant")];
            fp16 const_1345_to_fp16 = const()[name = string("const_1345_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_71_cast_fp16 = pad(constant_val = const_1345_to_fp16, mode = x_71_mode_1, pad = x_71_pad_1, x = x_69_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<int32, [4]> var_1182 = const()[name = string("op_1182"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_73_cast_fp16 = reshape(shape = var_1182, x = x_71_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_1186_begin_1 = const()[name = string("op_1186_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1186_end_1 = const()[name = string("op_1186_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_1186_end_mask_1 = const()[name = string("op_1186_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_1186_cast_fp16 = slice_by_index(begin = var_1186_begin_1, end = var_1186_end_1, end_mask = var_1186_end_mask_1, x = x_73_cast_fp16)[name = string("op_1186_cast_fp16")];
            tensor<int32, [4]> var_1188 = const()[name = string("op_1188"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_1189_cast_fp16 = reshape(shape = var_1188, x = var_1186_cast_fp16)[name = string("op_1189_cast_fp16")];
            tensor<int32, [4]> var_1194_begin_1 = const()[name = string("op_1194_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1194_end_1 = const()[name = string("op_1194_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_1194_end_mask_1 = const()[name = string("op_1194_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_1194_cast_fp16 = slice_by_index(begin = var_1194_begin_1, end = var_1194_end_1, end_mask = var_1194_end_mask_1, x = var_1189_cast_fp16)[name = string("op_1194_cast_fp16")];
            fp16 var_1195_to_fp16 = const()[name = string("op_1195_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_13_cast_fp16 = mul(x = var_1194_cast_fp16, y = var_1195_to_fp16)[name = string("scores_pos_13_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_13_cast_fp16 = add(x = scores_content_13_cast_fp16, y = scores_pos_13_cast_fp16)[name = string("logits_13_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_1198_cast_fp16 = softmax(axis = var_197, x = logits_13_cast_fp16)[name = string("op_1198_cast_fp16")];
            bool var_1200_transpose_x_1 = const()[name = string("op_1200_transpose_x_1"), val = bool(false)];
            bool var_1200_transpose_y_1 = const()[name = string("op_1200_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_1200_cast_fp16 = matmul(transpose_x = var_1200_transpose_x_1, transpose_y = var_1200_transpose_y_1, x = var_1198_cast_fp16, y = v_head_27_cast_fp16)[name = string("op_1200_cast_fp16")];
            tensor<int32, [1]> var_1201_axes_1 = const()[name = string("op_1201_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_1201_cast_fp16 = squeeze(axes = var_1201_axes_1, x = var_1200_cast_fp16)[name = string("op_1201_cast_fp16")];
            string dense_output_73_pad_type_1 = const()[name = string("dense_output_73_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_73_strides_1 = const()[name = string("dense_output_73_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_73_pad_1 = const()[name = string("dense_output_73_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_73_dilations_1 = const()[name = string("dense_output_73_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_73_groups_1 = const()[name = string("dense_output_73_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18963136))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19094272))))[name = string("layers_0_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_73_cast_fp16 = conv(dilations = dense_output_73_dilations_1, groups = dense_output_73_groups_1, pad = dense_output_73_pad_1, pad_type = dense_output_73_pad_type_1, strides = dense_output_73_strides_1, weight = layers_0_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_73_cast_fp16")];
            string sparse_output_73_pad_type_1 = const()[name = string("sparse_output_73_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_73_strides_1 = const()[name = string("sparse_output_73_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_73_pad_1 = const()[name = string("sparse_output_73_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_73_dilations_1 = const()[name = string("sparse_output_73_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_73_groups_1 = const()[name = string("sparse_output_73_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19097536))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19094848))))[name = string("layers_0_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_73_cast_fp16 = conv(dilations = sparse_output_73_dilations_1, groups = sparse_output_73_groups_1, pad = sparse_output_73_pad_1, pad_type = sparse_output_73_pad_type_1, strides = sparse_output_73_strides_1, weight = layers_0_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_73_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1216_cast_fp16 = add(x = dense_output_73_cast_fp16, y = sparse_output_73_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> var_1217 = const()[name = string("op_1217"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1219 = const()[name = string("op_1219"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1218_cast_fp16 = transpose(perm = var_1217, x = var_1216_cast_fp16)[name = string("transpose_952")];
            tensor<fp16, [1, 51, 128]> q_head_15_cast_fp16 = reshape(shape = var_1219, x = var_1218_cast_fp16)[name = string("q_head_15_cast_fp16")];
            string dense_output_75_pad_type_1 = const()[name = string("dense_output_75_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_75_strides_1 = const()[name = string("dense_output_75_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_75_pad_1 = const()[name = string("dense_output_75_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_75_dilations_1 = const()[name = string("dense_output_75_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_75_groups_1 = const()[name = string("dense_output_75_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19113984))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19245120))))[name = string("layers_0_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_75_cast_fp16 = conv(dilations = dense_output_75_dilations_1, groups = dense_output_75_groups_1, pad = dense_output_75_pad_1, pad_type = dense_output_75_pad_type_1, strides = dense_output_75_strides_1, weight = layers_0_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_75_cast_fp16")];
            string sparse_output_75_pad_type_1 = const()[name = string("sparse_output_75_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_75_strides_1 = const()[name = string("sparse_output_75_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_75_pad_1 = const()[name = string("sparse_output_75_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_75_dilations_1 = const()[name = string("sparse_output_75_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_75_groups_1 = const()[name = string("sparse_output_75_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19248384))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19245696))))[name = string("layers_0_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_75_cast_fp16 = conv(dilations = sparse_output_75_dilations_1, groups = sparse_output_75_groups_1, pad = sparse_output_75_pad_1, pad_type = sparse_output_75_pad_type_1, strides = sparse_output_75_strides_1, weight = layers_0_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_75_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1235_cast_fp16 = add(x = dense_output_75_cast_fp16, y = sparse_output_75_cast_fp16)[name = string("op_1235_cast_fp16")];
            tensor<int32, [4]> var_1236 = const()[name = string("op_1236"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1238 = const()[name = string("op_1238"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1237_cast_fp16 = transpose(perm = var_1236, x = var_1235_cast_fp16)[name = string("transpose_951")];
            tensor<fp16, [1, 51, 128]> k_head_29_cast_fp16 = reshape(shape = var_1238, x = var_1237_cast_fp16)[name = string("k_head_29_cast_fp16")];
            string dense_output_77_pad_type_1 = const()[name = string("dense_output_77_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_77_strides_1 = const()[name = string("dense_output_77_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_77_pad_1 = const()[name = string("dense_output_77_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_77_dilations_1 = const()[name = string("dense_output_77_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_77_groups_1 = const()[name = string("dense_output_77_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19264832))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19395968))))[name = string("layers_0_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_77_cast_fp16 = conv(dilations = dense_output_77_dilations_1, groups = dense_output_77_groups_1, pad = dense_output_77_pad_1, pad_type = dense_output_77_pad_type_1, strides = dense_output_77_strides_1, weight = layers_0_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("dense_output_77_cast_fp16")];
            string sparse_output_77_pad_type_1 = const()[name = string("sparse_output_77_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_77_strides_1 = const()[name = string("sparse_output_77_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_77_pad_1 = const()[name = string("sparse_output_77_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_77_dilations_1 = const()[name = string("sparse_output_77_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_77_groups_1 = const()[name = string("sparse_output_77_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19399232))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19396544))))[name = string("layers_0_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_77_cast_fp16 = conv(dilations = sparse_output_77_dilations_1, groups = sparse_output_77_groups_1, pad = sparse_output_77_pad_1, pad_type = sparse_output_77_pad_type_1, strides = sparse_output_77_strides_1, weight = layers_0_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = string("sparse_output_77_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1254_cast_fp16 = add(x = dense_output_77_cast_fp16, y = sparse_output_77_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<int32, [4]> var_1255 = const()[name = string("op_1255"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1257 = const()[name = string("op_1257"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1256_cast_fp16 = transpose(perm = var_1255, x = var_1254_cast_fp16)[name = string("transpose_950")];
            tensor<fp16, [1, 51, 128]> v_head_29_cast_fp16 = reshape(shape = var_1257, x = var_1256_cast_fp16)[name = string("v_head_29_cast_fp16")];
            string dense_output_79_pad_type_1 = const()[name = string("dense_output_79_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_79_strides_1 = const()[name = string("dense_output_79_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_79_pad_1 = const()[name = string("dense_output_79_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_79_dilations_1 = const()[name = string("dense_output_79_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_79_groups_1 = const()[name = string("dense_output_79_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19415680))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19546816))))[name = string("layers_0_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_79_cast_fp16 = conv(dilations = dense_output_79_dilations_1, groups = dense_output_79_groups_1, pad = dense_output_79_pad_1, pad_type = dense_output_79_pad_type_1, strides = dense_output_79_strides_1, weight = layers_0_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_79_cast_fp16")];
            string sparse_output_79_pad_type_1 = const()[name = string("sparse_output_79_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_79_strides_1 = const()[name = string("sparse_output_79_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_79_pad_1 = const()[name = string("sparse_output_79_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_79_dilations_1 = const()[name = string("sparse_output_79_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_79_groups_1 = const()[name = string("sparse_output_79_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_0_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19550080))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19547392))))[name = string("layers_0_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_79_cast_fp16 = conv(dilations = sparse_output_79_dilations_1, groups = sparse_output_79_groups_1, pad = sparse_output_79_pad_1, pad_type = sparse_output_79_pad_type_1, strides = sparse_output_79_strides_1, weight = layers_0_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_79_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_1273_cast_fp16 = add(x = dense_output_79_cast_fp16, y = sparse_output_79_cast_fp16)[name = string("op_1273_cast_fp16")];
            tensor<int32, [4]> var_1274 = const()[name = string("op_1274"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1276 = const()[name = string("op_1276"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_1275_cast_fp16 = transpose(perm = var_1274, x = var_1273_cast_fp16)[name = string("transpose_949")];
            tensor<fp16, [1, 101, 128]> p_head_29_cast_fp16 = reshape(shape = var_1276, x = var_1275_cast_fp16)[name = string("p_head_29_cast_fp16")];
            tensor<fp16, [128]> var_1278_to_fp16 = const()[name = string("op_1278_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19566528)))];
            tensor<fp16, [1, 51, 128]> var_1279_cast_fp16 = add(x = q_head_15_cast_fp16, y = var_1278_to_fp16)[name = string("op_1279_cast_fp16")];
            tensor<int32, [1]> q_u_15_axes_1 = const()[name = string("q_u_15_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_15_cast_fp16 = expand_dims(axes = q_u_15_axes_1, x = var_1279_cast_fp16)[name = string("q_u_15_cast_fp16")];
            tensor<fp16, [128]> var_1281_to_fp16 = const()[name = string("op_1281_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19566848)))];
            tensor<fp16, [1, 51, 128]> var_1282_cast_fp16 = add(x = q_head_15_cast_fp16, y = var_1281_to_fp16)[name = string("op_1282_cast_fp16")];
            tensor<int32, [1]> q_v_15_axes_1 = const()[name = string("q_v_15_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_15_cast_fp16 = expand_dims(axes = q_v_15_axes_1, x = var_1282_cast_fp16)[name = string("q_v_15_cast_fp16")];
            tensor<int32, [1]> k_head_31_axes_1 = const()[name = string("k_head_31_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_31_cast_fp16 = expand_dims(axes = k_head_31_axes_1, x = k_head_29_cast_fp16)[name = string("k_head_31_cast_fp16")];
            tensor<int32, [1]> v_head_31_axes_1 = const()[name = string("v_head_31_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_31_cast_fp16 = expand_dims(axes = v_head_31_axes_1, x = v_head_29_cast_fp16)[name = string("v_head_31_cast_fp16")];
            tensor<int32, [1]> p_head_31_axes_1 = const()[name = string("p_head_31_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_31_cast_fp16 = expand_dims(axes = p_head_31_axes_1, x = p_head_29_cast_fp16)[name = string("p_head_31_cast_fp16")];
            bool var_1288_transpose_x_3 = const()[name = string("op_1288_transpose_x_3"), val = bool(false)];
            bool var_1288_transpose_y_3 = const()[name = string("op_1288_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_1288_cast_fp16 = matmul(transpose_x = var_1288_transpose_x_3, transpose_y = var_1288_transpose_y_3, x = q_u_15_cast_fp16, y = k_head_31_cast_fp16)[name = string("op_1288_cast_fp16")];
            fp16 var_1289_to_fp16 = const()[name = string("op_1289_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_15_cast_fp16 = mul(x = var_1288_cast_fp16, y = var_1289_to_fp16)[name = string("scores_content_15_cast_fp16")];
            bool x_77_transpose_x_3 = const()[name = string("x_77_transpose_x_3"), val = bool(false)];
            bool x_77_transpose_y_3 = const()[name = string("x_77_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_77_cast_fp16 = matmul(transpose_x = x_77_transpose_x_3, transpose_y = x_77_transpose_y_3, x = q_v_15_cast_fp16, y = p_head_31_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<int32, [8]> x_79_pad_1 = const()[name = string("x_79_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_79_mode_1 = const()[name = string("x_79_mode_1"), val = string("constant")];
            fp16 const_1351_to_fp16 = const()[name = string("const_1351_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_79_cast_fp16 = pad(constant_val = const_1351_to_fp16, mode = x_79_mode_1, pad = x_79_pad_1, x = x_77_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<int32, [4]> var_1303 = const()[name = string("op_1303"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_81_cast_fp16 = reshape(shape = var_1303, x = x_79_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> var_1307_begin_1 = const()[name = string("op_1307_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1307_end_1 = const()[name = string("op_1307_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_1307_end_mask_1 = const()[name = string("op_1307_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_1307_cast_fp16 = slice_by_index(begin = var_1307_begin_1, end = var_1307_end_1, end_mask = var_1307_end_mask_1, x = x_81_cast_fp16)[name = string("op_1307_cast_fp16")];
            tensor<int32, [4]> var_1309 = const()[name = string("op_1309"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_1310_cast_fp16 = reshape(shape = var_1309, x = var_1307_cast_fp16)[name = string("op_1310_cast_fp16")];
            tensor<int32, [4]> var_1315_begin_1 = const()[name = string("op_1315_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1315_end_1 = const()[name = string("op_1315_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_1315_end_mask_1 = const()[name = string("op_1315_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_1315_cast_fp16 = slice_by_index(begin = var_1315_begin_1, end = var_1315_end_1, end_mask = var_1315_end_mask_1, x = var_1310_cast_fp16)[name = string("op_1315_cast_fp16")];
            fp16 var_1316_to_fp16 = const()[name = string("op_1316_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_15_cast_fp16 = mul(x = var_1315_cast_fp16, y = var_1316_to_fp16)[name = string("scores_pos_15_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_15_cast_fp16 = add(x = scores_content_15_cast_fp16, y = scores_pos_15_cast_fp16)[name = string("logits_15_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_1319_cast_fp16 = softmax(axis = var_197, x = logits_15_cast_fp16)[name = string("op_1319_cast_fp16")];
            bool var_1321_transpose_x_1 = const()[name = string("op_1321_transpose_x_1"), val = bool(false)];
            bool var_1321_transpose_y_1 = const()[name = string("op_1321_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_1321_cast_fp16 = matmul(transpose_x = var_1321_transpose_x_1, transpose_y = var_1321_transpose_y_1, x = var_1319_cast_fp16, y = v_head_31_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<int32, [1]> o_head_1_axes_1 = const()[name = string("o_head_1_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_1_cast_fp16 = squeeze(axes = o_head_1_axes_1, x = var_1321_cast_fp16)[name = string("o_head_1_cast_fp16")];
            bool out_1_interleave_1 = const()[name = string("out_1_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_1_cast_fp16 = concat(axis = var_197, interleave = out_1_interleave_1, values = (var_475_cast_fp16, var_596_cast_fp16, var_717_cast_fp16, var_838_cast_fp16, var_959_cast_fp16, var_1080_cast_fp16, var_1201_cast_fp16, o_head_1_cast_fp16))[name = string("out_1_cast_fp16")];
            tensor<int32, [3]> var_1325_perm_1 = const()[name = string("op_1325_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_37_axes_1 = const()[name = string("input_37_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_1325_cast_fp16 = transpose(perm = var_1325_perm_1, x = out_1_cast_fp16)[name = string("transpose_948")];
            tensor<fp16, [1, 1024, 51, 1]> input_37_cast_fp16 = expand_dims(axes = input_37_axes_1, x = var_1325_cast_fp16)[name = string("input_37_cast_fp16")];
            string dense_output_81_pad_type_1 = const()[name = string("dense_output_81_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_81_strides_1 = const()[name = string("dense_output_81_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_81_pad_1 = const()[name = string("dense_output_81_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_81_dilations_1 = const()[name = string("dense_output_81_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_81_groups_1 = const()[name = string("dense_output_81_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19567168))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20615808))))[name = string("layers_0_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_81_cast_fp16 = conv(dilations = dense_output_81_dilations_1, groups = dense_output_81_groups_1, pad = dense_output_81_pad_1, pad_type = dense_output_81_pad_type_1, strides = dense_output_81_strides_1, weight = layers_0_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = string("dense_output_81_cast_fp16")];
            string sparse_output_81_pad_type_1 = const()[name = string("sparse_output_81_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_81_strides_1 = const()[name = string("sparse_output_81_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_81_pad_1 = const()[name = string("sparse_output_81_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_81_dilations_1 = const()[name = string("sparse_output_81_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_81_groups_1 = const()[name = string("sparse_output_81_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20637440))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20616384))))[name = string("layers_0_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_81_cast_fp16 = conv(dilations = sparse_output_81_dilations_1, groups = sparse_output_81_groups_1, pad = sparse_output_81_pad_1, pad_type = sparse_output_81_pad_type_1, strides = sparse_output_81_strides_1, weight = layers_0_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_37_cast_fp16)[name = string("sparse_output_81_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_1_cast_fp16 = add(x = dense_output_81_cast_fp16, y = sparse_output_81_cast_fp16)[name = string("out_conv_1_cast_fp16")];
            tensor<int32, [1]> var_1342_axes_1 = const()[name = string("op_1342_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_1342_cast_fp16 = squeeze(axes = var_1342_axes_1, x = out_conv_1_cast_fp16)[name = string("op_1342_cast_fp16")];
            tensor<int32, [3]> var_1343_perm_1 = const()[name = string("op_1343_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_1343_cast_fp16 = transpose(perm = var_1343_perm_1, x = var_1342_cast_fp16)[name = string("transpose_947")];
            tensor<fp16, [1, 51, 1024]> input_39_cast_fp16 = add(x = input_27_cast_fp16, y = var_1343_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<int32, [1]> x_85_axes_1 = const()[name = string("x_85_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_0_norm_conv_weight_to_fp16 = const()[name = string("layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20768576)))];
            tensor<fp16, [1024]> layers_0_norm_conv_bias_to_fp16 = const()[name = string("layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20770688)))];
            tensor<fp16, [1, 51, 1024]> x_85_cast_fp16 = layer_norm(axes = x_85_axes_1, beta = layers_0_norm_conv_bias_to_fp16, epsilon = var_212_to_fp16, gamma = layers_0_norm_conv_weight_to_fp16, x = input_39_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<int32, [3]> var_1353_perm_1 = const()[name = string("op_1353_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_41_axes_1 = const()[name = string("input_41_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_1353_cast_fp16 = transpose(perm = var_1353_perm_1, x = x_85_cast_fp16)[name = string("transpose_946")];
            tensor<fp16, [1, 1024, 51, 1]> input_41_cast_fp16 = expand_dims(axes = input_41_axes_1, x = var_1353_cast_fp16)[name = string("input_41_cast_fp16")];
            string dense_output_83_pad_type_1 = const()[name = string("dense_output_83_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_83_strides_1 = const()[name = string("dense_output_83_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_83_pad_1 = const()[name = string("dense_output_83_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_83_dilations_1 = const()[name = string("dense_output_83_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_83_groups_1 = const()[name = string("dense_output_83_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_0_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20772800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22870016))))[name = string("layers_0_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_83_cast_fp16 = conv(dilations = dense_output_83_dilations_1, groups = dense_output_83_groups_1, pad = dense_output_83_pad_1, pad_type = dense_output_83_pad_type_1, strides = dense_output_83_strides_1, weight = layers_0_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = string("dense_output_83_cast_fp16")];
            string sparse_output_83_pad_type_1 = const()[name = string("sparse_output_83_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_83_strides_1 = const()[name = string("sparse_output_83_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_83_pad_1 = const()[name = string("sparse_output_83_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_83_dilations_1 = const()[name = string("sparse_output_83_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_83_groups_1 = const()[name = string("sparse_output_83_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_0_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22912640))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22870592))))[name = string("layers_0_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_83_cast_fp16 = conv(dilations = sparse_output_83_dilations_1, groups = sparse_output_83_groups_1, pad = sparse_output_83_pad_1, pad_type = sparse_output_83_pad_type_1, strides = sparse_output_83_strides_1, weight = layers_0_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_41_cast_fp16)[name = string("sparse_output_83_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_43_cast_fp16 = add(x = dense_output_83_cast_fp16, y = sparse_output_83_cast_fp16)[name = string("input_43_cast_fp16")];
            int32 input_45_split_num_splits_1 = const()[name = string("input_45_split_num_splits_1"), val = int32(2)];
            int32 input_45_split_axis_1 = const()[name = string("input_45_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_45_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_45_split_cast_fp16_1 = split(axis = input_45_split_axis_1, num_splits = input_45_split_num_splits_1, x = input_43_cast_fp16)[name = string("input_45_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_45_split_1_sigmoid_cast_fp16 = sigmoid(x = input_45_split_cast_fp16_1)[name = string("input_45_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_45_cast_fp16 = mul(x = input_45_split_cast_fp16_0, y = input_45_split_1_sigmoid_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<int32, [8]> input_47_pad_1 = const()[name = string("input_47_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_47_mode_1 = const()[name = string("input_47_mode_1"), val = string("constant")];
            fp16 const_1353_to_fp16 = const()[name = string("const_1353_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_47_cast_fp16 = pad(constant_val = const_1353_to_fp16, mode = input_47_mode_1, pad = input_47_pad_1, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
            string dense_output_85_pad_type_1 = const()[name = string("dense_output_85_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_85_strides_1 = const()[name = string("dense_output_85_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_85_pad_1 = const()[name = string("dense_output_85_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_85_dilations_1 = const()[name = string("dense_output_85_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_85_groups_1 = const()[name = string("dense_output_85_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23174848))), nonzero_data = tensor<fp16, [0]>([]))[name = string("layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_85_cast_fp16 = conv(dilations = dense_output_85_dilations_1, groups = dense_output_85_groups_1, pad = dense_output_85_pad_1, pad_type = dense_output_85_pad_type_1, strides = dense_output_85_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_47_cast_fp16)[name = string("dense_output_85_cast_fp16")];
            string sparse_output_85_pad_type_1 = const()[name = string("sparse_output_85_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_85_strides_1 = const()[name = string("sparse_output_85_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_85_pad_1 = const()[name = string("sparse_output_85_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_85_dilations_1 = const()[name = string("sparse_output_85_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_85_groups_1 = const()[name = string("sparse_output_85_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_0_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24354560))))[name = string("layers_0_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_85_cast_fp16 = conv(dilations = sparse_output_85_dilations_1, groups = sparse_output_85_groups_1, pad = sparse_output_85_pad_1, pad_type = sparse_output_85_pad_type_1, strides = sparse_output_85_strides_1, weight = layers_0_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_47_cast_fp16)[name = string("sparse_output_85_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_49_cast_fp16 = add(x = dense_output_85_cast_fp16, y = sparse_output_85_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024]> layers_0_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_0_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25552768)))];
            tensor<fp16, [1024]> layers_0_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_0_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25554880)))];
            tensor<fp16, [1024]> layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25556992)))];
            tensor<fp16, [1024]> layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25559104)))];
            tensor<fp16, [1, 1024, 51, 1]> input_51_cast_fp16 = batch_norm(beta = layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_212_to_fp16, gamma = layers_0_conv_batch_norm_weight_to_fp16, mean = layers_0_conv_batch_norm_running_mean_to_fp16, variance = layers_0_conv_batch_norm_running_var_to_fp16, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_53_cast_fp16 = silu(x = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
            string dense_output_87_pad_type_1 = const()[name = string("dense_output_87_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_87_strides_1 = const()[name = string("dense_output_87_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_87_pad_1 = const()[name = string("dense_output_87_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_87_dilations_1 = const()[name = string("dense_output_87_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_87_groups_1 = const()[name = string("dense_output_87_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25561216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26609856))))[name = string("layers_0_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_87_cast_fp16 = conv(dilations = dense_output_87_dilations_1, groups = dense_output_87_groups_1, pad = dense_output_87_pad_1, pad_type = dense_output_87_pad_type_1, strides = dense_output_87_strides_1, weight = layers_0_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_53_cast_fp16)[name = string("dense_output_87_cast_fp16")];
            string sparse_output_87_pad_type_1 = const()[name = string("sparse_output_87_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_87_strides_1 = const()[name = string("sparse_output_87_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_87_pad_1 = const()[name = string("sparse_output_87_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_87_dilations_1 = const()[name = string("sparse_output_87_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_87_groups_1 = const()[name = string("sparse_output_87_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_0_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26631488))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26610432))))[name = string("layers_0_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_87_cast_fp16 = conv(dilations = sparse_output_87_dilations_1, groups = sparse_output_87_groups_1, pad = sparse_output_87_pad_1, pad_type = sparse_output_87_pad_type_1, strides = sparse_output_87_strides_1, weight = layers_0_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_53_cast_fp16)[name = string("sparse_output_87_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_87_cast_fp16 = add(x = dense_output_87_cast_fp16, y = sparse_output_87_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<int32, [1]> var_1409_axes_1 = const()[name = string("op_1409_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_1409_cast_fp16 = squeeze(axes = var_1409_axes_1, x = x_87_cast_fp16)[name = string("op_1409_cast_fp16")];
            tensor<int32, [3]> var_1410_perm_1 = const()[name = string("op_1410_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_1410_cast_fp16 = transpose(perm = var_1410_perm_1, x = var_1409_cast_fp16)[name = string("transpose_945")];
            tensor<fp16, [1, 51, 1024]> input_55_cast_fp16 = add(x = input_39_cast_fp16, y = var_1410_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<int32, [1]> x_89_axes_1 = const()[name = string("x_89_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26762624)))];
            tensor<fp16, [1024]> layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26764736)))];
            tensor<fp16, [1, 51, 1024]> x_89_cast_fp16 = layer_norm(axes = x_89_axes_1, beta = layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_212_to_fp16, gamma = layers_0_norm_feed_forward2_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [4]> var_1420 = const()[name = string("op_1420"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_91_cast_fp16 = reshape(shape = var_1420, x = x_89_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<int32, [4]> input_57_perm_1 = const()[name = string("input_57_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_89_pad_type_1 = const()[name = string("dense_output_89_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_89_strides_1 = const()[name = string("dense_output_89_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_89_pad_1 = const()[name = string("dense_output_89_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_89_dilations_1 = const()[name = string("dense_output_89_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_89_groups_1 = const()[name = string("dense_output_89_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_0_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26766848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30961216))))[name = string("layers_0_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_57_cast_fp16 = transpose(perm = input_57_perm_1, x = x_91_cast_fp16)[name = string("transpose_944")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_89_cast_fp16 = conv(dilations = dense_output_89_dilations_1, groups = dense_output_89_groups_1, pad = dense_output_89_pad_1, pad_type = dense_output_89_pad_type_1, strides = dense_output_89_strides_1, weight = layers_0_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = string("dense_output_89_cast_fp16")];
            string sparse_output_89_pad_type_1 = const()[name = string("sparse_output_89_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_89_strides_1 = const()[name = string("sparse_output_89_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_89_pad_1 = const()[name = string("sparse_output_89_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_89_dilations_1 = const()[name = string("sparse_output_89_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_89_groups_1 = const()[name = string("sparse_output_89_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_0_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31045760))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30961792))))[name = string("layers_0_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_89_cast_fp16 = conv(dilations = sparse_output_89_dilations_1, groups = sparse_output_89_groups_1, pad = sparse_output_89_pad_1, pad_type = sparse_output_89_pad_type_1, strides = sparse_output_89_strides_1, weight = layers_0_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_57_cast_fp16)[name = string("sparse_output_89_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_59_cast_fp16 = add(x = dense_output_89_cast_fp16, y = sparse_output_89_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_61_cast_fp16 = silu(x = input_59_cast_fp16)[name = string("input_61_cast_fp16")];
            string dense_output_91_pad_type_1 = const()[name = string("dense_output_91_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_91_strides_1 = const()[name = string("dense_output_91_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_91_pad_1 = const()[name = string("dense_output_91_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_91_dilations_1 = const()[name = string("dense_output_91_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_91_groups_1 = const()[name = string("dense_output_91_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_0_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31570112))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35764480))))[name = string("layers_0_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_91_cast_fp16 = conv(dilations = dense_output_91_dilations_1, groups = dense_output_91_groups_1, pad = dense_output_91_pad_1, pad_type = dense_output_91_pad_type_1, strides = dense_output_91_strides_1, weight = layers_0_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_61_cast_fp16)[name = string("dense_output_91_cast_fp16")];
            string sparse_output_91_pad_type_1 = const()[name = string("sparse_output_91_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_91_strides_1 = const()[name = string("sparse_output_91_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_91_pad_1 = const()[name = string("sparse_output_91_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_91_dilations_1 = const()[name = string("sparse_output_91_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_91_groups_1 = const()[name = string("sparse_output_91_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_0_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35849024))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35765056))))[name = string("layers_0_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_91_cast_fp16 = conv(dilations = sparse_output_91_dilations_1, groups = sparse_output_91_groups_1, pad = sparse_output_91_pad_1, pad_type = sparse_output_91_pad_type_1, strides = sparse_output_91_strides_1, weight = layers_0_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_61_cast_fp16)[name = string("sparse_output_91_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_93_cast_fp16 = add(x = dense_output_91_cast_fp16, y = sparse_output_91_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<int32, [4]> x_95_perm_1 = const()[name = string("x_95_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_1455 = const()[name = string("op_1455"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_95_cast_fp16 = transpose(perm = x_95_perm_1, x = x_93_cast_fp16)[name = string("transpose_943")];
            tensor<fp16, [1, 51, 1024]> var_1456_cast_fp16 = reshape(shape = var_1455, x = x_95_cast_fp16)[name = string("op_1456_cast_fp16")];
            fp16 var_1457_to_fp16 = const()[name = string("op_1457_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_1458_cast_fp16 = mul(x = var_1456_cast_fp16, y = var_1457_to_fp16)[name = string("op_1458_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_63_cast_fp16 = add(x = input_55_cast_fp16, y = var_1458_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<int32, [1]> input_65_axes_1 = const()[name = string("input_65_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_0_norm_out_weight_to_fp16 = const()[name = string("layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36373376)))];
            tensor<fp16, [1024]> layers_0_norm_out_bias_to_fp16 = const()[name = string("layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36375488)))];
            tensor<fp16, [1, 51, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_1, beta = layers_0_norm_out_bias_to_fp16, epsilon = var_212_to_fp16, gamma = layers_0_norm_out_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")];
            int32 var_1466 = const()[name = string("op_1466"), val = int32(-1)];
            tensor<int32, [1]> x_97_axes_1 = const()[name = string("x_97_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36377600)))];
            tensor<fp16, [1024]> layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36379712)))];
            fp16 var_1481_to_fp16 = const()[name = string("op_1481_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_97_cast_fp16 = layer_norm(axes = x_97_axes_1, beta = layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_1481_to_fp16, gamma = layers_1_norm_feed_forward1_weight_to_fp16, x = input_65_cast_fp16)[name = string("x_97_cast_fp16")];
            tensor<int32, [4]> var_1500 = const()[name = string("op_1500"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_99_cast_fp16 = reshape(shape = var_1500, x = x_97_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<int32, [4]> input_67_perm_1 = const()[name = string("input_67_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_93_pad_type_1 = const()[name = string("dense_output_93_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_93_strides_1 = const()[name = string("dense_output_93_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_93_pad_1 = const()[name = string("dense_output_93_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_93_dilations_1 = const()[name = string("dense_output_93_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_93_groups_1 = const()[name = string("dense_output_93_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_1_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36381824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40576192))))[name = string("layers_1_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_67_cast_fp16 = transpose(perm = input_67_perm_1, x = x_99_cast_fp16)[name = string("transpose_942")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_93_cast_fp16 = conv(dilations = dense_output_93_dilations_1, groups = dense_output_93_groups_1, pad = dense_output_93_pad_1, pad_type = dense_output_93_pad_type_1, strides = dense_output_93_strides_1, weight = layers_1_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = string("dense_output_93_cast_fp16")];
            string sparse_output_93_pad_type_1 = const()[name = string("sparse_output_93_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_93_strides_1 = const()[name = string("sparse_output_93_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_93_pad_1 = const()[name = string("sparse_output_93_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_93_dilations_1 = const()[name = string("sparse_output_93_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_93_groups_1 = const()[name = string("sparse_output_93_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_1_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40660736))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40576768))))[name = string("layers_1_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_93_cast_fp16 = conv(dilations = sparse_output_93_dilations_1, groups = sparse_output_93_groups_1, pad = sparse_output_93_pad_1, pad_type = sparse_output_93_pad_type_1, strides = sparse_output_93_strides_1, weight = layers_1_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_67_cast_fp16)[name = string("sparse_output_93_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_69_cast_fp16 = add(x = dense_output_93_cast_fp16, y = sparse_output_93_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_71_cast_fp16 = silu(x = input_69_cast_fp16)[name = string("input_71_cast_fp16")];
            string dense_output_95_pad_type_1 = const()[name = string("dense_output_95_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_95_strides_1 = const()[name = string("dense_output_95_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_95_pad_1 = const()[name = string("dense_output_95_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_95_dilations_1 = const()[name = string("dense_output_95_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_95_groups_1 = const()[name = string("dense_output_95_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_1_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41185088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45379456))))[name = string("layers_1_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_95_cast_fp16 = conv(dilations = dense_output_95_dilations_1, groups = dense_output_95_groups_1, pad = dense_output_95_pad_1, pad_type = dense_output_95_pad_type_1, strides = dense_output_95_strides_1, weight = layers_1_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = string("dense_output_95_cast_fp16")];
            string sparse_output_95_pad_type_1 = const()[name = string("sparse_output_95_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_95_strides_1 = const()[name = string("sparse_output_95_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_95_pad_1 = const()[name = string("sparse_output_95_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_95_dilations_1 = const()[name = string("sparse_output_95_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_95_groups_1 = const()[name = string("sparse_output_95_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_1_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45464000))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45380032))))[name = string("layers_1_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_95_cast_fp16 = conv(dilations = sparse_output_95_dilations_1, groups = sparse_output_95_groups_1, pad = sparse_output_95_pad_1, pad_type = sparse_output_95_pad_type_1, strides = sparse_output_95_strides_1, weight = layers_1_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_71_cast_fp16)[name = string("sparse_output_95_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_101_cast_fp16 = add(x = dense_output_95_cast_fp16, y = sparse_output_95_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> x_103_perm_1 = const()[name = string("x_103_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_1535 = const()[name = string("op_1535"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_103_cast_fp16 = transpose(perm = x_103_perm_1, x = x_101_cast_fp16)[name = string("transpose_941")];
            tensor<fp16, [1, 51, 1024]> var_1536_cast_fp16 = reshape(shape = var_1535, x = x_103_cast_fp16)[name = string("op_1536_cast_fp16")];
            fp16 var_1537_to_fp16 = const()[name = string("op_1537_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_1538_cast_fp16 = mul(x = var_1536_cast_fp16, y = var_1537_to_fp16)[name = string("op_1538_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_73_cast_fp16 = add(x = input_65_cast_fp16, y = var_1538_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<int32, [1]> q_3_axes_1 = const()[name = string("q_3_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_1_norm_self_att_weight_to_fp16 = const()[name = string("layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45988352)))];
            tensor<fp16, [1024]> layers_1_norm_self_att_bias_to_fp16 = const()[name = string("layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45990464)))];
            tensor<fp16, [1, 51, 1024]> q_3_cast_fp16 = layer_norm(axes = q_3_axes_1, beta = layers_1_norm_self_att_bias_to_fp16, epsilon = var_1481_to_fp16, gamma = layers_1_norm_self_att_weight_to_fp16, x = input_73_cast_fp16)[name = string("q_3_cast_fp16")];
            tensor<int32, [3]> var_1612 = const()[name = string("op_1612"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_75_axes_1 = const()[name = string("input_75_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_1613_cast_fp16 = transpose(perm = var_1612, x = q_3_cast_fp16)[name = string("transpose_940")];
            tensor<fp16, [1, 1024, 51, 1]> input_75_cast_fp16 = expand_dims(axes = input_75_axes_1, x = var_1613_cast_fp16)[name = string("input_75_cast_fp16")];
            string dense_output_97_pad_type_1 = const()[name = string("dense_output_97_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_97_strides_1 = const()[name = string("dense_output_97_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_97_pad_1 = const()[name = string("dense_output_97_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_97_dilations_1 = const()[name = string("dense_output_97_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_97_groups_1 = const()[name = string("dense_output_97_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45992576))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46123712))))[name = string("layers_1_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_97_cast_fp16 = conv(dilations = dense_output_97_dilations_1, groups = dense_output_97_groups_1, pad = dense_output_97_pad_1, pad_type = dense_output_97_pad_type_1, strides = dense_output_97_strides_1, weight = layers_1_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_97_cast_fp16")];
            string sparse_output_97_pad_type_1 = const()[name = string("sparse_output_97_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_97_strides_1 = const()[name = string("sparse_output_97_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_97_pad_1 = const()[name = string("sparse_output_97_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_97_dilations_1 = const()[name = string("sparse_output_97_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_97_groups_1 = const()[name = string("sparse_output_97_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46126976))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46124288))))[name = string("layers_1_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_97_cast_fp16 = conv(dilations = sparse_output_97_dilations_1, groups = sparse_output_97_groups_1, pad = sparse_output_97_pad_1, pad_type = sparse_output_97_pad_type_1, strides = sparse_output_97_strides_1, weight = layers_1_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_97_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1638_cast_fp16 = add(x = dense_output_97_cast_fp16, y = sparse_output_97_cast_fp16)[name = string("op_1638_cast_fp16")];
            tensor<int32, [4]> var_1639 = const()[name = string("op_1639"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1641 = const()[name = string("op_1641"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1640_cast_fp16 = transpose(perm = var_1639, x = var_1638_cast_fp16)[name = string("transpose_939")];
            tensor<fp16, [1, 51, 128]> q_head_17_cast_fp16 = reshape(shape = var_1641, x = var_1640_cast_fp16)[name = string("q_head_17_cast_fp16")];
            string dense_output_99_pad_type_1 = const()[name = string("dense_output_99_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_99_strides_1 = const()[name = string("dense_output_99_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_99_pad_1 = const()[name = string("dense_output_99_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_99_dilations_1 = const()[name = string("dense_output_99_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_99_groups_1 = const()[name = string("dense_output_99_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46143424))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46274560))))[name = string("layers_1_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_99_cast_fp16 = conv(dilations = dense_output_99_dilations_1, groups = dense_output_99_groups_1, pad = dense_output_99_pad_1, pad_type = dense_output_99_pad_type_1, strides = dense_output_99_strides_1, weight = layers_1_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_99_cast_fp16")];
            string sparse_output_99_pad_type_1 = const()[name = string("sparse_output_99_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_99_strides_1 = const()[name = string("sparse_output_99_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_99_pad_1 = const()[name = string("sparse_output_99_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_99_dilations_1 = const()[name = string("sparse_output_99_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_99_groups_1 = const()[name = string("sparse_output_99_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46277824))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46275136))))[name = string("layers_1_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_99_cast_fp16 = conv(dilations = sparse_output_99_dilations_1, groups = sparse_output_99_groups_1, pad = sparse_output_99_pad_1, pad_type = sparse_output_99_pad_type_1, strides = sparse_output_99_strides_1, weight = layers_1_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_99_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1657_cast_fp16 = add(x = dense_output_99_cast_fp16, y = sparse_output_99_cast_fp16)[name = string("op_1657_cast_fp16")];
            tensor<int32, [4]> var_1658 = const()[name = string("op_1658"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1660 = const()[name = string("op_1660"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1659_cast_fp16 = transpose(perm = var_1658, x = var_1657_cast_fp16)[name = string("transpose_938")];
            tensor<fp16, [1, 51, 128]> k_head_33_cast_fp16 = reshape(shape = var_1660, x = var_1659_cast_fp16)[name = string("k_head_33_cast_fp16")];
            string dense_output_101_pad_type_1 = const()[name = string("dense_output_101_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_101_strides_1 = const()[name = string("dense_output_101_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_101_pad_1 = const()[name = string("dense_output_101_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_101_dilations_1 = const()[name = string("dense_output_101_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_101_groups_1 = const()[name = string("dense_output_101_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46294272))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46425408))))[name = string("layers_1_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_101_cast_fp16 = conv(dilations = dense_output_101_dilations_1, groups = dense_output_101_groups_1, pad = dense_output_101_pad_1, pad_type = dense_output_101_pad_type_1, strides = dense_output_101_strides_1, weight = layers_1_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_101_cast_fp16")];
            string sparse_output_101_pad_type_1 = const()[name = string("sparse_output_101_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_101_strides_1 = const()[name = string("sparse_output_101_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_101_pad_1 = const()[name = string("sparse_output_101_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_101_dilations_1 = const()[name = string("sparse_output_101_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_101_groups_1 = const()[name = string("sparse_output_101_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46428672))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46425984))))[name = string("layers_1_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_101_cast_fp16 = conv(dilations = sparse_output_101_dilations_1, groups = sparse_output_101_groups_1, pad = sparse_output_101_pad_1, pad_type = sparse_output_101_pad_type_1, strides = sparse_output_101_strides_1, weight = layers_1_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_101_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1676_cast_fp16 = add(x = dense_output_101_cast_fp16, y = sparse_output_101_cast_fp16)[name = string("op_1676_cast_fp16")];
            tensor<int32, [4]> var_1677 = const()[name = string("op_1677"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1679 = const()[name = string("op_1679"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1678_cast_fp16 = transpose(perm = var_1677, x = var_1676_cast_fp16)[name = string("transpose_937")];
            tensor<fp16, [1, 51, 128]> v_head_33_cast_fp16 = reshape(shape = var_1679, x = var_1678_cast_fp16)[name = string("v_head_33_cast_fp16")];
            string dense_output_103_pad_type_1 = const()[name = string("dense_output_103_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_103_strides_1 = const()[name = string("dense_output_103_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_103_pad_1 = const()[name = string("dense_output_103_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_103_dilations_1 = const()[name = string("dense_output_103_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_103_groups_1 = const()[name = string("dense_output_103_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46445120))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46576256))))[name = string("layers_1_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_103_cast_fp16 = conv(dilations = dense_output_103_dilations_1, groups = dense_output_103_groups_1, pad = dense_output_103_pad_1, pad_type = dense_output_103_pad_type_1, strides = dense_output_103_strides_1, weight = layers_1_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_103_cast_fp16")];
            string sparse_output_103_pad_type_1 = const()[name = string("sparse_output_103_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_103_strides_1 = const()[name = string("sparse_output_103_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_103_pad_1 = const()[name = string("sparse_output_103_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_103_dilations_1 = const()[name = string("sparse_output_103_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_103_groups_1 = const()[name = string("sparse_output_103_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46579520))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46576832))))[name = string("layers_1_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_103_cast_fp16 = conv(dilations = sparse_output_103_dilations_1, groups = sparse_output_103_groups_1, pad = sparse_output_103_pad_1, pad_type = sparse_output_103_pad_type_1, strides = sparse_output_103_strides_1, weight = layers_1_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_103_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_1695_cast_fp16 = add(x = dense_output_103_cast_fp16, y = sparse_output_103_cast_fp16)[name = string("op_1695_cast_fp16")];
            tensor<int32, [4]> var_1696 = const()[name = string("op_1696"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1698 = const()[name = string("op_1698"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_1697_cast_fp16 = transpose(perm = var_1696, x = var_1695_cast_fp16)[name = string("transpose_936")];
            tensor<fp16, [1, 101, 128]> p_head_33_cast_fp16 = reshape(shape = var_1698, x = var_1697_cast_fp16)[name = string("p_head_33_cast_fp16")];
            tensor<fp16, [128]> var_1700_to_fp16 = const()[name = string("op_1700_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46595968)))];
            tensor<fp16, [1, 51, 128]> var_1701_cast_fp16 = add(x = q_head_17_cast_fp16, y = var_1700_to_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [1]> q_u_17_axes_1 = const()[name = string("q_u_17_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_17_cast_fp16 = expand_dims(axes = q_u_17_axes_1, x = var_1701_cast_fp16)[name = string("q_u_17_cast_fp16")];
            tensor<fp16, [128]> var_1703_to_fp16 = const()[name = string("op_1703_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46596288)))];
            tensor<fp16, [1, 51, 128]> var_1704_cast_fp16 = add(x = q_head_17_cast_fp16, y = var_1703_to_fp16)[name = string("op_1704_cast_fp16")];
            tensor<int32, [1]> q_v_17_axes_1 = const()[name = string("q_v_17_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_17_cast_fp16 = expand_dims(axes = q_v_17_axes_1, x = var_1704_cast_fp16)[name = string("q_v_17_cast_fp16")];
            tensor<int32, [1]> k_head_35_axes_1 = const()[name = string("k_head_35_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_35_cast_fp16 = expand_dims(axes = k_head_35_axes_1, x = k_head_33_cast_fp16)[name = string("k_head_35_cast_fp16")];
            tensor<int32, [1]> v_head_35_axes_1 = const()[name = string("v_head_35_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_35_cast_fp16 = expand_dims(axes = v_head_35_axes_1, x = v_head_33_cast_fp16)[name = string("v_head_35_cast_fp16")];
            tensor<int32, [1]> p_head_35_axes_1 = const()[name = string("p_head_35_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_35_cast_fp16 = expand_dims(axes = p_head_35_axes_1, x = p_head_33_cast_fp16)[name = string("p_head_35_cast_fp16")];
            bool var_1710_transpose_x_3 = const()[name = string("op_1710_transpose_x_3"), val = bool(false)];
            bool var_1710_transpose_y_3 = const()[name = string("op_1710_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_1710_cast_fp16 = matmul(transpose_x = var_1710_transpose_x_3, transpose_y = var_1710_transpose_y_3, x = q_u_17_cast_fp16, y = k_head_35_cast_fp16)[name = string("op_1710_cast_fp16")];
            fp16 var_1711_to_fp16 = const()[name = string("op_1711_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_17_cast_fp16 = mul(x = var_1710_cast_fp16, y = var_1711_to_fp16)[name = string("scores_content_17_cast_fp16")];
            bool x_105_transpose_x_3 = const()[name = string("x_105_transpose_x_3"), val = bool(false)];
            bool x_105_transpose_y_3 = const()[name = string("x_105_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_105_cast_fp16 = matmul(transpose_x = x_105_transpose_x_3, transpose_y = x_105_transpose_y_3, x = q_v_17_cast_fp16, y = p_head_35_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [8]> x_107_pad_1 = const()[name = string("x_107_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_107_mode_1 = const()[name = string("x_107_mode_1"), val = string("constant")];
            fp16 const_1363_to_fp16 = const()[name = string("const_1363_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_107_cast_fp16 = pad(constant_val = const_1363_to_fp16, mode = x_107_mode_1, pad = x_107_pad_1, x = x_105_cast_fp16)[name = string("x_107_cast_fp16")];
            tensor<int32, [4]> var_1725 = const()[name = string("op_1725"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_109_cast_fp16 = reshape(shape = var_1725, x = x_107_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [4]> var_1729_begin_1 = const()[name = string("op_1729_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1729_end_1 = const()[name = string("op_1729_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_1729_end_mask_1 = const()[name = string("op_1729_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_1729_cast_fp16 = slice_by_index(begin = var_1729_begin_1, end = var_1729_end_1, end_mask = var_1729_end_mask_1, x = x_109_cast_fp16)[name = string("op_1729_cast_fp16")];
            tensor<int32, [4]> var_1731 = const()[name = string("op_1731"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_1732_cast_fp16 = reshape(shape = var_1731, x = var_1729_cast_fp16)[name = string("op_1732_cast_fp16")];
            tensor<int32, [4]> var_1737_begin_1 = const()[name = string("op_1737_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1737_end_1 = const()[name = string("op_1737_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_1737_end_mask_1 = const()[name = string("op_1737_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_1737_cast_fp16 = slice_by_index(begin = var_1737_begin_1, end = var_1737_end_1, end_mask = var_1737_end_mask_1, x = var_1732_cast_fp16)[name = string("op_1737_cast_fp16")];
            fp16 var_1738_to_fp16 = const()[name = string("op_1738_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_17_cast_fp16 = mul(x = var_1737_cast_fp16, y = var_1738_to_fp16)[name = string("scores_pos_17_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_17_cast_fp16 = add(x = scores_content_17_cast_fp16, y = scores_pos_17_cast_fp16)[name = string("logits_17_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_1741_cast_fp16 = softmax(axis = var_1466, x = logits_17_cast_fp16)[name = string("op_1741_cast_fp16")];
            bool var_1743_transpose_x_1 = const()[name = string("op_1743_transpose_x_1"), val = bool(false)];
            bool var_1743_transpose_y_1 = const()[name = string("op_1743_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_1743_cast_fp16 = matmul(transpose_x = var_1743_transpose_x_1, transpose_y = var_1743_transpose_y_1, x = var_1741_cast_fp16, y = v_head_35_cast_fp16)[name = string("op_1743_cast_fp16")];
            tensor<int32, [1]> var_1744_axes_1 = const()[name = string("op_1744_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_1744_cast_fp16 = squeeze(axes = var_1744_axes_1, x = var_1743_cast_fp16)[name = string("op_1744_cast_fp16")];
            string dense_output_105_pad_type_1 = const()[name = string("dense_output_105_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_105_strides_1 = const()[name = string("dense_output_105_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_105_pad_1 = const()[name = string("dense_output_105_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_105_dilations_1 = const()[name = string("dense_output_105_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_105_groups_1 = const()[name = string("dense_output_105_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46596608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46727744))))[name = string("layers_1_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_105_cast_fp16 = conv(dilations = dense_output_105_dilations_1, groups = dense_output_105_groups_1, pad = dense_output_105_pad_1, pad_type = dense_output_105_pad_type_1, strides = dense_output_105_strides_1, weight = layers_1_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_105_cast_fp16")];
            string sparse_output_105_pad_type_1 = const()[name = string("sparse_output_105_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_105_strides_1 = const()[name = string("sparse_output_105_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_105_pad_1 = const()[name = string("sparse_output_105_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_105_dilations_1 = const()[name = string("sparse_output_105_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_105_groups_1 = const()[name = string("sparse_output_105_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46731008))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46728320))))[name = string("layers_1_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_105_cast_fp16 = conv(dilations = sparse_output_105_dilations_1, groups = sparse_output_105_groups_1, pad = sparse_output_105_pad_1, pad_type = sparse_output_105_pad_type_1, strides = sparse_output_105_strides_1, weight = layers_1_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_105_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1759_cast_fp16 = add(x = dense_output_105_cast_fp16, y = sparse_output_105_cast_fp16)[name = string("op_1759_cast_fp16")];
            tensor<int32, [4]> var_1760 = const()[name = string("op_1760"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1762 = const()[name = string("op_1762"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1761_cast_fp16 = transpose(perm = var_1760, x = var_1759_cast_fp16)[name = string("transpose_935")];
            tensor<fp16, [1, 51, 128]> q_head_19_cast_fp16 = reshape(shape = var_1762, x = var_1761_cast_fp16)[name = string("q_head_19_cast_fp16")];
            string dense_output_107_pad_type_1 = const()[name = string("dense_output_107_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_107_strides_1 = const()[name = string("dense_output_107_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_107_pad_1 = const()[name = string("dense_output_107_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_107_dilations_1 = const()[name = string("dense_output_107_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_107_groups_1 = const()[name = string("dense_output_107_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46747456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46878592))))[name = string("layers_1_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_107_cast_fp16 = conv(dilations = dense_output_107_dilations_1, groups = dense_output_107_groups_1, pad = dense_output_107_pad_1, pad_type = dense_output_107_pad_type_1, strides = dense_output_107_strides_1, weight = layers_1_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_107_cast_fp16")];
            string sparse_output_107_pad_type_1 = const()[name = string("sparse_output_107_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_107_strides_1 = const()[name = string("sparse_output_107_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_107_pad_1 = const()[name = string("sparse_output_107_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_107_dilations_1 = const()[name = string("sparse_output_107_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_107_groups_1 = const()[name = string("sparse_output_107_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46881856))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46879168))))[name = string("layers_1_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_107_cast_fp16 = conv(dilations = sparse_output_107_dilations_1, groups = sparse_output_107_groups_1, pad = sparse_output_107_pad_1, pad_type = sparse_output_107_pad_type_1, strides = sparse_output_107_strides_1, weight = layers_1_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_107_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1778_cast_fp16 = add(x = dense_output_107_cast_fp16, y = sparse_output_107_cast_fp16)[name = string("op_1778_cast_fp16")];
            tensor<int32, [4]> var_1779 = const()[name = string("op_1779"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1781 = const()[name = string("op_1781"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1780_cast_fp16 = transpose(perm = var_1779, x = var_1778_cast_fp16)[name = string("transpose_934")];
            tensor<fp16, [1, 51, 128]> k_head_37_cast_fp16 = reshape(shape = var_1781, x = var_1780_cast_fp16)[name = string("k_head_37_cast_fp16")];
            string dense_output_109_pad_type_1 = const()[name = string("dense_output_109_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_109_strides_1 = const()[name = string("dense_output_109_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_109_pad_1 = const()[name = string("dense_output_109_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_109_dilations_1 = const()[name = string("dense_output_109_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_109_groups_1 = const()[name = string("dense_output_109_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46898304))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47029440))))[name = string("layers_1_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_109_cast_fp16 = conv(dilations = dense_output_109_dilations_1, groups = dense_output_109_groups_1, pad = dense_output_109_pad_1, pad_type = dense_output_109_pad_type_1, strides = dense_output_109_strides_1, weight = layers_1_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_109_cast_fp16")];
            string sparse_output_109_pad_type_1 = const()[name = string("sparse_output_109_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_109_strides_1 = const()[name = string("sparse_output_109_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_109_pad_1 = const()[name = string("sparse_output_109_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_109_dilations_1 = const()[name = string("sparse_output_109_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_109_groups_1 = const()[name = string("sparse_output_109_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47032704))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47030016))))[name = string("layers_1_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_109_cast_fp16 = conv(dilations = sparse_output_109_dilations_1, groups = sparse_output_109_groups_1, pad = sparse_output_109_pad_1, pad_type = sparse_output_109_pad_type_1, strides = sparse_output_109_strides_1, weight = layers_1_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_109_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1797_cast_fp16 = add(x = dense_output_109_cast_fp16, y = sparse_output_109_cast_fp16)[name = string("op_1797_cast_fp16")];
            tensor<int32, [4]> var_1798 = const()[name = string("op_1798"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1800 = const()[name = string("op_1800"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1799_cast_fp16 = transpose(perm = var_1798, x = var_1797_cast_fp16)[name = string("transpose_933")];
            tensor<fp16, [1, 51, 128]> v_head_37_cast_fp16 = reshape(shape = var_1800, x = var_1799_cast_fp16)[name = string("v_head_37_cast_fp16")];
            string dense_output_111_pad_type_1 = const()[name = string("dense_output_111_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_111_strides_1 = const()[name = string("dense_output_111_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_111_pad_1 = const()[name = string("dense_output_111_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_111_dilations_1 = const()[name = string("dense_output_111_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_111_groups_1 = const()[name = string("dense_output_111_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47049152))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47180288))))[name = string("layers_1_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_111_cast_fp16 = conv(dilations = dense_output_111_dilations_1, groups = dense_output_111_groups_1, pad = dense_output_111_pad_1, pad_type = dense_output_111_pad_type_1, strides = dense_output_111_strides_1, weight = layers_1_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_111_cast_fp16")];
            string sparse_output_111_pad_type_1 = const()[name = string("sparse_output_111_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_111_strides_1 = const()[name = string("sparse_output_111_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_111_pad_1 = const()[name = string("sparse_output_111_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_111_dilations_1 = const()[name = string("sparse_output_111_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_111_groups_1 = const()[name = string("sparse_output_111_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47183552))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47180864))))[name = string("layers_1_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_111_cast_fp16 = conv(dilations = sparse_output_111_dilations_1, groups = sparse_output_111_groups_1, pad = sparse_output_111_pad_1, pad_type = sparse_output_111_pad_type_1, strides = sparse_output_111_strides_1, weight = layers_1_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_111_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_1816_cast_fp16 = add(x = dense_output_111_cast_fp16, y = sparse_output_111_cast_fp16)[name = string("op_1816_cast_fp16")];
            tensor<int32, [4]> var_1817 = const()[name = string("op_1817"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_1818_cast_fp16 = transpose(perm = var_1817, x = var_1816_cast_fp16)[name = string("transpose_932")];
            tensor<fp16, [1, 101, 128]> p_head_37_cast_fp16 = reshape(shape = var_1819, x = var_1818_cast_fp16)[name = string("p_head_37_cast_fp16")];
            tensor<fp16, [128]> var_1821_to_fp16 = const()[name = string("op_1821_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47200000)))];
            tensor<fp16, [1, 51, 128]> var_1822_cast_fp16 = add(x = q_head_19_cast_fp16, y = var_1821_to_fp16)[name = string("op_1822_cast_fp16")];
            tensor<int32, [1]> q_u_19_axes_1 = const()[name = string("q_u_19_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_19_cast_fp16 = expand_dims(axes = q_u_19_axes_1, x = var_1822_cast_fp16)[name = string("q_u_19_cast_fp16")];
            tensor<fp16, [128]> var_1824_to_fp16 = const()[name = string("op_1824_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47200320)))];
            tensor<fp16, [1, 51, 128]> var_1825_cast_fp16 = add(x = q_head_19_cast_fp16, y = var_1824_to_fp16)[name = string("op_1825_cast_fp16")];
            tensor<int32, [1]> q_v_19_axes_1 = const()[name = string("q_v_19_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_19_cast_fp16 = expand_dims(axes = q_v_19_axes_1, x = var_1825_cast_fp16)[name = string("q_v_19_cast_fp16")];
            tensor<int32, [1]> k_head_39_axes_1 = const()[name = string("k_head_39_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_39_cast_fp16 = expand_dims(axes = k_head_39_axes_1, x = k_head_37_cast_fp16)[name = string("k_head_39_cast_fp16")];
            tensor<int32, [1]> v_head_39_axes_1 = const()[name = string("v_head_39_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_39_cast_fp16 = expand_dims(axes = v_head_39_axes_1, x = v_head_37_cast_fp16)[name = string("v_head_39_cast_fp16")];
            tensor<int32, [1]> p_head_39_axes_1 = const()[name = string("p_head_39_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_39_cast_fp16 = expand_dims(axes = p_head_39_axes_1, x = p_head_37_cast_fp16)[name = string("p_head_39_cast_fp16")];
            bool var_1831_transpose_x_3 = const()[name = string("op_1831_transpose_x_3"), val = bool(false)];
            bool var_1831_transpose_y_3 = const()[name = string("op_1831_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_1831_cast_fp16 = matmul(transpose_x = var_1831_transpose_x_3, transpose_y = var_1831_transpose_y_3, x = q_u_19_cast_fp16, y = k_head_39_cast_fp16)[name = string("op_1831_cast_fp16")];
            fp16 var_1832_to_fp16 = const()[name = string("op_1832_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_19_cast_fp16 = mul(x = var_1831_cast_fp16, y = var_1832_to_fp16)[name = string("scores_content_19_cast_fp16")];
            bool x_113_transpose_x_3 = const()[name = string("x_113_transpose_x_3"), val = bool(false)];
            bool x_113_transpose_y_3 = const()[name = string("x_113_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_113_cast_fp16 = matmul(transpose_x = x_113_transpose_x_3, transpose_y = x_113_transpose_y_3, x = q_v_19_cast_fp16, y = p_head_39_cast_fp16)[name = string("x_113_cast_fp16")];
            tensor<int32, [8]> x_115_pad_1 = const()[name = string("x_115_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_115_mode_1 = const()[name = string("x_115_mode_1"), val = string("constant")];
            fp16 const_1369_to_fp16 = const()[name = string("const_1369_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_115_cast_fp16 = pad(constant_val = const_1369_to_fp16, mode = x_115_mode_1, pad = x_115_pad_1, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")];
            tensor<int32, [4]> var_1846 = const()[name = string("op_1846"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_117_cast_fp16 = reshape(shape = var_1846, x = x_115_cast_fp16)[name = string("x_117_cast_fp16")];
            tensor<int32, [4]> var_1850_begin_1 = const()[name = string("op_1850_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1850_end_1 = const()[name = string("op_1850_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_1850_end_mask_1 = const()[name = string("op_1850_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_1850_cast_fp16 = slice_by_index(begin = var_1850_begin_1, end = var_1850_end_1, end_mask = var_1850_end_mask_1, x = x_117_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<int32, [4]> var_1852 = const()[name = string("op_1852"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_1853_cast_fp16 = reshape(shape = var_1852, x = var_1850_cast_fp16)[name = string("op_1853_cast_fp16")];
            tensor<int32, [4]> var_1858_begin_1 = const()[name = string("op_1858_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1858_end_1 = const()[name = string("op_1858_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_1858_end_mask_1 = const()[name = string("op_1858_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_1858_cast_fp16 = slice_by_index(begin = var_1858_begin_1, end = var_1858_end_1, end_mask = var_1858_end_mask_1, x = var_1853_cast_fp16)[name = string("op_1858_cast_fp16")];
            fp16 var_1859_to_fp16 = const()[name = string("op_1859_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_19_cast_fp16 = mul(x = var_1858_cast_fp16, y = var_1859_to_fp16)[name = string("scores_pos_19_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_19_cast_fp16 = add(x = scores_content_19_cast_fp16, y = scores_pos_19_cast_fp16)[name = string("logits_19_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_1862_cast_fp16 = softmax(axis = var_1466, x = logits_19_cast_fp16)[name = string("op_1862_cast_fp16")];
            bool var_1864_transpose_x_1 = const()[name = string("op_1864_transpose_x_1"), val = bool(false)];
            bool var_1864_transpose_y_1 = const()[name = string("op_1864_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_1864_cast_fp16 = matmul(transpose_x = var_1864_transpose_x_1, transpose_y = var_1864_transpose_y_1, x = var_1862_cast_fp16, y = v_head_39_cast_fp16)[name = string("op_1864_cast_fp16")];
            tensor<int32, [1]> var_1865_axes_1 = const()[name = string("op_1865_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_1865_cast_fp16 = squeeze(axes = var_1865_axes_1, x = var_1864_cast_fp16)[name = string("op_1865_cast_fp16")];
            string dense_output_113_pad_type_1 = const()[name = string("dense_output_113_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_113_strides_1 = const()[name = string("dense_output_113_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_113_pad_1 = const()[name = string("dense_output_113_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_113_dilations_1 = const()[name = string("dense_output_113_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_113_groups_1 = const()[name = string("dense_output_113_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47200640))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47331776))))[name = string("layers_1_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_113_cast_fp16 = conv(dilations = dense_output_113_dilations_1, groups = dense_output_113_groups_1, pad = dense_output_113_pad_1, pad_type = dense_output_113_pad_type_1, strides = dense_output_113_strides_1, weight = layers_1_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_113_cast_fp16")];
            string sparse_output_113_pad_type_1 = const()[name = string("sparse_output_113_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_113_strides_1 = const()[name = string("sparse_output_113_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_113_pad_1 = const()[name = string("sparse_output_113_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_113_dilations_1 = const()[name = string("sparse_output_113_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_113_groups_1 = const()[name = string("sparse_output_113_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47335040))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47332352))))[name = string("layers_1_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_113_cast_fp16 = conv(dilations = sparse_output_113_dilations_1, groups = sparse_output_113_groups_1, pad = sparse_output_113_pad_1, pad_type = sparse_output_113_pad_type_1, strides = sparse_output_113_strides_1, weight = layers_1_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_113_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1880_cast_fp16 = add(x = dense_output_113_cast_fp16, y = sparse_output_113_cast_fp16)[name = string("op_1880_cast_fp16")];
            tensor<int32, [4]> var_1881 = const()[name = string("op_1881"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1882_cast_fp16 = transpose(perm = var_1881, x = var_1880_cast_fp16)[name = string("transpose_931")];
            tensor<fp16, [1, 51, 128]> q_head_21_cast_fp16 = reshape(shape = var_1883, x = var_1882_cast_fp16)[name = string("q_head_21_cast_fp16")];
            string dense_output_115_pad_type_1 = const()[name = string("dense_output_115_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_115_strides_1 = const()[name = string("dense_output_115_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_115_pad_1 = const()[name = string("dense_output_115_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_115_dilations_1 = const()[name = string("dense_output_115_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_115_groups_1 = const()[name = string("dense_output_115_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47351488))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47482624))))[name = string("layers_1_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_115_cast_fp16 = conv(dilations = dense_output_115_dilations_1, groups = dense_output_115_groups_1, pad = dense_output_115_pad_1, pad_type = dense_output_115_pad_type_1, strides = dense_output_115_strides_1, weight = layers_1_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_115_cast_fp16")];
            string sparse_output_115_pad_type_1 = const()[name = string("sparse_output_115_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_115_strides_1 = const()[name = string("sparse_output_115_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_115_pad_1 = const()[name = string("sparse_output_115_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_115_dilations_1 = const()[name = string("sparse_output_115_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_115_groups_1 = const()[name = string("sparse_output_115_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47485888))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47483200))))[name = string("layers_1_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_115_cast_fp16 = conv(dilations = sparse_output_115_dilations_1, groups = sparse_output_115_groups_1, pad = sparse_output_115_pad_1, pad_type = sparse_output_115_pad_type_1, strides = sparse_output_115_strides_1, weight = layers_1_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_115_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1899_cast_fp16 = add(x = dense_output_115_cast_fp16, y = sparse_output_115_cast_fp16)[name = string("op_1899_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1902 = const()[name = string("op_1902"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1901_cast_fp16 = transpose(perm = var_1900, x = var_1899_cast_fp16)[name = string("transpose_930")];
            tensor<fp16, [1, 51, 128]> k_head_41_cast_fp16 = reshape(shape = var_1902, x = var_1901_cast_fp16)[name = string("k_head_41_cast_fp16")];
            string dense_output_117_pad_type_1 = const()[name = string("dense_output_117_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_117_strides_1 = const()[name = string("dense_output_117_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_117_pad_1 = const()[name = string("dense_output_117_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_117_dilations_1 = const()[name = string("dense_output_117_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_117_groups_1 = const()[name = string("dense_output_117_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47502336))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47633472))))[name = string("layers_1_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_117_cast_fp16 = conv(dilations = dense_output_117_dilations_1, groups = dense_output_117_groups_1, pad = dense_output_117_pad_1, pad_type = dense_output_117_pad_type_1, strides = dense_output_117_strides_1, weight = layers_1_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_117_cast_fp16")];
            string sparse_output_117_pad_type_1 = const()[name = string("sparse_output_117_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_117_strides_1 = const()[name = string("sparse_output_117_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_117_pad_1 = const()[name = string("sparse_output_117_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_117_dilations_1 = const()[name = string("sparse_output_117_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_117_groups_1 = const()[name = string("sparse_output_117_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47636736))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47634048))))[name = string("layers_1_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_117_cast_fp16 = conv(dilations = sparse_output_117_dilations_1, groups = sparse_output_117_groups_1, pad = sparse_output_117_pad_1, pad_type = sparse_output_117_pad_type_1, strides = sparse_output_117_strides_1, weight = layers_1_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_117_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_1918_cast_fp16 = add(x = dense_output_117_cast_fp16, y = sparse_output_117_cast_fp16)[name = string("op_1918_cast_fp16")];
            tensor<int32, [4]> var_1919 = const()[name = string("op_1919"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1921 = const()[name = string("op_1921"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_1920_cast_fp16 = transpose(perm = var_1919, x = var_1918_cast_fp16)[name = string("transpose_929")];
            tensor<fp16, [1, 51, 128]> v_head_41_cast_fp16 = reshape(shape = var_1921, x = var_1920_cast_fp16)[name = string("v_head_41_cast_fp16")];
            string dense_output_119_pad_type_1 = const()[name = string("dense_output_119_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_119_strides_1 = const()[name = string("dense_output_119_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_119_pad_1 = const()[name = string("dense_output_119_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_119_dilations_1 = const()[name = string("dense_output_119_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_119_groups_1 = const()[name = string("dense_output_119_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47653184))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47784320))))[name = string("layers_1_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_119_cast_fp16 = conv(dilations = dense_output_119_dilations_1, groups = dense_output_119_groups_1, pad = dense_output_119_pad_1, pad_type = dense_output_119_pad_type_1, strides = dense_output_119_strides_1, weight = layers_1_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_119_cast_fp16")];
            string sparse_output_119_pad_type_1 = const()[name = string("sparse_output_119_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_119_strides_1 = const()[name = string("sparse_output_119_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_119_pad_1 = const()[name = string("sparse_output_119_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_119_dilations_1 = const()[name = string("sparse_output_119_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_119_groups_1 = const()[name = string("sparse_output_119_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47787584))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47784896))))[name = string("layers_1_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_119_cast_fp16 = conv(dilations = sparse_output_119_dilations_1, groups = sparse_output_119_groups_1, pad = sparse_output_119_pad_1, pad_type = sparse_output_119_pad_type_1, strides = sparse_output_119_strides_1, weight = layers_1_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_119_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_1937_cast_fp16 = add(x = dense_output_119_cast_fp16, y = sparse_output_119_cast_fp16)[name = string("op_1937_cast_fp16")];
            tensor<int32, [4]> var_1938 = const()[name = string("op_1938"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_1940 = const()[name = string("op_1940"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_1939_cast_fp16 = transpose(perm = var_1938, x = var_1937_cast_fp16)[name = string("transpose_928")];
            tensor<fp16, [1, 101, 128]> p_head_41_cast_fp16 = reshape(shape = var_1940, x = var_1939_cast_fp16)[name = string("p_head_41_cast_fp16")];
            tensor<fp16, [128]> var_1942_to_fp16 = const()[name = string("op_1942_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47804032)))];
            tensor<fp16, [1, 51, 128]> var_1943_cast_fp16 = add(x = q_head_21_cast_fp16, y = var_1942_to_fp16)[name = string("op_1943_cast_fp16")];
            tensor<int32, [1]> q_u_21_axes_1 = const()[name = string("q_u_21_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_21_cast_fp16 = expand_dims(axes = q_u_21_axes_1, x = var_1943_cast_fp16)[name = string("q_u_21_cast_fp16")];
            tensor<fp16, [128]> var_1945_to_fp16 = const()[name = string("op_1945_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47804352)))];
            tensor<fp16, [1, 51, 128]> var_1946_cast_fp16 = add(x = q_head_21_cast_fp16, y = var_1945_to_fp16)[name = string("op_1946_cast_fp16")];
            tensor<int32, [1]> q_v_21_axes_1 = const()[name = string("q_v_21_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_21_cast_fp16 = expand_dims(axes = q_v_21_axes_1, x = var_1946_cast_fp16)[name = string("q_v_21_cast_fp16")];
            tensor<int32, [1]> k_head_43_axes_1 = const()[name = string("k_head_43_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_43_cast_fp16 = expand_dims(axes = k_head_43_axes_1, x = k_head_41_cast_fp16)[name = string("k_head_43_cast_fp16")];
            tensor<int32, [1]> v_head_43_axes_1 = const()[name = string("v_head_43_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_43_cast_fp16 = expand_dims(axes = v_head_43_axes_1, x = v_head_41_cast_fp16)[name = string("v_head_43_cast_fp16")];
            tensor<int32, [1]> p_head_43_axes_1 = const()[name = string("p_head_43_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_43_cast_fp16 = expand_dims(axes = p_head_43_axes_1, x = p_head_41_cast_fp16)[name = string("p_head_43_cast_fp16")];
            bool var_1952_transpose_x_3 = const()[name = string("op_1952_transpose_x_3"), val = bool(false)];
            bool var_1952_transpose_y_3 = const()[name = string("op_1952_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_1952_cast_fp16 = matmul(transpose_x = var_1952_transpose_x_3, transpose_y = var_1952_transpose_y_3, x = q_u_21_cast_fp16, y = k_head_43_cast_fp16)[name = string("op_1952_cast_fp16")];
            fp16 var_1953_to_fp16 = const()[name = string("op_1953_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_21_cast_fp16 = mul(x = var_1952_cast_fp16, y = var_1953_to_fp16)[name = string("scores_content_21_cast_fp16")];
            bool x_121_transpose_x_3 = const()[name = string("x_121_transpose_x_3"), val = bool(false)];
            bool x_121_transpose_y_3 = const()[name = string("x_121_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_121_cast_fp16 = matmul(transpose_x = x_121_transpose_x_3, transpose_y = x_121_transpose_y_3, x = q_v_21_cast_fp16, y = p_head_43_cast_fp16)[name = string("x_121_cast_fp16")];
            tensor<int32, [8]> x_123_pad_1 = const()[name = string("x_123_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_123_mode_1 = const()[name = string("x_123_mode_1"), val = string("constant")];
            fp16 const_1375_to_fp16 = const()[name = string("const_1375_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_123_cast_fp16 = pad(constant_val = const_1375_to_fp16, mode = x_123_mode_1, pad = x_123_pad_1, x = x_121_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<int32, [4]> var_1967 = const()[name = string("op_1967"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_125_cast_fp16 = reshape(shape = var_1967, x = x_123_cast_fp16)[name = string("x_125_cast_fp16")];
            tensor<int32, [4]> var_1971_begin_1 = const()[name = string("op_1971_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_1971_end_1 = const()[name = string("op_1971_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_1971_end_mask_1 = const()[name = string("op_1971_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_1971_cast_fp16 = slice_by_index(begin = var_1971_begin_1, end = var_1971_end_1, end_mask = var_1971_end_mask_1, x = x_125_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> var_1973 = const()[name = string("op_1973"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_1974_cast_fp16 = reshape(shape = var_1973, x = var_1971_cast_fp16)[name = string("op_1974_cast_fp16")];
            tensor<int32, [4]> var_1979_begin_1 = const()[name = string("op_1979_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1979_end_1 = const()[name = string("op_1979_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_1979_end_mask_1 = const()[name = string("op_1979_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_1979_cast_fp16 = slice_by_index(begin = var_1979_begin_1, end = var_1979_end_1, end_mask = var_1979_end_mask_1, x = var_1974_cast_fp16)[name = string("op_1979_cast_fp16")];
            fp16 var_1980_to_fp16 = const()[name = string("op_1980_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_21_cast_fp16 = mul(x = var_1979_cast_fp16, y = var_1980_to_fp16)[name = string("scores_pos_21_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_21_cast_fp16 = add(x = scores_content_21_cast_fp16, y = scores_pos_21_cast_fp16)[name = string("logits_21_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_1983_cast_fp16 = softmax(axis = var_1466, x = logits_21_cast_fp16)[name = string("op_1983_cast_fp16")];
            bool var_1985_transpose_x_1 = const()[name = string("op_1985_transpose_x_1"), val = bool(false)];
            bool var_1985_transpose_y_1 = const()[name = string("op_1985_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_1985_cast_fp16 = matmul(transpose_x = var_1985_transpose_x_1, transpose_y = var_1985_transpose_y_1, x = var_1983_cast_fp16, y = v_head_43_cast_fp16)[name = string("op_1985_cast_fp16")];
            tensor<int32, [1]> var_1986_axes_1 = const()[name = string("op_1986_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_1986_cast_fp16 = squeeze(axes = var_1986_axes_1, x = var_1985_cast_fp16)[name = string("op_1986_cast_fp16")];
            string dense_output_121_pad_type_1 = const()[name = string("dense_output_121_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_121_strides_1 = const()[name = string("dense_output_121_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_121_pad_1 = const()[name = string("dense_output_121_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_121_dilations_1 = const()[name = string("dense_output_121_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_121_groups_1 = const()[name = string("dense_output_121_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47804672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47935808))))[name = string("layers_1_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_121_cast_fp16 = conv(dilations = dense_output_121_dilations_1, groups = dense_output_121_groups_1, pad = dense_output_121_pad_1, pad_type = dense_output_121_pad_type_1, strides = dense_output_121_strides_1, weight = layers_1_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_121_cast_fp16")];
            string sparse_output_121_pad_type_1 = const()[name = string("sparse_output_121_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_121_strides_1 = const()[name = string("sparse_output_121_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_121_pad_1 = const()[name = string("sparse_output_121_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_121_dilations_1 = const()[name = string("sparse_output_121_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_121_groups_1 = const()[name = string("sparse_output_121_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47939072))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47936384))))[name = string("layers_1_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_121_cast_fp16 = conv(dilations = sparse_output_121_dilations_1, groups = sparse_output_121_groups_1, pad = sparse_output_121_pad_1, pad_type = sparse_output_121_pad_type_1, strides = sparse_output_121_strides_1, weight = layers_1_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_121_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2001_cast_fp16 = add(x = dense_output_121_cast_fp16, y = sparse_output_121_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<int32, [4]> var_2002 = const()[name = string("op_2002"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2004 = const()[name = string("op_2004"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2003_cast_fp16 = transpose(perm = var_2002, x = var_2001_cast_fp16)[name = string("transpose_927")];
            tensor<fp16, [1, 51, 128]> q_head_23_cast_fp16 = reshape(shape = var_2004, x = var_2003_cast_fp16)[name = string("q_head_23_cast_fp16")];
            string dense_output_123_pad_type_1 = const()[name = string("dense_output_123_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_123_strides_1 = const()[name = string("dense_output_123_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_123_pad_1 = const()[name = string("dense_output_123_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_123_dilations_1 = const()[name = string("dense_output_123_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_123_groups_1 = const()[name = string("dense_output_123_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47955520))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48086656))))[name = string("layers_1_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_123_cast_fp16 = conv(dilations = dense_output_123_dilations_1, groups = dense_output_123_groups_1, pad = dense_output_123_pad_1, pad_type = dense_output_123_pad_type_1, strides = dense_output_123_strides_1, weight = layers_1_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_123_cast_fp16")];
            string sparse_output_123_pad_type_1 = const()[name = string("sparse_output_123_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_123_strides_1 = const()[name = string("sparse_output_123_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_123_pad_1 = const()[name = string("sparse_output_123_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_123_dilations_1 = const()[name = string("sparse_output_123_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_123_groups_1 = const()[name = string("sparse_output_123_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48089920))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48087232))))[name = string("layers_1_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_123_cast_fp16 = conv(dilations = sparse_output_123_dilations_1, groups = sparse_output_123_groups_1, pad = sparse_output_123_pad_1, pad_type = sparse_output_123_pad_type_1, strides = sparse_output_123_strides_1, weight = layers_1_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_123_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2020_cast_fp16 = add(x = dense_output_123_cast_fp16, y = sparse_output_123_cast_fp16)[name = string("op_2020_cast_fp16")];
            tensor<int32, [4]> var_2021 = const()[name = string("op_2021"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2023 = const()[name = string("op_2023"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2022_cast_fp16 = transpose(perm = var_2021, x = var_2020_cast_fp16)[name = string("transpose_926")];
            tensor<fp16, [1, 51, 128]> k_head_45_cast_fp16 = reshape(shape = var_2023, x = var_2022_cast_fp16)[name = string("k_head_45_cast_fp16")];
            string dense_output_125_pad_type_1 = const()[name = string("dense_output_125_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_125_strides_1 = const()[name = string("dense_output_125_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_125_pad_1 = const()[name = string("dense_output_125_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_125_dilations_1 = const()[name = string("dense_output_125_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_125_groups_1 = const()[name = string("dense_output_125_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48106368))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48237504))))[name = string("layers_1_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_125_cast_fp16 = conv(dilations = dense_output_125_dilations_1, groups = dense_output_125_groups_1, pad = dense_output_125_pad_1, pad_type = dense_output_125_pad_type_1, strides = dense_output_125_strides_1, weight = layers_1_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_125_cast_fp16")];
            string sparse_output_125_pad_type_1 = const()[name = string("sparse_output_125_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_125_strides_1 = const()[name = string("sparse_output_125_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_125_pad_1 = const()[name = string("sparse_output_125_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_125_dilations_1 = const()[name = string("sparse_output_125_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_125_groups_1 = const()[name = string("sparse_output_125_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48240768))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48238080))))[name = string("layers_1_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_125_cast_fp16 = conv(dilations = sparse_output_125_dilations_1, groups = sparse_output_125_groups_1, pad = sparse_output_125_pad_1, pad_type = sparse_output_125_pad_type_1, strides = sparse_output_125_strides_1, weight = layers_1_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_125_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2039_cast_fp16 = add(x = dense_output_125_cast_fp16, y = sparse_output_125_cast_fp16)[name = string("op_2039_cast_fp16")];
            tensor<int32, [4]> var_2040 = const()[name = string("op_2040"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2042 = const()[name = string("op_2042"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2041_cast_fp16 = transpose(perm = var_2040, x = var_2039_cast_fp16)[name = string("transpose_925")];
            tensor<fp16, [1, 51, 128]> v_head_45_cast_fp16 = reshape(shape = var_2042, x = var_2041_cast_fp16)[name = string("v_head_45_cast_fp16")];
            string dense_output_127_pad_type_1 = const()[name = string("dense_output_127_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_127_strides_1 = const()[name = string("dense_output_127_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_127_pad_1 = const()[name = string("dense_output_127_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_127_dilations_1 = const()[name = string("dense_output_127_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_127_groups_1 = const()[name = string("dense_output_127_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48257216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48388352))))[name = string("layers_1_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_127_cast_fp16 = conv(dilations = dense_output_127_dilations_1, groups = dense_output_127_groups_1, pad = dense_output_127_pad_1, pad_type = dense_output_127_pad_type_1, strides = dense_output_127_strides_1, weight = layers_1_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_127_cast_fp16")];
            string sparse_output_127_pad_type_1 = const()[name = string("sparse_output_127_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_127_strides_1 = const()[name = string("sparse_output_127_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_127_pad_1 = const()[name = string("sparse_output_127_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_127_dilations_1 = const()[name = string("sparse_output_127_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_127_groups_1 = const()[name = string("sparse_output_127_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48391616))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48388928))))[name = string("layers_1_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_127_cast_fp16 = conv(dilations = sparse_output_127_dilations_1, groups = sparse_output_127_groups_1, pad = sparse_output_127_pad_1, pad_type = sparse_output_127_pad_type_1, strides = sparse_output_127_strides_1, weight = layers_1_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_127_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_2058_cast_fp16 = add(x = dense_output_127_cast_fp16, y = sparse_output_127_cast_fp16)[name = string("op_2058_cast_fp16")];
            tensor<int32, [4]> var_2059 = const()[name = string("op_2059"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2061 = const()[name = string("op_2061"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_2060_cast_fp16 = transpose(perm = var_2059, x = var_2058_cast_fp16)[name = string("transpose_924")];
            tensor<fp16, [1, 101, 128]> p_head_45_cast_fp16 = reshape(shape = var_2061, x = var_2060_cast_fp16)[name = string("p_head_45_cast_fp16")];
            tensor<fp16, [128]> var_2063_to_fp16 = const()[name = string("op_2063_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48408064)))];
            tensor<fp16, [1, 51, 128]> var_2064_cast_fp16 = add(x = q_head_23_cast_fp16, y = var_2063_to_fp16)[name = string("op_2064_cast_fp16")];
            tensor<int32, [1]> q_u_23_axes_1 = const()[name = string("q_u_23_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_23_cast_fp16 = expand_dims(axes = q_u_23_axes_1, x = var_2064_cast_fp16)[name = string("q_u_23_cast_fp16")];
            tensor<fp16, [128]> var_2066_to_fp16 = const()[name = string("op_2066_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48408384)))];
            tensor<fp16, [1, 51, 128]> var_2067_cast_fp16 = add(x = q_head_23_cast_fp16, y = var_2066_to_fp16)[name = string("op_2067_cast_fp16")];
            tensor<int32, [1]> q_v_23_axes_1 = const()[name = string("q_v_23_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_23_cast_fp16 = expand_dims(axes = q_v_23_axes_1, x = var_2067_cast_fp16)[name = string("q_v_23_cast_fp16")];
            tensor<int32, [1]> k_head_47_axes_1 = const()[name = string("k_head_47_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_47_cast_fp16 = expand_dims(axes = k_head_47_axes_1, x = k_head_45_cast_fp16)[name = string("k_head_47_cast_fp16")];
            tensor<int32, [1]> v_head_47_axes_1 = const()[name = string("v_head_47_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_47_cast_fp16 = expand_dims(axes = v_head_47_axes_1, x = v_head_45_cast_fp16)[name = string("v_head_47_cast_fp16")];
            tensor<int32, [1]> p_head_47_axes_1 = const()[name = string("p_head_47_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_47_cast_fp16 = expand_dims(axes = p_head_47_axes_1, x = p_head_45_cast_fp16)[name = string("p_head_47_cast_fp16")];
            bool var_2073_transpose_x_3 = const()[name = string("op_2073_transpose_x_3"), val = bool(false)];
            bool var_2073_transpose_y_3 = const()[name = string("op_2073_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_2073_cast_fp16 = matmul(transpose_x = var_2073_transpose_x_3, transpose_y = var_2073_transpose_y_3, x = q_u_23_cast_fp16, y = k_head_47_cast_fp16)[name = string("op_2073_cast_fp16")];
            fp16 var_2074_to_fp16 = const()[name = string("op_2074_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_23_cast_fp16 = mul(x = var_2073_cast_fp16, y = var_2074_to_fp16)[name = string("scores_content_23_cast_fp16")];
            bool x_129_transpose_x_3 = const()[name = string("x_129_transpose_x_3"), val = bool(false)];
            bool x_129_transpose_y_3 = const()[name = string("x_129_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_129_cast_fp16 = matmul(transpose_x = x_129_transpose_x_3, transpose_y = x_129_transpose_y_3, x = q_v_23_cast_fp16, y = p_head_47_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<int32, [8]> x_131_pad_1 = const()[name = string("x_131_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_131_mode_1 = const()[name = string("x_131_mode_1"), val = string("constant")];
            fp16 const_1381_to_fp16 = const()[name = string("const_1381_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_131_cast_fp16 = pad(constant_val = const_1381_to_fp16, mode = x_131_mode_1, pad = x_131_pad_1, x = x_129_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<int32, [4]> var_2088 = const()[name = string("op_2088"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_133_cast_fp16 = reshape(shape = var_2088, x = x_131_cast_fp16)[name = string("x_133_cast_fp16")];
            tensor<int32, [4]> var_2092_begin_1 = const()[name = string("op_2092_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2092_end_1 = const()[name = string("op_2092_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_2092_end_mask_1 = const()[name = string("op_2092_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_2092_cast_fp16 = slice_by_index(begin = var_2092_begin_1, end = var_2092_end_1, end_mask = var_2092_end_mask_1, x = x_133_cast_fp16)[name = string("op_2092_cast_fp16")];
            tensor<int32, [4]> var_2094 = const()[name = string("op_2094"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_2095_cast_fp16 = reshape(shape = var_2094, x = var_2092_cast_fp16)[name = string("op_2095_cast_fp16")];
            tensor<int32, [4]> var_2100_begin_1 = const()[name = string("op_2100_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2100_end_1 = const()[name = string("op_2100_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_2100_end_mask_1 = const()[name = string("op_2100_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_2100_cast_fp16 = slice_by_index(begin = var_2100_begin_1, end = var_2100_end_1, end_mask = var_2100_end_mask_1, x = var_2095_cast_fp16)[name = string("op_2100_cast_fp16")];
            fp16 var_2101_to_fp16 = const()[name = string("op_2101_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_23_cast_fp16 = mul(x = var_2100_cast_fp16, y = var_2101_to_fp16)[name = string("scores_pos_23_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_23_cast_fp16 = add(x = scores_content_23_cast_fp16, y = scores_pos_23_cast_fp16)[name = string("logits_23_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_2104_cast_fp16 = softmax(axis = var_1466, x = logits_23_cast_fp16)[name = string("op_2104_cast_fp16")];
            bool var_2106_transpose_x_1 = const()[name = string("op_2106_transpose_x_1"), val = bool(false)];
            bool var_2106_transpose_y_1 = const()[name = string("op_2106_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_2106_cast_fp16 = matmul(transpose_x = var_2106_transpose_x_1, transpose_y = var_2106_transpose_y_1, x = var_2104_cast_fp16, y = v_head_47_cast_fp16)[name = string("op_2106_cast_fp16")];
            tensor<int32, [1]> var_2107_axes_1 = const()[name = string("op_2107_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_2107_cast_fp16 = squeeze(axes = var_2107_axes_1, x = var_2106_cast_fp16)[name = string("op_2107_cast_fp16")];
            string dense_output_129_pad_type_1 = const()[name = string("dense_output_129_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_129_strides_1 = const()[name = string("dense_output_129_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_129_pad_1 = const()[name = string("dense_output_129_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_129_dilations_1 = const()[name = string("dense_output_129_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_129_groups_1 = const()[name = string("dense_output_129_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48408704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48539840))))[name = string("layers_1_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_129_cast_fp16 = conv(dilations = dense_output_129_dilations_1, groups = dense_output_129_groups_1, pad = dense_output_129_pad_1, pad_type = dense_output_129_pad_type_1, strides = dense_output_129_strides_1, weight = layers_1_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_129_cast_fp16")];
            string sparse_output_129_pad_type_1 = const()[name = string("sparse_output_129_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_129_strides_1 = const()[name = string("sparse_output_129_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_129_pad_1 = const()[name = string("sparse_output_129_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_129_dilations_1 = const()[name = string("sparse_output_129_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_129_groups_1 = const()[name = string("sparse_output_129_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48543104))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48540416))))[name = string("layers_1_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_129_cast_fp16 = conv(dilations = sparse_output_129_dilations_1, groups = sparse_output_129_groups_1, pad = sparse_output_129_pad_1, pad_type = sparse_output_129_pad_type_1, strides = sparse_output_129_strides_1, weight = layers_1_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_129_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2122_cast_fp16 = add(x = dense_output_129_cast_fp16, y = sparse_output_129_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> var_2123 = const()[name = string("op_2123"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2125 = const()[name = string("op_2125"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2124_cast_fp16 = transpose(perm = var_2123, x = var_2122_cast_fp16)[name = string("transpose_923")];
            tensor<fp16, [1, 51, 128]> q_head_25_cast_fp16 = reshape(shape = var_2125, x = var_2124_cast_fp16)[name = string("q_head_25_cast_fp16")];
            string dense_output_131_pad_type_1 = const()[name = string("dense_output_131_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_131_strides_1 = const()[name = string("dense_output_131_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_131_pad_1 = const()[name = string("dense_output_131_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_131_dilations_1 = const()[name = string("dense_output_131_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_131_groups_1 = const()[name = string("dense_output_131_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48559552))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48690688))))[name = string("layers_1_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_131_cast_fp16 = conv(dilations = dense_output_131_dilations_1, groups = dense_output_131_groups_1, pad = dense_output_131_pad_1, pad_type = dense_output_131_pad_type_1, strides = dense_output_131_strides_1, weight = layers_1_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_131_cast_fp16")];
            string sparse_output_131_pad_type_1 = const()[name = string("sparse_output_131_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_131_strides_1 = const()[name = string("sparse_output_131_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_131_pad_1 = const()[name = string("sparse_output_131_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_131_dilations_1 = const()[name = string("sparse_output_131_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_131_groups_1 = const()[name = string("sparse_output_131_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48693952))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48691264))))[name = string("layers_1_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_131_cast_fp16 = conv(dilations = sparse_output_131_dilations_1, groups = sparse_output_131_groups_1, pad = sparse_output_131_pad_1, pad_type = sparse_output_131_pad_type_1, strides = sparse_output_131_strides_1, weight = layers_1_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_131_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2141_cast_fp16 = add(x = dense_output_131_cast_fp16, y = sparse_output_131_cast_fp16)[name = string("op_2141_cast_fp16")];
            tensor<int32, [4]> var_2142 = const()[name = string("op_2142"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2144 = const()[name = string("op_2144"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2143_cast_fp16 = transpose(perm = var_2142, x = var_2141_cast_fp16)[name = string("transpose_922")];
            tensor<fp16, [1, 51, 128]> k_head_49_cast_fp16 = reshape(shape = var_2144, x = var_2143_cast_fp16)[name = string("k_head_49_cast_fp16")];
            string dense_output_133_pad_type_1 = const()[name = string("dense_output_133_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_133_strides_1 = const()[name = string("dense_output_133_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_133_pad_1 = const()[name = string("dense_output_133_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_133_dilations_1 = const()[name = string("dense_output_133_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_133_groups_1 = const()[name = string("dense_output_133_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48710400))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48841536))))[name = string("layers_1_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_133_cast_fp16 = conv(dilations = dense_output_133_dilations_1, groups = dense_output_133_groups_1, pad = dense_output_133_pad_1, pad_type = dense_output_133_pad_type_1, strides = dense_output_133_strides_1, weight = layers_1_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_133_cast_fp16")];
            string sparse_output_133_pad_type_1 = const()[name = string("sparse_output_133_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_133_strides_1 = const()[name = string("sparse_output_133_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_133_pad_1 = const()[name = string("sparse_output_133_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_133_dilations_1 = const()[name = string("sparse_output_133_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_133_groups_1 = const()[name = string("sparse_output_133_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48844800))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48842112))))[name = string("layers_1_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_133_cast_fp16 = conv(dilations = sparse_output_133_dilations_1, groups = sparse_output_133_groups_1, pad = sparse_output_133_pad_1, pad_type = sparse_output_133_pad_type_1, strides = sparse_output_133_strides_1, weight = layers_1_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_133_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2160_cast_fp16 = add(x = dense_output_133_cast_fp16, y = sparse_output_133_cast_fp16)[name = string("op_2160_cast_fp16")];
            tensor<int32, [4]> var_2161 = const()[name = string("op_2161"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2163 = const()[name = string("op_2163"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2162_cast_fp16 = transpose(perm = var_2161, x = var_2160_cast_fp16)[name = string("transpose_921")];
            tensor<fp16, [1, 51, 128]> v_head_49_cast_fp16 = reshape(shape = var_2163, x = var_2162_cast_fp16)[name = string("v_head_49_cast_fp16")];
            string dense_output_135_pad_type_1 = const()[name = string("dense_output_135_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_135_strides_1 = const()[name = string("dense_output_135_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_135_pad_1 = const()[name = string("dense_output_135_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_135_dilations_1 = const()[name = string("dense_output_135_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_135_groups_1 = const()[name = string("dense_output_135_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48861248))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48992384))))[name = string("layers_1_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_135_cast_fp16 = conv(dilations = dense_output_135_dilations_1, groups = dense_output_135_groups_1, pad = dense_output_135_pad_1, pad_type = dense_output_135_pad_type_1, strides = dense_output_135_strides_1, weight = layers_1_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_135_cast_fp16")];
            string sparse_output_135_pad_type_1 = const()[name = string("sparse_output_135_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_135_strides_1 = const()[name = string("sparse_output_135_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_135_pad_1 = const()[name = string("sparse_output_135_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_135_dilations_1 = const()[name = string("sparse_output_135_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_135_groups_1 = const()[name = string("sparse_output_135_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48995648))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48992960))))[name = string("layers_1_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_135_cast_fp16 = conv(dilations = sparse_output_135_dilations_1, groups = sparse_output_135_groups_1, pad = sparse_output_135_pad_1, pad_type = sparse_output_135_pad_type_1, strides = sparse_output_135_strides_1, weight = layers_1_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_135_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_2179_cast_fp16 = add(x = dense_output_135_cast_fp16, y = sparse_output_135_cast_fp16)[name = string("op_2179_cast_fp16")];
            tensor<int32, [4]> var_2180 = const()[name = string("op_2180"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2182 = const()[name = string("op_2182"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_2181_cast_fp16 = transpose(perm = var_2180, x = var_2179_cast_fp16)[name = string("transpose_920")];
            tensor<fp16, [1, 101, 128]> p_head_49_cast_fp16 = reshape(shape = var_2182, x = var_2181_cast_fp16)[name = string("p_head_49_cast_fp16")];
            tensor<fp16, [128]> var_2184_to_fp16 = const()[name = string("op_2184_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49012096)))];
            tensor<fp16, [1, 51, 128]> var_2185_cast_fp16 = add(x = q_head_25_cast_fp16, y = var_2184_to_fp16)[name = string("op_2185_cast_fp16")];
            tensor<int32, [1]> q_u_25_axes_1 = const()[name = string("q_u_25_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_25_cast_fp16 = expand_dims(axes = q_u_25_axes_1, x = var_2185_cast_fp16)[name = string("q_u_25_cast_fp16")];
            tensor<fp16, [128]> var_2187_to_fp16 = const()[name = string("op_2187_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49012416)))];
            tensor<fp16, [1, 51, 128]> var_2188_cast_fp16 = add(x = q_head_25_cast_fp16, y = var_2187_to_fp16)[name = string("op_2188_cast_fp16")];
            tensor<int32, [1]> q_v_25_axes_1 = const()[name = string("q_v_25_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_25_cast_fp16 = expand_dims(axes = q_v_25_axes_1, x = var_2188_cast_fp16)[name = string("q_v_25_cast_fp16")];
            tensor<int32, [1]> k_head_51_axes_1 = const()[name = string("k_head_51_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_51_cast_fp16 = expand_dims(axes = k_head_51_axes_1, x = k_head_49_cast_fp16)[name = string("k_head_51_cast_fp16")];
            tensor<int32, [1]> v_head_51_axes_1 = const()[name = string("v_head_51_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_51_cast_fp16 = expand_dims(axes = v_head_51_axes_1, x = v_head_49_cast_fp16)[name = string("v_head_51_cast_fp16")];
            tensor<int32, [1]> p_head_51_axes_1 = const()[name = string("p_head_51_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_51_cast_fp16 = expand_dims(axes = p_head_51_axes_1, x = p_head_49_cast_fp16)[name = string("p_head_51_cast_fp16")];
            bool var_2194_transpose_x_3 = const()[name = string("op_2194_transpose_x_3"), val = bool(false)];
            bool var_2194_transpose_y_3 = const()[name = string("op_2194_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_2194_cast_fp16 = matmul(transpose_x = var_2194_transpose_x_3, transpose_y = var_2194_transpose_y_3, x = q_u_25_cast_fp16, y = k_head_51_cast_fp16)[name = string("op_2194_cast_fp16")];
            fp16 var_2195_to_fp16 = const()[name = string("op_2195_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_25_cast_fp16 = mul(x = var_2194_cast_fp16, y = var_2195_to_fp16)[name = string("scores_content_25_cast_fp16")];
            bool x_137_transpose_x_3 = const()[name = string("x_137_transpose_x_3"), val = bool(false)];
            bool x_137_transpose_y_3 = const()[name = string("x_137_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_3, transpose_y = x_137_transpose_y_3, x = q_v_25_cast_fp16, y = p_head_51_cast_fp16)[name = string("x_137_cast_fp16")];
            tensor<int32, [8]> x_139_pad_1 = const()[name = string("x_139_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_139_mode_1 = const()[name = string("x_139_mode_1"), val = string("constant")];
            fp16 const_1387_to_fp16 = const()[name = string("const_1387_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_139_cast_fp16 = pad(constant_val = const_1387_to_fp16, mode = x_139_mode_1, pad = x_139_pad_1, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")];
            tensor<int32, [4]> var_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_141_cast_fp16 = reshape(shape = var_2209, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<int32, [4]> var_2213_begin_1 = const()[name = string("op_2213_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2213_end_1 = const()[name = string("op_2213_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_2213_end_mask_1 = const()[name = string("op_2213_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_2213_cast_fp16 = slice_by_index(begin = var_2213_begin_1, end = var_2213_end_1, end_mask = var_2213_end_mask_1, x = x_141_cast_fp16)[name = string("op_2213_cast_fp16")];
            tensor<int32, [4]> var_2215 = const()[name = string("op_2215"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_2216_cast_fp16 = reshape(shape = var_2215, x = var_2213_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<int32, [4]> var_2221_begin_1 = const()[name = string("op_2221_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2221_end_1 = const()[name = string("op_2221_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_2221_end_mask_1 = const()[name = string("op_2221_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_2221_cast_fp16 = slice_by_index(begin = var_2221_begin_1, end = var_2221_end_1, end_mask = var_2221_end_mask_1, x = var_2216_cast_fp16)[name = string("op_2221_cast_fp16")];
            fp16 var_2222_to_fp16 = const()[name = string("op_2222_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_25_cast_fp16 = mul(x = var_2221_cast_fp16, y = var_2222_to_fp16)[name = string("scores_pos_25_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_25_cast_fp16 = add(x = scores_content_25_cast_fp16, y = scores_pos_25_cast_fp16)[name = string("logits_25_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_2225_cast_fp16 = softmax(axis = var_1466, x = logits_25_cast_fp16)[name = string("op_2225_cast_fp16")];
            bool var_2227_transpose_x_1 = const()[name = string("op_2227_transpose_x_1"), val = bool(false)];
            bool var_2227_transpose_y_1 = const()[name = string("op_2227_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_2227_cast_fp16 = matmul(transpose_x = var_2227_transpose_x_1, transpose_y = var_2227_transpose_y_1, x = var_2225_cast_fp16, y = v_head_51_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<int32, [1]> var_2228_axes_1 = const()[name = string("op_2228_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_2228_cast_fp16 = squeeze(axes = var_2228_axes_1, x = var_2227_cast_fp16)[name = string("op_2228_cast_fp16")];
            string dense_output_137_pad_type_1 = const()[name = string("dense_output_137_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_137_strides_1 = const()[name = string("dense_output_137_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_137_pad_1 = const()[name = string("dense_output_137_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_137_dilations_1 = const()[name = string("dense_output_137_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_137_groups_1 = const()[name = string("dense_output_137_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49012736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49143872))))[name = string("layers_1_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_137_cast_fp16 = conv(dilations = dense_output_137_dilations_1, groups = dense_output_137_groups_1, pad = dense_output_137_pad_1, pad_type = dense_output_137_pad_type_1, strides = dense_output_137_strides_1, weight = layers_1_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_137_cast_fp16")];
            string sparse_output_137_pad_type_1 = const()[name = string("sparse_output_137_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_137_strides_1 = const()[name = string("sparse_output_137_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_137_pad_1 = const()[name = string("sparse_output_137_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_137_dilations_1 = const()[name = string("sparse_output_137_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_137_groups_1 = const()[name = string("sparse_output_137_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49147136))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49144448))))[name = string("layers_1_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_137_cast_fp16 = conv(dilations = sparse_output_137_dilations_1, groups = sparse_output_137_groups_1, pad = sparse_output_137_pad_1, pad_type = sparse_output_137_pad_type_1, strides = sparse_output_137_strides_1, weight = layers_1_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_137_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2243_cast_fp16 = add(x = dense_output_137_cast_fp16, y = sparse_output_137_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [4]> var_2244 = const()[name = string("op_2244"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2246 = const()[name = string("op_2246"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2245_cast_fp16 = transpose(perm = var_2244, x = var_2243_cast_fp16)[name = string("transpose_919")];
            tensor<fp16, [1, 51, 128]> q_head_27_cast_fp16 = reshape(shape = var_2246, x = var_2245_cast_fp16)[name = string("q_head_27_cast_fp16")];
            string dense_output_139_pad_type_1 = const()[name = string("dense_output_139_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_139_strides_1 = const()[name = string("dense_output_139_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_139_pad_1 = const()[name = string("dense_output_139_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_139_dilations_1 = const()[name = string("dense_output_139_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_139_groups_1 = const()[name = string("dense_output_139_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49163584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49294720))))[name = string("layers_1_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_139_cast_fp16 = conv(dilations = dense_output_139_dilations_1, groups = dense_output_139_groups_1, pad = dense_output_139_pad_1, pad_type = dense_output_139_pad_type_1, strides = dense_output_139_strides_1, weight = layers_1_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_139_cast_fp16")];
            string sparse_output_139_pad_type_1 = const()[name = string("sparse_output_139_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_139_strides_1 = const()[name = string("sparse_output_139_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_139_pad_1 = const()[name = string("sparse_output_139_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_139_dilations_1 = const()[name = string("sparse_output_139_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_139_groups_1 = const()[name = string("sparse_output_139_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49297984))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49295296))))[name = string("layers_1_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_139_cast_fp16 = conv(dilations = sparse_output_139_dilations_1, groups = sparse_output_139_groups_1, pad = sparse_output_139_pad_1, pad_type = sparse_output_139_pad_type_1, strides = sparse_output_139_strides_1, weight = layers_1_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_139_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2262_cast_fp16 = add(x = dense_output_139_cast_fp16, y = sparse_output_139_cast_fp16)[name = string("op_2262_cast_fp16")];
            tensor<int32, [4]> var_2263 = const()[name = string("op_2263"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2265 = const()[name = string("op_2265"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2264_cast_fp16 = transpose(perm = var_2263, x = var_2262_cast_fp16)[name = string("transpose_918")];
            tensor<fp16, [1, 51, 128]> k_head_53_cast_fp16 = reshape(shape = var_2265, x = var_2264_cast_fp16)[name = string("k_head_53_cast_fp16")];
            string dense_output_141_pad_type_1 = const()[name = string("dense_output_141_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_141_strides_1 = const()[name = string("dense_output_141_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_141_pad_1 = const()[name = string("dense_output_141_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_141_dilations_1 = const()[name = string("dense_output_141_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_141_groups_1 = const()[name = string("dense_output_141_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49314432))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49445568))))[name = string("layers_1_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_141_cast_fp16 = conv(dilations = dense_output_141_dilations_1, groups = dense_output_141_groups_1, pad = dense_output_141_pad_1, pad_type = dense_output_141_pad_type_1, strides = dense_output_141_strides_1, weight = layers_1_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_141_cast_fp16")];
            string sparse_output_141_pad_type_1 = const()[name = string("sparse_output_141_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_141_strides_1 = const()[name = string("sparse_output_141_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_141_pad_1 = const()[name = string("sparse_output_141_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_141_dilations_1 = const()[name = string("sparse_output_141_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_141_groups_1 = const()[name = string("sparse_output_141_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49448832))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49446144))))[name = string("layers_1_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_141_cast_fp16 = conv(dilations = sparse_output_141_dilations_1, groups = sparse_output_141_groups_1, pad = sparse_output_141_pad_1, pad_type = sparse_output_141_pad_type_1, strides = sparse_output_141_strides_1, weight = layers_1_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_141_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2281_cast_fp16 = add(x = dense_output_141_cast_fp16, y = sparse_output_141_cast_fp16)[name = string("op_2281_cast_fp16")];
            tensor<int32, [4]> var_2282 = const()[name = string("op_2282"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2283_cast_fp16 = transpose(perm = var_2282, x = var_2281_cast_fp16)[name = string("transpose_917")];
            tensor<fp16, [1, 51, 128]> v_head_53_cast_fp16 = reshape(shape = var_2284, x = var_2283_cast_fp16)[name = string("v_head_53_cast_fp16")];
            string dense_output_143_pad_type_1 = const()[name = string("dense_output_143_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_143_strides_1 = const()[name = string("dense_output_143_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_143_pad_1 = const()[name = string("dense_output_143_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_143_dilations_1 = const()[name = string("dense_output_143_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_143_groups_1 = const()[name = string("dense_output_143_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49465280))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49596416))))[name = string("layers_1_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_143_cast_fp16 = conv(dilations = dense_output_143_dilations_1, groups = dense_output_143_groups_1, pad = dense_output_143_pad_1, pad_type = dense_output_143_pad_type_1, strides = dense_output_143_strides_1, weight = layers_1_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_143_cast_fp16")];
            string sparse_output_143_pad_type_1 = const()[name = string("sparse_output_143_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_143_strides_1 = const()[name = string("sparse_output_143_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_143_pad_1 = const()[name = string("sparse_output_143_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_143_dilations_1 = const()[name = string("sparse_output_143_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_143_groups_1 = const()[name = string("sparse_output_143_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49599680))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49596992))))[name = string("layers_1_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_143_cast_fp16 = conv(dilations = sparse_output_143_dilations_1, groups = sparse_output_143_groups_1, pad = sparse_output_143_pad_1, pad_type = sparse_output_143_pad_type_1, strides = sparse_output_143_strides_1, weight = layers_1_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_143_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_2300_cast_fp16 = add(x = dense_output_143_cast_fp16, y = sparse_output_143_cast_fp16)[name = string("op_2300_cast_fp16")];
            tensor<int32, [4]> var_2301 = const()[name = string("op_2301"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2303 = const()[name = string("op_2303"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_2302_cast_fp16 = transpose(perm = var_2301, x = var_2300_cast_fp16)[name = string("transpose_916")];
            tensor<fp16, [1, 101, 128]> p_head_53_cast_fp16 = reshape(shape = var_2303, x = var_2302_cast_fp16)[name = string("p_head_53_cast_fp16")];
            tensor<fp16, [128]> var_2305_to_fp16 = const()[name = string("op_2305_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49616128)))];
            tensor<fp16, [1, 51, 128]> var_2306_cast_fp16 = add(x = q_head_27_cast_fp16, y = var_2305_to_fp16)[name = string("op_2306_cast_fp16")];
            tensor<int32, [1]> q_u_27_axes_1 = const()[name = string("q_u_27_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_27_cast_fp16 = expand_dims(axes = q_u_27_axes_1, x = var_2306_cast_fp16)[name = string("q_u_27_cast_fp16")];
            tensor<fp16, [128]> var_2308_to_fp16 = const()[name = string("op_2308_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49616448)))];
            tensor<fp16, [1, 51, 128]> var_2309_cast_fp16 = add(x = q_head_27_cast_fp16, y = var_2308_to_fp16)[name = string("op_2309_cast_fp16")];
            tensor<int32, [1]> q_v_27_axes_1 = const()[name = string("q_v_27_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_27_cast_fp16 = expand_dims(axes = q_v_27_axes_1, x = var_2309_cast_fp16)[name = string("q_v_27_cast_fp16")];
            tensor<int32, [1]> k_head_55_axes_1 = const()[name = string("k_head_55_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_55_cast_fp16 = expand_dims(axes = k_head_55_axes_1, x = k_head_53_cast_fp16)[name = string("k_head_55_cast_fp16")];
            tensor<int32, [1]> v_head_55_axes_1 = const()[name = string("v_head_55_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_55_cast_fp16 = expand_dims(axes = v_head_55_axes_1, x = v_head_53_cast_fp16)[name = string("v_head_55_cast_fp16")];
            tensor<int32, [1]> p_head_55_axes_1 = const()[name = string("p_head_55_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_55_cast_fp16 = expand_dims(axes = p_head_55_axes_1, x = p_head_53_cast_fp16)[name = string("p_head_55_cast_fp16")];
            bool var_2315_transpose_x_3 = const()[name = string("op_2315_transpose_x_3"), val = bool(false)];
            bool var_2315_transpose_y_3 = const()[name = string("op_2315_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_2315_cast_fp16 = matmul(transpose_x = var_2315_transpose_x_3, transpose_y = var_2315_transpose_y_3, x = q_u_27_cast_fp16, y = k_head_55_cast_fp16)[name = string("op_2315_cast_fp16")];
            fp16 var_2316_to_fp16 = const()[name = string("op_2316_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_27_cast_fp16 = mul(x = var_2315_cast_fp16, y = var_2316_to_fp16)[name = string("scores_content_27_cast_fp16")];
            bool x_145_transpose_x_3 = const()[name = string("x_145_transpose_x_3"), val = bool(false)];
            bool x_145_transpose_y_3 = const()[name = string("x_145_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_145_cast_fp16 = matmul(transpose_x = x_145_transpose_x_3, transpose_y = x_145_transpose_y_3, x = q_v_27_cast_fp16, y = p_head_55_cast_fp16)[name = string("x_145_cast_fp16")];
            tensor<int32, [8]> x_147_pad_1 = const()[name = string("x_147_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_147_mode_1 = const()[name = string("x_147_mode_1"), val = string("constant")];
            fp16 const_1393_to_fp16 = const()[name = string("const_1393_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_147_cast_fp16 = pad(constant_val = const_1393_to_fp16, mode = x_147_mode_1, pad = x_147_pad_1, x = x_145_cast_fp16)[name = string("x_147_cast_fp16")];
            tensor<int32, [4]> var_2330 = const()[name = string("op_2330"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_149_cast_fp16 = reshape(shape = var_2330, x = x_147_cast_fp16)[name = string("x_149_cast_fp16")];
            tensor<int32, [4]> var_2334_begin_1 = const()[name = string("op_2334_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2334_end_1 = const()[name = string("op_2334_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_2334_end_mask_1 = const()[name = string("op_2334_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_2334_cast_fp16 = slice_by_index(begin = var_2334_begin_1, end = var_2334_end_1, end_mask = var_2334_end_mask_1, x = x_149_cast_fp16)[name = string("op_2334_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_2337_cast_fp16 = reshape(shape = var_2336, x = var_2334_cast_fp16)[name = string("op_2337_cast_fp16")];
            tensor<int32, [4]> var_2342_begin_1 = const()[name = string("op_2342_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2342_end_1 = const()[name = string("op_2342_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_2342_end_mask_1 = const()[name = string("op_2342_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_2342_cast_fp16 = slice_by_index(begin = var_2342_begin_1, end = var_2342_end_1, end_mask = var_2342_end_mask_1, x = var_2337_cast_fp16)[name = string("op_2342_cast_fp16")];
            fp16 var_2343_to_fp16 = const()[name = string("op_2343_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_27_cast_fp16 = mul(x = var_2342_cast_fp16, y = var_2343_to_fp16)[name = string("scores_pos_27_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_27_cast_fp16 = add(x = scores_content_27_cast_fp16, y = scores_pos_27_cast_fp16)[name = string("logits_27_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_2346_cast_fp16 = softmax(axis = var_1466, x = logits_27_cast_fp16)[name = string("op_2346_cast_fp16")];
            bool var_2348_transpose_x_1 = const()[name = string("op_2348_transpose_x_1"), val = bool(false)];
            bool var_2348_transpose_y_1 = const()[name = string("op_2348_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_2348_cast_fp16 = matmul(transpose_x = var_2348_transpose_x_1, transpose_y = var_2348_transpose_y_1, x = var_2346_cast_fp16, y = v_head_55_cast_fp16)[name = string("op_2348_cast_fp16")];
            tensor<int32, [1]> var_2349_axes_1 = const()[name = string("op_2349_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_2349_cast_fp16 = squeeze(axes = var_2349_axes_1, x = var_2348_cast_fp16)[name = string("op_2349_cast_fp16")];
            string dense_output_145_pad_type_1 = const()[name = string("dense_output_145_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_145_strides_1 = const()[name = string("dense_output_145_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_145_pad_1 = const()[name = string("dense_output_145_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_145_dilations_1 = const()[name = string("dense_output_145_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_145_groups_1 = const()[name = string("dense_output_145_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49616768))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49747904))))[name = string("layers_1_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_145_cast_fp16 = conv(dilations = dense_output_145_dilations_1, groups = dense_output_145_groups_1, pad = dense_output_145_pad_1, pad_type = dense_output_145_pad_type_1, strides = dense_output_145_strides_1, weight = layers_1_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_145_cast_fp16")];
            string sparse_output_145_pad_type_1 = const()[name = string("sparse_output_145_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_145_strides_1 = const()[name = string("sparse_output_145_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_145_pad_1 = const()[name = string("sparse_output_145_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_145_dilations_1 = const()[name = string("sparse_output_145_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_145_groups_1 = const()[name = string("sparse_output_145_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49751168))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49748480))))[name = string("layers_1_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_145_cast_fp16 = conv(dilations = sparse_output_145_dilations_1, groups = sparse_output_145_groups_1, pad = sparse_output_145_pad_1, pad_type = sparse_output_145_pad_type_1, strides = sparse_output_145_strides_1, weight = layers_1_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_145_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2364_cast_fp16 = add(x = dense_output_145_cast_fp16, y = sparse_output_145_cast_fp16)[name = string("op_2364_cast_fp16")];
            tensor<int32, [4]> var_2365 = const()[name = string("op_2365"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2367 = const()[name = string("op_2367"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2366_cast_fp16 = transpose(perm = var_2365, x = var_2364_cast_fp16)[name = string("transpose_915")];
            tensor<fp16, [1, 51, 128]> q_head_29_cast_fp16 = reshape(shape = var_2367, x = var_2366_cast_fp16)[name = string("q_head_29_cast_fp16")];
            string dense_output_147_pad_type_1 = const()[name = string("dense_output_147_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_147_strides_1 = const()[name = string("dense_output_147_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_147_pad_1 = const()[name = string("dense_output_147_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_147_dilations_1 = const()[name = string("dense_output_147_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_147_groups_1 = const()[name = string("dense_output_147_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49767616))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49898752))))[name = string("layers_1_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_147_cast_fp16 = conv(dilations = dense_output_147_dilations_1, groups = dense_output_147_groups_1, pad = dense_output_147_pad_1, pad_type = dense_output_147_pad_type_1, strides = dense_output_147_strides_1, weight = layers_1_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_147_cast_fp16")];
            string sparse_output_147_pad_type_1 = const()[name = string("sparse_output_147_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_147_strides_1 = const()[name = string("sparse_output_147_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_147_pad_1 = const()[name = string("sparse_output_147_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_147_dilations_1 = const()[name = string("sparse_output_147_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_147_groups_1 = const()[name = string("sparse_output_147_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49902016))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49899328))))[name = string("layers_1_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_147_cast_fp16 = conv(dilations = sparse_output_147_dilations_1, groups = sparse_output_147_groups_1, pad = sparse_output_147_pad_1, pad_type = sparse_output_147_pad_type_1, strides = sparse_output_147_strides_1, weight = layers_1_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_147_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2383_cast_fp16 = add(x = dense_output_147_cast_fp16, y = sparse_output_147_cast_fp16)[name = string("op_2383_cast_fp16")];
            tensor<int32, [4]> var_2384 = const()[name = string("op_2384"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2385_cast_fp16 = transpose(perm = var_2384, x = var_2383_cast_fp16)[name = string("transpose_914")];
            tensor<fp16, [1, 51, 128]> k_head_57_cast_fp16 = reshape(shape = var_2386, x = var_2385_cast_fp16)[name = string("k_head_57_cast_fp16")];
            string dense_output_149_pad_type_1 = const()[name = string("dense_output_149_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_149_strides_1 = const()[name = string("dense_output_149_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_149_pad_1 = const()[name = string("dense_output_149_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_149_dilations_1 = const()[name = string("dense_output_149_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_149_groups_1 = const()[name = string("dense_output_149_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49918464))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50049600))))[name = string("layers_1_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_149_cast_fp16 = conv(dilations = dense_output_149_dilations_1, groups = dense_output_149_groups_1, pad = dense_output_149_pad_1, pad_type = dense_output_149_pad_type_1, strides = dense_output_149_strides_1, weight = layers_1_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_149_cast_fp16")];
            string sparse_output_149_pad_type_1 = const()[name = string("sparse_output_149_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_149_strides_1 = const()[name = string("sparse_output_149_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_149_pad_1 = const()[name = string("sparse_output_149_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_149_dilations_1 = const()[name = string("sparse_output_149_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_149_groups_1 = const()[name = string("sparse_output_149_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50052864))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50050176))))[name = string("layers_1_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_149_cast_fp16 = conv(dilations = sparse_output_149_dilations_1, groups = sparse_output_149_groups_1, pad = sparse_output_149_pad_1, pad_type = sparse_output_149_pad_type_1, strides = sparse_output_149_strides_1, weight = layers_1_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_149_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2402_cast_fp16 = add(x = dense_output_149_cast_fp16, y = sparse_output_149_cast_fp16)[name = string("op_2402_cast_fp16")];
            tensor<int32, [4]> var_2403 = const()[name = string("op_2403"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2405 = const()[name = string("op_2405"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2404_cast_fp16 = transpose(perm = var_2403, x = var_2402_cast_fp16)[name = string("transpose_913")];
            tensor<fp16, [1, 51, 128]> v_head_57_cast_fp16 = reshape(shape = var_2405, x = var_2404_cast_fp16)[name = string("v_head_57_cast_fp16")];
            string dense_output_151_pad_type_1 = const()[name = string("dense_output_151_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_151_strides_1 = const()[name = string("dense_output_151_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_151_pad_1 = const()[name = string("dense_output_151_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_151_dilations_1 = const()[name = string("dense_output_151_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_151_groups_1 = const()[name = string("dense_output_151_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50069312))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50200448))))[name = string("layers_1_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_151_cast_fp16 = conv(dilations = dense_output_151_dilations_1, groups = dense_output_151_groups_1, pad = dense_output_151_pad_1, pad_type = dense_output_151_pad_type_1, strides = dense_output_151_strides_1, weight = layers_1_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_151_cast_fp16")];
            string sparse_output_151_pad_type_1 = const()[name = string("sparse_output_151_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_151_strides_1 = const()[name = string("sparse_output_151_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_151_pad_1 = const()[name = string("sparse_output_151_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_151_dilations_1 = const()[name = string("sparse_output_151_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_151_groups_1 = const()[name = string("sparse_output_151_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50203712))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50201024))))[name = string("layers_1_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_151_cast_fp16 = conv(dilations = sparse_output_151_dilations_1, groups = sparse_output_151_groups_1, pad = sparse_output_151_pad_1, pad_type = sparse_output_151_pad_type_1, strides = sparse_output_151_strides_1, weight = layers_1_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_151_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_2421_cast_fp16 = add(x = dense_output_151_cast_fp16, y = sparse_output_151_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<int32, [4]> var_2422 = const()[name = string("op_2422"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2424 = const()[name = string("op_2424"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_2423_cast_fp16 = transpose(perm = var_2422, x = var_2421_cast_fp16)[name = string("transpose_912")];
            tensor<fp16, [1, 101, 128]> p_head_57_cast_fp16 = reshape(shape = var_2424, x = var_2423_cast_fp16)[name = string("p_head_57_cast_fp16")];
            tensor<fp16, [128]> var_2426_to_fp16 = const()[name = string("op_2426_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50220160)))];
            tensor<fp16, [1, 51, 128]> var_2427_cast_fp16 = add(x = q_head_29_cast_fp16, y = var_2426_to_fp16)[name = string("op_2427_cast_fp16")];
            tensor<int32, [1]> q_u_29_axes_1 = const()[name = string("q_u_29_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_29_cast_fp16 = expand_dims(axes = q_u_29_axes_1, x = var_2427_cast_fp16)[name = string("q_u_29_cast_fp16")];
            tensor<fp16, [128]> var_2429_to_fp16 = const()[name = string("op_2429_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50220480)))];
            tensor<fp16, [1, 51, 128]> var_2430_cast_fp16 = add(x = q_head_29_cast_fp16, y = var_2429_to_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [1]> q_v_29_axes_1 = const()[name = string("q_v_29_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_29_cast_fp16 = expand_dims(axes = q_v_29_axes_1, x = var_2430_cast_fp16)[name = string("q_v_29_cast_fp16")];
            tensor<int32, [1]> k_head_59_axes_1 = const()[name = string("k_head_59_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_59_cast_fp16 = expand_dims(axes = k_head_59_axes_1, x = k_head_57_cast_fp16)[name = string("k_head_59_cast_fp16")];
            tensor<int32, [1]> v_head_59_axes_1 = const()[name = string("v_head_59_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_59_cast_fp16 = expand_dims(axes = v_head_59_axes_1, x = v_head_57_cast_fp16)[name = string("v_head_59_cast_fp16")];
            tensor<int32, [1]> p_head_59_axes_1 = const()[name = string("p_head_59_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_59_cast_fp16 = expand_dims(axes = p_head_59_axes_1, x = p_head_57_cast_fp16)[name = string("p_head_59_cast_fp16")];
            bool var_2436_transpose_x_3 = const()[name = string("op_2436_transpose_x_3"), val = bool(false)];
            bool var_2436_transpose_y_3 = const()[name = string("op_2436_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_2436_cast_fp16 = matmul(transpose_x = var_2436_transpose_x_3, transpose_y = var_2436_transpose_y_3, x = q_u_29_cast_fp16, y = k_head_59_cast_fp16)[name = string("op_2436_cast_fp16")];
            fp16 var_2437_to_fp16 = const()[name = string("op_2437_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_29_cast_fp16 = mul(x = var_2436_cast_fp16, y = var_2437_to_fp16)[name = string("scores_content_29_cast_fp16")];
            bool x_153_transpose_x_3 = const()[name = string("x_153_transpose_x_3"), val = bool(false)];
            bool x_153_transpose_y_3 = const()[name = string("x_153_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_153_cast_fp16 = matmul(transpose_x = x_153_transpose_x_3, transpose_y = x_153_transpose_y_3, x = q_v_29_cast_fp16, y = p_head_59_cast_fp16)[name = string("x_153_cast_fp16")];
            tensor<int32, [8]> x_155_pad_1 = const()[name = string("x_155_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_155_mode_1 = const()[name = string("x_155_mode_1"), val = string("constant")];
            fp16 const_1399_to_fp16 = const()[name = string("const_1399_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_155_cast_fp16 = pad(constant_val = const_1399_to_fp16, mode = x_155_mode_1, pad = x_155_pad_1, x = x_153_cast_fp16)[name = string("x_155_cast_fp16")];
            tensor<int32, [4]> var_2451 = const()[name = string("op_2451"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_157_cast_fp16 = reshape(shape = var_2451, x = x_155_cast_fp16)[name = string("x_157_cast_fp16")];
            tensor<int32, [4]> var_2455_begin_1 = const()[name = string("op_2455_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2455_end_1 = const()[name = string("op_2455_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_2455_end_mask_1 = const()[name = string("op_2455_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_2455_cast_fp16 = slice_by_index(begin = var_2455_begin_1, end = var_2455_end_1, end_mask = var_2455_end_mask_1, x = x_157_cast_fp16)[name = string("op_2455_cast_fp16")];
            tensor<int32, [4]> var_2457 = const()[name = string("op_2457"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_2458_cast_fp16 = reshape(shape = var_2457, x = var_2455_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<int32, [4]> var_2463_begin_1 = const()[name = string("op_2463_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2463_end_1 = const()[name = string("op_2463_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_2463_end_mask_1 = const()[name = string("op_2463_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_2463_cast_fp16 = slice_by_index(begin = var_2463_begin_1, end = var_2463_end_1, end_mask = var_2463_end_mask_1, x = var_2458_cast_fp16)[name = string("op_2463_cast_fp16")];
            fp16 var_2464_to_fp16 = const()[name = string("op_2464_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_29_cast_fp16 = mul(x = var_2463_cast_fp16, y = var_2464_to_fp16)[name = string("scores_pos_29_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_29_cast_fp16 = add(x = scores_content_29_cast_fp16, y = scores_pos_29_cast_fp16)[name = string("logits_29_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_2467_cast_fp16 = softmax(axis = var_1466, x = logits_29_cast_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_transpose_x_1 = const()[name = string("op_2469_transpose_x_1"), val = bool(false)];
            bool var_2469_transpose_y_1 = const()[name = string("op_2469_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_2469_cast_fp16 = matmul(transpose_x = var_2469_transpose_x_1, transpose_y = var_2469_transpose_y_1, x = var_2467_cast_fp16, y = v_head_59_cast_fp16)[name = string("op_2469_cast_fp16")];
            tensor<int32, [1]> var_2470_axes_1 = const()[name = string("op_2470_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_2470_cast_fp16 = squeeze(axes = var_2470_axes_1, x = var_2469_cast_fp16)[name = string("op_2470_cast_fp16")];
            string dense_output_153_pad_type_1 = const()[name = string("dense_output_153_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_153_strides_1 = const()[name = string("dense_output_153_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_153_pad_1 = const()[name = string("dense_output_153_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_153_dilations_1 = const()[name = string("dense_output_153_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_153_groups_1 = const()[name = string("dense_output_153_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50220800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50351936))))[name = string("layers_1_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_153_cast_fp16 = conv(dilations = dense_output_153_dilations_1, groups = dense_output_153_groups_1, pad = dense_output_153_pad_1, pad_type = dense_output_153_pad_type_1, strides = dense_output_153_strides_1, weight = layers_1_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_153_cast_fp16")];
            string sparse_output_153_pad_type_1 = const()[name = string("sparse_output_153_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_153_strides_1 = const()[name = string("sparse_output_153_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_153_pad_1 = const()[name = string("sparse_output_153_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_153_dilations_1 = const()[name = string("sparse_output_153_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_153_groups_1 = const()[name = string("sparse_output_153_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50355200))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50352512))))[name = string("layers_1_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_153_cast_fp16 = conv(dilations = sparse_output_153_dilations_1, groups = sparse_output_153_groups_1, pad = sparse_output_153_pad_1, pad_type = sparse_output_153_pad_type_1, strides = sparse_output_153_strides_1, weight = layers_1_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_153_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2485_cast_fp16 = add(x = dense_output_153_cast_fp16, y = sparse_output_153_cast_fp16)[name = string("op_2485_cast_fp16")];
            tensor<int32, [4]> var_2486 = const()[name = string("op_2486"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2488 = const()[name = string("op_2488"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2487_cast_fp16 = transpose(perm = var_2486, x = var_2485_cast_fp16)[name = string("transpose_911")];
            tensor<fp16, [1, 51, 128]> q_head_31_cast_fp16 = reshape(shape = var_2488, x = var_2487_cast_fp16)[name = string("q_head_31_cast_fp16")];
            string dense_output_155_pad_type_1 = const()[name = string("dense_output_155_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_155_strides_1 = const()[name = string("dense_output_155_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_155_pad_1 = const()[name = string("dense_output_155_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_155_dilations_1 = const()[name = string("dense_output_155_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_155_groups_1 = const()[name = string("dense_output_155_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50371648))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50502784))))[name = string("layers_1_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_155_cast_fp16 = conv(dilations = dense_output_155_dilations_1, groups = dense_output_155_groups_1, pad = dense_output_155_pad_1, pad_type = dense_output_155_pad_type_1, strides = dense_output_155_strides_1, weight = layers_1_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_155_cast_fp16")];
            string sparse_output_155_pad_type_1 = const()[name = string("sparse_output_155_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_155_strides_1 = const()[name = string("sparse_output_155_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_155_pad_1 = const()[name = string("sparse_output_155_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_155_dilations_1 = const()[name = string("sparse_output_155_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_155_groups_1 = const()[name = string("sparse_output_155_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50506048))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50503360))))[name = string("layers_1_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_155_cast_fp16 = conv(dilations = sparse_output_155_dilations_1, groups = sparse_output_155_groups_1, pad = sparse_output_155_pad_1, pad_type = sparse_output_155_pad_type_1, strides = sparse_output_155_strides_1, weight = layers_1_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_155_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2504_cast_fp16 = add(x = dense_output_155_cast_fp16, y = sparse_output_155_cast_fp16)[name = string("op_2504_cast_fp16")];
            tensor<int32, [4]> var_2505 = const()[name = string("op_2505"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2507 = const()[name = string("op_2507"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2506_cast_fp16 = transpose(perm = var_2505, x = var_2504_cast_fp16)[name = string("transpose_910")];
            tensor<fp16, [1, 51, 128]> k_head_61_cast_fp16 = reshape(shape = var_2507, x = var_2506_cast_fp16)[name = string("k_head_61_cast_fp16")];
            string dense_output_157_pad_type_1 = const()[name = string("dense_output_157_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_157_strides_1 = const()[name = string("dense_output_157_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_157_pad_1 = const()[name = string("dense_output_157_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_157_dilations_1 = const()[name = string("dense_output_157_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_157_groups_1 = const()[name = string("dense_output_157_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50522496))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50653632))))[name = string("layers_1_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_157_cast_fp16 = conv(dilations = dense_output_157_dilations_1, groups = dense_output_157_groups_1, pad = dense_output_157_pad_1, pad_type = dense_output_157_pad_type_1, strides = dense_output_157_strides_1, weight = layers_1_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("dense_output_157_cast_fp16")];
            string sparse_output_157_pad_type_1 = const()[name = string("sparse_output_157_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_157_strides_1 = const()[name = string("sparse_output_157_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_157_pad_1 = const()[name = string("sparse_output_157_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_157_dilations_1 = const()[name = string("sparse_output_157_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_157_groups_1 = const()[name = string("sparse_output_157_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50656896))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50654208))))[name = string("layers_1_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_157_cast_fp16 = conv(dilations = sparse_output_157_dilations_1, groups = sparse_output_157_groups_1, pad = sparse_output_157_pad_1, pad_type = sparse_output_157_pad_type_1, strides = sparse_output_157_strides_1, weight = layers_1_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_75_cast_fp16)[name = string("sparse_output_157_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2523_cast_fp16 = add(x = dense_output_157_cast_fp16, y = sparse_output_157_cast_fp16)[name = string("op_2523_cast_fp16")];
            tensor<int32, [4]> var_2524 = const()[name = string("op_2524"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2526 = const()[name = string("op_2526"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2525_cast_fp16 = transpose(perm = var_2524, x = var_2523_cast_fp16)[name = string("transpose_909")];
            tensor<fp16, [1, 51, 128]> v_head_61_cast_fp16 = reshape(shape = var_2526, x = var_2525_cast_fp16)[name = string("v_head_61_cast_fp16")];
            string dense_output_159_pad_type_1 = const()[name = string("dense_output_159_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_159_strides_1 = const()[name = string("dense_output_159_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_159_pad_1 = const()[name = string("dense_output_159_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_159_dilations_1 = const()[name = string("dense_output_159_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_159_groups_1 = const()[name = string("dense_output_159_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50673344))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50804480))))[name = string("layers_1_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_159_cast_fp16 = conv(dilations = dense_output_159_dilations_1, groups = dense_output_159_groups_1, pad = dense_output_159_pad_1, pad_type = dense_output_159_pad_type_1, strides = dense_output_159_strides_1, weight = layers_1_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_159_cast_fp16")];
            string sparse_output_159_pad_type_1 = const()[name = string("sparse_output_159_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_159_strides_1 = const()[name = string("sparse_output_159_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_159_pad_1 = const()[name = string("sparse_output_159_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_159_dilations_1 = const()[name = string("sparse_output_159_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_159_groups_1 = const()[name = string("sparse_output_159_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_1_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50807744))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50805056))))[name = string("layers_1_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_159_cast_fp16 = conv(dilations = sparse_output_159_dilations_1, groups = sparse_output_159_groups_1, pad = sparse_output_159_pad_1, pad_type = sparse_output_159_pad_type_1, strides = sparse_output_159_strides_1, weight = layers_1_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_159_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_2542_cast_fp16 = add(x = dense_output_159_cast_fp16, y = sparse_output_159_cast_fp16)[name = string("op_2542_cast_fp16")];
            tensor<int32, [4]> var_2543 = const()[name = string("op_2543"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2545 = const()[name = string("op_2545"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_2544_cast_fp16 = transpose(perm = var_2543, x = var_2542_cast_fp16)[name = string("transpose_908")];
            tensor<fp16, [1, 101, 128]> p_head_61_cast_fp16 = reshape(shape = var_2545, x = var_2544_cast_fp16)[name = string("p_head_61_cast_fp16")];
            tensor<fp16, [128]> var_2547_to_fp16 = const()[name = string("op_2547_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50824192)))];
            tensor<fp16, [1, 51, 128]> var_2548_cast_fp16 = add(x = q_head_31_cast_fp16, y = var_2547_to_fp16)[name = string("op_2548_cast_fp16")];
            tensor<int32, [1]> q_u_31_axes_1 = const()[name = string("q_u_31_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_31_cast_fp16 = expand_dims(axes = q_u_31_axes_1, x = var_2548_cast_fp16)[name = string("q_u_31_cast_fp16")];
            tensor<fp16, [128]> var_2550_to_fp16 = const()[name = string("op_2550_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50824512)))];
            tensor<fp16, [1, 51, 128]> var_2551_cast_fp16 = add(x = q_head_31_cast_fp16, y = var_2550_to_fp16)[name = string("op_2551_cast_fp16")];
            tensor<int32, [1]> q_v_31_axes_1 = const()[name = string("q_v_31_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_31_cast_fp16 = expand_dims(axes = q_v_31_axes_1, x = var_2551_cast_fp16)[name = string("q_v_31_cast_fp16")];
            tensor<int32, [1]> k_head_63_axes_1 = const()[name = string("k_head_63_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_63_cast_fp16 = expand_dims(axes = k_head_63_axes_1, x = k_head_61_cast_fp16)[name = string("k_head_63_cast_fp16")];
            tensor<int32, [1]> v_head_63_axes_1 = const()[name = string("v_head_63_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_63_cast_fp16 = expand_dims(axes = v_head_63_axes_1, x = v_head_61_cast_fp16)[name = string("v_head_63_cast_fp16")];
            tensor<int32, [1]> p_head_63_axes_1 = const()[name = string("p_head_63_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_63_cast_fp16 = expand_dims(axes = p_head_63_axes_1, x = p_head_61_cast_fp16)[name = string("p_head_63_cast_fp16")];
            bool var_2557_transpose_x_3 = const()[name = string("op_2557_transpose_x_3"), val = bool(false)];
            bool var_2557_transpose_y_3 = const()[name = string("op_2557_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_2557_cast_fp16 = matmul(transpose_x = var_2557_transpose_x_3, transpose_y = var_2557_transpose_y_3, x = q_u_31_cast_fp16, y = k_head_63_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp16 var_2558_to_fp16 = const()[name = string("op_2558_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_31_cast_fp16 = mul(x = var_2557_cast_fp16, y = var_2558_to_fp16)[name = string("scores_content_31_cast_fp16")];
            bool x_161_transpose_x_3 = const()[name = string("x_161_transpose_x_3"), val = bool(false)];
            bool x_161_transpose_y_3 = const()[name = string("x_161_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_161_cast_fp16 = matmul(transpose_x = x_161_transpose_x_3, transpose_y = x_161_transpose_y_3, x = q_v_31_cast_fp16, y = p_head_63_cast_fp16)[name = string("x_161_cast_fp16")];
            tensor<int32, [8]> x_163_pad_1 = const()[name = string("x_163_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_163_mode_1 = const()[name = string("x_163_mode_1"), val = string("constant")];
            fp16 const_1405_to_fp16 = const()[name = string("const_1405_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_163_cast_fp16 = pad(constant_val = const_1405_to_fp16, mode = x_163_mode_1, pad = x_163_pad_1, x = x_161_cast_fp16)[name = string("x_163_cast_fp16")];
            tensor<int32, [4]> var_2572 = const()[name = string("op_2572"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_165_cast_fp16 = reshape(shape = var_2572, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")];
            tensor<int32, [4]> var_2576_begin_1 = const()[name = string("op_2576_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2576_end_1 = const()[name = string("op_2576_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_2576_end_mask_1 = const()[name = string("op_2576_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_2576_cast_fp16 = slice_by_index(begin = var_2576_begin_1, end = var_2576_end_1, end_mask = var_2576_end_mask_1, x = x_165_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [4]> var_2578 = const()[name = string("op_2578"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_2579_cast_fp16 = reshape(shape = var_2578, x = var_2576_cast_fp16)[name = string("op_2579_cast_fp16")];
            tensor<int32, [4]> var_2584_begin_1 = const()[name = string("op_2584_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2584_end_1 = const()[name = string("op_2584_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_2584_end_mask_1 = const()[name = string("op_2584_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_2584_cast_fp16 = slice_by_index(begin = var_2584_begin_1, end = var_2584_end_1, end_mask = var_2584_end_mask_1, x = var_2579_cast_fp16)[name = string("op_2584_cast_fp16")];
            fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_31_cast_fp16 = mul(x = var_2584_cast_fp16, y = var_2585_to_fp16)[name = string("scores_pos_31_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_31_cast_fp16 = add(x = scores_content_31_cast_fp16, y = scores_pos_31_cast_fp16)[name = string("logits_31_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_2588_cast_fp16 = softmax(axis = var_1466, x = logits_31_cast_fp16)[name = string("op_2588_cast_fp16")];
            bool var_2590_transpose_x_1 = const()[name = string("op_2590_transpose_x_1"), val = bool(false)];
            bool var_2590_transpose_y_1 = const()[name = string("op_2590_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_2590_cast_fp16 = matmul(transpose_x = var_2590_transpose_x_1, transpose_y = var_2590_transpose_y_1, x = var_2588_cast_fp16, y = v_head_63_cast_fp16)[name = string("op_2590_cast_fp16")];
            tensor<int32, [1]> o_head_3_axes_1 = const()[name = string("o_head_3_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_3_cast_fp16 = squeeze(axes = o_head_3_axes_1, x = var_2590_cast_fp16)[name = string("o_head_3_cast_fp16")];
            bool out_3_interleave_1 = const()[name = string("out_3_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_3_cast_fp16 = concat(axis = var_1466, interleave = out_3_interleave_1, values = (var_1744_cast_fp16, var_1865_cast_fp16, var_1986_cast_fp16, var_2107_cast_fp16, var_2228_cast_fp16, var_2349_cast_fp16, var_2470_cast_fp16, o_head_3_cast_fp16))[name = string("out_3_cast_fp16")];
            tensor<int32, [3]> var_2594_perm_1 = const()[name = string("op_2594_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_83_axes_1 = const()[name = string("input_83_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_2594_cast_fp16 = transpose(perm = var_2594_perm_1, x = out_3_cast_fp16)[name = string("transpose_907")];
            tensor<fp16, [1, 1024, 51, 1]> input_83_cast_fp16 = expand_dims(axes = input_83_axes_1, x = var_2594_cast_fp16)[name = string("input_83_cast_fp16")];
            string dense_output_161_pad_type_1 = const()[name = string("dense_output_161_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_161_strides_1 = const()[name = string("dense_output_161_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_161_pad_1 = const()[name = string("dense_output_161_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_161_dilations_1 = const()[name = string("dense_output_161_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_161_groups_1 = const()[name = string("dense_output_161_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50824832))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51873472))))[name = string("layers_1_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_161_cast_fp16 = conv(dilations = dense_output_161_dilations_1, groups = dense_output_161_groups_1, pad = dense_output_161_pad_1, pad_type = dense_output_161_pad_type_1, strides = dense_output_161_strides_1, weight = layers_1_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_83_cast_fp16)[name = string("dense_output_161_cast_fp16")];
            string sparse_output_161_pad_type_1 = const()[name = string("sparse_output_161_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_161_strides_1 = const()[name = string("sparse_output_161_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_161_pad_1 = const()[name = string("sparse_output_161_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_161_dilations_1 = const()[name = string("sparse_output_161_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_161_groups_1 = const()[name = string("sparse_output_161_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51895104))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51874048))))[name = string("layers_1_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_161_cast_fp16 = conv(dilations = sparse_output_161_dilations_1, groups = sparse_output_161_groups_1, pad = sparse_output_161_pad_1, pad_type = sparse_output_161_pad_type_1, strides = sparse_output_161_strides_1, weight = layers_1_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_83_cast_fp16)[name = string("sparse_output_161_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_3_cast_fp16 = add(x = dense_output_161_cast_fp16, y = sparse_output_161_cast_fp16)[name = string("out_conv_3_cast_fp16")];
            tensor<int32, [1]> var_2611_axes_1 = const()[name = string("op_2611_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_2611_cast_fp16 = squeeze(axes = var_2611_axes_1, x = out_conv_3_cast_fp16)[name = string("op_2611_cast_fp16")];
            tensor<int32, [3]> var_2612_perm_1 = const()[name = string("op_2612_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_2612_cast_fp16 = transpose(perm = var_2612_perm_1, x = var_2611_cast_fp16)[name = string("transpose_906")];
            tensor<fp16, [1, 51, 1024]> input_85_cast_fp16 = add(x = input_73_cast_fp16, y = var_2612_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<int32, [1]> x_169_axes_1 = const()[name = string("x_169_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_1_norm_conv_weight_to_fp16 = const()[name = string("layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52026240)))];
            tensor<fp16, [1024]> layers_1_norm_conv_bias_to_fp16 = const()[name = string("layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52028352)))];
            tensor<fp16, [1, 51, 1024]> x_169_cast_fp16 = layer_norm(axes = x_169_axes_1, beta = layers_1_norm_conv_bias_to_fp16, epsilon = var_1481_to_fp16, gamma = layers_1_norm_conv_weight_to_fp16, x = input_85_cast_fp16)[name = string("x_169_cast_fp16")];
            tensor<int32, [3]> var_2622_perm_1 = const()[name = string("op_2622_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_87_axes_1 = const()[name = string("input_87_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_2622_cast_fp16 = transpose(perm = var_2622_perm_1, x = x_169_cast_fp16)[name = string("transpose_905")];
            tensor<fp16, [1, 1024, 51, 1]> input_87_cast_fp16 = expand_dims(axes = input_87_axes_1, x = var_2622_cast_fp16)[name = string("input_87_cast_fp16")];
            string dense_output_163_pad_type_1 = const()[name = string("dense_output_163_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_163_strides_1 = const()[name = string("dense_output_163_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_163_pad_1 = const()[name = string("dense_output_163_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_163_dilations_1 = const()[name = string("dense_output_163_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_163_groups_1 = const()[name = string("dense_output_163_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_1_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52030464))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54127680))))[name = string("layers_1_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_163_cast_fp16 = conv(dilations = dense_output_163_dilations_1, groups = dense_output_163_groups_1, pad = dense_output_163_pad_1, pad_type = dense_output_163_pad_type_1, strides = dense_output_163_strides_1, weight = layers_1_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = string("dense_output_163_cast_fp16")];
            string sparse_output_163_pad_type_1 = const()[name = string("sparse_output_163_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_163_strides_1 = const()[name = string("sparse_output_163_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_163_pad_1 = const()[name = string("sparse_output_163_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_163_dilations_1 = const()[name = string("sparse_output_163_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_163_groups_1 = const()[name = string("sparse_output_163_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_1_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54170304))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54128256))))[name = string("layers_1_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_163_cast_fp16 = conv(dilations = sparse_output_163_dilations_1, groups = sparse_output_163_groups_1, pad = sparse_output_163_pad_1, pad_type = sparse_output_163_pad_type_1, strides = sparse_output_163_strides_1, weight = layers_1_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_87_cast_fp16)[name = string("sparse_output_163_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_89_cast_fp16 = add(x = dense_output_163_cast_fp16, y = sparse_output_163_cast_fp16)[name = string("input_89_cast_fp16")];
            int32 input_91_split_num_splits_1 = const()[name = string("input_91_split_num_splits_1"), val = int32(2)];
            int32 input_91_split_axis_1 = const()[name = string("input_91_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_91_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_91_split_cast_fp16_1 = split(axis = input_91_split_axis_1, num_splits = input_91_split_num_splits_1, x = input_89_cast_fp16)[name = string("input_91_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_91_split_1_sigmoid_cast_fp16 = sigmoid(x = input_91_split_cast_fp16_1)[name = string("input_91_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_91_cast_fp16 = mul(x = input_91_split_cast_fp16_0, y = input_91_split_1_sigmoid_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<int32, [8]> input_93_pad_1 = const()[name = string("input_93_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_93_mode_1 = const()[name = string("input_93_mode_1"), val = string("constant")];
            fp16 const_1407_to_fp16 = const()[name = string("const_1407_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_93_cast_fp16 = pad(constant_val = const_1407_to_fp16, mode = input_93_mode_1, pad = input_93_pad_1, x = input_91_cast_fp16)[name = string("input_93_cast_fp16")];
            string dense_output_165_pad_type_1 = const()[name = string("dense_output_165_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_165_strides_1 = const()[name = string("dense_output_165_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_165_pad_1 = const()[name = string("dense_output_165_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_165_dilations_1 = const()[name = string("dense_output_165_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_165_groups_1 = const()[name = string("dense_output_165_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_165_cast_fp16 = conv(dilations = dense_output_165_dilations_1, groups = dense_output_165_groups_1, pad = dense_output_165_pad_1, pad_type = dense_output_165_pad_type_1, strides = dense_output_165_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_93_cast_fp16)[name = string("dense_output_165_cast_fp16")];
            string sparse_output_165_pad_type_1 = const()[name = string("sparse_output_165_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_165_strides_1 = const()[name = string("sparse_output_165_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_165_pad_1 = const()[name = string("sparse_output_165_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_165_dilations_1 = const()[name = string("sparse_output_165_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_165_groups_1 = const()[name = string("sparse_output_165_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_1_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54432512))))[name = string("layers_1_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_165_cast_fp16 = conv(dilations = sparse_output_165_dilations_1, groups = sparse_output_165_groups_1, pad = sparse_output_165_pad_1, pad_type = sparse_output_165_pad_type_1, strides = sparse_output_165_strides_1, weight = layers_1_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_93_cast_fp16)[name = string("sparse_output_165_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_95_cast_fp16 = add(x = dense_output_165_cast_fp16, y = sparse_output_165_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1024]> layers_1_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_1_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54451008)))];
            tensor<fp16, [1024]> layers_1_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_1_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54453120)))];
            tensor<fp16, [1024]> layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54455232)))];
            tensor<fp16, [1024]> layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54457344)))];
            tensor<fp16, [1, 1024, 51, 1]> input_97_cast_fp16 = batch_norm(beta = layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_1481_to_fp16, gamma = layers_1_conv_batch_norm_weight_to_fp16, mean = layers_1_conv_batch_norm_running_mean_to_fp16, variance = layers_1_conv_batch_norm_running_var_to_fp16, x = input_95_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_99_cast_fp16 = silu(x = input_97_cast_fp16)[name = string("input_99_cast_fp16")];
            string dense_output_167_pad_type_1 = const()[name = string("dense_output_167_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_167_strides_1 = const()[name = string("dense_output_167_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_167_pad_1 = const()[name = string("dense_output_167_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_167_dilations_1 = const()[name = string("dense_output_167_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_167_groups_1 = const()[name = string("dense_output_167_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54459456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55508096))))[name = string("layers_1_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_167_cast_fp16 = conv(dilations = dense_output_167_dilations_1, groups = dense_output_167_groups_1, pad = dense_output_167_pad_1, pad_type = dense_output_167_pad_type_1, strides = dense_output_167_strides_1, weight = layers_1_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = string("dense_output_167_cast_fp16")];
            string sparse_output_167_pad_type_1 = const()[name = string("sparse_output_167_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_167_strides_1 = const()[name = string("sparse_output_167_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_167_pad_1 = const()[name = string("sparse_output_167_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_167_dilations_1 = const()[name = string("sparse_output_167_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_167_groups_1 = const()[name = string("sparse_output_167_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_1_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55529728))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55508672))))[name = string("layers_1_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_167_cast_fp16 = conv(dilations = sparse_output_167_dilations_1, groups = sparse_output_167_groups_1, pad = sparse_output_167_pad_1, pad_type = sparse_output_167_pad_type_1, strides = sparse_output_167_strides_1, weight = layers_1_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_99_cast_fp16)[name = string("sparse_output_167_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_171_cast_fp16 = add(x = dense_output_167_cast_fp16, y = sparse_output_167_cast_fp16)[name = string("x_171_cast_fp16")];
            tensor<int32, [1]> var_2678_axes_1 = const()[name = string("op_2678_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_2678_cast_fp16 = squeeze(axes = var_2678_axes_1, x = x_171_cast_fp16)[name = string("op_2678_cast_fp16")];
            tensor<int32, [3]> var_2679_perm_1 = const()[name = string("op_2679_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_2679_cast_fp16 = transpose(perm = var_2679_perm_1, x = var_2678_cast_fp16)[name = string("transpose_904")];
            tensor<fp16, [1, 51, 1024]> input_101_cast_fp16 = add(x = input_85_cast_fp16, y = var_2679_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<int32, [1]> x_173_axes_1 = const()[name = string("x_173_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55660864)))];
            tensor<fp16, [1024]> layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55662976)))];
            tensor<fp16, [1, 51, 1024]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_1, beta = layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_1481_to_fp16, gamma = layers_1_norm_feed_forward2_weight_to_fp16, x = input_101_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<int32, [4]> var_2689 = const()[name = string("op_2689"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_175_cast_fp16 = reshape(shape = var_2689, x = x_173_cast_fp16)[name = string("x_175_cast_fp16")];
            tensor<int32, [4]> input_103_perm_1 = const()[name = string("input_103_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_169_pad_type_1 = const()[name = string("dense_output_169_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_169_strides_1 = const()[name = string("dense_output_169_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_169_pad_1 = const()[name = string("dense_output_169_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_169_dilations_1 = const()[name = string("dense_output_169_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_169_groups_1 = const()[name = string("dense_output_169_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_1_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55665088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59859456))))[name = string("layers_1_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_103_cast_fp16 = transpose(perm = input_103_perm_1, x = x_175_cast_fp16)[name = string("transpose_903")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_169_cast_fp16 = conv(dilations = dense_output_169_dilations_1, groups = dense_output_169_groups_1, pad = dense_output_169_pad_1, pad_type = dense_output_169_pad_type_1, strides = dense_output_169_strides_1, weight = layers_1_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_103_cast_fp16)[name = string("dense_output_169_cast_fp16")];
            string sparse_output_169_pad_type_1 = const()[name = string("sparse_output_169_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_169_strides_1 = const()[name = string("sparse_output_169_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_169_pad_1 = const()[name = string("sparse_output_169_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_169_dilations_1 = const()[name = string("sparse_output_169_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_169_groups_1 = const()[name = string("sparse_output_169_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_1_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59944000))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59860032))))[name = string("layers_1_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_169_cast_fp16 = conv(dilations = sparse_output_169_dilations_1, groups = sparse_output_169_groups_1, pad = sparse_output_169_pad_1, pad_type = sparse_output_169_pad_type_1, strides = sparse_output_169_strides_1, weight = layers_1_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_103_cast_fp16)[name = string("sparse_output_169_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_105_cast_fp16 = add(x = dense_output_169_cast_fp16, y = sparse_output_169_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_107_cast_fp16 = silu(x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
            string dense_output_171_pad_type_1 = const()[name = string("dense_output_171_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_171_strides_1 = const()[name = string("dense_output_171_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_171_pad_1 = const()[name = string("dense_output_171_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_171_dilations_1 = const()[name = string("dense_output_171_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_171_groups_1 = const()[name = string("dense_output_171_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_1_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60468352))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64662720))))[name = string("layers_1_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_171_cast_fp16 = conv(dilations = dense_output_171_dilations_1, groups = dense_output_171_groups_1, pad = dense_output_171_pad_1, pad_type = dense_output_171_pad_type_1, strides = dense_output_171_strides_1, weight = layers_1_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = string("dense_output_171_cast_fp16")];
            string sparse_output_171_pad_type_1 = const()[name = string("sparse_output_171_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_171_strides_1 = const()[name = string("sparse_output_171_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_171_pad_1 = const()[name = string("sparse_output_171_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_171_dilations_1 = const()[name = string("sparse_output_171_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_171_groups_1 = const()[name = string("sparse_output_171_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_1_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64747264))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64663296))))[name = string("layers_1_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_171_cast_fp16 = conv(dilations = sparse_output_171_dilations_1, groups = sparse_output_171_groups_1, pad = sparse_output_171_pad_1, pad_type = sparse_output_171_pad_type_1, strides = sparse_output_171_strides_1, weight = layers_1_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_107_cast_fp16)[name = string("sparse_output_171_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_177_cast_fp16 = add(x = dense_output_171_cast_fp16, y = sparse_output_171_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<int32, [4]> x_179_perm_1 = const()[name = string("x_179_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_2724 = const()[name = string("op_2724"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_179_cast_fp16 = transpose(perm = x_179_perm_1, x = x_177_cast_fp16)[name = string("transpose_902")];
            tensor<fp16, [1, 51, 1024]> var_2725_cast_fp16 = reshape(shape = var_2724, x = x_179_cast_fp16)[name = string("op_2725_cast_fp16")];
            fp16 var_2726_to_fp16 = const()[name = string("op_2726_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_2727_cast_fp16 = mul(x = var_2725_cast_fp16, y = var_2726_to_fp16)[name = string("op_2727_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_109_cast_fp16 = add(x = input_101_cast_fp16, y = var_2727_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<int32, [1]> input_111_axes_1 = const()[name = string("input_111_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_1_norm_out_weight_to_fp16 = const()[name = string("layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65271616)))];
            tensor<fp16, [1024]> layers_1_norm_out_bias_to_fp16 = const()[name = string("layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65273728)))];
            tensor<fp16, [1, 51, 1024]> input_111_cast_fp16 = layer_norm(axes = input_111_axes_1, beta = layers_1_norm_out_bias_to_fp16, epsilon = var_1481_to_fp16, gamma = layers_1_norm_out_weight_to_fp16, x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
            int32 var_2735 = const()[name = string("op_2735"), val = int32(-1)];
            tensor<int32, [1]> x_181_axes_1 = const()[name = string("x_181_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65275840)))];
            tensor<fp16, [1024]> layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65277952)))];
            fp16 var_2750_to_fp16 = const()[name = string("op_2750_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_181_cast_fp16 = layer_norm(axes = x_181_axes_1, beta = layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_2750_to_fp16, gamma = layers_2_norm_feed_forward1_weight_to_fp16, x = input_111_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<int32, [4]> var_2769 = const()[name = string("op_2769"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_183_cast_fp16 = reshape(shape = var_2769, x = x_181_cast_fp16)[name = string("x_183_cast_fp16")];
            tensor<int32, [4]> input_113_perm_1 = const()[name = string("input_113_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_173_pad_type_1 = const()[name = string("dense_output_173_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_173_strides_1 = const()[name = string("dense_output_173_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_173_pad_1 = const()[name = string("dense_output_173_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_173_dilations_1 = const()[name = string("dense_output_173_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_173_groups_1 = const()[name = string("dense_output_173_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_2_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65280064))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69474432))))[name = string("layers_2_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_113_cast_fp16 = transpose(perm = input_113_perm_1, x = x_183_cast_fp16)[name = string("transpose_901")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_173_cast_fp16 = conv(dilations = dense_output_173_dilations_1, groups = dense_output_173_groups_1, pad = dense_output_173_pad_1, pad_type = dense_output_173_pad_type_1, strides = dense_output_173_strides_1, weight = layers_2_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = string("dense_output_173_cast_fp16")];
            string sparse_output_173_pad_type_1 = const()[name = string("sparse_output_173_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_173_strides_1 = const()[name = string("sparse_output_173_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_173_pad_1 = const()[name = string("sparse_output_173_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_173_dilations_1 = const()[name = string("sparse_output_173_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_173_groups_1 = const()[name = string("sparse_output_173_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_2_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69558976))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69475008))))[name = string("layers_2_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_173_cast_fp16 = conv(dilations = sparse_output_173_dilations_1, groups = sparse_output_173_groups_1, pad = sparse_output_173_pad_1, pad_type = sparse_output_173_pad_type_1, strides = sparse_output_173_strides_1, weight = layers_2_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_113_cast_fp16)[name = string("sparse_output_173_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_115_cast_fp16 = add(x = dense_output_173_cast_fp16, y = sparse_output_173_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_117_cast_fp16 = silu(x = input_115_cast_fp16)[name = string("input_117_cast_fp16")];
            string dense_output_175_pad_type_1 = const()[name = string("dense_output_175_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_175_strides_1 = const()[name = string("dense_output_175_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_175_pad_1 = const()[name = string("dense_output_175_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_175_dilations_1 = const()[name = string("dense_output_175_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_175_groups_1 = const()[name = string("dense_output_175_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_2_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70083328))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74277696))))[name = string("layers_2_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_175_cast_fp16 = conv(dilations = dense_output_175_dilations_1, groups = dense_output_175_groups_1, pad = dense_output_175_pad_1, pad_type = dense_output_175_pad_type_1, strides = dense_output_175_strides_1, weight = layers_2_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = string("dense_output_175_cast_fp16")];
            string sparse_output_175_pad_type_1 = const()[name = string("sparse_output_175_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_175_strides_1 = const()[name = string("sparse_output_175_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_175_pad_1 = const()[name = string("sparse_output_175_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_175_dilations_1 = const()[name = string("sparse_output_175_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_175_groups_1 = const()[name = string("sparse_output_175_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_2_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74362240))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74278272))))[name = string("layers_2_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_175_cast_fp16 = conv(dilations = sparse_output_175_dilations_1, groups = sparse_output_175_groups_1, pad = sparse_output_175_pad_1, pad_type = sparse_output_175_pad_type_1, strides = sparse_output_175_strides_1, weight = layers_2_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_117_cast_fp16)[name = string("sparse_output_175_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_185_cast_fp16 = add(x = dense_output_175_cast_fp16, y = sparse_output_175_cast_fp16)[name = string("x_185_cast_fp16")];
            tensor<int32, [4]> x_187_perm_1 = const()[name = string("x_187_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_2804 = const()[name = string("op_2804"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_187_cast_fp16 = transpose(perm = x_187_perm_1, x = x_185_cast_fp16)[name = string("transpose_900")];
            tensor<fp16, [1, 51, 1024]> var_2805_cast_fp16 = reshape(shape = var_2804, x = x_187_cast_fp16)[name = string("op_2805_cast_fp16")];
            fp16 var_2806_to_fp16 = const()[name = string("op_2806_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_2807_cast_fp16 = mul(x = var_2805_cast_fp16, y = var_2806_to_fp16)[name = string("op_2807_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_119_cast_fp16 = add(x = input_111_cast_fp16, y = var_2807_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<int32, [1]> q_5_axes_1 = const()[name = string("q_5_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_2_norm_self_att_weight_to_fp16 = const()[name = string("layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74886592)))];
            tensor<fp16, [1024]> layers_2_norm_self_att_bias_to_fp16 = const()[name = string("layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74888704)))];
            tensor<fp16, [1, 51, 1024]> q_5_cast_fp16 = layer_norm(axes = q_5_axes_1, beta = layers_2_norm_self_att_bias_to_fp16, epsilon = var_2750_to_fp16, gamma = layers_2_norm_self_att_weight_to_fp16, x = input_119_cast_fp16)[name = string("q_5_cast_fp16")];
            tensor<int32, [3]> var_2881 = const()[name = string("op_2881"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_121_axes_1 = const()[name = string("input_121_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_2882_cast_fp16 = transpose(perm = var_2881, x = q_5_cast_fp16)[name = string("transpose_899")];
            tensor<fp16, [1, 1024, 51, 1]> input_121_cast_fp16 = expand_dims(axes = input_121_axes_1, x = var_2882_cast_fp16)[name = string("input_121_cast_fp16")];
            string dense_output_177_pad_type_1 = const()[name = string("dense_output_177_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_177_strides_1 = const()[name = string("dense_output_177_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_177_pad_1 = const()[name = string("dense_output_177_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_177_dilations_1 = const()[name = string("dense_output_177_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_177_groups_1 = const()[name = string("dense_output_177_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74890816))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75021952))))[name = string("layers_2_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_177_cast_fp16 = conv(dilations = dense_output_177_dilations_1, groups = dense_output_177_groups_1, pad = dense_output_177_pad_1, pad_type = dense_output_177_pad_type_1, strides = dense_output_177_strides_1, weight = layers_2_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_177_cast_fp16")];
            string sparse_output_177_pad_type_1 = const()[name = string("sparse_output_177_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_177_strides_1 = const()[name = string("sparse_output_177_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_177_pad_1 = const()[name = string("sparse_output_177_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_177_dilations_1 = const()[name = string("sparse_output_177_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_177_groups_1 = const()[name = string("sparse_output_177_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75025216))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75022528))))[name = string("layers_2_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_177_cast_fp16 = conv(dilations = sparse_output_177_dilations_1, groups = sparse_output_177_groups_1, pad = sparse_output_177_pad_1, pad_type = sparse_output_177_pad_type_1, strides = sparse_output_177_strides_1, weight = layers_2_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_177_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2907_cast_fp16 = add(x = dense_output_177_cast_fp16, y = sparse_output_177_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<int32, [4]> var_2908 = const()[name = string("op_2908"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2910 = const()[name = string("op_2910"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2909_cast_fp16 = transpose(perm = var_2908, x = var_2907_cast_fp16)[name = string("transpose_898")];
            tensor<fp16, [1, 51, 128]> q_head_33_cast_fp16 = reshape(shape = var_2910, x = var_2909_cast_fp16)[name = string("q_head_33_cast_fp16")];
            string dense_output_179_pad_type_1 = const()[name = string("dense_output_179_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_179_strides_1 = const()[name = string("dense_output_179_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_179_pad_1 = const()[name = string("dense_output_179_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_179_dilations_1 = const()[name = string("dense_output_179_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_179_groups_1 = const()[name = string("dense_output_179_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75041664))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75172800))))[name = string("layers_2_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_179_cast_fp16 = conv(dilations = dense_output_179_dilations_1, groups = dense_output_179_groups_1, pad = dense_output_179_pad_1, pad_type = dense_output_179_pad_type_1, strides = dense_output_179_strides_1, weight = layers_2_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_179_cast_fp16")];
            string sparse_output_179_pad_type_1 = const()[name = string("sparse_output_179_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_179_strides_1 = const()[name = string("sparse_output_179_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_179_pad_1 = const()[name = string("sparse_output_179_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_179_dilations_1 = const()[name = string("sparse_output_179_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_179_groups_1 = const()[name = string("sparse_output_179_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75176064))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75173376))))[name = string("layers_2_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_179_cast_fp16 = conv(dilations = sparse_output_179_dilations_1, groups = sparse_output_179_groups_1, pad = sparse_output_179_pad_1, pad_type = sparse_output_179_pad_type_1, strides = sparse_output_179_strides_1, weight = layers_2_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_179_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2926_cast_fp16 = add(x = dense_output_179_cast_fp16, y = sparse_output_179_cast_fp16)[name = string("op_2926_cast_fp16")];
            tensor<int32, [4]> var_2927 = const()[name = string("op_2927"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2929 = const()[name = string("op_2929"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2928_cast_fp16 = transpose(perm = var_2927, x = var_2926_cast_fp16)[name = string("transpose_897")];
            tensor<fp16, [1, 51, 128]> k_head_65_cast_fp16 = reshape(shape = var_2929, x = var_2928_cast_fp16)[name = string("k_head_65_cast_fp16")];
            string dense_output_181_pad_type_1 = const()[name = string("dense_output_181_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_181_strides_1 = const()[name = string("dense_output_181_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_181_pad_1 = const()[name = string("dense_output_181_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_181_dilations_1 = const()[name = string("dense_output_181_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_181_groups_1 = const()[name = string("dense_output_181_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75192512))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75323648))))[name = string("layers_2_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_181_cast_fp16 = conv(dilations = dense_output_181_dilations_1, groups = dense_output_181_groups_1, pad = dense_output_181_pad_1, pad_type = dense_output_181_pad_type_1, strides = dense_output_181_strides_1, weight = layers_2_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_181_cast_fp16")];
            string sparse_output_181_pad_type_1 = const()[name = string("sparse_output_181_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_181_strides_1 = const()[name = string("sparse_output_181_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_181_pad_1 = const()[name = string("sparse_output_181_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_181_dilations_1 = const()[name = string("sparse_output_181_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_181_groups_1 = const()[name = string("sparse_output_181_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75326912))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75324224))))[name = string("layers_2_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_181_cast_fp16 = conv(dilations = sparse_output_181_dilations_1, groups = sparse_output_181_groups_1, pad = sparse_output_181_pad_1, pad_type = sparse_output_181_pad_type_1, strides = sparse_output_181_strides_1, weight = layers_2_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_181_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_2945_cast_fp16 = add(x = dense_output_181_cast_fp16, y = sparse_output_181_cast_fp16)[name = string("op_2945_cast_fp16")];
            tensor<int32, [4]> var_2946 = const()[name = string("op_2946"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2948 = const()[name = string("op_2948"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_2947_cast_fp16 = transpose(perm = var_2946, x = var_2945_cast_fp16)[name = string("transpose_896")];
            tensor<fp16, [1, 51, 128]> v_head_65_cast_fp16 = reshape(shape = var_2948, x = var_2947_cast_fp16)[name = string("v_head_65_cast_fp16")];
            string dense_output_183_pad_type_1 = const()[name = string("dense_output_183_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_183_strides_1 = const()[name = string("dense_output_183_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_183_pad_1 = const()[name = string("dense_output_183_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_183_dilations_1 = const()[name = string("dense_output_183_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_183_groups_1 = const()[name = string("dense_output_183_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75343360))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75474496))))[name = string("layers_2_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_183_cast_fp16 = conv(dilations = dense_output_183_dilations_1, groups = dense_output_183_groups_1, pad = dense_output_183_pad_1, pad_type = dense_output_183_pad_type_1, strides = dense_output_183_strides_1, weight = layers_2_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_183_cast_fp16")];
            string sparse_output_183_pad_type_1 = const()[name = string("sparse_output_183_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_183_strides_1 = const()[name = string("sparse_output_183_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_183_pad_1 = const()[name = string("sparse_output_183_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_183_dilations_1 = const()[name = string("sparse_output_183_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_183_groups_1 = const()[name = string("sparse_output_183_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75477760))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75475072))))[name = string("layers_2_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_183_cast_fp16 = conv(dilations = sparse_output_183_dilations_1, groups = sparse_output_183_groups_1, pad = sparse_output_183_pad_1, pad_type = sparse_output_183_pad_type_1, strides = sparse_output_183_strides_1, weight = layers_2_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_183_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_2964_cast_fp16 = add(x = dense_output_183_cast_fp16, y = sparse_output_183_cast_fp16)[name = string("op_2964_cast_fp16")];
            tensor<int32, [4]> var_2965 = const()[name = string("op_2965"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_2967 = const()[name = string("op_2967"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_2966_cast_fp16 = transpose(perm = var_2965, x = var_2964_cast_fp16)[name = string("transpose_895")];
            tensor<fp16, [1, 101, 128]> p_head_65_cast_fp16 = reshape(shape = var_2967, x = var_2966_cast_fp16)[name = string("p_head_65_cast_fp16")];
            tensor<fp16, [128]> var_2969_to_fp16 = const()[name = string("op_2969_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75494208)))];
            tensor<fp16, [1, 51, 128]> var_2970_cast_fp16 = add(x = q_head_33_cast_fp16, y = var_2969_to_fp16)[name = string("op_2970_cast_fp16")];
            tensor<int32, [1]> q_u_33_axes_1 = const()[name = string("q_u_33_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_33_cast_fp16 = expand_dims(axes = q_u_33_axes_1, x = var_2970_cast_fp16)[name = string("q_u_33_cast_fp16")];
            tensor<fp16, [128]> var_2972_to_fp16 = const()[name = string("op_2972_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75494528)))];
            tensor<fp16, [1, 51, 128]> var_2973_cast_fp16 = add(x = q_head_33_cast_fp16, y = var_2972_to_fp16)[name = string("op_2973_cast_fp16")];
            tensor<int32, [1]> q_v_33_axes_1 = const()[name = string("q_v_33_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_33_cast_fp16 = expand_dims(axes = q_v_33_axes_1, x = var_2973_cast_fp16)[name = string("q_v_33_cast_fp16")];
            tensor<int32, [1]> k_head_67_axes_1 = const()[name = string("k_head_67_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_67_cast_fp16 = expand_dims(axes = k_head_67_axes_1, x = k_head_65_cast_fp16)[name = string("k_head_67_cast_fp16")];
            tensor<int32, [1]> v_head_67_axes_1 = const()[name = string("v_head_67_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_67_cast_fp16 = expand_dims(axes = v_head_67_axes_1, x = v_head_65_cast_fp16)[name = string("v_head_67_cast_fp16")];
            tensor<int32, [1]> p_head_67_axes_1 = const()[name = string("p_head_67_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_67_cast_fp16 = expand_dims(axes = p_head_67_axes_1, x = p_head_65_cast_fp16)[name = string("p_head_67_cast_fp16")];
            bool var_2979_transpose_x_3 = const()[name = string("op_2979_transpose_x_3"), val = bool(false)];
            bool var_2979_transpose_y_3 = const()[name = string("op_2979_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_2979_cast_fp16 = matmul(transpose_x = var_2979_transpose_x_3, transpose_y = var_2979_transpose_y_3, x = q_u_33_cast_fp16, y = k_head_67_cast_fp16)[name = string("op_2979_cast_fp16")];
            fp16 var_2980_to_fp16 = const()[name = string("op_2980_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_33_cast_fp16 = mul(x = var_2979_cast_fp16, y = var_2980_to_fp16)[name = string("scores_content_33_cast_fp16")];
            bool x_189_transpose_x_3 = const()[name = string("x_189_transpose_x_3"), val = bool(false)];
            bool x_189_transpose_y_3 = const()[name = string("x_189_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_3, transpose_y = x_189_transpose_y_3, x = q_v_33_cast_fp16, y = p_head_67_cast_fp16)[name = string("x_189_cast_fp16")];
            tensor<int32, [8]> x_191_pad_1 = const()[name = string("x_191_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_191_mode_1 = const()[name = string("x_191_mode_1"), val = string("constant")];
            fp16 const_1417_to_fp16 = const()[name = string("const_1417_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_191_cast_fp16 = pad(constant_val = const_1417_to_fp16, mode = x_191_mode_1, pad = x_191_pad_1, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")];
            tensor<int32, [4]> var_2994 = const()[name = string("op_2994"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_193_cast_fp16 = reshape(shape = var_2994, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")];
            tensor<int32, [4]> var_2998_begin_1 = const()[name = string("op_2998_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2998_end_1 = const()[name = string("op_2998_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_2998_end_mask_1 = const()[name = string("op_2998_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_2998_cast_fp16 = slice_by_index(begin = var_2998_begin_1, end = var_2998_end_1, end_mask = var_2998_end_mask_1, x = x_193_cast_fp16)[name = string("op_2998_cast_fp16")];
            tensor<int32, [4]> var_3000 = const()[name = string("op_3000"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_3001_cast_fp16 = reshape(shape = var_3000, x = var_2998_cast_fp16)[name = string("op_3001_cast_fp16")];
            tensor<int32, [4]> var_3006_begin_1 = const()[name = string("op_3006_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3006_end_1 = const()[name = string("op_3006_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_3006_end_mask_1 = const()[name = string("op_3006_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_3006_cast_fp16 = slice_by_index(begin = var_3006_begin_1, end = var_3006_end_1, end_mask = var_3006_end_mask_1, x = var_3001_cast_fp16)[name = string("op_3006_cast_fp16")];
            fp16 var_3007_to_fp16 = const()[name = string("op_3007_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_33_cast_fp16 = mul(x = var_3006_cast_fp16, y = var_3007_to_fp16)[name = string("scores_pos_33_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_33_cast_fp16 = add(x = scores_content_33_cast_fp16, y = scores_pos_33_cast_fp16)[name = string("logits_33_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_3010_cast_fp16 = softmax(axis = var_2735, x = logits_33_cast_fp16)[name = string("op_3010_cast_fp16")];
            bool var_3012_transpose_x_1 = const()[name = string("op_3012_transpose_x_1"), val = bool(false)];
            bool var_3012_transpose_y_1 = const()[name = string("op_3012_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_3012_cast_fp16 = matmul(transpose_x = var_3012_transpose_x_1, transpose_y = var_3012_transpose_y_1, x = var_3010_cast_fp16, y = v_head_67_cast_fp16)[name = string("op_3012_cast_fp16")];
            tensor<int32, [1]> var_3013_axes_1 = const()[name = string("op_3013_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_3013_cast_fp16 = squeeze(axes = var_3013_axes_1, x = var_3012_cast_fp16)[name = string("op_3013_cast_fp16")];
            string dense_output_185_pad_type_1 = const()[name = string("dense_output_185_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_185_strides_1 = const()[name = string("dense_output_185_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_185_pad_1 = const()[name = string("dense_output_185_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_185_dilations_1 = const()[name = string("dense_output_185_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_185_groups_1 = const()[name = string("dense_output_185_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75494848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75625984))))[name = string("layers_2_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_185_cast_fp16 = conv(dilations = dense_output_185_dilations_1, groups = dense_output_185_groups_1, pad = dense_output_185_pad_1, pad_type = dense_output_185_pad_type_1, strides = dense_output_185_strides_1, weight = layers_2_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_185_cast_fp16")];
            string sparse_output_185_pad_type_1 = const()[name = string("sparse_output_185_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_185_strides_1 = const()[name = string("sparse_output_185_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_185_pad_1 = const()[name = string("sparse_output_185_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_185_dilations_1 = const()[name = string("sparse_output_185_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_185_groups_1 = const()[name = string("sparse_output_185_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75629248))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75626560))))[name = string("layers_2_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_185_cast_fp16 = conv(dilations = sparse_output_185_dilations_1, groups = sparse_output_185_groups_1, pad = sparse_output_185_pad_1, pad_type = sparse_output_185_pad_type_1, strides = sparse_output_185_strides_1, weight = layers_2_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_185_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3028_cast_fp16 = add(x = dense_output_185_cast_fp16, y = sparse_output_185_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> var_3029 = const()[name = string("op_3029"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3031 = const()[name = string("op_3031"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3030_cast_fp16 = transpose(perm = var_3029, x = var_3028_cast_fp16)[name = string("transpose_894")];
            tensor<fp16, [1, 51, 128]> q_head_35_cast_fp16 = reshape(shape = var_3031, x = var_3030_cast_fp16)[name = string("q_head_35_cast_fp16")];
            string dense_output_187_pad_type_1 = const()[name = string("dense_output_187_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_187_strides_1 = const()[name = string("dense_output_187_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_187_pad_1 = const()[name = string("dense_output_187_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_187_dilations_1 = const()[name = string("dense_output_187_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_187_groups_1 = const()[name = string("dense_output_187_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75645696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75776832))))[name = string("layers_2_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_187_cast_fp16 = conv(dilations = dense_output_187_dilations_1, groups = dense_output_187_groups_1, pad = dense_output_187_pad_1, pad_type = dense_output_187_pad_type_1, strides = dense_output_187_strides_1, weight = layers_2_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_187_cast_fp16")];
            string sparse_output_187_pad_type_1 = const()[name = string("sparse_output_187_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_187_strides_1 = const()[name = string("sparse_output_187_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_187_pad_1 = const()[name = string("sparse_output_187_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_187_dilations_1 = const()[name = string("sparse_output_187_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_187_groups_1 = const()[name = string("sparse_output_187_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75780096))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75777408))))[name = string("layers_2_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_187_cast_fp16 = conv(dilations = sparse_output_187_dilations_1, groups = sparse_output_187_groups_1, pad = sparse_output_187_pad_1, pad_type = sparse_output_187_pad_type_1, strides = sparse_output_187_strides_1, weight = layers_2_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_187_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3047_cast_fp16 = add(x = dense_output_187_cast_fp16, y = sparse_output_187_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<int32, [4]> var_3048 = const()[name = string("op_3048"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3050 = const()[name = string("op_3050"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3049_cast_fp16 = transpose(perm = var_3048, x = var_3047_cast_fp16)[name = string("transpose_893")];
            tensor<fp16, [1, 51, 128]> k_head_69_cast_fp16 = reshape(shape = var_3050, x = var_3049_cast_fp16)[name = string("k_head_69_cast_fp16")];
            string dense_output_189_pad_type_1 = const()[name = string("dense_output_189_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_189_strides_1 = const()[name = string("dense_output_189_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_189_pad_1 = const()[name = string("dense_output_189_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_189_dilations_1 = const()[name = string("dense_output_189_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_189_groups_1 = const()[name = string("dense_output_189_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75796544))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75927680))))[name = string("layers_2_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_189_cast_fp16 = conv(dilations = dense_output_189_dilations_1, groups = dense_output_189_groups_1, pad = dense_output_189_pad_1, pad_type = dense_output_189_pad_type_1, strides = dense_output_189_strides_1, weight = layers_2_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_189_cast_fp16")];
            string sparse_output_189_pad_type_1 = const()[name = string("sparse_output_189_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_189_strides_1 = const()[name = string("sparse_output_189_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_189_pad_1 = const()[name = string("sparse_output_189_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_189_dilations_1 = const()[name = string("sparse_output_189_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_189_groups_1 = const()[name = string("sparse_output_189_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75930944))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75928256))))[name = string("layers_2_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_189_cast_fp16 = conv(dilations = sparse_output_189_dilations_1, groups = sparse_output_189_groups_1, pad = sparse_output_189_pad_1, pad_type = sparse_output_189_pad_type_1, strides = sparse_output_189_strides_1, weight = layers_2_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_189_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3066_cast_fp16 = add(x = dense_output_189_cast_fp16, y = sparse_output_189_cast_fp16)[name = string("op_3066_cast_fp16")];
            tensor<int32, [4]> var_3067 = const()[name = string("op_3067"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3069 = const()[name = string("op_3069"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3068_cast_fp16 = transpose(perm = var_3067, x = var_3066_cast_fp16)[name = string("transpose_892")];
            tensor<fp16, [1, 51, 128]> v_head_69_cast_fp16 = reshape(shape = var_3069, x = var_3068_cast_fp16)[name = string("v_head_69_cast_fp16")];
            string dense_output_191_pad_type_1 = const()[name = string("dense_output_191_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_191_strides_1 = const()[name = string("dense_output_191_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_191_pad_1 = const()[name = string("dense_output_191_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_191_dilations_1 = const()[name = string("dense_output_191_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_191_groups_1 = const()[name = string("dense_output_191_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75947392))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76078528))))[name = string("layers_2_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_191_cast_fp16 = conv(dilations = dense_output_191_dilations_1, groups = dense_output_191_groups_1, pad = dense_output_191_pad_1, pad_type = dense_output_191_pad_type_1, strides = dense_output_191_strides_1, weight = layers_2_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_191_cast_fp16")];
            string sparse_output_191_pad_type_1 = const()[name = string("sparse_output_191_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_191_strides_1 = const()[name = string("sparse_output_191_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_191_pad_1 = const()[name = string("sparse_output_191_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_191_dilations_1 = const()[name = string("sparse_output_191_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_191_groups_1 = const()[name = string("sparse_output_191_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76081792))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76079104))))[name = string("layers_2_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_191_cast_fp16 = conv(dilations = sparse_output_191_dilations_1, groups = sparse_output_191_groups_1, pad = sparse_output_191_pad_1, pad_type = sparse_output_191_pad_type_1, strides = sparse_output_191_strides_1, weight = layers_2_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_191_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_3085_cast_fp16 = add(x = dense_output_191_cast_fp16, y = sparse_output_191_cast_fp16)[name = string("op_3085_cast_fp16")];
            tensor<int32, [4]> var_3086 = const()[name = string("op_3086"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3088 = const()[name = string("op_3088"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_3087_cast_fp16 = transpose(perm = var_3086, x = var_3085_cast_fp16)[name = string("transpose_891")];
            tensor<fp16, [1, 101, 128]> p_head_69_cast_fp16 = reshape(shape = var_3088, x = var_3087_cast_fp16)[name = string("p_head_69_cast_fp16")];
            tensor<fp16, [128]> var_3090_to_fp16 = const()[name = string("op_3090_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76098240)))];
            tensor<fp16, [1, 51, 128]> var_3091_cast_fp16 = add(x = q_head_35_cast_fp16, y = var_3090_to_fp16)[name = string("op_3091_cast_fp16")];
            tensor<int32, [1]> q_u_35_axes_1 = const()[name = string("q_u_35_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_35_cast_fp16 = expand_dims(axes = q_u_35_axes_1, x = var_3091_cast_fp16)[name = string("q_u_35_cast_fp16")];
            tensor<fp16, [128]> var_3093_to_fp16 = const()[name = string("op_3093_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76098560)))];
            tensor<fp16, [1, 51, 128]> var_3094_cast_fp16 = add(x = q_head_35_cast_fp16, y = var_3093_to_fp16)[name = string("op_3094_cast_fp16")];
            tensor<int32, [1]> q_v_35_axes_1 = const()[name = string("q_v_35_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_35_cast_fp16 = expand_dims(axes = q_v_35_axes_1, x = var_3094_cast_fp16)[name = string("q_v_35_cast_fp16")];
            tensor<int32, [1]> k_head_71_axes_1 = const()[name = string("k_head_71_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_71_cast_fp16 = expand_dims(axes = k_head_71_axes_1, x = k_head_69_cast_fp16)[name = string("k_head_71_cast_fp16")];
            tensor<int32, [1]> v_head_71_axes_1 = const()[name = string("v_head_71_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_71_cast_fp16 = expand_dims(axes = v_head_71_axes_1, x = v_head_69_cast_fp16)[name = string("v_head_71_cast_fp16")];
            tensor<int32, [1]> p_head_71_axes_1 = const()[name = string("p_head_71_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_71_cast_fp16 = expand_dims(axes = p_head_71_axes_1, x = p_head_69_cast_fp16)[name = string("p_head_71_cast_fp16")];
            bool var_3100_transpose_x_3 = const()[name = string("op_3100_transpose_x_3"), val = bool(false)];
            bool var_3100_transpose_y_3 = const()[name = string("op_3100_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_3100_cast_fp16 = matmul(transpose_x = var_3100_transpose_x_3, transpose_y = var_3100_transpose_y_3, x = q_u_35_cast_fp16, y = k_head_71_cast_fp16)[name = string("op_3100_cast_fp16")];
            fp16 var_3101_to_fp16 = const()[name = string("op_3101_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_35_cast_fp16 = mul(x = var_3100_cast_fp16, y = var_3101_to_fp16)[name = string("scores_content_35_cast_fp16")];
            bool x_197_transpose_x_3 = const()[name = string("x_197_transpose_x_3"), val = bool(false)];
            bool x_197_transpose_y_3 = const()[name = string("x_197_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_197_cast_fp16 = matmul(transpose_x = x_197_transpose_x_3, transpose_y = x_197_transpose_y_3, x = q_v_35_cast_fp16, y = p_head_71_cast_fp16)[name = string("x_197_cast_fp16")];
            tensor<int32, [8]> x_199_pad_1 = const()[name = string("x_199_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_199_mode_1 = const()[name = string("x_199_mode_1"), val = string("constant")];
            fp16 const_1423_to_fp16 = const()[name = string("const_1423_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_199_cast_fp16 = pad(constant_val = const_1423_to_fp16, mode = x_199_mode_1, pad = x_199_pad_1, x = x_197_cast_fp16)[name = string("x_199_cast_fp16")];
            tensor<int32, [4]> var_3115 = const()[name = string("op_3115"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_201_cast_fp16 = reshape(shape = var_3115, x = x_199_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<int32, [4]> var_3119_begin_1 = const()[name = string("op_3119_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3119_end_1 = const()[name = string("op_3119_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_3119_end_mask_1 = const()[name = string("op_3119_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_3119_cast_fp16 = slice_by_index(begin = var_3119_begin_1, end = var_3119_end_1, end_mask = var_3119_end_mask_1, x = x_201_cast_fp16)[name = string("op_3119_cast_fp16")];
            tensor<int32, [4]> var_3121 = const()[name = string("op_3121"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_3122_cast_fp16 = reshape(shape = var_3121, x = var_3119_cast_fp16)[name = string("op_3122_cast_fp16")];
            tensor<int32, [4]> var_3127_begin_1 = const()[name = string("op_3127_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3127_end_1 = const()[name = string("op_3127_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_3127_end_mask_1 = const()[name = string("op_3127_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_3127_cast_fp16 = slice_by_index(begin = var_3127_begin_1, end = var_3127_end_1, end_mask = var_3127_end_mask_1, x = var_3122_cast_fp16)[name = string("op_3127_cast_fp16")];
            fp16 var_3128_to_fp16 = const()[name = string("op_3128_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_35_cast_fp16 = mul(x = var_3127_cast_fp16, y = var_3128_to_fp16)[name = string("scores_pos_35_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_35_cast_fp16 = add(x = scores_content_35_cast_fp16, y = scores_pos_35_cast_fp16)[name = string("logits_35_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_3131_cast_fp16 = softmax(axis = var_2735, x = logits_35_cast_fp16)[name = string("op_3131_cast_fp16")];
            bool var_3133_transpose_x_1 = const()[name = string("op_3133_transpose_x_1"), val = bool(false)];
            bool var_3133_transpose_y_1 = const()[name = string("op_3133_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_3133_cast_fp16 = matmul(transpose_x = var_3133_transpose_x_1, transpose_y = var_3133_transpose_y_1, x = var_3131_cast_fp16, y = v_head_71_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<int32, [1]> var_3134_axes_1 = const()[name = string("op_3134_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_3134_cast_fp16 = squeeze(axes = var_3134_axes_1, x = var_3133_cast_fp16)[name = string("op_3134_cast_fp16")];
            string dense_output_193_pad_type_1 = const()[name = string("dense_output_193_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_193_strides_1 = const()[name = string("dense_output_193_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_193_pad_1 = const()[name = string("dense_output_193_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_193_dilations_1 = const()[name = string("dense_output_193_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_193_groups_1 = const()[name = string("dense_output_193_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76098880))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76230016))))[name = string("layers_2_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_193_cast_fp16 = conv(dilations = dense_output_193_dilations_1, groups = dense_output_193_groups_1, pad = dense_output_193_pad_1, pad_type = dense_output_193_pad_type_1, strides = dense_output_193_strides_1, weight = layers_2_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_193_cast_fp16")];
            string sparse_output_193_pad_type_1 = const()[name = string("sparse_output_193_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_193_strides_1 = const()[name = string("sparse_output_193_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_193_pad_1 = const()[name = string("sparse_output_193_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_193_dilations_1 = const()[name = string("sparse_output_193_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_193_groups_1 = const()[name = string("sparse_output_193_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76233280))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76230592))))[name = string("layers_2_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_193_cast_fp16 = conv(dilations = sparse_output_193_dilations_1, groups = sparse_output_193_groups_1, pad = sparse_output_193_pad_1, pad_type = sparse_output_193_pad_type_1, strides = sparse_output_193_strides_1, weight = layers_2_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_193_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3149_cast_fp16 = add(x = dense_output_193_cast_fp16, y = sparse_output_193_cast_fp16)[name = string("op_3149_cast_fp16")];
            tensor<int32, [4]> var_3150 = const()[name = string("op_3150"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3152 = const()[name = string("op_3152"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3151_cast_fp16 = transpose(perm = var_3150, x = var_3149_cast_fp16)[name = string("transpose_890")];
            tensor<fp16, [1, 51, 128]> q_head_37_cast_fp16 = reshape(shape = var_3152, x = var_3151_cast_fp16)[name = string("q_head_37_cast_fp16")];
            string dense_output_195_pad_type_1 = const()[name = string("dense_output_195_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_195_strides_1 = const()[name = string("dense_output_195_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_195_pad_1 = const()[name = string("dense_output_195_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_195_dilations_1 = const()[name = string("dense_output_195_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_195_groups_1 = const()[name = string("dense_output_195_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76249728))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76380864))))[name = string("layers_2_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_195_cast_fp16 = conv(dilations = dense_output_195_dilations_1, groups = dense_output_195_groups_1, pad = dense_output_195_pad_1, pad_type = dense_output_195_pad_type_1, strides = dense_output_195_strides_1, weight = layers_2_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_195_cast_fp16")];
            string sparse_output_195_pad_type_1 = const()[name = string("sparse_output_195_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_195_strides_1 = const()[name = string("sparse_output_195_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_195_pad_1 = const()[name = string("sparse_output_195_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_195_dilations_1 = const()[name = string("sparse_output_195_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_195_groups_1 = const()[name = string("sparse_output_195_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76384128))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76381440))))[name = string("layers_2_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_195_cast_fp16 = conv(dilations = sparse_output_195_dilations_1, groups = sparse_output_195_groups_1, pad = sparse_output_195_pad_1, pad_type = sparse_output_195_pad_type_1, strides = sparse_output_195_strides_1, weight = layers_2_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_195_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3168_cast_fp16 = add(x = dense_output_195_cast_fp16, y = sparse_output_195_cast_fp16)[name = string("op_3168_cast_fp16")];
            tensor<int32, [4]> var_3169 = const()[name = string("op_3169"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3171 = const()[name = string("op_3171"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3170_cast_fp16 = transpose(perm = var_3169, x = var_3168_cast_fp16)[name = string("transpose_889")];
            tensor<fp16, [1, 51, 128]> k_head_73_cast_fp16 = reshape(shape = var_3171, x = var_3170_cast_fp16)[name = string("k_head_73_cast_fp16")];
            string dense_output_197_pad_type_1 = const()[name = string("dense_output_197_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_197_strides_1 = const()[name = string("dense_output_197_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_197_pad_1 = const()[name = string("dense_output_197_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_197_dilations_1 = const()[name = string("dense_output_197_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_197_groups_1 = const()[name = string("dense_output_197_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76400576))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76531712))))[name = string("layers_2_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_197_cast_fp16 = conv(dilations = dense_output_197_dilations_1, groups = dense_output_197_groups_1, pad = dense_output_197_pad_1, pad_type = dense_output_197_pad_type_1, strides = dense_output_197_strides_1, weight = layers_2_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_197_cast_fp16")];
            string sparse_output_197_pad_type_1 = const()[name = string("sparse_output_197_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_197_strides_1 = const()[name = string("sparse_output_197_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_197_pad_1 = const()[name = string("sparse_output_197_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_197_dilations_1 = const()[name = string("sparse_output_197_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_197_groups_1 = const()[name = string("sparse_output_197_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76534976))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76532288))))[name = string("layers_2_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_197_cast_fp16 = conv(dilations = sparse_output_197_dilations_1, groups = sparse_output_197_groups_1, pad = sparse_output_197_pad_1, pad_type = sparse_output_197_pad_type_1, strides = sparse_output_197_strides_1, weight = layers_2_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_197_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3187_cast_fp16 = add(x = dense_output_197_cast_fp16, y = sparse_output_197_cast_fp16)[name = string("op_3187_cast_fp16")];
            tensor<int32, [4]> var_3188 = const()[name = string("op_3188"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3189_cast_fp16 = transpose(perm = var_3188, x = var_3187_cast_fp16)[name = string("transpose_888")];
            tensor<fp16, [1, 51, 128]> v_head_73_cast_fp16 = reshape(shape = var_3190, x = var_3189_cast_fp16)[name = string("v_head_73_cast_fp16")];
            string dense_output_199_pad_type_1 = const()[name = string("dense_output_199_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_199_strides_1 = const()[name = string("dense_output_199_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_199_pad_1 = const()[name = string("dense_output_199_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_199_dilations_1 = const()[name = string("dense_output_199_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_199_groups_1 = const()[name = string("dense_output_199_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76551424))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76682560))))[name = string("layers_2_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_199_cast_fp16 = conv(dilations = dense_output_199_dilations_1, groups = dense_output_199_groups_1, pad = dense_output_199_pad_1, pad_type = dense_output_199_pad_type_1, strides = dense_output_199_strides_1, weight = layers_2_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_199_cast_fp16")];
            string sparse_output_199_pad_type_1 = const()[name = string("sparse_output_199_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_199_strides_1 = const()[name = string("sparse_output_199_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_199_pad_1 = const()[name = string("sparse_output_199_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_199_dilations_1 = const()[name = string("sparse_output_199_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_199_groups_1 = const()[name = string("sparse_output_199_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76685824))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76683136))))[name = string("layers_2_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_199_cast_fp16 = conv(dilations = sparse_output_199_dilations_1, groups = sparse_output_199_groups_1, pad = sparse_output_199_pad_1, pad_type = sparse_output_199_pad_type_1, strides = sparse_output_199_strides_1, weight = layers_2_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_199_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_3206_cast_fp16 = add(x = dense_output_199_cast_fp16, y = sparse_output_199_cast_fp16)[name = string("op_3206_cast_fp16")];
            tensor<int32, [4]> var_3207 = const()[name = string("op_3207"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3209 = const()[name = string("op_3209"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_3208_cast_fp16 = transpose(perm = var_3207, x = var_3206_cast_fp16)[name = string("transpose_887")];
            tensor<fp16, [1, 101, 128]> p_head_73_cast_fp16 = reshape(shape = var_3209, x = var_3208_cast_fp16)[name = string("p_head_73_cast_fp16")];
            tensor<fp16, [128]> var_3211_to_fp16 = const()[name = string("op_3211_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76702272)))];
            tensor<fp16, [1, 51, 128]> var_3212_cast_fp16 = add(x = q_head_37_cast_fp16, y = var_3211_to_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [1]> q_u_37_axes_1 = const()[name = string("q_u_37_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_37_cast_fp16 = expand_dims(axes = q_u_37_axes_1, x = var_3212_cast_fp16)[name = string("q_u_37_cast_fp16")];
            tensor<fp16, [128]> var_3214_to_fp16 = const()[name = string("op_3214_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76702592)))];
            tensor<fp16, [1, 51, 128]> var_3215_cast_fp16 = add(x = q_head_37_cast_fp16, y = var_3214_to_fp16)[name = string("op_3215_cast_fp16")];
            tensor<int32, [1]> q_v_37_axes_1 = const()[name = string("q_v_37_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_37_cast_fp16 = expand_dims(axes = q_v_37_axes_1, x = var_3215_cast_fp16)[name = string("q_v_37_cast_fp16")];
            tensor<int32, [1]> k_head_75_axes_1 = const()[name = string("k_head_75_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_75_cast_fp16 = expand_dims(axes = k_head_75_axes_1, x = k_head_73_cast_fp16)[name = string("k_head_75_cast_fp16")];
            tensor<int32, [1]> v_head_75_axes_1 = const()[name = string("v_head_75_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_75_cast_fp16 = expand_dims(axes = v_head_75_axes_1, x = v_head_73_cast_fp16)[name = string("v_head_75_cast_fp16")];
            tensor<int32, [1]> p_head_75_axes_1 = const()[name = string("p_head_75_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_75_cast_fp16 = expand_dims(axes = p_head_75_axes_1, x = p_head_73_cast_fp16)[name = string("p_head_75_cast_fp16")];
            bool var_3221_transpose_x_3 = const()[name = string("op_3221_transpose_x_3"), val = bool(false)];
            bool var_3221_transpose_y_3 = const()[name = string("op_3221_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_3221_cast_fp16 = matmul(transpose_x = var_3221_transpose_x_3, transpose_y = var_3221_transpose_y_3, x = q_u_37_cast_fp16, y = k_head_75_cast_fp16)[name = string("op_3221_cast_fp16")];
            fp16 var_3222_to_fp16 = const()[name = string("op_3222_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_37_cast_fp16 = mul(x = var_3221_cast_fp16, y = var_3222_to_fp16)[name = string("scores_content_37_cast_fp16")];
            bool x_205_transpose_x_3 = const()[name = string("x_205_transpose_x_3"), val = bool(false)];
            bool x_205_transpose_y_3 = const()[name = string("x_205_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_205_cast_fp16 = matmul(transpose_x = x_205_transpose_x_3, transpose_y = x_205_transpose_y_3, x = q_v_37_cast_fp16, y = p_head_75_cast_fp16)[name = string("x_205_cast_fp16")];
            tensor<int32, [8]> x_207_pad_1 = const()[name = string("x_207_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_207_mode_1 = const()[name = string("x_207_mode_1"), val = string("constant")];
            fp16 const_1429_to_fp16 = const()[name = string("const_1429_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_207_cast_fp16 = pad(constant_val = const_1429_to_fp16, mode = x_207_mode_1, pad = x_207_pad_1, x = x_205_cast_fp16)[name = string("x_207_cast_fp16")];
            tensor<int32, [4]> var_3236 = const()[name = string("op_3236"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_209_cast_fp16 = reshape(shape = var_3236, x = x_207_cast_fp16)[name = string("x_209_cast_fp16")];
            tensor<int32, [4]> var_3240_begin_1 = const()[name = string("op_3240_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3240_end_1 = const()[name = string("op_3240_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_3240_end_mask_1 = const()[name = string("op_3240_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_3240_cast_fp16 = slice_by_index(begin = var_3240_begin_1, end = var_3240_end_1, end_mask = var_3240_end_mask_1, x = x_209_cast_fp16)[name = string("op_3240_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_3243_cast_fp16 = reshape(shape = var_3242, x = var_3240_cast_fp16)[name = string("op_3243_cast_fp16")];
            tensor<int32, [4]> var_3248_begin_1 = const()[name = string("op_3248_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3248_end_1 = const()[name = string("op_3248_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_3248_end_mask_1 = const()[name = string("op_3248_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_3248_cast_fp16 = slice_by_index(begin = var_3248_begin_1, end = var_3248_end_1, end_mask = var_3248_end_mask_1, x = var_3243_cast_fp16)[name = string("op_3248_cast_fp16")];
            fp16 var_3249_to_fp16 = const()[name = string("op_3249_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_37_cast_fp16 = mul(x = var_3248_cast_fp16, y = var_3249_to_fp16)[name = string("scores_pos_37_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_37_cast_fp16 = add(x = scores_content_37_cast_fp16, y = scores_pos_37_cast_fp16)[name = string("logits_37_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_3252_cast_fp16 = softmax(axis = var_2735, x = logits_37_cast_fp16)[name = string("op_3252_cast_fp16")];
            bool var_3254_transpose_x_1 = const()[name = string("op_3254_transpose_x_1"), val = bool(false)];
            bool var_3254_transpose_y_1 = const()[name = string("op_3254_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_3254_cast_fp16 = matmul(transpose_x = var_3254_transpose_x_1, transpose_y = var_3254_transpose_y_1, x = var_3252_cast_fp16, y = v_head_75_cast_fp16)[name = string("op_3254_cast_fp16")];
            tensor<int32, [1]> var_3255_axes_1 = const()[name = string("op_3255_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_3255_cast_fp16 = squeeze(axes = var_3255_axes_1, x = var_3254_cast_fp16)[name = string("op_3255_cast_fp16")];
            string dense_output_201_pad_type_1 = const()[name = string("dense_output_201_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_201_strides_1 = const()[name = string("dense_output_201_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_201_pad_1 = const()[name = string("dense_output_201_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_201_dilations_1 = const()[name = string("dense_output_201_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_201_groups_1 = const()[name = string("dense_output_201_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76702912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76834048))))[name = string("layers_2_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_201_cast_fp16 = conv(dilations = dense_output_201_dilations_1, groups = dense_output_201_groups_1, pad = dense_output_201_pad_1, pad_type = dense_output_201_pad_type_1, strides = dense_output_201_strides_1, weight = layers_2_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_201_cast_fp16")];
            string sparse_output_201_pad_type_1 = const()[name = string("sparse_output_201_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_201_strides_1 = const()[name = string("sparse_output_201_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_201_pad_1 = const()[name = string("sparse_output_201_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_201_dilations_1 = const()[name = string("sparse_output_201_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_201_groups_1 = const()[name = string("sparse_output_201_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76837312))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76834624))))[name = string("layers_2_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_201_cast_fp16 = conv(dilations = sparse_output_201_dilations_1, groups = sparse_output_201_groups_1, pad = sparse_output_201_pad_1, pad_type = sparse_output_201_pad_type_1, strides = sparse_output_201_strides_1, weight = layers_2_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_201_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3270_cast_fp16 = add(x = dense_output_201_cast_fp16, y = sparse_output_201_cast_fp16)[name = string("op_3270_cast_fp16")];
            tensor<int32, [4]> var_3271 = const()[name = string("op_3271"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3273 = const()[name = string("op_3273"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3272_cast_fp16 = transpose(perm = var_3271, x = var_3270_cast_fp16)[name = string("transpose_886")];
            tensor<fp16, [1, 51, 128]> q_head_39_cast_fp16 = reshape(shape = var_3273, x = var_3272_cast_fp16)[name = string("q_head_39_cast_fp16")];
            string dense_output_203_pad_type_1 = const()[name = string("dense_output_203_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_203_strides_1 = const()[name = string("dense_output_203_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_203_pad_1 = const()[name = string("dense_output_203_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_203_dilations_1 = const()[name = string("dense_output_203_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_203_groups_1 = const()[name = string("dense_output_203_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76853760))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76984896))))[name = string("layers_2_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_203_cast_fp16 = conv(dilations = dense_output_203_dilations_1, groups = dense_output_203_groups_1, pad = dense_output_203_pad_1, pad_type = dense_output_203_pad_type_1, strides = dense_output_203_strides_1, weight = layers_2_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_203_cast_fp16")];
            string sparse_output_203_pad_type_1 = const()[name = string("sparse_output_203_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_203_strides_1 = const()[name = string("sparse_output_203_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_203_pad_1 = const()[name = string("sparse_output_203_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_203_dilations_1 = const()[name = string("sparse_output_203_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_203_groups_1 = const()[name = string("sparse_output_203_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76988160))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76985472))))[name = string("layers_2_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_203_cast_fp16 = conv(dilations = sparse_output_203_dilations_1, groups = sparse_output_203_groups_1, pad = sparse_output_203_pad_1, pad_type = sparse_output_203_pad_type_1, strides = sparse_output_203_strides_1, weight = layers_2_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_203_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3289_cast_fp16 = add(x = dense_output_203_cast_fp16, y = sparse_output_203_cast_fp16)[name = string("op_3289_cast_fp16")];
            tensor<int32, [4]> var_3290 = const()[name = string("op_3290"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3292 = const()[name = string("op_3292"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3291_cast_fp16 = transpose(perm = var_3290, x = var_3289_cast_fp16)[name = string("transpose_885")];
            tensor<fp16, [1, 51, 128]> k_head_77_cast_fp16 = reshape(shape = var_3292, x = var_3291_cast_fp16)[name = string("k_head_77_cast_fp16")];
            string dense_output_205_pad_type_1 = const()[name = string("dense_output_205_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_205_strides_1 = const()[name = string("dense_output_205_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_205_pad_1 = const()[name = string("dense_output_205_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_205_dilations_1 = const()[name = string("dense_output_205_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_205_groups_1 = const()[name = string("dense_output_205_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77004608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77135744))))[name = string("layers_2_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_205_cast_fp16 = conv(dilations = dense_output_205_dilations_1, groups = dense_output_205_groups_1, pad = dense_output_205_pad_1, pad_type = dense_output_205_pad_type_1, strides = dense_output_205_strides_1, weight = layers_2_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_205_cast_fp16")];
            string sparse_output_205_pad_type_1 = const()[name = string("sparse_output_205_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_205_strides_1 = const()[name = string("sparse_output_205_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_205_pad_1 = const()[name = string("sparse_output_205_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_205_dilations_1 = const()[name = string("sparse_output_205_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_205_groups_1 = const()[name = string("sparse_output_205_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77139008))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77136320))))[name = string("layers_2_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_205_cast_fp16 = conv(dilations = sparse_output_205_dilations_1, groups = sparse_output_205_groups_1, pad = sparse_output_205_pad_1, pad_type = sparse_output_205_pad_type_1, strides = sparse_output_205_strides_1, weight = layers_2_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_205_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3308_cast_fp16 = add(x = dense_output_205_cast_fp16, y = sparse_output_205_cast_fp16)[name = string("op_3308_cast_fp16")];
            tensor<int32, [4]> var_3309 = const()[name = string("op_3309"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3311 = const()[name = string("op_3311"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3310_cast_fp16 = transpose(perm = var_3309, x = var_3308_cast_fp16)[name = string("transpose_884")];
            tensor<fp16, [1, 51, 128]> v_head_77_cast_fp16 = reshape(shape = var_3311, x = var_3310_cast_fp16)[name = string("v_head_77_cast_fp16")];
            string dense_output_207_pad_type_1 = const()[name = string("dense_output_207_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_207_strides_1 = const()[name = string("dense_output_207_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_207_pad_1 = const()[name = string("dense_output_207_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_207_dilations_1 = const()[name = string("dense_output_207_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_207_groups_1 = const()[name = string("dense_output_207_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77155456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77286592))))[name = string("layers_2_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_207_cast_fp16 = conv(dilations = dense_output_207_dilations_1, groups = dense_output_207_groups_1, pad = dense_output_207_pad_1, pad_type = dense_output_207_pad_type_1, strides = dense_output_207_strides_1, weight = layers_2_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_207_cast_fp16")];
            string sparse_output_207_pad_type_1 = const()[name = string("sparse_output_207_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_207_strides_1 = const()[name = string("sparse_output_207_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_207_pad_1 = const()[name = string("sparse_output_207_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_207_dilations_1 = const()[name = string("sparse_output_207_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_207_groups_1 = const()[name = string("sparse_output_207_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77289856))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77287168))))[name = string("layers_2_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_207_cast_fp16 = conv(dilations = sparse_output_207_dilations_1, groups = sparse_output_207_groups_1, pad = sparse_output_207_pad_1, pad_type = sparse_output_207_pad_type_1, strides = sparse_output_207_strides_1, weight = layers_2_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_207_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_3327_cast_fp16 = add(x = dense_output_207_cast_fp16, y = sparse_output_207_cast_fp16)[name = string("op_3327_cast_fp16")];
            tensor<int32, [4]> var_3328 = const()[name = string("op_3328"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3330 = const()[name = string("op_3330"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_3329_cast_fp16 = transpose(perm = var_3328, x = var_3327_cast_fp16)[name = string("transpose_883")];
            tensor<fp16, [1, 101, 128]> p_head_77_cast_fp16 = reshape(shape = var_3330, x = var_3329_cast_fp16)[name = string("p_head_77_cast_fp16")];
            tensor<fp16, [128]> var_3332_to_fp16 = const()[name = string("op_3332_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77306304)))];
            tensor<fp16, [1, 51, 128]> var_3333_cast_fp16 = add(x = q_head_39_cast_fp16, y = var_3332_to_fp16)[name = string("op_3333_cast_fp16")];
            tensor<int32, [1]> q_u_39_axes_1 = const()[name = string("q_u_39_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_39_cast_fp16 = expand_dims(axes = q_u_39_axes_1, x = var_3333_cast_fp16)[name = string("q_u_39_cast_fp16")];
            tensor<fp16, [128]> var_3335_to_fp16 = const()[name = string("op_3335_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77306624)))];
            tensor<fp16, [1, 51, 128]> var_3336_cast_fp16 = add(x = q_head_39_cast_fp16, y = var_3335_to_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [1]> q_v_39_axes_1 = const()[name = string("q_v_39_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_39_cast_fp16 = expand_dims(axes = q_v_39_axes_1, x = var_3336_cast_fp16)[name = string("q_v_39_cast_fp16")];
            tensor<int32, [1]> k_head_79_axes_1 = const()[name = string("k_head_79_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_79_cast_fp16 = expand_dims(axes = k_head_79_axes_1, x = k_head_77_cast_fp16)[name = string("k_head_79_cast_fp16")];
            tensor<int32, [1]> v_head_79_axes_1 = const()[name = string("v_head_79_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_79_cast_fp16 = expand_dims(axes = v_head_79_axes_1, x = v_head_77_cast_fp16)[name = string("v_head_79_cast_fp16")];
            tensor<int32, [1]> p_head_79_axes_1 = const()[name = string("p_head_79_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_79_cast_fp16 = expand_dims(axes = p_head_79_axes_1, x = p_head_77_cast_fp16)[name = string("p_head_79_cast_fp16")];
            bool var_3342_transpose_x_3 = const()[name = string("op_3342_transpose_x_3"), val = bool(false)];
            bool var_3342_transpose_y_3 = const()[name = string("op_3342_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_3342_cast_fp16 = matmul(transpose_x = var_3342_transpose_x_3, transpose_y = var_3342_transpose_y_3, x = q_u_39_cast_fp16, y = k_head_79_cast_fp16)[name = string("op_3342_cast_fp16")];
            fp16 var_3343_to_fp16 = const()[name = string("op_3343_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_39_cast_fp16 = mul(x = var_3342_cast_fp16, y = var_3343_to_fp16)[name = string("scores_content_39_cast_fp16")];
            bool x_213_transpose_x_3 = const()[name = string("x_213_transpose_x_3"), val = bool(false)];
            bool x_213_transpose_y_3 = const()[name = string("x_213_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_213_cast_fp16 = matmul(transpose_x = x_213_transpose_x_3, transpose_y = x_213_transpose_y_3, x = q_v_39_cast_fp16, y = p_head_79_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<int32, [8]> x_215_pad_1 = const()[name = string("x_215_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_215_mode_1 = const()[name = string("x_215_mode_1"), val = string("constant")];
            fp16 const_1435_to_fp16 = const()[name = string("const_1435_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_215_cast_fp16 = pad(constant_val = const_1435_to_fp16, mode = x_215_mode_1, pad = x_215_pad_1, x = x_213_cast_fp16)[name = string("x_215_cast_fp16")];
            tensor<int32, [4]> var_3357 = const()[name = string("op_3357"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_217_cast_fp16 = reshape(shape = var_3357, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")];
            tensor<int32, [4]> var_3361_begin_1 = const()[name = string("op_3361_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3361_end_1 = const()[name = string("op_3361_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_3361_end_mask_1 = const()[name = string("op_3361_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_3361_cast_fp16 = slice_by_index(begin = var_3361_begin_1, end = var_3361_end_1, end_mask = var_3361_end_mask_1, x = x_217_cast_fp16)[name = string("op_3361_cast_fp16")];
            tensor<int32, [4]> var_3363 = const()[name = string("op_3363"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_3364_cast_fp16 = reshape(shape = var_3363, x = var_3361_cast_fp16)[name = string("op_3364_cast_fp16")];
            tensor<int32, [4]> var_3369_begin_1 = const()[name = string("op_3369_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3369_end_1 = const()[name = string("op_3369_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_3369_end_mask_1 = const()[name = string("op_3369_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_3369_cast_fp16 = slice_by_index(begin = var_3369_begin_1, end = var_3369_end_1, end_mask = var_3369_end_mask_1, x = var_3364_cast_fp16)[name = string("op_3369_cast_fp16")];
            fp16 var_3370_to_fp16 = const()[name = string("op_3370_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_39_cast_fp16 = mul(x = var_3369_cast_fp16, y = var_3370_to_fp16)[name = string("scores_pos_39_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_39_cast_fp16 = add(x = scores_content_39_cast_fp16, y = scores_pos_39_cast_fp16)[name = string("logits_39_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_3373_cast_fp16 = softmax(axis = var_2735, x = logits_39_cast_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_transpose_x_1 = const()[name = string("op_3375_transpose_x_1"), val = bool(false)];
            bool var_3375_transpose_y_1 = const()[name = string("op_3375_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_3375_cast_fp16 = matmul(transpose_x = var_3375_transpose_x_1, transpose_y = var_3375_transpose_y_1, x = var_3373_cast_fp16, y = v_head_79_cast_fp16)[name = string("op_3375_cast_fp16")];
            tensor<int32, [1]> var_3376_axes_1 = const()[name = string("op_3376_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_3376_cast_fp16 = squeeze(axes = var_3376_axes_1, x = var_3375_cast_fp16)[name = string("op_3376_cast_fp16")];
            string dense_output_209_pad_type_1 = const()[name = string("dense_output_209_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_209_strides_1 = const()[name = string("dense_output_209_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_209_pad_1 = const()[name = string("dense_output_209_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_209_dilations_1 = const()[name = string("dense_output_209_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_209_groups_1 = const()[name = string("dense_output_209_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77306944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77438080))))[name = string("layers_2_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_209_cast_fp16 = conv(dilations = dense_output_209_dilations_1, groups = dense_output_209_groups_1, pad = dense_output_209_pad_1, pad_type = dense_output_209_pad_type_1, strides = dense_output_209_strides_1, weight = layers_2_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_209_cast_fp16")];
            string sparse_output_209_pad_type_1 = const()[name = string("sparse_output_209_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_209_strides_1 = const()[name = string("sparse_output_209_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_209_pad_1 = const()[name = string("sparse_output_209_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_209_dilations_1 = const()[name = string("sparse_output_209_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_209_groups_1 = const()[name = string("sparse_output_209_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77441344))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77438656))))[name = string("layers_2_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_209_cast_fp16 = conv(dilations = sparse_output_209_dilations_1, groups = sparse_output_209_groups_1, pad = sparse_output_209_pad_1, pad_type = sparse_output_209_pad_type_1, strides = sparse_output_209_strides_1, weight = layers_2_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_209_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3391_cast_fp16 = add(x = dense_output_209_cast_fp16, y = sparse_output_209_cast_fp16)[name = string("op_3391_cast_fp16")];
            tensor<int32, [4]> var_3392 = const()[name = string("op_3392"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3394 = const()[name = string("op_3394"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3393_cast_fp16 = transpose(perm = var_3392, x = var_3391_cast_fp16)[name = string("transpose_882")];
            tensor<fp16, [1, 51, 128]> q_head_41_cast_fp16 = reshape(shape = var_3394, x = var_3393_cast_fp16)[name = string("q_head_41_cast_fp16")];
            string dense_output_211_pad_type_1 = const()[name = string("dense_output_211_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_211_strides_1 = const()[name = string("dense_output_211_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_211_pad_1 = const()[name = string("dense_output_211_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_211_dilations_1 = const()[name = string("dense_output_211_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_211_groups_1 = const()[name = string("dense_output_211_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77457792))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77588928))))[name = string("layers_2_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_211_cast_fp16 = conv(dilations = dense_output_211_dilations_1, groups = dense_output_211_groups_1, pad = dense_output_211_pad_1, pad_type = dense_output_211_pad_type_1, strides = dense_output_211_strides_1, weight = layers_2_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_211_cast_fp16")];
            string sparse_output_211_pad_type_1 = const()[name = string("sparse_output_211_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_211_strides_1 = const()[name = string("sparse_output_211_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_211_pad_1 = const()[name = string("sparse_output_211_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_211_dilations_1 = const()[name = string("sparse_output_211_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_211_groups_1 = const()[name = string("sparse_output_211_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77592192))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77589504))))[name = string("layers_2_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_211_cast_fp16 = conv(dilations = sparse_output_211_dilations_1, groups = sparse_output_211_groups_1, pad = sparse_output_211_pad_1, pad_type = sparse_output_211_pad_type_1, strides = sparse_output_211_strides_1, weight = layers_2_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_211_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3410_cast_fp16 = add(x = dense_output_211_cast_fp16, y = sparse_output_211_cast_fp16)[name = string("op_3410_cast_fp16")];
            tensor<int32, [4]> var_3411 = const()[name = string("op_3411"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3413 = const()[name = string("op_3413"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3412_cast_fp16 = transpose(perm = var_3411, x = var_3410_cast_fp16)[name = string("transpose_881")];
            tensor<fp16, [1, 51, 128]> k_head_81_cast_fp16 = reshape(shape = var_3413, x = var_3412_cast_fp16)[name = string("k_head_81_cast_fp16")];
            string dense_output_213_pad_type_1 = const()[name = string("dense_output_213_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_213_strides_1 = const()[name = string("dense_output_213_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_213_pad_1 = const()[name = string("dense_output_213_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_213_dilations_1 = const()[name = string("dense_output_213_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_213_groups_1 = const()[name = string("dense_output_213_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77608640))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77739776))))[name = string("layers_2_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_213_cast_fp16 = conv(dilations = dense_output_213_dilations_1, groups = dense_output_213_groups_1, pad = dense_output_213_pad_1, pad_type = dense_output_213_pad_type_1, strides = dense_output_213_strides_1, weight = layers_2_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_213_cast_fp16")];
            string sparse_output_213_pad_type_1 = const()[name = string("sparse_output_213_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_213_strides_1 = const()[name = string("sparse_output_213_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_213_pad_1 = const()[name = string("sparse_output_213_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_213_dilations_1 = const()[name = string("sparse_output_213_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_213_groups_1 = const()[name = string("sparse_output_213_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77743040))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77740352))))[name = string("layers_2_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_213_cast_fp16 = conv(dilations = sparse_output_213_dilations_1, groups = sparse_output_213_groups_1, pad = sparse_output_213_pad_1, pad_type = sparse_output_213_pad_type_1, strides = sparse_output_213_strides_1, weight = layers_2_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_213_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3429_cast_fp16 = add(x = dense_output_213_cast_fp16, y = sparse_output_213_cast_fp16)[name = string("op_3429_cast_fp16")];
            tensor<int32, [4]> var_3430 = const()[name = string("op_3430"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3432 = const()[name = string("op_3432"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3431_cast_fp16 = transpose(perm = var_3430, x = var_3429_cast_fp16)[name = string("transpose_880")];
            tensor<fp16, [1, 51, 128]> v_head_81_cast_fp16 = reshape(shape = var_3432, x = var_3431_cast_fp16)[name = string("v_head_81_cast_fp16")];
            string dense_output_215_pad_type_1 = const()[name = string("dense_output_215_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_215_strides_1 = const()[name = string("dense_output_215_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_215_pad_1 = const()[name = string("dense_output_215_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_215_dilations_1 = const()[name = string("dense_output_215_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_215_groups_1 = const()[name = string("dense_output_215_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77759488))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77890624))))[name = string("layers_2_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_215_cast_fp16 = conv(dilations = dense_output_215_dilations_1, groups = dense_output_215_groups_1, pad = dense_output_215_pad_1, pad_type = dense_output_215_pad_type_1, strides = dense_output_215_strides_1, weight = layers_2_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_215_cast_fp16")];
            string sparse_output_215_pad_type_1 = const()[name = string("sparse_output_215_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_215_strides_1 = const()[name = string("sparse_output_215_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_215_pad_1 = const()[name = string("sparse_output_215_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_215_dilations_1 = const()[name = string("sparse_output_215_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_215_groups_1 = const()[name = string("sparse_output_215_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77893888))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77891200))))[name = string("layers_2_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_215_cast_fp16 = conv(dilations = sparse_output_215_dilations_1, groups = sparse_output_215_groups_1, pad = sparse_output_215_pad_1, pad_type = sparse_output_215_pad_type_1, strides = sparse_output_215_strides_1, weight = layers_2_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_215_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_3448_cast_fp16 = add(x = dense_output_215_cast_fp16, y = sparse_output_215_cast_fp16)[name = string("op_3448_cast_fp16")];
            tensor<int32, [4]> var_3449 = const()[name = string("op_3449"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3451 = const()[name = string("op_3451"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_3450_cast_fp16 = transpose(perm = var_3449, x = var_3448_cast_fp16)[name = string("transpose_879")];
            tensor<fp16, [1, 101, 128]> p_head_81_cast_fp16 = reshape(shape = var_3451, x = var_3450_cast_fp16)[name = string("p_head_81_cast_fp16")];
            tensor<fp16, [128]> var_3453_to_fp16 = const()[name = string("op_3453_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77910336)))];
            tensor<fp16, [1, 51, 128]> var_3454_cast_fp16 = add(x = q_head_41_cast_fp16, y = var_3453_to_fp16)[name = string("op_3454_cast_fp16")];
            tensor<int32, [1]> q_u_41_axes_1 = const()[name = string("q_u_41_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_41_cast_fp16 = expand_dims(axes = q_u_41_axes_1, x = var_3454_cast_fp16)[name = string("q_u_41_cast_fp16")];
            tensor<fp16, [128]> var_3456_to_fp16 = const()[name = string("op_3456_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77910656)))];
            tensor<fp16, [1, 51, 128]> var_3457_cast_fp16 = add(x = q_head_41_cast_fp16, y = var_3456_to_fp16)[name = string("op_3457_cast_fp16")];
            tensor<int32, [1]> q_v_41_axes_1 = const()[name = string("q_v_41_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_41_cast_fp16 = expand_dims(axes = q_v_41_axes_1, x = var_3457_cast_fp16)[name = string("q_v_41_cast_fp16")];
            tensor<int32, [1]> k_head_83_axes_1 = const()[name = string("k_head_83_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_83_cast_fp16 = expand_dims(axes = k_head_83_axes_1, x = k_head_81_cast_fp16)[name = string("k_head_83_cast_fp16")];
            tensor<int32, [1]> v_head_83_axes_1 = const()[name = string("v_head_83_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_83_cast_fp16 = expand_dims(axes = v_head_83_axes_1, x = v_head_81_cast_fp16)[name = string("v_head_83_cast_fp16")];
            tensor<int32, [1]> p_head_83_axes_1 = const()[name = string("p_head_83_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_83_cast_fp16 = expand_dims(axes = p_head_83_axes_1, x = p_head_81_cast_fp16)[name = string("p_head_83_cast_fp16")];
            bool var_3463_transpose_x_3 = const()[name = string("op_3463_transpose_x_3"), val = bool(false)];
            bool var_3463_transpose_y_3 = const()[name = string("op_3463_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_3463_cast_fp16 = matmul(transpose_x = var_3463_transpose_x_3, transpose_y = var_3463_transpose_y_3, x = q_u_41_cast_fp16, y = k_head_83_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp16 var_3464_to_fp16 = const()[name = string("op_3464_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_41_cast_fp16 = mul(x = var_3463_cast_fp16, y = var_3464_to_fp16)[name = string("scores_content_41_cast_fp16")];
            bool x_221_transpose_x_3 = const()[name = string("x_221_transpose_x_3"), val = bool(false)];
            bool x_221_transpose_y_3 = const()[name = string("x_221_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_3, transpose_y = x_221_transpose_y_3, x = q_v_41_cast_fp16, y = p_head_83_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<int32, [8]> x_223_pad_1 = const()[name = string("x_223_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_223_mode_1 = const()[name = string("x_223_mode_1"), val = string("constant")];
            fp16 const_1441_to_fp16 = const()[name = string("const_1441_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_223_cast_fp16 = pad(constant_val = const_1441_to_fp16, mode = x_223_mode_1, pad = x_223_pad_1, x = x_221_cast_fp16)[name = string("x_223_cast_fp16")];
            tensor<int32, [4]> var_3478 = const()[name = string("op_3478"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_225_cast_fp16 = reshape(shape = var_3478, x = x_223_cast_fp16)[name = string("x_225_cast_fp16")];
            tensor<int32, [4]> var_3482_begin_1 = const()[name = string("op_3482_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3482_end_1 = const()[name = string("op_3482_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_3482_end_mask_1 = const()[name = string("op_3482_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_3482_cast_fp16 = slice_by_index(begin = var_3482_begin_1, end = var_3482_end_1, end_mask = var_3482_end_mask_1, x = x_225_cast_fp16)[name = string("op_3482_cast_fp16")];
            tensor<int32, [4]> var_3484 = const()[name = string("op_3484"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_3485_cast_fp16 = reshape(shape = var_3484, x = var_3482_cast_fp16)[name = string("op_3485_cast_fp16")];
            tensor<int32, [4]> var_3490_begin_1 = const()[name = string("op_3490_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3490_end_1 = const()[name = string("op_3490_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_3490_end_mask_1 = const()[name = string("op_3490_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_3490_cast_fp16 = slice_by_index(begin = var_3490_begin_1, end = var_3490_end_1, end_mask = var_3490_end_mask_1, x = var_3485_cast_fp16)[name = string("op_3490_cast_fp16")];
            fp16 var_3491_to_fp16 = const()[name = string("op_3491_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_41_cast_fp16 = mul(x = var_3490_cast_fp16, y = var_3491_to_fp16)[name = string("scores_pos_41_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_41_cast_fp16 = add(x = scores_content_41_cast_fp16, y = scores_pos_41_cast_fp16)[name = string("logits_41_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_3494_cast_fp16 = softmax(axis = var_2735, x = logits_41_cast_fp16)[name = string("op_3494_cast_fp16")];
            bool var_3496_transpose_x_1 = const()[name = string("op_3496_transpose_x_1"), val = bool(false)];
            bool var_3496_transpose_y_1 = const()[name = string("op_3496_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_3496_cast_fp16 = matmul(transpose_x = var_3496_transpose_x_1, transpose_y = var_3496_transpose_y_1, x = var_3494_cast_fp16, y = v_head_83_cast_fp16)[name = string("op_3496_cast_fp16")];
            tensor<int32, [1]> var_3497_axes_1 = const()[name = string("op_3497_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_3497_cast_fp16 = squeeze(axes = var_3497_axes_1, x = var_3496_cast_fp16)[name = string("op_3497_cast_fp16")];
            string dense_output_217_pad_type_1 = const()[name = string("dense_output_217_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_217_strides_1 = const()[name = string("dense_output_217_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_217_pad_1 = const()[name = string("dense_output_217_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_217_dilations_1 = const()[name = string("dense_output_217_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_217_groups_1 = const()[name = string("dense_output_217_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77910976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78042112))))[name = string("layers_2_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_217_cast_fp16 = conv(dilations = dense_output_217_dilations_1, groups = dense_output_217_groups_1, pad = dense_output_217_pad_1, pad_type = dense_output_217_pad_type_1, strides = dense_output_217_strides_1, weight = layers_2_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_217_cast_fp16")];
            string sparse_output_217_pad_type_1 = const()[name = string("sparse_output_217_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_217_strides_1 = const()[name = string("sparse_output_217_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_217_pad_1 = const()[name = string("sparse_output_217_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_217_dilations_1 = const()[name = string("sparse_output_217_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_217_groups_1 = const()[name = string("sparse_output_217_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78045376))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78042688))))[name = string("layers_2_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_217_cast_fp16 = conv(dilations = sparse_output_217_dilations_1, groups = sparse_output_217_groups_1, pad = sparse_output_217_pad_1, pad_type = sparse_output_217_pad_type_1, strides = sparse_output_217_strides_1, weight = layers_2_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_217_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3512_cast_fp16 = add(x = dense_output_217_cast_fp16, y = sparse_output_217_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<int32, [4]> var_3513 = const()[name = string("op_3513"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3515 = const()[name = string("op_3515"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3514_cast_fp16 = transpose(perm = var_3513, x = var_3512_cast_fp16)[name = string("transpose_878")];
            tensor<fp16, [1, 51, 128]> q_head_43_cast_fp16 = reshape(shape = var_3515, x = var_3514_cast_fp16)[name = string("q_head_43_cast_fp16")];
            string dense_output_219_pad_type_1 = const()[name = string("dense_output_219_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_219_strides_1 = const()[name = string("dense_output_219_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_219_pad_1 = const()[name = string("dense_output_219_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_219_dilations_1 = const()[name = string("dense_output_219_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_219_groups_1 = const()[name = string("dense_output_219_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78061824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78192960))))[name = string("layers_2_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_219_cast_fp16 = conv(dilations = dense_output_219_dilations_1, groups = dense_output_219_groups_1, pad = dense_output_219_pad_1, pad_type = dense_output_219_pad_type_1, strides = dense_output_219_strides_1, weight = layers_2_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_219_cast_fp16")];
            string sparse_output_219_pad_type_1 = const()[name = string("sparse_output_219_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_219_strides_1 = const()[name = string("sparse_output_219_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_219_pad_1 = const()[name = string("sparse_output_219_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_219_dilations_1 = const()[name = string("sparse_output_219_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_219_groups_1 = const()[name = string("sparse_output_219_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78196224))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78193536))))[name = string("layers_2_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_219_cast_fp16 = conv(dilations = sparse_output_219_dilations_1, groups = sparse_output_219_groups_1, pad = sparse_output_219_pad_1, pad_type = sparse_output_219_pad_type_1, strides = sparse_output_219_strides_1, weight = layers_2_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_219_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3531_cast_fp16 = add(x = dense_output_219_cast_fp16, y = sparse_output_219_cast_fp16)[name = string("op_3531_cast_fp16")];
            tensor<int32, [4]> var_3532 = const()[name = string("op_3532"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3534 = const()[name = string("op_3534"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3533_cast_fp16 = transpose(perm = var_3532, x = var_3531_cast_fp16)[name = string("transpose_877")];
            tensor<fp16, [1, 51, 128]> k_head_85_cast_fp16 = reshape(shape = var_3534, x = var_3533_cast_fp16)[name = string("k_head_85_cast_fp16")];
            string dense_output_221_pad_type_1 = const()[name = string("dense_output_221_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_221_strides_1 = const()[name = string("dense_output_221_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_221_pad_1 = const()[name = string("dense_output_221_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_221_dilations_1 = const()[name = string("dense_output_221_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_221_groups_1 = const()[name = string("dense_output_221_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78212672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78343808))))[name = string("layers_2_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_221_cast_fp16 = conv(dilations = dense_output_221_dilations_1, groups = dense_output_221_groups_1, pad = dense_output_221_pad_1, pad_type = dense_output_221_pad_type_1, strides = dense_output_221_strides_1, weight = layers_2_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_221_cast_fp16")];
            string sparse_output_221_pad_type_1 = const()[name = string("sparse_output_221_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_221_strides_1 = const()[name = string("sparse_output_221_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_221_pad_1 = const()[name = string("sparse_output_221_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_221_dilations_1 = const()[name = string("sparse_output_221_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_221_groups_1 = const()[name = string("sparse_output_221_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78347072))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78344384))))[name = string("layers_2_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_221_cast_fp16 = conv(dilations = sparse_output_221_dilations_1, groups = sparse_output_221_groups_1, pad = sparse_output_221_pad_1, pad_type = sparse_output_221_pad_type_1, strides = sparse_output_221_strides_1, weight = layers_2_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_221_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3550_cast_fp16 = add(x = dense_output_221_cast_fp16, y = sparse_output_221_cast_fp16)[name = string("op_3550_cast_fp16")];
            tensor<int32, [4]> var_3551 = const()[name = string("op_3551"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3553 = const()[name = string("op_3553"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3552_cast_fp16 = transpose(perm = var_3551, x = var_3550_cast_fp16)[name = string("transpose_876")];
            tensor<fp16, [1, 51, 128]> v_head_85_cast_fp16 = reshape(shape = var_3553, x = var_3552_cast_fp16)[name = string("v_head_85_cast_fp16")];
            string dense_output_223_pad_type_1 = const()[name = string("dense_output_223_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_223_strides_1 = const()[name = string("dense_output_223_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_223_pad_1 = const()[name = string("dense_output_223_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_223_dilations_1 = const()[name = string("dense_output_223_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_223_groups_1 = const()[name = string("dense_output_223_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78363520))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78494656))))[name = string("layers_2_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_223_cast_fp16 = conv(dilations = dense_output_223_dilations_1, groups = dense_output_223_groups_1, pad = dense_output_223_pad_1, pad_type = dense_output_223_pad_type_1, strides = dense_output_223_strides_1, weight = layers_2_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_223_cast_fp16")];
            string sparse_output_223_pad_type_1 = const()[name = string("sparse_output_223_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_223_strides_1 = const()[name = string("sparse_output_223_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_223_pad_1 = const()[name = string("sparse_output_223_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_223_dilations_1 = const()[name = string("sparse_output_223_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_223_groups_1 = const()[name = string("sparse_output_223_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78497920))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78495232))))[name = string("layers_2_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_223_cast_fp16 = conv(dilations = sparse_output_223_dilations_1, groups = sparse_output_223_groups_1, pad = sparse_output_223_pad_1, pad_type = sparse_output_223_pad_type_1, strides = sparse_output_223_strides_1, weight = layers_2_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_223_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_3569_cast_fp16 = add(x = dense_output_223_cast_fp16, y = sparse_output_223_cast_fp16)[name = string("op_3569_cast_fp16")];
            tensor<int32, [4]> var_3570 = const()[name = string("op_3570"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3572 = const()[name = string("op_3572"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_3571_cast_fp16 = transpose(perm = var_3570, x = var_3569_cast_fp16)[name = string("transpose_875")];
            tensor<fp16, [1, 101, 128]> p_head_85_cast_fp16 = reshape(shape = var_3572, x = var_3571_cast_fp16)[name = string("p_head_85_cast_fp16")];
            tensor<fp16, [128]> var_3574_to_fp16 = const()[name = string("op_3574_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78514368)))];
            tensor<fp16, [1, 51, 128]> var_3575_cast_fp16 = add(x = q_head_43_cast_fp16, y = var_3574_to_fp16)[name = string("op_3575_cast_fp16")];
            tensor<int32, [1]> q_u_43_axes_1 = const()[name = string("q_u_43_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_43_cast_fp16 = expand_dims(axes = q_u_43_axes_1, x = var_3575_cast_fp16)[name = string("q_u_43_cast_fp16")];
            tensor<fp16, [128]> var_3577_to_fp16 = const()[name = string("op_3577_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78514688)))];
            tensor<fp16, [1, 51, 128]> var_3578_cast_fp16 = add(x = q_head_43_cast_fp16, y = var_3577_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> q_v_43_axes_1 = const()[name = string("q_v_43_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_43_cast_fp16 = expand_dims(axes = q_v_43_axes_1, x = var_3578_cast_fp16)[name = string("q_v_43_cast_fp16")];
            tensor<int32, [1]> k_head_87_axes_1 = const()[name = string("k_head_87_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_87_cast_fp16 = expand_dims(axes = k_head_87_axes_1, x = k_head_85_cast_fp16)[name = string("k_head_87_cast_fp16")];
            tensor<int32, [1]> v_head_87_axes_1 = const()[name = string("v_head_87_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_87_cast_fp16 = expand_dims(axes = v_head_87_axes_1, x = v_head_85_cast_fp16)[name = string("v_head_87_cast_fp16")];
            tensor<int32, [1]> p_head_87_axes_1 = const()[name = string("p_head_87_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_87_cast_fp16 = expand_dims(axes = p_head_87_axes_1, x = p_head_85_cast_fp16)[name = string("p_head_87_cast_fp16")];
            bool var_3584_transpose_x_3 = const()[name = string("op_3584_transpose_x_3"), val = bool(false)];
            bool var_3584_transpose_y_3 = const()[name = string("op_3584_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_3584_cast_fp16 = matmul(transpose_x = var_3584_transpose_x_3, transpose_y = var_3584_transpose_y_3, x = q_u_43_cast_fp16, y = k_head_87_cast_fp16)[name = string("op_3584_cast_fp16")];
            fp16 var_3585_to_fp16 = const()[name = string("op_3585_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_43_cast_fp16 = mul(x = var_3584_cast_fp16, y = var_3585_to_fp16)[name = string("scores_content_43_cast_fp16")];
            bool x_229_transpose_x_3 = const()[name = string("x_229_transpose_x_3"), val = bool(false)];
            bool x_229_transpose_y_3 = const()[name = string("x_229_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_229_cast_fp16 = matmul(transpose_x = x_229_transpose_x_3, transpose_y = x_229_transpose_y_3, x = q_v_43_cast_fp16, y = p_head_87_cast_fp16)[name = string("x_229_cast_fp16")];
            tensor<int32, [8]> x_231_pad_1 = const()[name = string("x_231_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_231_mode_1 = const()[name = string("x_231_mode_1"), val = string("constant")];
            fp16 const_1447_to_fp16 = const()[name = string("const_1447_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_231_cast_fp16 = pad(constant_val = const_1447_to_fp16, mode = x_231_mode_1, pad = x_231_pad_1, x = x_229_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<int32, [4]> var_3599 = const()[name = string("op_3599"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_233_cast_fp16 = reshape(shape = var_3599, x = x_231_cast_fp16)[name = string("x_233_cast_fp16")];
            tensor<int32, [4]> var_3603_begin_1 = const()[name = string("op_3603_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3603_end_1 = const()[name = string("op_3603_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_3603_end_mask_1 = const()[name = string("op_3603_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_3603_cast_fp16 = slice_by_index(begin = var_3603_begin_1, end = var_3603_end_1, end_mask = var_3603_end_mask_1, x = x_233_cast_fp16)[name = string("op_3603_cast_fp16")];
            tensor<int32, [4]> var_3605 = const()[name = string("op_3605"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_3606_cast_fp16 = reshape(shape = var_3605, x = var_3603_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [4]> var_3611_begin_1 = const()[name = string("op_3611_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3611_end_1 = const()[name = string("op_3611_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_3611_end_mask_1 = const()[name = string("op_3611_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_3611_cast_fp16 = slice_by_index(begin = var_3611_begin_1, end = var_3611_end_1, end_mask = var_3611_end_mask_1, x = var_3606_cast_fp16)[name = string("op_3611_cast_fp16")];
            fp16 var_3612_to_fp16 = const()[name = string("op_3612_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_43_cast_fp16 = mul(x = var_3611_cast_fp16, y = var_3612_to_fp16)[name = string("scores_pos_43_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_43_cast_fp16 = add(x = scores_content_43_cast_fp16, y = scores_pos_43_cast_fp16)[name = string("logits_43_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_3615_cast_fp16 = softmax(axis = var_2735, x = logits_43_cast_fp16)[name = string("op_3615_cast_fp16")];
            bool var_3617_transpose_x_1 = const()[name = string("op_3617_transpose_x_1"), val = bool(false)];
            bool var_3617_transpose_y_1 = const()[name = string("op_3617_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_3617_cast_fp16 = matmul(transpose_x = var_3617_transpose_x_1, transpose_y = var_3617_transpose_y_1, x = var_3615_cast_fp16, y = v_head_87_cast_fp16)[name = string("op_3617_cast_fp16")];
            tensor<int32, [1]> var_3618_axes_1 = const()[name = string("op_3618_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_3618_cast_fp16 = squeeze(axes = var_3618_axes_1, x = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            string dense_output_225_pad_type_1 = const()[name = string("dense_output_225_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_225_strides_1 = const()[name = string("dense_output_225_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_225_pad_1 = const()[name = string("dense_output_225_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_225_dilations_1 = const()[name = string("dense_output_225_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_225_groups_1 = const()[name = string("dense_output_225_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78515008))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78646144))))[name = string("layers_2_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_225_cast_fp16 = conv(dilations = dense_output_225_dilations_1, groups = dense_output_225_groups_1, pad = dense_output_225_pad_1, pad_type = dense_output_225_pad_type_1, strides = dense_output_225_strides_1, weight = layers_2_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_225_cast_fp16")];
            string sparse_output_225_pad_type_1 = const()[name = string("sparse_output_225_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_225_strides_1 = const()[name = string("sparse_output_225_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_225_pad_1 = const()[name = string("sparse_output_225_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_225_dilations_1 = const()[name = string("sparse_output_225_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_225_groups_1 = const()[name = string("sparse_output_225_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78649408))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78646720))))[name = string("layers_2_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_225_cast_fp16 = conv(dilations = sparse_output_225_dilations_1, groups = sparse_output_225_groups_1, pad = sparse_output_225_pad_1, pad_type = sparse_output_225_pad_type_1, strides = sparse_output_225_strides_1, weight = layers_2_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_225_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3633_cast_fp16 = add(x = dense_output_225_cast_fp16, y = sparse_output_225_cast_fp16)[name = string("op_3633_cast_fp16")];
            tensor<int32, [4]> var_3634 = const()[name = string("op_3634"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3636 = const()[name = string("op_3636"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3635_cast_fp16 = transpose(perm = var_3634, x = var_3633_cast_fp16)[name = string("transpose_874")];
            tensor<fp16, [1, 51, 128]> q_head_45_cast_fp16 = reshape(shape = var_3636, x = var_3635_cast_fp16)[name = string("q_head_45_cast_fp16")];
            string dense_output_227_pad_type_1 = const()[name = string("dense_output_227_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_227_strides_1 = const()[name = string("dense_output_227_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_227_pad_1 = const()[name = string("dense_output_227_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_227_dilations_1 = const()[name = string("dense_output_227_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_227_groups_1 = const()[name = string("dense_output_227_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78665856))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78796992))))[name = string("layers_2_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_227_cast_fp16 = conv(dilations = dense_output_227_dilations_1, groups = dense_output_227_groups_1, pad = dense_output_227_pad_1, pad_type = dense_output_227_pad_type_1, strides = dense_output_227_strides_1, weight = layers_2_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_227_cast_fp16")];
            string sparse_output_227_pad_type_1 = const()[name = string("sparse_output_227_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_227_strides_1 = const()[name = string("sparse_output_227_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_227_pad_1 = const()[name = string("sparse_output_227_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_227_dilations_1 = const()[name = string("sparse_output_227_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_227_groups_1 = const()[name = string("sparse_output_227_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78800256))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78797568))))[name = string("layers_2_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_227_cast_fp16 = conv(dilations = sparse_output_227_dilations_1, groups = sparse_output_227_groups_1, pad = sparse_output_227_pad_1, pad_type = sparse_output_227_pad_type_1, strides = sparse_output_227_strides_1, weight = layers_2_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_227_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3652_cast_fp16 = add(x = dense_output_227_cast_fp16, y = sparse_output_227_cast_fp16)[name = string("op_3652_cast_fp16")];
            tensor<int32, [4]> var_3653 = const()[name = string("op_3653"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3655 = const()[name = string("op_3655"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3654_cast_fp16 = transpose(perm = var_3653, x = var_3652_cast_fp16)[name = string("transpose_873")];
            tensor<fp16, [1, 51, 128]> k_head_89_cast_fp16 = reshape(shape = var_3655, x = var_3654_cast_fp16)[name = string("k_head_89_cast_fp16")];
            string dense_output_229_pad_type_1 = const()[name = string("dense_output_229_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_229_strides_1 = const()[name = string("dense_output_229_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_229_pad_1 = const()[name = string("dense_output_229_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_229_dilations_1 = const()[name = string("dense_output_229_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_229_groups_1 = const()[name = string("dense_output_229_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78816704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78947840))))[name = string("layers_2_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_229_cast_fp16 = conv(dilations = dense_output_229_dilations_1, groups = dense_output_229_groups_1, pad = dense_output_229_pad_1, pad_type = dense_output_229_pad_type_1, strides = dense_output_229_strides_1, weight = layers_2_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_229_cast_fp16")];
            string sparse_output_229_pad_type_1 = const()[name = string("sparse_output_229_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_229_strides_1 = const()[name = string("sparse_output_229_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_229_pad_1 = const()[name = string("sparse_output_229_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_229_dilations_1 = const()[name = string("sparse_output_229_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_229_groups_1 = const()[name = string("sparse_output_229_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78951104))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78948416))))[name = string("layers_2_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_229_cast_fp16 = conv(dilations = sparse_output_229_dilations_1, groups = sparse_output_229_groups_1, pad = sparse_output_229_pad_1, pad_type = sparse_output_229_pad_type_1, strides = sparse_output_229_strides_1, weight = layers_2_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_229_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3671_cast_fp16 = add(x = dense_output_229_cast_fp16, y = sparse_output_229_cast_fp16)[name = string("op_3671_cast_fp16")];
            tensor<int32, [4]> var_3672 = const()[name = string("op_3672"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3674 = const()[name = string("op_3674"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3673_cast_fp16 = transpose(perm = var_3672, x = var_3671_cast_fp16)[name = string("transpose_872")];
            tensor<fp16, [1, 51, 128]> v_head_89_cast_fp16 = reshape(shape = var_3674, x = var_3673_cast_fp16)[name = string("v_head_89_cast_fp16")];
            string dense_output_231_pad_type_1 = const()[name = string("dense_output_231_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_231_strides_1 = const()[name = string("dense_output_231_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_231_pad_1 = const()[name = string("dense_output_231_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_231_dilations_1 = const()[name = string("dense_output_231_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_231_groups_1 = const()[name = string("dense_output_231_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78967552))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79098688))))[name = string("layers_2_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_231_cast_fp16 = conv(dilations = dense_output_231_dilations_1, groups = dense_output_231_groups_1, pad = dense_output_231_pad_1, pad_type = dense_output_231_pad_type_1, strides = dense_output_231_strides_1, weight = layers_2_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_231_cast_fp16")];
            string sparse_output_231_pad_type_1 = const()[name = string("sparse_output_231_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_231_strides_1 = const()[name = string("sparse_output_231_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_231_pad_1 = const()[name = string("sparse_output_231_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_231_dilations_1 = const()[name = string("sparse_output_231_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_231_groups_1 = const()[name = string("sparse_output_231_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79101952))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79099264))))[name = string("layers_2_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_231_cast_fp16 = conv(dilations = sparse_output_231_dilations_1, groups = sparse_output_231_groups_1, pad = sparse_output_231_pad_1, pad_type = sparse_output_231_pad_type_1, strides = sparse_output_231_strides_1, weight = layers_2_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_231_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_3690_cast_fp16 = add(x = dense_output_231_cast_fp16, y = sparse_output_231_cast_fp16)[name = string("op_3690_cast_fp16")];
            tensor<int32, [4]> var_3691 = const()[name = string("op_3691"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3693 = const()[name = string("op_3693"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_3692_cast_fp16 = transpose(perm = var_3691, x = var_3690_cast_fp16)[name = string("transpose_871")];
            tensor<fp16, [1, 101, 128]> p_head_89_cast_fp16 = reshape(shape = var_3693, x = var_3692_cast_fp16)[name = string("p_head_89_cast_fp16")];
            tensor<fp16, [128]> var_3695_to_fp16 = const()[name = string("op_3695_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79118400)))];
            tensor<fp16, [1, 51, 128]> var_3696_cast_fp16 = add(x = q_head_45_cast_fp16, y = var_3695_to_fp16)[name = string("op_3696_cast_fp16")];
            tensor<int32, [1]> q_u_45_axes_1 = const()[name = string("q_u_45_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_45_cast_fp16 = expand_dims(axes = q_u_45_axes_1, x = var_3696_cast_fp16)[name = string("q_u_45_cast_fp16")];
            tensor<fp16, [128]> var_3698_to_fp16 = const()[name = string("op_3698_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79118720)))];
            tensor<fp16, [1, 51, 128]> var_3699_cast_fp16 = add(x = q_head_45_cast_fp16, y = var_3698_to_fp16)[name = string("op_3699_cast_fp16")];
            tensor<int32, [1]> q_v_45_axes_1 = const()[name = string("q_v_45_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_45_cast_fp16 = expand_dims(axes = q_v_45_axes_1, x = var_3699_cast_fp16)[name = string("q_v_45_cast_fp16")];
            tensor<int32, [1]> k_head_91_axes_1 = const()[name = string("k_head_91_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_91_cast_fp16 = expand_dims(axes = k_head_91_axes_1, x = k_head_89_cast_fp16)[name = string("k_head_91_cast_fp16")];
            tensor<int32, [1]> v_head_91_axes_1 = const()[name = string("v_head_91_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_91_cast_fp16 = expand_dims(axes = v_head_91_axes_1, x = v_head_89_cast_fp16)[name = string("v_head_91_cast_fp16")];
            tensor<int32, [1]> p_head_91_axes_1 = const()[name = string("p_head_91_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_91_cast_fp16 = expand_dims(axes = p_head_91_axes_1, x = p_head_89_cast_fp16)[name = string("p_head_91_cast_fp16")];
            bool var_3705_transpose_x_3 = const()[name = string("op_3705_transpose_x_3"), val = bool(false)];
            bool var_3705_transpose_y_3 = const()[name = string("op_3705_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_3705_cast_fp16 = matmul(transpose_x = var_3705_transpose_x_3, transpose_y = var_3705_transpose_y_3, x = q_u_45_cast_fp16, y = k_head_91_cast_fp16)[name = string("op_3705_cast_fp16")];
            fp16 var_3706_to_fp16 = const()[name = string("op_3706_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_45_cast_fp16 = mul(x = var_3705_cast_fp16, y = var_3706_to_fp16)[name = string("scores_content_45_cast_fp16")];
            bool x_237_transpose_x_3 = const()[name = string("x_237_transpose_x_3"), val = bool(false)];
            bool x_237_transpose_y_3 = const()[name = string("x_237_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_237_cast_fp16 = matmul(transpose_x = x_237_transpose_x_3, transpose_y = x_237_transpose_y_3, x = q_v_45_cast_fp16, y = p_head_91_cast_fp16)[name = string("x_237_cast_fp16")];
            tensor<int32, [8]> x_239_pad_1 = const()[name = string("x_239_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_239_mode_1 = const()[name = string("x_239_mode_1"), val = string("constant")];
            fp16 const_1453_to_fp16 = const()[name = string("const_1453_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_239_cast_fp16 = pad(constant_val = const_1453_to_fp16, mode = x_239_mode_1, pad = x_239_pad_1, x = x_237_cast_fp16)[name = string("x_239_cast_fp16")];
            tensor<int32, [4]> var_3720 = const()[name = string("op_3720"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_241_cast_fp16 = reshape(shape = var_3720, x = x_239_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<int32, [4]> var_3724_begin_1 = const()[name = string("op_3724_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3724_end_1 = const()[name = string("op_3724_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_3724_end_mask_1 = const()[name = string("op_3724_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_1, end = var_3724_end_1, end_mask = var_3724_end_mask_1, x = x_241_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [4]> var_3726 = const()[name = string("op_3726"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_3727_cast_fp16 = reshape(shape = var_3726, x = var_3724_cast_fp16)[name = string("op_3727_cast_fp16")];
            tensor<int32, [4]> var_3732_begin_1 = const()[name = string("op_3732_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3732_end_1 = const()[name = string("op_3732_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_3732_end_mask_1 = const()[name = string("op_3732_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_3732_cast_fp16 = slice_by_index(begin = var_3732_begin_1, end = var_3732_end_1, end_mask = var_3732_end_mask_1, x = var_3727_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp16 var_3733_to_fp16 = const()[name = string("op_3733_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_45_cast_fp16 = mul(x = var_3732_cast_fp16, y = var_3733_to_fp16)[name = string("scores_pos_45_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_45_cast_fp16 = add(x = scores_content_45_cast_fp16, y = scores_pos_45_cast_fp16)[name = string("logits_45_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_3736_cast_fp16 = softmax(axis = var_2735, x = logits_45_cast_fp16)[name = string("op_3736_cast_fp16")];
            bool var_3738_transpose_x_1 = const()[name = string("op_3738_transpose_x_1"), val = bool(false)];
            bool var_3738_transpose_y_1 = const()[name = string("op_3738_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_3738_cast_fp16 = matmul(transpose_x = var_3738_transpose_x_1, transpose_y = var_3738_transpose_y_1, x = var_3736_cast_fp16, y = v_head_91_cast_fp16)[name = string("op_3738_cast_fp16")];
            tensor<int32, [1]> var_3739_axes_1 = const()[name = string("op_3739_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_3739_cast_fp16 = squeeze(axes = var_3739_axes_1, x = var_3738_cast_fp16)[name = string("op_3739_cast_fp16")];
            string dense_output_233_pad_type_1 = const()[name = string("dense_output_233_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_233_strides_1 = const()[name = string("dense_output_233_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_233_pad_1 = const()[name = string("dense_output_233_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_233_dilations_1 = const()[name = string("dense_output_233_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_233_groups_1 = const()[name = string("dense_output_233_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79119040))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79250176))))[name = string("layers_2_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_233_cast_fp16 = conv(dilations = dense_output_233_dilations_1, groups = dense_output_233_groups_1, pad = dense_output_233_pad_1, pad_type = dense_output_233_pad_type_1, strides = dense_output_233_strides_1, weight = layers_2_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_233_cast_fp16")];
            string sparse_output_233_pad_type_1 = const()[name = string("sparse_output_233_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_233_strides_1 = const()[name = string("sparse_output_233_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_233_pad_1 = const()[name = string("sparse_output_233_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_233_dilations_1 = const()[name = string("sparse_output_233_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_233_groups_1 = const()[name = string("sparse_output_233_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79253440))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79250752))))[name = string("layers_2_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_233_cast_fp16 = conv(dilations = sparse_output_233_dilations_1, groups = sparse_output_233_groups_1, pad = sparse_output_233_pad_1, pad_type = sparse_output_233_pad_type_1, strides = sparse_output_233_strides_1, weight = layers_2_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_233_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3754_cast_fp16 = add(x = dense_output_233_cast_fp16, y = sparse_output_233_cast_fp16)[name = string("op_3754_cast_fp16")];
            tensor<int32, [4]> var_3755 = const()[name = string("op_3755"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3757 = const()[name = string("op_3757"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3756_cast_fp16 = transpose(perm = var_3755, x = var_3754_cast_fp16)[name = string("transpose_870")];
            tensor<fp16, [1, 51, 128]> q_head_47_cast_fp16 = reshape(shape = var_3757, x = var_3756_cast_fp16)[name = string("q_head_47_cast_fp16")];
            string dense_output_235_pad_type_1 = const()[name = string("dense_output_235_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_235_strides_1 = const()[name = string("dense_output_235_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_235_pad_1 = const()[name = string("dense_output_235_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_235_dilations_1 = const()[name = string("dense_output_235_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_235_groups_1 = const()[name = string("dense_output_235_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79269888))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79401024))))[name = string("layers_2_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_235_cast_fp16 = conv(dilations = dense_output_235_dilations_1, groups = dense_output_235_groups_1, pad = dense_output_235_pad_1, pad_type = dense_output_235_pad_type_1, strides = dense_output_235_strides_1, weight = layers_2_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_235_cast_fp16")];
            string sparse_output_235_pad_type_1 = const()[name = string("sparse_output_235_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_235_strides_1 = const()[name = string("sparse_output_235_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_235_pad_1 = const()[name = string("sparse_output_235_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_235_dilations_1 = const()[name = string("sparse_output_235_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_235_groups_1 = const()[name = string("sparse_output_235_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79404288))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79401600))))[name = string("layers_2_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_235_cast_fp16 = conv(dilations = sparse_output_235_dilations_1, groups = sparse_output_235_groups_1, pad = sparse_output_235_pad_1, pad_type = sparse_output_235_pad_type_1, strides = sparse_output_235_strides_1, weight = layers_2_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_235_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3773_cast_fp16 = add(x = dense_output_235_cast_fp16, y = sparse_output_235_cast_fp16)[name = string("op_3773_cast_fp16")];
            tensor<int32, [4]> var_3774 = const()[name = string("op_3774"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3776 = const()[name = string("op_3776"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3775_cast_fp16 = transpose(perm = var_3774, x = var_3773_cast_fp16)[name = string("transpose_869")];
            tensor<fp16, [1, 51, 128]> k_head_93_cast_fp16 = reshape(shape = var_3776, x = var_3775_cast_fp16)[name = string("k_head_93_cast_fp16")];
            string dense_output_237_pad_type_1 = const()[name = string("dense_output_237_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_237_strides_1 = const()[name = string("dense_output_237_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_237_pad_1 = const()[name = string("dense_output_237_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_237_dilations_1 = const()[name = string("dense_output_237_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_237_groups_1 = const()[name = string("dense_output_237_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79420736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79551872))))[name = string("layers_2_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_237_cast_fp16 = conv(dilations = dense_output_237_dilations_1, groups = dense_output_237_groups_1, pad = dense_output_237_pad_1, pad_type = dense_output_237_pad_type_1, strides = dense_output_237_strides_1, weight = layers_2_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("dense_output_237_cast_fp16")];
            string sparse_output_237_pad_type_1 = const()[name = string("sparse_output_237_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_237_strides_1 = const()[name = string("sparse_output_237_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_237_pad_1 = const()[name = string("sparse_output_237_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_237_dilations_1 = const()[name = string("sparse_output_237_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_237_groups_1 = const()[name = string("sparse_output_237_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79555136))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79552448))))[name = string("layers_2_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_237_cast_fp16 = conv(dilations = sparse_output_237_dilations_1, groups = sparse_output_237_groups_1, pad = sparse_output_237_pad_1, pad_type = sparse_output_237_pad_type_1, strides = sparse_output_237_strides_1, weight = layers_2_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_121_cast_fp16)[name = string("sparse_output_237_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_3792_cast_fp16 = add(x = dense_output_237_cast_fp16, y = sparse_output_237_cast_fp16)[name = string("op_3792_cast_fp16")];
            tensor<int32, [4]> var_3793 = const()[name = string("op_3793"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3795 = const()[name = string("op_3795"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_3794_cast_fp16 = transpose(perm = var_3793, x = var_3792_cast_fp16)[name = string("transpose_868")];
            tensor<fp16, [1, 51, 128]> v_head_93_cast_fp16 = reshape(shape = var_3795, x = var_3794_cast_fp16)[name = string("v_head_93_cast_fp16")];
            string dense_output_239_pad_type_1 = const()[name = string("dense_output_239_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_239_strides_1 = const()[name = string("dense_output_239_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_239_pad_1 = const()[name = string("dense_output_239_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_239_dilations_1 = const()[name = string("dense_output_239_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_239_groups_1 = const()[name = string("dense_output_239_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79571584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79702720))))[name = string("layers_2_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_239_cast_fp16 = conv(dilations = dense_output_239_dilations_1, groups = dense_output_239_groups_1, pad = dense_output_239_pad_1, pad_type = dense_output_239_pad_type_1, strides = dense_output_239_strides_1, weight = layers_2_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_239_cast_fp16")];
            string sparse_output_239_pad_type_1 = const()[name = string("sparse_output_239_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_239_strides_1 = const()[name = string("sparse_output_239_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_239_pad_1 = const()[name = string("sparse_output_239_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_239_dilations_1 = const()[name = string("sparse_output_239_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_239_groups_1 = const()[name = string("sparse_output_239_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_2_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79705984))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79703296))))[name = string("layers_2_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_239_cast_fp16 = conv(dilations = sparse_output_239_dilations_1, groups = sparse_output_239_groups_1, pad = sparse_output_239_pad_1, pad_type = sparse_output_239_pad_type_1, strides = sparse_output_239_strides_1, weight = layers_2_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_239_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_3811_cast_fp16 = add(x = dense_output_239_cast_fp16, y = sparse_output_239_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<int32, [4]> var_3812 = const()[name = string("op_3812"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_3814 = const()[name = string("op_3814"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_3813_cast_fp16 = transpose(perm = var_3812, x = var_3811_cast_fp16)[name = string("transpose_867")];
            tensor<fp16, [1, 101, 128]> p_head_93_cast_fp16 = reshape(shape = var_3814, x = var_3813_cast_fp16)[name = string("p_head_93_cast_fp16")];
            tensor<fp16, [128]> var_3816_to_fp16 = const()[name = string("op_3816_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79722432)))];
            tensor<fp16, [1, 51, 128]> var_3817_cast_fp16 = add(x = q_head_47_cast_fp16, y = var_3816_to_fp16)[name = string("op_3817_cast_fp16")];
            tensor<int32, [1]> q_u_47_axes_1 = const()[name = string("q_u_47_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_47_cast_fp16 = expand_dims(axes = q_u_47_axes_1, x = var_3817_cast_fp16)[name = string("q_u_47_cast_fp16")];
            tensor<fp16, [128]> var_3819_to_fp16 = const()[name = string("op_3819_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79722752)))];
            tensor<fp16, [1, 51, 128]> var_3820_cast_fp16 = add(x = q_head_47_cast_fp16, y = var_3819_to_fp16)[name = string("op_3820_cast_fp16")];
            tensor<int32, [1]> q_v_47_axes_1 = const()[name = string("q_v_47_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_47_cast_fp16 = expand_dims(axes = q_v_47_axes_1, x = var_3820_cast_fp16)[name = string("q_v_47_cast_fp16")];
            tensor<int32, [1]> k_head_95_axes_1 = const()[name = string("k_head_95_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_95_cast_fp16 = expand_dims(axes = k_head_95_axes_1, x = k_head_93_cast_fp16)[name = string("k_head_95_cast_fp16")];
            tensor<int32, [1]> v_head_95_axes_1 = const()[name = string("v_head_95_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_95_cast_fp16 = expand_dims(axes = v_head_95_axes_1, x = v_head_93_cast_fp16)[name = string("v_head_95_cast_fp16")];
            tensor<int32, [1]> p_head_95_axes_1 = const()[name = string("p_head_95_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_95_cast_fp16 = expand_dims(axes = p_head_95_axes_1, x = p_head_93_cast_fp16)[name = string("p_head_95_cast_fp16")];
            bool var_3826_transpose_x_3 = const()[name = string("op_3826_transpose_x_3"), val = bool(false)];
            bool var_3826_transpose_y_3 = const()[name = string("op_3826_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_3826_cast_fp16 = matmul(transpose_x = var_3826_transpose_x_3, transpose_y = var_3826_transpose_y_3, x = q_u_47_cast_fp16, y = k_head_95_cast_fp16)[name = string("op_3826_cast_fp16")];
            fp16 var_3827_to_fp16 = const()[name = string("op_3827_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_47_cast_fp16 = mul(x = var_3826_cast_fp16, y = var_3827_to_fp16)[name = string("scores_content_47_cast_fp16")];
            bool x_245_transpose_x_3 = const()[name = string("x_245_transpose_x_3"), val = bool(false)];
            bool x_245_transpose_y_3 = const()[name = string("x_245_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_245_cast_fp16 = matmul(transpose_x = x_245_transpose_x_3, transpose_y = x_245_transpose_y_3, x = q_v_47_cast_fp16, y = p_head_95_cast_fp16)[name = string("x_245_cast_fp16")];
            tensor<int32, [8]> x_247_pad_1 = const()[name = string("x_247_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_247_mode_1 = const()[name = string("x_247_mode_1"), val = string("constant")];
            fp16 const_1459_to_fp16 = const()[name = string("const_1459_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_247_cast_fp16 = pad(constant_val = const_1459_to_fp16, mode = x_247_mode_1, pad = x_247_pad_1, x = x_245_cast_fp16)[name = string("x_247_cast_fp16")];
            tensor<int32, [4]> var_3841 = const()[name = string("op_3841"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_249_cast_fp16 = reshape(shape = var_3841, x = x_247_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<int32, [4]> var_3845_begin_1 = const()[name = string("op_3845_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3845_end_1 = const()[name = string("op_3845_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_3845_end_mask_1 = const()[name = string("op_3845_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_3845_cast_fp16 = slice_by_index(begin = var_3845_begin_1, end = var_3845_end_1, end_mask = var_3845_end_mask_1, x = x_249_cast_fp16)[name = string("op_3845_cast_fp16")];
            tensor<int32, [4]> var_3847 = const()[name = string("op_3847"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_3848_cast_fp16 = reshape(shape = var_3847, x = var_3845_cast_fp16)[name = string("op_3848_cast_fp16")];
            tensor<int32, [4]> var_3853_begin_1 = const()[name = string("op_3853_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3853_end_1 = const()[name = string("op_3853_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_3853_end_mask_1 = const()[name = string("op_3853_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_3853_cast_fp16 = slice_by_index(begin = var_3853_begin_1, end = var_3853_end_1, end_mask = var_3853_end_mask_1, x = var_3848_cast_fp16)[name = string("op_3853_cast_fp16")];
            fp16 var_3854_to_fp16 = const()[name = string("op_3854_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_47_cast_fp16 = mul(x = var_3853_cast_fp16, y = var_3854_to_fp16)[name = string("scores_pos_47_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_47_cast_fp16 = add(x = scores_content_47_cast_fp16, y = scores_pos_47_cast_fp16)[name = string("logits_47_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_3857_cast_fp16 = softmax(axis = var_2735, x = logits_47_cast_fp16)[name = string("op_3857_cast_fp16")];
            bool var_3859_transpose_x_1 = const()[name = string("op_3859_transpose_x_1"), val = bool(false)];
            bool var_3859_transpose_y_1 = const()[name = string("op_3859_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_3859_cast_fp16 = matmul(transpose_x = var_3859_transpose_x_1, transpose_y = var_3859_transpose_y_1, x = var_3857_cast_fp16, y = v_head_95_cast_fp16)[name = string("op_3859_cast_fp16")];
            tensor<int32, [1]> o_head_5_axes_1 = const()[name = string("o_head_5_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_5_cast_fp16 = squeeze(axes = o_head_5_axes_1, x = var_3859_cast_fp16)[name = string("o_head_5_cast_fp16")];
            bool out_5_interleave_1 = const()[name = string("out_5_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_5_cast_fp16 = concat(axis = var_2735, interleave = out_5_interleave_1, values = (var_3013_cast_fp16, var_3134_cast_fp16, var_3255_cast_fp16, var_3376_cast_fp16, var_3497_cast_fp16, var_3618_cast_fp16, var_3739_cast_fp16, o_head_5_cast_fp16))[name = string("out_5_cast_fp16")];
            tensor<int32, [3]> var_3863_perm_1 = const()[name = string("op_3863_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_129_axes_1 = const()[name = string("input_129_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_3863_cast_fp16 = transpose(perm = var_3863_perm_1, x = out_5_cast_fp16)[name = string("transpose_866")];
            tensor<fp16, [1, 1024, 51, 1]> input_129_cast_fp16 = expand_dims(axes = input_129_axes_1, x = var_3863_cast_fp16)[name = string("input_129_cast_fp16")];
            string dense_output_241_pad_type_1 = const()[name = string("dense_output_241_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_241_strides_1 = const()[name = string("dense_output_241_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_241_pad_1 = const()[name = string("dense_output_241_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_241_dilations_1 = const()[name = string("dense_output_241_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_241_groups_1 = const()[name = string("dense_output_241_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79723072))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80771712))))[name = string("layers_2_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_241_cast_fp16 = conv(dilations = dense_output_241_dilations_1, groups = dense_output_241_groups_1, pad = dense_output_241_pad_1, pad_type = dense_output_241_pad_type_1, strides = dense_output_241_strides_1, weight = layers_2_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_129_cast_fp16)[name = string("dense_output_241_cast_fp16")];
            string sparse_output_241_pad_type_1 = const()[name = string("sparse_output_241_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_241_strides_1 = const()[name = string("sparse_output_241_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_241_pad_1 = const()[name = string("sparse_output_241_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_241_dilations_1 = const()[name = string("sparse_output_241_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_241_groups_1 = const()[name = string("sparse_output_241_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80793344))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80772288))))[name = string("layers_2_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_241_cast_fp16 = conv(dilations = sparse_output_241_dilations_1, groups = sparse_output_241_groups_1, pad = sparse_output_241_pad_1, pad_type = sparse_output_241_pad_type_1, strides = sparse_output_241_strides_1, weight = layers_2_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_129_cast_fp16)[name = string("sparse_output_241_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_5_cast_fp16 = add(x = dense_output_241_cast_fp16, y = sparse_output_241_cast_fp16)[name = string("out_conv_5_cast_fp16")];
            tensor<int32, [1]> var_3880_axes_1 = const()[name = string("op_3880_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_3880_cast_fp16 = squeeze(axes = var_3880_axes_1, x = out_conv_5_cast_fp16)[name = string("op_3880_cast_fp16")];
            tensor<int32, [3]> var_3881_perm_1 = const()[name = string("op_3881_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_3881_cast_fp16 = transpose(perm = var_3881_perm_1, x = var_3880_cast_fp16)[name = string("transpose_865")];
            tensor<fp16, [1, 51, 1024]> input_131_cast_fp16 = add(x = input_119_cast_fp16, y = var_3881_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<int32, [1]> x_253_axes_1 = const()[name = string("x_253_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_2_norm_conv_weight_to_fp16 = const()[name = string("layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80924480)))];
            tensor<fp16, [1024]> layers_2_norm_conv_bias_to_fp16 = const()[name = string("layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80926592)))];
            tensor<fp16, [1, 51, 1024]> x_253_cast_fp16 = layer_norm(axes = x_253_axes_1, beta = layers_2_norm_conv_bias_to_fp16, epsilon = var_2750_to_fp16, gamma = layers_2_norm_conv_weight_to_fp16, x = input_131_cast_fp16)[name = string("x_253_cast_fp16")];
            tensor<int32, [3]> var_3891_perm_1 = const()[name = string("op_3891_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_133_axes_1 = const()[name = string("input_133_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_3891_cast_fp16 = transpose(perm = var_3891_perm_1, x = x_253_cast_fp16)[name = string("transpose_864")];
            tensor<fp16, [1, 1024, 51, 1]> input_133_cast_fp16 = expand_dims(axes = input_133_axes_1, x = var_3891_cast_fp16)[name = string("input_133_cast_fp16")];
            string dense_output_243_pad_type_1 = const()[name = string("dense_output_243_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_243_strides_1 = const()[name = string("dense_output_243_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_243_pad_1 = const()[name = string("dense_output_243_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_243_dilations_1 = const()[name = string("dense_output_243_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_243_groups_1 = const()[name = string("dense_output_243_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_2_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80928704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83025920))))[name = string("layers_2_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_243_cast_fp16 = conv(dilations = dense_output_243_dilations_1, groups = dense_output_243_groups_1, pad = dense_output_243_pad_1, pad_type = dense_output_243_pad_type_1, strides = dense_output_243_strides_1, weight = layers_2_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_133_cast_fp16)[name = string("dense_output_243_cast_fp16")];
            string sparse_output_243_pad_type_1 = const()[name = string("sparse_output_243_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_243_strides_1 = const()[name = string("sparse_output_243_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_243_pad_1 = const()[name = string("sparse_output_243_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_243_dilations_1 = const()[name = string("sparse_output_243_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_243_groups_1 = const()[name = string("sparse_output_243_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_2_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83068544))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83026496))))[name = string("layers_2_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_243_cast_fp16 = conv(dilations = sparse_output_243_dilations_1, groups = sparse_output_243_groups_1, pad = sparse_output_243_pad_1, pad_type = sparse_output_243_pad_type_1, strides = sparse_output_243_strides_1, weight = layers_2_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_133_cast_fp16)[name = string("sparse_output_243_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_135_cast_fp16 = add(x = dense_output_243_cast_fp16, y = sparse_output_243_cast_fp16)[name = string("input_135_cast_fp16")];
            int32 input_137_split_num_splits_1 = const()[name = string("input_137_split_num_splits_1"), val = int32(2)];
            int32 input_137_split_axis_1 = const()[name = string("input_137_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_137_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_137_split_cast_fp16_1 = split(axis = input_137_split_axis_1, num_splits = input_137_split_num_splits_1, x = input_135_cast_fp16)[name = string("input_137_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_137_split_1_sigmoid_cast_fp16 = sigmoid(x = input_137_split_cast_fp16_1)[name = string("input_137_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_137_cast_fp16 = mul(x = input_137_split_cast_fp16_0, y = input_137_split_1_sigmoid_cast_fp16)[name = string("input_137_cast_fp16")];
            tensor<int32, [8]> input_139_pad_1 = const()[name = string("input_139_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_139_mode_1 = const()[name = string("input_139_mode_1"), val = string("constant")];
            fp16 const_1461_to_fp16 = const()[name = string("const_1461_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_139_cast_fp16 = pad(constant_val = const_1461_to_fp16, mode = input_139_mode_1, pad = input_139_pad_1, x = input_137_cast_fp16)[name = string("input_139_cast_fp16")];
            string dense_output_245_pad_type_1 = const()[name = string("dense_output_245_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_245_strides_1 = const()[name = string("dense_output_245_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_245_pad_1 = const()[name = string("dense_output_245_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_245_dilations_1 = const()[name = string("dense_output_245_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_245_groups_1 = const()[name = string("dense_output_245_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_245_cast_fp16 = conv(dilations = dense_output_245_dilations_1, groups = dense_output_245_groups_1, pad = dense_output_245_pad_1, pad_type = dense_output_245_pad_type_1, strides = dense_output_245_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_139_cast_fp16)[name = string("dense_output_245_cast_fp16")];
            string sparse_output_245_pad_type_1 = const()[name = string("sparse_output_245_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_245_strides_1 = const()[name = string("sparse_output_245_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_245_pad_1 = const()[name = string("sparse_output_245_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_245_dilations_1 = const()[name = string("sparse_output_245_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_245_groups_1 = const()[name = string("sparse_output_245_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_2_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83330752))))[name = string("layers_2_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_245_cast_fp16 = conv(dilations = sparse_output_245_dilations_1, groups = sparse_output_245_groups_1, pad = sparse_output_245_pad_1, pad_type = sparse_output_245_pad_type_1, strides = sparse_output_245_strides_1, weight = layers_2_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_139_cast_fp16)[name = string("sparse_output_245_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_141_cast_fp16 = add(x = dense_output_245_cast_fp16, y = sparse_output_245_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [1024]> layers_2_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_2_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83349248)))];
            tensor<fp16, [1024]> layers_2_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_2_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83351360)))];
            tensor<fp16, [1024]> layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83353472)))];
            tensor<fp16, [1024]> layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83355584)))];
            tensor<fp16, [1, 1024, 51, 1]> input_143_cast_fp16 = batch_norm(beta = layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_2750_to_fp16, gamma = layers_2_conv_batch_norm_weight_to_fp16, mean = layers_2_conv_batch_norm_running_mean_to_fp16, variance = layers_2_conv_batch_norm_running_var_to_fp16, x = input_141_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_145_cast_fp16 = silu(x = input_143_cast_fp16)[name = string("input_145_cast_fp16")];
            string dense_output_247_pad_type_1 = const()[name = string("dense_output_247_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_247_strides_1 = const()[name = string("dense_output_247_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_247_pad_1 = const()[name = string("dense_output_247_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_247_dilations_1 = const()[name = string("dense_output_247_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_247_groups_1 = const()[name = string("dense_output_247_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83357696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84406336))))[name = string("layers_2_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_247_cast_fp16 = conv(dilations = dense_output_247_dilations_1, groups = dense_output_247_groups_1, pad = dense_output_247_pad_1, pad_type = dense_output_247_pad_type_1, strides = dense_output_247_strides_1, weight = layers_2_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_145_cast_fp16)[name = string("dense_output_247_cast_fp16")];
            string sparse_output_247_pad_type_1 = const()[name = string("sparse_output_247_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_247_strides_1 = const()[name = string("sparse_output_247_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_247_pad_1 = const()[name = string("sparse_output_247_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_247_dilations_1 = const()[name = string("sparse_output_247_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_247_groups_1 = const()[name = string("sparse_output_247_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_2_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84427968))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84406912))))[name = string("layers_2_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_247_cast_fp16 = conv(dilations = sparse_output_247_dilations_1, groups = sparse_output_247_groups_1, pad = sparse_output_247_pad_1, pad_type = sparse_output_247_pad_type_1, strides = sparse_output_247_strides_1, weight = layers_2_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_145_cast_fp16)[name = string("sparse_output_247_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_255_cast_fp16 = add(x = dense_output_247_cast_fp16, y = sparse_output_247_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<int32, [1]> var_3947_axes_1 = const()[name = string("op_3947_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_3947_cast_fp16 = squeeze(axes = var_3947_axes_1, x = x_255_cast_fp16)[name = string("op_3947_cast_fp16")];
            tensor<int32, [3]> var_3948_perm_1 = const()[name = string("op_3948_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_3948_cast_fp16 = transpose(perm = var_3948_perm_1, x = var_3947_cast_fp16)[name = string("transpose_863")];
            tensor<fp16, [1, 51, 1024]> input_147_cast_fp16 = add(x = input_131_cast_fp16, y = var_3948_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<int32, [1]> x_257_axes_1 = const()[name = string("x_257_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84559104)))];
            tensor<fp16, [1024]> layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84561216)))];
            tensor<fp16, [1, 51, 1024]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_1, beta = layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_2750_to_fp16, gamma = layers_2_norm_feed_forward2_weight_to_fp16, x = input_147_cast_fp16)[name = string("x_257_cast_fp16")];
            tensor<int32, [4]> var_3958 = const()[name = string("op_3958"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_259_cast_fp16 = reshape(shape = var_3958, x = x_257_cast_fp16)[name = string("x_259_cast_fp16")];
            tensor<int32, [4]> input_149_perm_1 = const()[name = string("input_149_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_249_pad_type_1 = const()[name = string("dense_output_249_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_249_strides_1 = const()[name = string("dense_output_249_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_249_pad_1 = const()[name = string("dense_output_249_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_249_dilations_1 = const()[name = string("dense_output_249_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_249_groups_1 = const()[name = string("dense_output_249_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_2_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84563328))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88757696))))[name = string("layers_2_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_149_cast_fp16 = transpose(perm = input_149_perm_1, x = x_259_cast_fp16)[name = string("transpose_862")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_249_cast_fp16 = conv(dilations = dense_output_249_dilations_1, groups = dense_output_249_groups_1, pad = dense_output_249_pad_1, pad_type = dense_output_249_pad_type_1, strides = dense_output_249_strides_1, weight = layers_2_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_149_cast_fp16)[name = string("dense_output_249_cast_fp16")];
            string sparse_output_249_pad_type_1 = const()[name = string("sparse_output_249_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_249_strides_1 = const()[name = string("sparse_output_249_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_249_pad_1 = const()[name = string("sparse_output_249_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_249_dilations_1 = const()[name = string("sparse_output_249_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_249_groups_1 = const()[name = string("sparse_output_249_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_2_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88842240))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88758272))))[name = string("layers_2_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_249_cast_fp16 = conv(dilations = sparse_output_249_dilations_1, groups = sparse_output_249_groups_1, pad = sparse_output_249_pad_1, pad_type = sparse_output_249_pad_type_1, strides = sparse_output_249_strides_1, weight = layers_2_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_149_cast_fp16)[name = string("sparse_output_249_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_151_cast_fp16 = add(x = dense_output_249_cast_fp16, y = sparse_output_249_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_153_cast_fp16 = silu(x = input_151_cast_fp16)[name = string("input_153_cast_fp16")];
            string dense_output_251_pad_type_1 = const()[name = string("dense_output_251_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_251_strides_1 = const()[name = string("dense_output_251_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_251_pad_1 = const()[name = string("dense_output_251_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_251_dilations_1 = const()[name = string("dense_output_251_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_251_groups_1 = const()[name = string("dense_output_251_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_2_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89366592))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93560960))))[name = string("layers_2_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_251_cast_fp16 = conv(dilations = dense_output_251_dilations_1, groups = dense_output_251_groups_1, pad = dense_output_251_pad_1, pad_type = dense_output_251_pad_type_1, strides = dense_output_251_strides_1, weight = layers_2_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_153_cast_fp16)[name = string("dense_output_251_cast_fp16")];
            string sparse_output_251_pad_type_1 = const()[name = string("sparse_output_251_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_251_strides_1 = const()[name = string("sparse_output_251_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_251_pad_1 = const()[name = string("sparse_output_251_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_251_dilations_1 = const()[name = string("sparse_output_251_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_251_groups_1 = const()[name = string("sparse_output_251_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_2_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93645504))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93561536))))[name = string("layers_2_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_251_cast_fp16 = conv(dilations = sparse_output_251_dilations_1, groups = sparse_output_251_groups_1, pad = sparse_output_251_pad_1, pad_type = sparse_output_251_pad_type_1, strides = sparse_output_251_strides_1, weight = layers_2_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_153_cast_fp16)[name = string("sparse_output_251_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_261_cast_fp16 = add(x = dense_output_251_cast_fp16, y = sparse_output_251_cast_fp16)[name = string("x_261_cast_fp16")];
            tensor<int32, [4]> x_263_perm_1 = const()[name = string("x_263_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_3993 = const()[name = string("op_3993"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_263_cast_fp16 = transpose(perm = x_263_perm_1, x = x_261_cast_fp16)[name = string("transpose_861")];
            tensor<fp16, [1, 51, 1024]> var_3994_cast_fp16 = reshape(shape = var_3993, x = x_263_cast_fp16)[name = string("op_3994_cast_fp16")];
            fp16 var_3995_to_fp16 = const()[name = string("op_3995_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_3996_cast_fp16 = mul(x = var_3994_cast_fp16, y = var_3995_to_fp16)[name = string("op_3996_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_155_cast_fp16 = add(x = input_147_cast_fp16, y = var_3996_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<int32, [1]> input_157_axes_1 = const()[name = string("input_157_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_2_norm_out_weight_to_fp16 = const()[name = string("layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94169856)))];
            tensor<fp16, [1024]> layers_2_norm_out_bias_to_fp16 = const()[name = string("layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94171968)))];
            tensor<fp16, [1, 51, 1024]> input_157_cast_fp16 = layer_norm(axes = input_157_axes_1, beta = layers_2_norm_out_bias_to_fp16, epsilon = var_2750_to_fp16, gamma = layers_2_norm_out_weight_to_fp16, x = input_155_cast_fp16)[name = string("input_157_cast_fp16")];
            int32 var_4004 = const()[name = string("op_4004"), val = int32(-1)];
            tensor<int32, [1]> x_265_axes_1 = const()[name = string("x_265_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94174080)))];
            tensor<fp16, [1024]> layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94176192)))];
            fp16 var_4019_to_fp16 = const()[name = string("op_4019_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_265_cast_fp16 = layer_norm(axes = x_265_axes_1, beta = layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_4019_to_fp16, gamma = layers_3_norm_feed_forward1_weight_to_fp16, x = input_157_cast_fp16)[name = string("x_265_cast_fp16")];
            tensor<int32, [4]> var_4038 = const()[name = string("op_4038"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_267_cast_fp16 = reshape(shape = var_4038, x = x_265_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<int32, [4]> input_159_perm_1 = const()[name = string("input_159_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_253_pad_type_1 = const()[name = string("dense_output_253_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_253_strides_1 = const()[name = string("dense_output_253_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_253_pad_1 = const()[name = string("dense_output_253_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_253_dilations_1 = const()[name = string("dense_output_253_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_253_groups_1 = const()[name = string("dense_output_253_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_3_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94178304))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98372672))))[name = string("layers_3_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_159_cast_fp16 = transpose(perm = input_159_perm_1, x = x_267_cast_fp16)[name = string("transpose_860")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_253_cast_fp16 = conv(dilations = dense_output_253_dilations_1, groups = dense_output_253_groups_1, pad = dense_output_253_pad_1, pad_type = dense_output_253_pad_type_1, strides = dense_output_253_strides_1, weight = layers_3_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_159_cast_fp16)[name = string("dense_output_253_cast_fp16")];
            string sparse_output_253_pad_type_1 = const()[name = string("sparse_output_253_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_253_strides_1 = const()[name = string("sparse_output_253_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_253_pad_1 = const()[name = string("sparse_output_253_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_253_dilations_1 = const()[name = string("sparse_output_253_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_253_groups_1 = const()[name = string("sparse_output_253_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_3_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98457216))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98373248))))[name = string("layers_3_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_253_cast_fp16 = conv(dilations = sparse_output_253_dilations_1, groups = sparse_output_253_groups_1, pad = sparse_output_253_pad_1, pad_type = sparse_output_253_pad_type_1, strides = sparse_output_253_strides_1, weight = layers_3_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_159_cast_fp16)[name = string("sparse_output_253_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_161_cast_fp16 = add(x = dense_output_253_cast_fp16, y = sparse_output_253_cast_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
            string dense_output_255_pad_type_1 = const()[name = string("dense_output_255_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_255_strides_1 = const()[name = string("dense_output_255_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_255_pad_1 = const()[name = string("dense_output_255_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_255_dilations_1 = const()[name = string("dense_output_255_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_255_groups_1 = const()[name = string("dense_output_255_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_3_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98981568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103175936))))[name = string("layers_3_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_255_cast_fp16 = conv(dilations = dense_output_255_dilations_1, groups = dense_output_255_groups_1, pad = dense_output_255_pad_1, pad_type = dense_output_255_pad_type_1, strides = dense_output_255_strides_1, weight = layers_3_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_163_cast_fp16)[name = string("dense_output_255_cast_fp16")];
            string sparse_output_255_pad_type_1 = const()[name = string("sparse_output_255_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_255_strides_1 = const()[name = string("sparse_output_255_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_255_pad_1 = const()[name = string("sparse_output_255_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_255_dilations_1 = const()[name = string("sparse_output_255_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_255_groups_1 = const()[name = string("sparse_output_255_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_3_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103260480))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103176512))))[name = string("layers_3_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_255_cast_fp16 = conv(dilations = sparse_output_255_dilations_1, groups = sparse_output_255_groups_1, pad = sparse_output_255_pad_1, pad_type = sparse_output_255_pad_type_1, strides = sparse_output_255_strides_1, weight = layers_3_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_163_cast_fp16)[name = string("sparse_output_255_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_269_cast_fp16 = add(x = dense_output_255_cast_fp16, y = sparse_output_255_cast_fp16)[name = string("x_269_cast_fp16")];
            tensor<int32, [4]> x_271_perm_1 = const()[name = string("x_271_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_4073 = const()[name = string("op_4073"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_271_cast_fp16 = transpose(perm = x_271_perm_1, x = x_269_cast_fp16)[name = string("transpose_859")];
            tensor<fp16, [1, 51, 1024]> var_4074_cast_fp16 = reshape(shape = var_4073, x = x_271_cast_fp16)[name = string("op_4074_cast_fp16")];
            fp16 var_4075_to_fp16 = const()[name = string("op_4075_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_4076_cast_fp16 = mul(x = var_4074_cast_fp16, y = var_4075_to_fp16)[name = string("op_4076_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_165_cast_fp16 = add(x = input_157_cast_fp16, y = var_4076_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<int32, [1]> q_7_axes_1 = const()[name = string("q_7_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_3_norm_self_att_weight_to_fp16 = const()[name = string("layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103784832)))];
            tensor<fp16, [1024]> layers_3_norm_self_att_bias_to_fp16 = const()[name = string("layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103786944)))];
            tensor<fp16, [1, 51, 1024]> q_7_cast_fp16 = layer_norm(axes = q_7_axes_1, beta = layers_3_norm_self_att_bias_to_fp16, epsilon = var_4019_to_fp16, gamma = layers_3_norm_self_att_weight_to_fp16, x = input_165_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<int32, [3]> var_4150 = const()[name = string("op_4150"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_167_axes_1 = const()[name = string("input_167_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_4151_cast_fp16 = transpose(perm = var_4150, x = q_7_cast_fp16)[name = string("transpose_858")];
            tensor<fp16, [1, 1024, 51, 1]> input_167_cast_fp16 = expand_dims(axes = input_167_axes_1, x = var_4151_cast_fp16)[name = string("input_167_cast_fp16")];
            string dense_output_257_pad_type_1 = const()[name = string("dense_output_257_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_257_strides_1 = const()[name = string("dense_output_257_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_257_pad_1 = const()[name = string("dense_output_257_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_257_dilations_1 = const()[name = string("dense_output_257_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_257_groups_1 = const()[name = string("dense_output_257_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103789056))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103920192))))[name = string("layers_3_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_257_cast_fp16 = conv(dilations = dense_output_257_dilations_1, groups = dense_output_257_groups_1, pad = dense_output_257_pad_1, pad_type = dense_output_257_pad_type_1, strides = dense_output_257_strides_1, weight = layers_3_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_257_cast_fp16")];
            string sparse_output_257_pad_type_1 = const()[name = string("sparse_output_257_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_257_strides_1 = const()[name = string("sparse_output_257_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_257_pad_1 = const()[name = string("sparse_output_257_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_257_dilations_1 = const()[name = string("sparse_output_257_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_257_groups_1 = const()[name = string("sparse_output_257_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103923456))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103920768))))[name = string("layers_3_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_257_cast_fp16 = conv(dilations = sparse_output_257_dilations_1, groups = sparse_output_257_groups_1, pad = sparse_output_257_pad_1, pad_type = sparse_output_257_pad_type_1, strides = sparse_output_257_strides_1, weight = layers_3_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_257_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4176_cast_fp16 = add(x = dense_output_257_cast_fp16, y = sparse_output_257_cast_fp16)[name = string("op_4176_cast_fp16")];
            tensor<int32, [4]> var_4177 = const()[name = string("op_4177"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4179 = const()[name = string("op_4179"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4178_cast_fp16 = transpose(perm = var_4177, x = var_4176_cast_fp16)[name = string("transpose_857")];
            tensor<fp16, [1, 51, 128]> q_head_49_cast_fp16 = reshape(shape = var_4179, x = var_4178_cast_fp16)[name = string("q_head_49_cast_fp16")];
            string dense_output_259_pad_type_1 = const()[name = string("dense_output_259_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_259_strides_1 = const()[name = string("dense_output_259_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_259_pad_1 = const()[name = string("dense_output_259_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_259_dilations_1 = const()[name = string("dense_output_259_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_259_groups_1 = const()[name = string("dense_output_259_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103939904))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104071040))))[name = string("layers_3_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_259_cast_fp16 = conv(dilations = dense_output_259_dilations_1, groups = dense_output_259_groups_1, pad = dense_output_259_pad_1, pad_type = dense_output_259_pad_type_1, strides = dense_output_259_strides_1, weight = layers_3_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_259_cast_fp16")];
            string sparse_output_259_pad_type_1 = const()[name = string("sparse_output_259_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_259_strides_1 = const()[name = string("sparse_output_259_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_259_pad_1 = const()[name = string("sparse_output_259_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_259_dilations_1 = const()[name = string("sparse_output_259_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_259_groups_1 = const()[name = string("sparse_output_259_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104074304))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104071616))))[name = string("layers_3_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_259_cast_fp16 = conv(dilations = sparse_output_259_dilations_1, groups = sparse_output_259_groups_1, pad = sparse_output_259_pad_1, pad_type = sparse_output_259_pad_type_1, strides = sparse_output_259_strides_1, weight = layers_3_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_259_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4195_cast_fp16 = add(x = dense_output_259_cast_fp16, y = sparse_output_259_cast_fp16)[name = string("op_4195_cast_fp16")];
            tensor<int32, [4]> var_4196 = const()[name = string("op_4196"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4198 = const()[name = string("op_4198"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4197_cast_fp16 = transpose(perm = var_4196, x = var_4195_cast_fp16)[name = string("transpose_856")];
            tensor<fp16, [1, 51, 128]> k_head_97_cast_fp16 = reshape(shape = var_4198, x = var_4197_cast_fp16)[name = string("k_head_97_cast_fp16")];
            string dense_output_261_pad_type_1 = const()[name = string("dense_output_261_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_261_strides_1 = const()[name = string("dense_output_261_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_261_pad_1 = const()[name = string("dense_output_261_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_261_dilations_1 = const()[name = string("dense_output_261_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_261_groups_1 = const()[name = string("dense_output_261_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104090752))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104221888))))[name = string("layers_3_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_261_cast_fp16 = conv(dilations = dense_output_261_dilations_1, groups = dense_output_261_groups_1, pad = dense_output_261_pad_1, pad_type = dense_output_261_pad_type_1, strides = dense_output_261_strides_1, weight = layers_3_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_261_cast_fp16")];
            string sparse_output_261_pad_type_1 = const()[name = string("sparse_output_261_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_261_strides_1 = const()[name = string("sparse_output_261_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_261_pad_1 = const()[name = string("sparse_output_261_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_261_dilations_1 = const()[name = string("sparse_output_261_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_261_groups_1 = const()[name = string("sparse_output_261_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104225152))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104222464))))[name = string("layers_3_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_261_cast_fp16 = conv(dilations = sparse_output_261_dilations_1, groups = sparse_output_261_groups_1, pad = sparse_output_261_pad_1, pad_type = sparse_output_261_pad_type_1, strides = sparse_output_261_strides_1, weight = layers_3_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_261_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4214_cast_fp16 = add(x = dense_output_261_cast_fp16, y = sparse_output_261_cast_fp16)[name = string("op_4214_cast_fp16")];
            tensor<int32, [4]> var_4215 = const()[name = string("op_4215"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4217 = const()[name = string("op_4217"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4216_cast_fp16 = transpose(perm = var_4215, x = var_4214_cast_fp16)[name = string("transpose_855")];
            tensor<fp16, [1, 51, 128]> v_head_97_cast_fp16 = reshape(shape = var_4217, x = var_4216_cast_fp16)[name = string("v_head_97_cast_fp16")];
            string dense_output_263_pad_type_1 = const()[name = string("dense_output_263_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_263_strides_1 = const()[name = string("dense_output_263_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_263_pad_1 = const()[name = string("dense_output_263_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_263_dilations_1 = const()[name = string("dense_output_263_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_263_groups_1 = const()[name = string("dense_output_263_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104241600))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104372736))))[name = string("layers_3_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_263_cast_fp16 = conv(dilations = dense_output_263_dilations_1, groups = dense_output_263_groups_1, pad = dense_output_263_pad_1, pad_type = dense_output_263_pad_type_1, strides = dense_output_263_strides_1, weight = layers_3_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_263_cast_fp16")];
            string sparse_output_263_pad_type_1 = const()[name = string("sparse_output_263_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_263_strides_1 = const()[name = string("sparse_output_263_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_263_pad_1 = const()[name = string("sparse_output_263_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_263_dilations_1 = const()[name = string("sparse_output_263_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_263_groups_1 = const()[name = string("sparse_output_263_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104376000))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104373312))))[name = string("layers_3_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_263_cast_fp16 = conv(dilations = sparse_output_263_dilations_1, groups = sparse_output_263_groups_1, pad = sparse_output_263_pad_1, pad_type = sparse_output_263_pad_type_1, strides = sparse_output_263_strides_1, weight = layers_3_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_263_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_4233_cast_fp16 = add(x = dense_output_263_cast_fp16, y = sparse_output_263_cast_fp16)[name = string("op_4233_cast_fp16")];
            tensor<int32, [4]> var_4234 = const()[name = string("op_4234"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4236 = const()[name = string("op_4236"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_4235_cast_fp16 = transpose(perm = var_4234, x = var_4233_cast_fp16)[name = string("transpose_854")];
            tensor<fp16, [1, 101, 128]> p_head_97_cast_fp16 = reshape(shape = var_4236, x = var_4235_cast_fp16)[name = string("p_head_97_cast_fp16")];
            tensor<fp16, [128]> var_4238_to_fp16 = const()[name = string("op_4238_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104392448)))];
            tensor<fp16, [1, 51, 128]> var_4239_cast_fp16 = add(x = q_head_49_cast_fp16, y = var_4238_to_fp16)[name = string("op_4239_cast_fp16")];
            tensor<int32, [1]> q_u_49_axes_1 = const()[name = string("q_u_49_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_49_cast_fp16 = expand_dims(axes = q_u_49_axes_1, x = var_4239_cast_fp16)[name = string("q_u_49_cast_fp16")];
            tensor<fp16, [128]> var_4241_to_fp16 = const()[name = string("op_4241_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104392768)))];
            tensor<fp16, [1, 51, 128]> var_4242_cast_fp16 = add(x = q_head_49_cast_fp16, y = var_4241_to_fp16)[name = string("op_4242_cast_fp16")];
            tensor<int32, [1]> q_v_49_axes_1 = const()[name = string("q_v_49_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_49_cast_fp16 = expand_dims(axes = q_v_49_axes_1, x = var_4242_cast_fp16)[name = string("q_v_49_cast_fp16")];
            tensor<int32, [1]> k_head_99_axes_1 = const()[name = string("k_head_99_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_99_cast_fp16 = expand_dims(axes = k_head_99_axes_1, x = k_head_97_cast_fp16)[name = string("k_head_99_cast_fp16")];
            tensor<int32, [1]> v_head_99_axes_1 = const()[name = string("v_head_99_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_99_cast_fp16 = expand_dims(axes = v_head_99_axes_1, x = v_head_97_cast_fp16)[name = string("v_head_99_cast_fp16")];
            tensor<int32, [1]> p_head_99_axes_1 = const()[name = string("p_head_99_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_99_cast_fp16 = expand_dims(axes = p_head_99_axes_1, x = p_head_97_cast_fp16)[name = string("p_head_99_cast_fp16")];
            bool var_4248_transpose_x_3 = const()[name = string("op_4248_transpose_x_3"), val = bool(false)];
            bool var_4248_transpose_y_3 = const()[name = string("op_4248_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_4248_cast_fp16 = matmul(transpose_x = var_4248_transpose_x_3, transpose_y = var_4248_transpose_y_3, x = q_u_49_cast_fp16, y = k_head_99_cast_fp16)[name = string("op_4248_cast_fp16")];
            fp16 var_4249_to_fp16 = const()[name = string("op_4249_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_49_cast_fp16 = mul(x = var_4248_cast_fp16, y = var_4249_to_fp16)[name = string("scores_content_49_cast_fp16")];
            bool x_273_transpose_x_3 = const()[name = string("x_273_transpose_x_3"), val = bool(false)];
            bool x_273_transpose_y_3 = const()[name = string("x_273_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_3, transpose_y = x_273_transpose_y_3, x = q_v_49_cast_fp16, y = p_head_99_cast_fp16)[name = string("x_273_cast_fp16")];
            tensor<int32, [8]> x_275_pad_1 = const()[name = string("x_275_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_275_mode_1 = const()[name = string("x_275_mode_1"), val = string("constant")];
            fp16 const_1471_to_fp16 = const()[name = string("const_1471_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_275_cast_fp16 = pad(constant_val = const_1471_to_fp16, mode = x_275_mode_1, pad = x_275_pad_1, x = x_273_cast_fp16)[name = string("x_275_cast_fp16")];
            tensor<int32, [4]> var_4263 = const()[name = string("op_4263"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_277_cast_fp16 = reshape(shape = var_4263, x = x_275_cast_fp16)[name = string("x_277_cast_fp16")];
            tensor<int32, [4]> var_4267_begin_1 = const()[name = string("op_4267_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4267_end_1 = const()[name = string("op_4267_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_4267_end_mask_1 = const()[name = string("op_4267_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_4267_cast_fp16 = slice_by_index(begin = var_4267_begin_1, end = var_4267_end_1, end_mask = var_4267_end_mask_1, x = x_277_cast_fp16)[name = string("op_4267_cast_fp16")];
            tensor<int32, [4]> var_4269 = const()[name = string("op_4269"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_4270_cast_fp16 = reshape(shape = var_4269, x = var_4267_cast_fp16)[name = string("op_4270_cast_fp16")];
            tensor<int32, [4]> var_4275_begin_1 = const()[name = string("op_4275_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4275_end_1 = const()[name = string("op_4275_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_4275_end_mask_1 = const()[name = string("op_4275_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_4275_cast_fp16 = slice_by_index(begin = var_4275_begin_1, end = var_4275_end_1, end_mask = var_4275_end_mask_1, x = var_4270_cast_fp16)[name = string("op_4275_cast_fp16")];
            fp16 var_4276_to_fp16 = const()[name = string("op_4276_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_49_cast_fp16 = mul(x = var_4275_cast_fp16, y = var_4276_to_fp16)[name = string("scores_pos_49_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_49_cast_fp16 = add(x = scores_content_49_cast_fp16, y = scores_pos_49_cast_fp16)[name = string("logits_49_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_4279_cast_fp16 = softmax(axis = var_4004, x = logits_49_cast_fp16)[name = string("op_4279_cast_fp16")];
            bool var_4281_transpose_x_1 = const()[name = string("op_4281_transpose_x_1"), val = bool(false)];
            bool var_4281_transpose_y_1 = const()[name = string("op_4281_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_4281_cast_fp16 = matmul(transpose_x = var_4281_transpose_x_1, transpose_y = var_4281_transpose_y_1, x = var_4279_cast_fp16, y = v_head_99_cast_fp16)[name = string("op_4281_cast_fp16")];
            tensor<int32, [1]> var_4282_axes_1 = const()[name = string("op_4282_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_4282_cast_fp16 = squeeze(axes = var_4282_axes_1, x = var_4281_cast_fp16)[name = string("op_4282_cast_fp16")];
            string dense_output_265_pad_type_1 = const()[name = string("dense_output_265_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_265_strides_1 = const()[name = string("dense_output_265_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_265_pad_1 = const()[name = string("dense_output_265_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_265_dilations_1 = const()[name = string("dense_output_265_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_265_groups_1 = const()[name = string("dense_output_265_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104393088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104524224))))[name = string("layers_3_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_265_cast_fp16 = conv(dilations = dense_output_265_dilations_1, groups = dense_output_265_groups_1, pad = dense_output_265_pad_1, pad_type = dense_output_265_pad_type_1, strides = dense_output_265_strides_1, weight = layers_3_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_265_cast_fp16")];
            string sparse_output_265_pad_type_1 = const()[name = string("sparse_output_265_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_265_strides_1 = const()[name = string("sparse_output_265_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_265_pad_1 = const()[name = string("sparse_output_265_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_265_dilations_1 = const()[name = string("sparse_output_265_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_265_groups_1 = const()[name = string("sparse_output_265_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104527488))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104524800))))[name = string("layers_3_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_265_cast_fp16 = conv(dilations = sparse_output_265_dilations_1, groups = sparse_output_265_groups_1, pad = sparse_output_265_pad_1, pad_type = sparse_output_265_pad_type_1, strides = sparse_output_265_strides_1, weight = layers_3_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_265_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4297_cast_fp16 = add(x = dense_output_265_cast_fp16, y = sparse_output_265_cast_fp16)[name = string("op_4297_cast_fp16")];
            tensor<int32, [4]> var_4298 = const()[name = string("op_4298"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4300 = const()[name = string("op_4300"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4299_cast_fp16 = transpose(perm = var_4298, x = var_4297_cast_fp16)[name = string("transpose_853")];
            tensor<fp16, [1, 51, 128]> q_head_51_cast_fp16 = reshape(shape = var_4300, x = var_4299_cast_fp16)[name = string("q_head_51_cast_fp16")];
            string dense_output_267_pad_type_1 = const()[name = string("dense_output_267_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_267_strides_1 = const()[name = string("dense_output_267_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_267_pad_1 = const()[name = string("dense_output_267_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_267_dilations_1 = const()[name = string("dense_output_267_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_267_groups_1 = const()[name = string("dense_output_267_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104543936))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104675072))))[name = string("layers_3_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_267_cast_fp16 = conv(dilations = dense_output_267_dilations_1, groups = dense_output_267_groups_1, pad = dense_output_267_pad_1, pad_type = dense_output_267_pad_type_1, strides = dense_output_267_strides_1, weight = layers_3_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_267_cast_fp16")];
            string sparse_output_267_pad_type_1 = const()[name = string("sparse_output_267_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_267_strides_1 = const()[name = string("sparse_output_267_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_267_pad_1 = const()[name = string("sparse_output_267_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_267_dilations_1 = const()[name = string("sparse_output_267_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_267_groups_1 = const()[name = string("sparse_output_267_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104678336))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104675648))))[name = string("layers_3_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_267_cast_fp16 = conv(dilations = sparse_output_267_dilations_1, groups = sparse_output_267_groups_1, pad = sparse_output_267_pad_1, pad_type = sparse_output_267_pad_type_1, strides = sparse_output_267_strides_1, weight = layers_3_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_267_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4316_cast_fp16 = add(x = dense_output_267_cast_fp16, y = sparse_output_267_cast_fp16)[name = string("op_4316_cast_fp16")];
            tensor<int32, [4]> var_4317 = const()[name = string("op_4317"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4319 = const()[name = string("op_4319"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4318_cast_fp16 = transpose(perm = var_4317, x = var_4316_cast_fp16)[name = string("transpose_852")];
            tensor<fp16, [1, 51, 128]> k_head_101_cast_fp16 = reshape(shape = var_4319, x = var_4318_cast_fp16)[name = string("k_head_101_cast_fp16")];
            string dense_output_269_pad_type_1 = const()[name = string("dense_output_269_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_269_strides_1 = const()[name = string("dense_output_269_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_269_pad_1 = const()[name = string("dense_output_269_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_269_dilations_1 = const()[name = string("dense_output_269_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_269_groups_1 = const()[name = string("dense_output_269_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104694784))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104825920))))[name = string("layers_3_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_269_cast_fp16 = conv(dilations = dense_output_269_dilations_1, groups = dense_output_269_groups_1, pad = dense_output_269_pad_1, pad_type = dense_output_269_pad_type_1, strides = dense_output_269_strides_1, weight = layers_3_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_269_cast_fp16")];
            string sparse_output_269_pad_type_1 = const()[name = string("sparse_output_269_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_269_strides_1 = const()[name = string("sparse_output_269_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_269_pad_1 = const()[name = string("sparse_output_269_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_269_dilations_1 = const()[name = string("sparse_output_269_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_269_groups_1 = const()[name = string("sparse_output_269_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104829184))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104826496))))[name = string("layers_3_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_269_cast_fp16 = conv(dilations = sparse_output_269_dilations_1, groups = sparse_output_269_groups_1, pad = sparse_output_269_pad_1, pad_type = sparse_output_269_pad_type_1, strides = sparse_output_269_strides_1, weight = layers_3_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_269_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4335_cast_fp16 = add(x = dense_output_269_cast_fp16, y = sparse_output_269_cast_fp16)[name = string("op_4335_cast_fp16")];
            tensor<int32, [4]> var_4336 = const()[name = string("op_4336"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4338 = const()[name = string("op_4338"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4337_cast_fp16 = transpose(perm = var_4336, x = var_4335_cast_fp16)[name = string("transpose_851")];
            tensor<fp16, [1, 51, 128]> v_head_101_cast_fp16 = reshape(shape = var_4338, x = var_4337_cast_fp16)[name = string("v_head_101_cast_fp16")];
            string dense_output_271_pad_type_1 = const()[name = string("dense_output_271_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_271_strides_1 = const()[name = string("dense_output_271_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_271_pad_1 = const()[name = string("dense_output_271_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_271_dilations_1 = const()[name = string("dense_output_271_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_271_groups_1 = const()[name = string("dense_output_271_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104845632))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104976768))))[name = string("layers_3_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_271_cast_fp16 = conv(dilations = dense_output_271_dilations_1, groups = dense_output_271_groups_1, pad = dense_output_271_pad_1, pad_type = dense_output_271_pad_type_1, strides = dense_output_271_strides_1, weight = layers_3_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_271_cast_fp16")];
            string sparse_output_271_pad_type_1 = const()[name = string("sparse_output_271_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_271_strides_1 = const()[name = string("sparse_output_271_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_271_pad_1 = const()[name = string("sparse_output_271_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_271_dilations_1 = const()[name = string("sparse_output_271_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_271_groups_1 = const()[name = string("sparse_output_271_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104980032))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104977344))))[name = string("layers_3_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_271_cast_fp16 = conv(dilations = sparse_output_271_dilations_1, groups = sparse_output_271_groups_1, pad = sparse_output_271_pad_1, pad_type = sparse_output_271_pad_type_1, strides = sparse_output_271_strides_1, weight = layers_3_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_271_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_4354_cast_fp16 = add(x = dense_output_271_cast_fp16, y = sparse_output_271_cast_fp16)[name = string("op_4354_cast_fp16")];
            tensor<int32, [4]> var_4355 = const()[name = string("op_4355"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4357 = const()[name = string("op_4357"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_4356_cast_fp16 = transpose(perm = var_4355, x = var_4354_cast_fp16)[name = string("transpose_850")];
            tensor<fp16, [1, 101, 128]> p_head_101_cast_fp16 = reshape(shape = var_4357, x = var_4356_cast_fp16)[name = string("p_head_101_cast_fp16")];
            tensor<fp16, [128]> var_4359_to_fp16 = const()[name = string("op_4359_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104996480)))];
            tensor<fp16, [1, 51, 128]> var_4360_cast_fp16 = add(x = q_head_51_cast_fp16, y = var_4359_to_fp16)[name = string("op_4360_cast_fp16")];
            tensor<int32, [1]> q_u_51_axes_1 = const()[name = string("q_u_51_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_51_cast_fp16 = expand_dims(axes = q_u_51_axes_1, x = var_4360_cast_fp16)[name = string("q_u_51_cast_fp16")];
            tensor<fp16, [128]> var_4362_to_fp16 = const()[name = string("op_4362_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104996800)))];
            tensor<fp16, [1, 51, 128]> var_4363_cast_fp16 = add(x = q_head_51_cast_fp16, y = var_4362_to_fp16)[name = string("op_4363_cast_fp16")];
            tensor<int32, [1]> q_v_51_axes_1 = const()[name = string("q_v_51_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_51_cast_fp16 = expand_dims(axes = q_v_51_axes_1, x = var_4363_cast_fp16)[name = string("q_v_51_cast_fp16")];
            tensor<int32, [1]> k_head_103_axes_1 = const()[name = string("k_head_103_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_103_cast_fp16 = expand_dims(axes = k_head_103_axes_1, x = k_head_101_cast_fp16)[name = string("k_head_103_cast_fp16")];
            tensor<int32, [1]> v_head_103_axes_1 = const()[name = string("v_head_103_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_103_cast_fp16 = expand_dims(axes = v_head_103_axes_1, x = v_head_101_cast_fp16)[name = string("v_head_103_cast_fp16")];
            tensor<int32, [1]> p_head_103_axes_1 = const()[name = string("p_head_103_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_103_cast_fp16 = expand_dims(axes = p_head_103_axes_1, x = p_head_101_cast_fp16)[name = string("p_head_103_cast_fp16")];
            bool var_4369_transpose_x_3 = const()[name = string("op_4369_transpose_x_3"), val = bool(false)];
            bool var_4369_transpose_y_3 = const()[name = string("op_4369_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_4369_cast_fp16 = matmul(transpose_x = var_4369_transpose_x_3, transpose_y = var_4369_transpose_y_3, x = q_u_51_cast_fp16, y = k_head_103_cast_fp16)[name = string("op_4369_cast_fp16")];
            fp16 var_4370_to_fp16 = const()[name = string("op_4370_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_51_cast_fp16 = mul(x = var_4369_cast_fp16, y = var_4370_to_fp16)[name = string("scores_content_51_cast_fp16")];
            bool x_281_transpose_x_3 = const()[name = string("x_281_transpose_x_3"), val = bool(false)];
            bool x_281_transpose_y_3 = const()[name = string("x_281_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_281_cast_fp16 = matmul(transpose_x = x_281_transpose_x_3, transpose_y = x_281_transpose_y_3, x = q_v_51_cast_fp16, y = p_head_103_cast_fp16)[name = string("x_281_cast_fp16")];
            tensor<int32, [8]> x_283_pad_1 = const()[name = string("x_283_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_283_mode_1 = const()[name = string("x_283_mode_1"), val = string("constant")];
            fp16 const_1477_to_fp16 = const()[name = string("const_1477_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_283_cast_fp16 = pad(constant_val = const_1477_to_fp16, mode = x_283_mode_1, pad = x_283_pad_1, x = x_281_cast_fp16)[name = string("x_283_cast_fp16")];
            tensor<int32, [4]> var_4384 = const()[name = string("op_4384"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_285_cast_fp16 = reshape(shape = var_4384, x = x_283_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<int32, [4]> var_4388_begin_1 = const()[name = string("op_4388_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4388_end_1 = const()[name = string("op_4388_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_4388_end_mask_1 = const()[name = string("op_4388_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_4388_cast_fp16 = slice_by_index(begin = var_4388_begin_1, end = var_4388_end_1, end_mask = var_4388_end_mask_1, x = x_285_cast_fp16)[name = string("op_4388_cast_fp16")];
            tensor<int32, [4]> var_4390 = const()[name = string("op_4390"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_4391_cast_fp16 = reshape(shape = var_4390, x = var_4388_cast_fp16)[name = string("op_4391_cast_fp16")];
            tensor<int32, [4]> var_4396_begin_1 = const()[name = string("op_4396_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4396_end_1 = const()[name = string("op_4396_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_4396_end_mask_1 = const()[name = string("op_4396_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_4396_cast_fp16 = slice_by_index(begin = var_4396_begin_1, end = var_4396_end_1, end_mask = var_4396_end_mask_1, x = var_4391_cast_fp16)[name = string("op_4396_cast_fp16")];
            fp16 var_4397_to_fp16 = const()[name = string("op_4397_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_51_cast_fp16 = mul(x = var_4396_cast_fp16, y = var_4397_to_fp16)[name = string("scores_pos_51_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_51_cast_fp16 = add(x = scores_content_51_cast_fp16, y = scores_pos_51_cast_fp16)[name = string("logits_51_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_4400_cast_fp16 = softmax(axis = var_4004, x = logits_51_cast_fp16)[name = string("op_4400_cast_fp16")];
            bool var_4402_transpose_x_1 = const()[name = string("op_4402_transpose_x_1"), val = bool(false)];
            bool var_4402_transpose_y_1 = const()[name = string("op_4402_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_4402_cast_fp16 = matmul(transpose_x = var_4402_transpose_x_1, transpose_y = var_4402_transpose_y_1, x = var_4400_cast_fp16, y = v_head_103_cast_fp16)[name = string("op_4402_cast_fp16")];
            tensor<int32, [1]> var_4403_axes_1 = const()[name = string("op_4403_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_4403_cast_fp16 = squeeze(axes = var_4403_axes_1, x = var_4402_cast_fp16)[name = string("op_4403_cast_fp16")];
            string dense_output_273_pad_type_1 = const()[name = string("dense_output_273_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_273_strides_1 = const()[name = string("dense_output_273_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_273_pad_1 = const()[name = string("dense_output_273_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_273_dilations_1 = const()[name = string("dense_output_273_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_273_groups_1 = const()[name = string("dense_output_273_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104997120))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105128256))))[name = string("layers_3_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_273_cast_fp16 = conv(dilations = dense_output_273_dilations_1, groups = dense_output_273_groups_1, pad = dense_output_273_pad_1, pad_type = dense_output_273_pad_type_1, strides = dense_output_273_strides_1, weight = layers_3_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_273_cast_fp16")];
            string sparse_output_273_pad_type_1 = const()[name = string("sparse_output_273_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_273_strides_1 = const()[name = string("sparse_output_273_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_273_pad_1 = const()[name = string("sparse_output_273_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_273_dilations_1 = const()[name = string("sparse_output_273_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_273_groups_1 = const()[name = string("sparse_output_273_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105131520))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105128832))))[name = string("layers_3_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_273_cast_fp16 = conv(dilations = sparse_output_273_dilations_1, groups = sparse_output_273_groups_1, pad = sparse_output_273_pad_1, pad_type = sparse_output_273_pad_type_1, strides = sparse_output_273_strides_1, weight = layers_3_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_273_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4418_cast_fp16 = add(x = dense_output_273_cast_fp16, y = sparse_output_273_cast_fp16)[name = string("op_4418_cast_fp16")];
            tensor<int32, [4]> var_4419 = const()[name = string("op_4419"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4421 = const()[name = string("op_4421"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4420_cast_fp16 = transpose(perm = var_4419, x = var_4418_cast_fp16)[name = string("transpose_849")];
            tensor<fp16, [1, 51, 128]> q_head_53_cast_fp16 = reshape(shape = var_4421, x = var_4420_cast_fp16)[name = string("q_head_53_cast_fp16")];
            string dense_output_275_pad_type_1 = const()[name = string("dense_output_275_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_275_strides_1 = const()[name = string("dense_output_275_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_275_pad_1 = const()[name = string("dense_output_275_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_275_dilations_1 = const()[name = string("dense_output_275_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_275_groups_1 = const()[name = string("dense_output_275_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105147968))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105279104))))[name = string("layers_3_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_275_cast_fp16 = conv(dilations = dense_output_275_dilations_1, groups = dense_output_275_groups_1, pad = dense_output_275_pad_1, pad_type = dense_output_275_pad_type_1, strides = dense_output_275_strides_1, weight = layers_3_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_275_cast_fp16")];
            string sparse_output_275_pad_type_1 = const()[name = string("sparse_output_275_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_275_strides_1 = const()[name = string("sparse_output_275_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_275_pad_1 = const()[name = string("sparse_output_275_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_275_dilations_1 = const()[name = string("sparse_output_275_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_275_groups_1 = const()[name = string("sparse_output_275_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105282368))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105279680))))[name = string("layers_3_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_275_cast_fp16 = conv(dilations = sparse_output_275_dilations_1, groups = sparse_output_275_groups_1, pad = sparse_output_275_pad_1, pad_type = sparse_output_275_pad_type_1, strides = sparse_output_275_strides_1, weight = layers_3_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_275_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4437_cast_fp16 = add(x = dense_output_275_cast_fp16, y = sparse_output_275_cast_fp16)[name = string("op_4437_cast_fp16")];
            tensor<int32, [4]> var_4438 = const()[name = string("op_4438"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4440 = const()[name = string("op_4440"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4439_cast_fp16 = transpose(perm = var_4438, x = var_4437_cast_fp16)[name = string("transpose_848")];
            tensor<fp16, [1, 51, 128]> k_head_105_cast_fp16 = reshape(shape = var_4440, x = var_4439_cast_fp16)[name = string("k_head_105_cast_fp16")];
            string dense_output_277_pad_type_1 = const()[name = string("dense_output_277_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_277_strides_1 = const()[name = string("dense_output_277_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_277_pad_1 = const()[name = string("dense_output_277_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_277_dilations_1 = const()[name = string("dense_output_277_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_277_groups_1 = const()[name = string("dense_output_277_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105298816))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105429952))))[name = string("layers_3_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_277_cast_fp16 = conv(dilations = dense_output_277_dilations_1, groups = dense_output_277_groups_1, pad = dense_output_277_pad_1, pad_type = dense_output_277_pad_type_1, strides = dense_output_277_strides_1, weight = layers_3_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_277_cast_fp16")];
            string sparse_output_277_pad_type_1 = const()[name = string("sparse_output_277_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_277_strides_1 = const()[name = string("sparse_output_277_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_277_pad_1 = const()[name = string("sparse_output_277_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_277_dilations_1 = const()[name = string("sparse_output_277_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_277_groups_1 = const()[name = string("sparse_output_277_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105433216))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105430528))))[name = string("layers_3_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_277_cast_fp16 = conv(dilations = sparse_output_277_dilations_1, groups = sparse_output_277_groups_1, pad = sparse_output_277_pad_1, pad_type = sparse_output_277_pad_type_1, strides = sparse_output_277_strides_1, weight = layers_3_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_277_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4456_cast_fp16 = add(x = dense_output_277_cast_fp16, y = sparse_output_277_cast_fp16)[name = string("op_4456_cast_fp16")];
            tensor<int32, [4]> var_4457 = const()[name = string("op_4457"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4459 = const()[name = string("op_4459"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4458_cast_fp16 = transpose(perm = var_4457, x = var_4456_cast_fp16)[name = string("transpose_847")];
            tensor<fp16, [1, 51, 128]> v_head_105_cast_fp16 = reshape(shape = var_4459, x = var_4458_cast_fp16)[name = string("v_head_105_cast_fp16")];
            string dense_output_279_pad_type_1 = const()[name = string("dense_output_279_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_279_strides_1 = const()[name = string("dense_output_279_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_279_pad_1 = const()[name = string("dense_output_279_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_279_dilations_1 = const()[name = string("dense_output_279_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_279_groups_1 = const()[name = string("dense_output_279_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105449664))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105580800))))[name = string("layers_3_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_279_cast_fp16 = conv(dilations = dense_output_279_dilations_1, groups = dense_output_279_groups_1, pad = dense_output_279_pad_1, pad_type = dense_output_279_pad_type_1, strides = dense_output_279_strides_1, weight = layers_3_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_279_cast_fp16")];
            string sparse_output_279_pad_type_1 = const()[name = string("sparse_output_279_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_279_strides_1 = const()[name = string("sparse_output_279_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_279_pad_1 = const()[name = string("sparse_output_279_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_279_dilations_1 = const()[name = string("sparse_output_279_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_279_groups_1 = const()[name = string("sparse_output_279_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105584064))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105581376))))[name = string("layers_3_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_279_cast_fp16 = conv(dilations = sparse_output_279_dilations_1, groups = sparse_output_279_groups_1, pad = sparse_output_279_pad_1, pad_type = sparse_output_279_pad_type_1, strides = sparse_output_279_strides_1, weight = layers_3_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_279_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_4475_cast_fp16 = add(x = dense_output_279_cast_fp16, y = sparse_output_279_cast_fp16)[name = string("op_4475_cast_fp16")];
            tensor<int32, [4]> var_4476 = const()[name = string("op_4476"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4478 = const()[name = string("op_4478"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_4477_cast_fp16 = transpose(perm = var_4476, x = var_4475_cast_fp16)[name = string("transpose_846")];
            tensor<fp16, [1, 101, 128]> p_head_105_cast_fp16 = reshape(shape = var_4478, x = var_4477_cast_fp16)[name = string("p_head_105_cast_fp16")];
            tensor<fp16, [128]> var_4480_to_fp16 = const()[name = string("op_4480_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105600512)))];
            tensor<fp16, [1, 51, 128]> var_4481_cast_fp16 = add(x = q_head_53_cast_fp16, y = var_4480_to_fp16)[name = string("op_4481_cast_fp16")];
            tensor<int32, [1]> q_u_53_axes_1 = const()[name = string("q_u_53_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_53_cast_fp16 = expand_dims(axes = q_u_53_axes_1, x = var_4481_cast_fp16)[name = string("q_u_53_cast_fp16")];
            tensor<fp16, [128]> var_4483_to_fp16 = const()[name = string("op_4483_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105600832)))];
            tensor<fp16, [1, 51, 128]> var_4484_cast_fp16 = add(x = q_head_53_cast_fp16, y = var_4483_to_fp16)[name = string("op_4484_cast_fp16")];
            tensor<int32, [1]> q_v_53_axes_1 = const()[name = string("q_v_53_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_53_cast_fp16 = expand_dims(axes = q_v_53_axes_1, x = var_4484_cast_fp16)[name = string("q_v_53_cast_fp16")];
            tensor<int32, [1]> k_head_107_axes_1 = const()[name = string("k_head_107_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_107_cast_fp16 = expand_dims(axes = k_head_107_axes_1, x = k_head_105_cast_fp16)[name = string("k_head_107_cast_fp16")];
            tensor<int32, [1]> v_head_107_axes_1 = const()[name = string("v_head_107_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_107_cast_fp16 = expand_dims(axes = v_head_107_axes_1, x = v_head_105_cast_fp16)[name = string("v_head_107_cast_fp16")];
            tensor<int32, [1]> p_head_107_axes_1 = const()[name = string("p_head_107_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_107_cast_fp16 = expand_dims(axes = p_head_107_axes_1, x = p_head_105_cast_fp16)[name = string("p_head_107_cast_fp16")];
            bool var_4490_transpose_x_3 = const()[name = string("op_4490_transpose_x_3"), val = bool(false)];
            bool var_4490_transpose_y_3 = const()[name = string("op_4490_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_4490_cast_fp16 = matmul(transpose_x = var_4490_transpose_x_3, transpose_y = var_4490_transpose_y_3, x = q_u_53_cast_fp16, y = k_head_107_cast_fp16)[name = string("op_4490_cast_fp16")];
            fp16 var_4491_to_fp16 = const()[name = string("op_4491_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_53_cast_fp16 = mul(x = var_4490_cast_fp16, y = var_4491_to_fp16)[name = string("scores_content_53_cast_fp16")];
            bool x_289_transpose_x_3 = const()[name = string("x_289_transpose_x_3"), val = bool(false)];
            bool x_289_transpose_y_3 = const()[name = string("x_289_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_289_cast_fp16 = matmul(transpose_x = x_289_transpose_x_3, transpose_y = x_289_transpose_y_3, x = q_v_53_cast_fp16, y = p_head_107_cast_fp16)[name = string("x_289_cast_fp16")];
            tensor<int32, [8]> x_291_pad_1 = const()[name = string("x_291_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_291_mode_1 = const()[name = string("x_291_mode_1"), val = string("constant")];
            fp16 const_1483_to_fp16 = const()[name = string("const_1483_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_291_cast_fp16 = pad(constant_val = const_1483_to_fp16, mode = x_291_mode_1, pad = x_291_pad_1, x = x_289_cast_fp16)[name = string("x_291_cast_fp16")];
            tensor<int32, [4]> var_4505 = const()[name = string("op_4505"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_293_cast_fp16 = reshape(shape = var_4505, x = x_291_cast_fp16)[name = string("x_293_cast_fp16")];
            tensor<int32, [4]> var_4509_begin_1 = const()[name = string("op_4509_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4509_end_1 = const()[name = string("op_4509_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_4509_end_mask_1 = const()[name = string("op_4509_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_4509_cast_fp16 = slice_by_index(begin = var_4509_begin_1, end = var_4509_end_1, end_mask = var_4509_end_mask_1, x = x_293_cast_fp16)[name = string("op_4509_cast_fp16")];
            tensor<int32, [4]> var_4511 = const()[name = string("op_4511"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_4512_cast_fp16 = reshape(shape = var_4511, x = var_4509_cast_fp16)[name = string("op_4512_cast_fp16")];
            tensor<int32, [4]> var_4517_begin_1 = const()[name = string("op_4517_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4517_end_1 = const()[name = string("op_4517_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_4517_end_mask_1 = const()[name = string("op_4517_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_4517_cast_fp16 = slice_by_index(begin = var_4517_begin_1, end = var_4517_end_1, end_mask = var_4517_end_mask_1, x = var_4512_cast_fp16)[name = string("op_4517_cast_fp16")];
            fp16 var_4518_to_fp16 = const()[name = string("op_4518_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_53_cast_fp16 = mul(x = var_4517_cast_fp16, y = var_4518_to_fp16)[name = string("scores_pos_53_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_53_cast_fp16 = add(x = scores_content_53_cast_fp16, y = scores_pos_53_cast_fp16)[name = string("logits_53_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_4521_cast_fp16 = softmax(axis = var_4004, x = logits_53_cast_fp16)[name = string("op_4521_cast_fp16")];
            bool var_4523_transpose_x_1 = const()[name = string("op_4523_transpose_x_1"), val = bool(false)];
            bool var_4523_transpose_y_1 = const()[name = string("op_4523_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_4523_cast_fp16 = matmul(transpose_x = var_4523_transpose_x_1, transpose_y = var_4523_transpose_y_1, x = var_4521_cast_fp16, y = v_head_107_cast_fp16)[name = string("op_4523_cast_fp16")];
            tensor<int32, [1]> var_4524_axes_1 = const()[name = string("op_4524_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_4524_cast_fp16 = squeeze(axes = var_4524_axes_1, x = var_4523_cast_fp16)[name = string("op_4524_cast_fp16")];
            string dense_output_281_pad_type_1 = const()[name = string("dense_output_281_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_281_strides_1 = const()[name = string("dense_output_281_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_281_pad_1 = const()[name = string("dense_output_281_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_281_dilations_1 = const()[name = string("dense_output_281_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_281_groups_1 = const()[name = string("dense_output_281_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105601152))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105732288))))[name = string("layers_3_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_281_cast_fp16 = conv(dilations = dense_output_281_dilations_1, groups = dense_output_281_groups_1, pad = dense_output_281_pad_1, pad_type = dense_output_281_pad_type_1, strides = dense_output_281_strides_1, weight = layers_3_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_281_cast_fp16")];
            string sparse_output_281_pad_type_1 = const()[name = string("sparse_output_281_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_281_strides_1 = const()[name = string("sparse_output_281_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_281_pad_1 = const()[name = string("sparse_output_281_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_281_dilations_1 = const()[name = string("sparse_output_281_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_281_groups_1 = const()[name = string("sparse_output_281_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105735552))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105732864))))[name = string("layers_3_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_281_cast_fp16 = conv(dilations = sparse_output_281_dilations_1, groups = sparse_output_281_groups_1, pad = sparse_output_281_pad_1, pad_type = sparse_output_281_pad_type_1, strides = sparse_output_281_strides_1, weight = layers_3_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_281_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4539_cast_fp16 = add(x = dense_output_281_cast_fp16, y = sparse_output_281_cast_fp16)[name = string("op_4539_cast_fp16")];
            tensor<int32, [4]> var_4540 = const()[name = string("op_4540"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4542 = const()[name = string("op_4542"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4541_cast_fp16 = transpose(perm = var_4540, x = var_4539_cast_fp16)[name = string("transpose_845")];
            tensor<fp16, [1, 51, 128]> q_head_55_cast_fp16 = reshape(shape = var_4542, x = var_4541_cast_fp16)[name = string("q_head_55_cast_fp16")];
            string dense_output_283_pad_type_1 = const()[name = string("dense_output_283_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_283_strides_1 = const()[name = string("dense_output_283_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_283_pad_1 = const()[name = string("dense_output_283_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_283_dilations_1 = const()[name = string("dense_output_283_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_283_groups_1 = const()[name = string("dense_output_283_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105752000))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105883136))))[name = string("layers_3_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_283_cast_fp16 = conv(dilations = dense_output_283_dilations_1, groups = dense_output_283_groups_1, pad = dense_output_283_pad_1, pad_type = dense_output_283_pad_type_1, strides = dense_output_283_strides_1, weight = layers_3_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_283_cast_fp16")];
            string sparse_output_283_pad_type_1 = const()[name = string("sparse_output_283_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_283_strides_1 = const()[name = string("sparse_output_283_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_283_pad_1 = const()[name = string("sparse_output_283_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_283_dilations_1 = const()[name = string("sparse_output_283_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_283_groups_1 = const()[name = string("sparse_output_283_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105886400))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105883712))))[name = string("layers_3_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_283_cast_fp16 = conv(dilations = sparse_output_283_dilations_1, groups = sparse_output_283_groups_1, pad = sparse_output_283_pad_1, pad_type = sparse_output_283_pad_type_1, strides = sparse_output_283_strides_1, weight = layers_3_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_283_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4558_cast_fp16 = add(x = dense_output_283_cast_fp16, y = sparse_output_283_cast_fp16)[name = string("op_4558_cast_fp16")];
            tensor<int32, [4]> var_4559 = const()[name = string("op_4559"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4561 = const()[name = string("op_4561"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4560_cast_fp16 = transpose(perm = var_4559, x = var_4558_cast_fp16)[name = string("transpose_844")];
            tensor<fp16, [1, 51, 128]> k_head_109_cast_fp16 = reshape(shape = var_4561, x = var_4560_cast_fp16)[name = string("k_head_109_cast_fp16")];
            string dense_output_285_pad_type_1 = const()[name = string("dense_output_285_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_285_strides_1 = const()[name = string("dense_output_285_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_285_pad_1 = const()[name = string("dense_output_285_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_285_dilations_1 = const()[name = string("dense_output_285_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_285_groups_1 = const()[name = string("dense_output_285_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105902848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106033984))))[name = string("layers_3_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_285_cast_fp16 = conv(dilations = dense_output_285_dilations_1, groups = dense_output_285_groups_1, pad = dense_output_285_pad_1, pad_type = dense_output_285_pad_type_1, strides = dense_output_285_strides_1, weight = layers_3_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_285_cast_fp16")];
            string sparse_output_285_pad_type_1 = const()[name = string("sparse_output_285_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_285_strides_1 = const()[name = string("sparse_output_285_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_285_pad_1 = const()[name = string("sparse_output_285_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_285_dilations_1 = const()[name = string("sparse_output_285_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_285_groups_1 = const()[name = string("sparse_output_285_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106037248))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106034560))))[name = string("layers_3_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_285_cast_fp16 = conv(dilations = sparse_output_285_dilations_1, groups = sparse_output_285_groups_1, pad = sparse_output_285_pad_1, pad_type = sparse_output_285_pad_type_1, strides = sparse_output_285_strides_1, weight = layers_3_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_285_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4577_cast_fp16 = add(x = dense_output_285_cast_fp16, y = sparse_output_285_cast_fp16)[name = string("op_4577_cast_fp16")];
            tensor<int32, [4]> var_4578 = const()[name = string("op_4578"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4580 = const()[name = string("op_4580"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4579_cast_fp16 = transpose(perm = var_4578, x = var_4577_cast_fp16)[name = string("transpose_843")];
            tensor<fp16, [1, 51, 128]> v_head_109_cast_fp16 = reshape(shape = var_4580, x = var_4579_cast_fp16)[name = string("v_head_109_cast_fp16")];
            string dense_output_287_pad_type_1 = const()[name = string("dense_output_287_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_287_strides_1 = const()[name = string("dense_output_287_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_287_pad_1 = const()[name = string("dense_output_287_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_287_dilations_1 = const()[name = string("dense_output_287_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_287_groups_1 = const()[name = string("dense_output_287_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106053696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106184832))))[name = string("layers_3_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_287_cast_fp16 = conv(dilations = dense_output_287_dilations_1, groups = dense_output_287_groups_1, pad = dense_output_287_pad_1, pad_type = dense_output_287_pad_type_1, strides = dense_output_287_strides_1, weight = layers_3_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_287_cast_fp16")];
            string sparse_output_287_pad_type_1 = const()[name = string("sparse_output_287_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_287_strides_1 = const()[name = string("sparse_output_287_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_287_pad_1 = const()[name = string("sparse_output_287_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_287_dilations_1 = const()[name = string("sparse_output_287_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_287_groups_1 = const()[name = string("sparse_output_287_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106188096))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106185408))))[name = string("layers_3_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_287_cast_fp16 = conv(dilations = sparse_output_287_dilations_1, groups = sparse_output_287_groups_1, pad = sparse_output_287_pad_1, pad_type = sparse_output_287_pad_type_1, strides = sparse_output_287_strides_1, weight = layers_3_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_287_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_4596_cast_fp16 = add(x = dense_output_287_cast_fp16, y = sparse_output_287_cast_fp16)[name = string("op_4596_cast_fp16")];
            tensor<int32, [4]> var_4597 = const()[name = string("op_4597"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4599 = const()[name = string("op_4599"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_4598_cast_fp16 = transpose(perm = var_4597, x = var_4596_cast_fp16)[name = string("transpose_842")];
            tensor<fp16, [1, 101, 128]> p_head_109_cast_fp16 = reshape(shape = var_4599, x = var_4598_cast_fp16)[name = string("p_head_109_cast_fp16")];
            tensor<fp16, [128]> var_4601_to_fp16 = const()[name = string("op_4601_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106204544)))];
            tensor<fp16, [1, 51, 128]> var_4602_cast_fp16 = add(x = q_head_55_cast_fp16, y = var_4601_to_fp16)[name = string("op_4602_cast_fp16")];
            tensor<int32, [1]> q_u_55_axes_1 = const()[name = string("q_u_55_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_55_cast_fp16 = expand_dims(axes = q_u_55_axes_1, x = var_4602_cast_fp16)[name = string("q_u_55_cast_fp16")];
            tensor<fp16, [128]> var_4604_to_fp16 = const()[name = string("op_4604_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106204864)))];
            tensor<fp16, [1, 51, 128]> var_4605_cast_fp16 = add(x = q_head_55_cast_fp16, y = var_4604_to_fp16)[name = string("op_4605_cast_fp16")];
            tensor<int32, [1]> q_v_55_axes_1 = const()[name = string("q_v_55_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_55_cast_fp16 = expand_dims(axes = q_v_55_axes_1, x = var_4605_cast_fp16)[name = string("q_v_55_cast_fp16")];
            tensor<int32, [1]> k_head_111_axes_1 = const()[name = string("k_head_111_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_111_cast_fp16 = expand_dims(axes = k_head_111_axes_1, x = k_head_109_cast_fp16)[name = string("k_head_111_cast_fp16")];
            tensor<int32, [1]> v_head_111_axes_1 = const()[name = string("v_head_111_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_111_cast_fp16 = expand_dims(axes = v_head_111_axes_1, x = v_head_109_cast_fp16)[name = string("v_head_111_cast_fp16")];
            tensor<int32, [1]> p_head_111_axes_1 = const()[name = string("p_head_111_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_111_cast_fp16 = expand_dims(axes = p_head_111_axes_1, x = p_head_109_cast_fp16)[name = string("p_head_111_cast_fp16")];
            bool var_4611_transpose_x_3 = const()[name = string("op_4611_transpose_x_3"), val = bool(false)];
            bool var_4611_transpose_y_3 = const()[name = string("op_4611_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_4611_cast_fp16 = matmul(transpose_x = var_4611_transpose_x_3, transpose_y = var_4611_transpose_y_3, x = q_u_55_cast_fp16, y = k_head_111_cast_fp16)[name = string("op_4611_cast_fp16")];
            fp16 var_4612_to_fp16 = const()[name = string("op_4612_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_55_cast_fp16 = mul(x = var_4611_cast_fp16, y = var_4612_to_fp16)[name = string("scores_content_55_cast_fp16")];
            bool x_297_transpose_x_3 = const()[name = string("x_297_transpose_x_3"), val = bool(false)];
            bool x_297_transpose_y_3 = const()[name = string("x_297_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_297_cast_fp16 = matmul(transpose_x = x_297_transpose_x_3, transpose_y = x_297_transpose_y_3, x = q_v_55_cast_fp16, y = p_head_111_cast_fp16)[name = string("x_297_cast_fp16")];
            tensor<int32, [8]> x_299_pad_1 = const()[name = string("x_299_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_299_mode_1 = const()[name = string("x_299_mode_1"), val = string("constant")];
            fp16 const_1489_to_fp16 = const()[name = string("const_1489_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_299_cast_fp16 = pad(constant_val = const_1489_to_fp16, mode = x_299_mode_1, pad = x_299_pad_1, x = x_297_cast_fp16)[name = string("x_299_cast_fp16")];
            tensor<int32, [4]> var_4626 = const()[name = string("op_4626"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_301_cast_fp16 = reshape(shape = var_4626, x = x_299_cast_fp16)[name = string("x_301_cast_fp16")];
            tensor<int32, [4]> var_4630_begin_1 = const()[name = string("op_4630_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4630_end_1 = const()[name = string("op_4630_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_4630_end_mask_1 = const()[name = string("op_4630_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_4630_cast_fp16 = slice_by_index(begin = var_4630_begin_1, end = var_4630_end_1, end_mask = var_4630_end_mask_1, x = x_301_cast_fp16)[name = string("op_4630_cast_fp16")];
            tensor<int32, [4]> var_4632 = const()[name = string("op_4632"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_4633_cast_fp16 = reshape(shape = var_4632, x = var_4630_cast_fp16)[name = string("op_4633_cast_fp16")];
            tensor<int32, [4]> var_4638_begin_1 = const()[name = string("op_4638_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4638_end_1 = const()[name = string("op_4638_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_4638_end_mask_1 = const()[name = string("op_4638_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_4638_cast_fp16 = slice_by_index(begin = var_4638_begin_1, end = var_4638_end_1, end_mask = var_4638_end_mask_1, x = var_4633_cast_fp16)[name = string("op_4638_cast_fp16")];
            fp16 var_4639_to_fp16 = const()[name = string("op_4639_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_55_cast_fp16 = mul(x = var_4638_cast_fp16, y = var_4639_to_fp16)[name = string("scores_pos_55_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_55_cast_fp16 = add(x = scores_content_55_cast_fp16, y = scores_pos_55_cast_fp16)[name = string("logits_55_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_4642_cast_fp16 = softmax(axis = var_4004, x = logits_55_cast_fp16)[name = string("op_4642_cast_fp16")];
            bool var_4644_transpose_x_1 = const()[name = string("op_4644_transpose_x_1"), val = bool(false)];
            bool var_4644_transpose_y_1 = const()[name = string("op_4644_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_4644_cast_fp16 = matmul(transpose_x = var_4644_transpose_x_1, transpose_y = var_4644_transpose_y_1, x = var_4642_cast_fp16, y = v_head_111_cast_fp16)[name = string("op_4644_cast_fp16")];
            tensor<int32, [1]> var_4645_axes_1 = const()[name = string("op_4645_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_4645_cast_fp16 = squeeze(axes = var_4645_axes_1, x = var_4644_cast_fp16)[name = string("op_4645_cast_fp16")];
            string dense_output_289_pad_type_1 = const()[name = string("dense_output_289_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_289_strides_1 = const()[name = string("dense_output_289_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_289_pad_1 = const()[name = string("dense_output_289_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_289_dilations_1 = const()[name = string("dense_output_289_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_289_groups_1 = const()[name = string("dense_output_289_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106205184))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106336320))))[name = string("layers_3_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_289_cast_fp16 = conv(dilations = dense_output_289_dilations_1, groups = dense_output_289_groups_1, pad = dense_output_289_pad_1, pad_type = dense_output_289_pad_type_1, strides = dense_output_289_strides_1, weight = layers_3_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_289_cast_fp16")];
            string sparse_output_289_pad_type_1 = const()[name = string("sparse_output_289_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_289_strides_1 = const()[name = string("sparse_output_289_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_289_pad_1 = const()[name = string("sparse_output_289_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_289_dilations_1 = const()[name = string("sparse_output_289_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_289_groups_1 = const()[name = string("sparse_output_289_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106339584))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106336896))))[name = string("layers_3_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_289_cast_fp16 = conv(dilations = sparse_output_289_dilations_1, groups = sparse_output_289_groups_1, pad = sparse_output_289_pad_1, pad_type = sparse_output_289_pad_type_1, strides = sparse_output_289_strides_1, weight = layers_3_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_289_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4660_cast_fp16 = add(x = dense_output_289_cast_fp16, y = sparse_output_289_cast_fp16)[name = string("op_4660_cast_fp16")];
            tensor<int32, [4]> var_4661 = const()[name = string("op_4661"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4663 = const()[name = string("op_4663"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4662_cast_fp16 = transpose(perm = var_4661, x = var_4660_cast_fp16)[name = string("transpose_841")];
            tensor<fp16, [1, 51, 128]> q_head_57_cast_fp16 = reshape(shape = var_4663, x = var_4662_cast_fp16)[name = string("q_head_57_cast_fp16")];
            string dense_output_291_pad_type_1 = const()[name = string("dense_output_291_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_291_strides_1 = const()[name = string("dense_output_291_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_291_pad_1 = const()[name = string("dense_output_291_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_291_dilations_1 = const()[name = string("dense_output_291_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_291_groups_1 = const()[name = string("dense_output_291_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106356032))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106487168))))[name = string("layers_3_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_291_cast_fp16 = conv(dilations = dense_output_291_dilations_1, groups = dense_output_291_groups_1, pad = dense_output_291_pad_1, pad_type = dense_output_291_pad_type_1, strides = dense_output_291_strides_1, weight = layers_3_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_291_cast_fp16")];
            string sparse_output_291_pad_type_1 = const()[name = string("sparse_output_291_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_291_strides_1 = const()[name = string("sparse_output_291_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_291_pad_1 = const()[name = string("sparse_output_291_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_291_dilations_1 = const()[name = string("sparse_output_291_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_291_groups_1 = const()[name = string("sparse_output_291_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106490432))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106487744))))[name = string("layers_3_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_291_cast_fp16 = conv(dilations = sparse_output_291_dilations_1, groups = sparse_output_291_groups_1, pad = sparse_output_291_pad_1, pad_type = sparse_output_291_pad_type_1, strides = sparse_output_291_strides_1, weight = layers_3_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_291_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4679_cast_fp16 = add(x = dense_output_291_cast_fp16, y = sparse_output_291_cast_fp16)[name = string("op_4679_cast_fp16")];
            tensor<int32, [4]> var_4680 = const()[name = string("op_4680"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4682 = const()[name = string("op_4682"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4681_cast_fp16 = transpose(perm = var_4680, x = var_4679_cast_fp16)[name = string("transpose_840")];
            tensor<fp16, [1, 51, 128]> k_head_113_cast_fp16 = reshape(shape = var_4682, x = var_4681_cast_fp16)[name = string("k_head_113_cast_fp16")];
            string dense_output_293_pad_type_1 = const()[name = string("dense_output_293_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_293_strides_1 = const()[name = string("dense_output_293_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_293_pad_1 = const()[name = string("dense_output_293_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_293_dilations_1 = const()[name = string("dense_output_293_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_293_groups_1 = const()[name = string("dense_output_293_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106506880))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106638016))))[name = string("layers_3_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_293_cast_fp16 = conv(dilations = dense_output_293_dilations_1, groups = dense_output_293_groups_1, pad = dense_output_293_pad_1, pad_type = dense_output_293_pad_type_1, strides = dense_output_293_strides_1, weight = layers_3_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_293_cast_fp16")];
            string sparse_output_293_pad_type_1 = const()[name = string("sparse_output_293_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_293_strides_1 = const()[name = string("sparse_output_293_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_293_pad_1 = const()[name = string("sparse_output_293_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_293_dilations_1 = const()[name = string("sparse_output_293_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_293_groups_1 = const()[name = string("sparse_output_293_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106641280))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106638592))))[name = string("layers_3_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_293_cast_fp16 = conv(dilations = sparse_output_293_dilations_1, groups = sparse_output_293_groups_1, pad = sparse_output_293_pad_1, pad_type = sparse_output_293_pad_type_1, strides = sparse_output_293_strides_1, weight = layers_3_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_293_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4698_cast_fp16 = add(x = dense_output_293_cast_fp16, y = sparse_output_293_cast_fp16)[name = string("op_4698_cast_fp16")];
            tensor<int32, [4]> var_4699 = const()[name = string("op_4699"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4701 = const()[name = string("op_4701"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4700_cast_fp16 = transpose(perm = var_4699, x = var_4698_cast_fp16)[name = string("transpose_839")];
            tensor<fp16, [1, 51, 128]> v_head_113_cast_fp16 = reshape(shape = var_4701, x = var_4700_cast_fp16)[name = string("v_head_113_cast_fp16")];
            string dense_output_295_pad_type_1 = const()[name = string("dense_output_295_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_295_strides_1 = const()[name = string("dense_output_295_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_295_pad_1 = const()[name = string("dense_output_295_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_295_dilations_1 = const()[name = string("dense_output_295_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_295_groups_1 = const()[name = string("dense_output_295_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106657728))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106788864))))[name = string("layers_3_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_295_cast_fp16 = conv(dilations = dense_output_295_dilations_1, groups = dense_output_295_groups_1, pad = dense_output_295_pad_1, pad_type = dense_output_295_pad_type_1, strides = dense_output_295_strides_1, weight = layers_3_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_295_cast_fp16")];
            string sparse_output_295_pad_type_1 = const()[name = string("sparse_output_295_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_295_strides_1 = const()[name = string("sparse_output_295_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_295_pad_1 = const()[name = string("sparse_output_295_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_295_dilations_1 = const()[name = string("sparse_output_295_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_295_groups_1 = const()[name = string("sparse_output_295_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106792128))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106789440))))[name = string("layers_3_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_295_cast_fp16 = conv(dilations = sparse_output_295_dilations_1, groups = sparse_output_295_groups_1, pad = sparse_output_295_pad_1, pad_type = sparse_output_295_pad_type_1, strides = sparse_output_295_strides_1, weight = layers_3_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_295_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_4717_cast_fp16 = add(x = dense_output_295_cast_fp16, y = sparse_output_295_cast_fp16)[name = string("op_4717_cast_fp16")];
            tensor<int32, [4]> var_4718 = const()[name = string("op_4718"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4720 = const()[name = string("op_4720"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_4719_cast_fp16 = transpose(perm = var_4718, x = var_4717_cast_fp16)[name = string("transpose_838")];
            tensor<fp16, [1, 101, 128]> p_head_113_cast_fp16 = reshape(shape = var_4720, x = var_4719_cast_fp16)[name = string("p_head_113_cast_fp16")];
            tensor<fp16, [128]> var_4722_to_fp16 = const()[name = string("op_4722_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106808576)))];
            tensor<fp16, [1, 51, 128]> var_4723_cast_fp16 = add(x = q_head_57_cast_fp16, y = var_4722_to_fp16)[name = string("op_4723_cast_fp16")];
            tensor<int32, [1]> q_u_57_axes_1 = const()[name = string("q_u_57_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_57_cast_fp16 = expand_dims(axes = q_u_57_axes_1, x = var_4723_cast_fp16)[name = string("q_u_57_cast_fp16")];
            tensor<fp16, [128]> var_4725_to_fp16 = const()[name = string("op_4725_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106808896)))];
            tensor<fp16, [1, 51, 128]> var_4726_cast_fp16 = add(x = q_head_57_cast_fp16, y = var_4725_to_fp16)[name = string("op_4726_cast_fp16")];
            tensor<int32, [1]> q_v_57_axes_1 = const()[name = string("q_v_57_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_57_cast_fp16 = expand_dims(axes = q_v_57_axes_1, x = var_4726_cast_fp16)[name = string("q_v_57_cast_fp16")];
            tensor<int32, [1]> k_head_115_axes_1 = const()[name = string("k_head_115_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_115_cast_fp16 = expand_dims(axes = k_head_115_axes_1, x = k_head_113_cast_fp16)[name = string("k_head_115_cast_fp16")];
            tensor<int32, [1]> v_head_115_axes_1 = const()[name = string("v_head_115_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_115_cast_fp16 = expand_dims(axes = v_head_115_axes_1, x = v_head_113_cast_fp16)[name = string("v_head_115_cast_fp16")];
            tensor<int32, [1]> p_head_115_axes_1 = const()[name = string("p_head_115_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_115_cast_fp16 = expand_dims(axes = p_head_115_axes_1, x = p_head_113_cast_fp16)[name = string("p_head_115_cast_fp16")];
            bool var_4732_transpose_x_3 = const()[name = string("op_4732_transpose_x_3"), val = bool(false)];
            bool var_4732_transpose_y_3 = const()[name = string("op_4732_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_4732_cast_fp16 = matmul(transpose_x = var_4732_transpose_x_3, transpose_y = var_4732_transpose_y_3, x = q_u_57_cast_fp16, y = k_head_115_cast_fp16)[name = string("op_4732_cast_fp16")];
            fp16 var_4733_to_fp16 = const()[name = string("op_4733_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_57_cast_fp16 = mul(x = var_4732_cast_fp16, y = var_4733_to_fp16)[name = string("scores_content_57_cast_fp16")];
            bool x_305_transpose_x_3 = const()[name = string("x_305_transpose_x_3"), val = bool(false)];
            bool x_305_transpose_y_3 = const()[name = string("x_305_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_305_cast_fp16 = matmul(transpose_x = x_305_transpose_x_3, transpose_y = x_305_transpose_y_3, x = q_v_57_cast_fp16, y = p_head_115_cast_fp16)[name = string("x_305_cast_fp16")];
            tensor<int32, [8]> x_307_pad_1 = const()[name = string("x_307_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_307_mode_1 = const()[name = string("x_307_mode_1"), val = string("constant")];
            fp16 const_1495_to_fp16 = const()[name = string("const_1495_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_307_cast_fp16 = pad(constant_val = const_1495_to_fp16, mode = x_307_mode_1, pad = x_307_pad_1, x = x_305_cast_fp16)[name = string("x_307_cast_fp16")];
            tensor<int32, [4]> var_4747 = const()[name = string("op_4747"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_309_cast_fp16 = reshape(shape = var_4747, x = x_307_cast_fp16)[name = string("x_309_cast_fp16")];
            tensor<int32, [4]> var_4751_begin_1 = const()[name = string("op_4751_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4751_end_1 = const()[name = string("op_4751_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_4751_end_mask_1 = const()[name = string("op_4751_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_4751_cast_fp16 = slice_by_index(begin = var_4751_begin_1, end = var_4751_end_1, end_mask = var_4751_end_mask_1, x = x_309_cast_fp16)[name = string("op_4751_cast_fp16")];
            tensor<int32, [4]> var_4753 = const()[name = string("op_4753"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_4754_cast_fp16 = reshape(shape = var_4753, x = var_4751_cast_fp16)[name = string("op_4754_cast_fp16")];
            tensor<int32, [4]> var_4759_begin_1 = const()[name = string("op_4759_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4759_end_1 = const()[name = string("op_4759_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_4759_end_mask_1 = const()[name = string("op_4759_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_4759_cast_fp16 = slice_by_index(begin = var_4759_begin_1, end = var_4759_end_1, end_mask = var_4759_end_mask_1, x = var_4754_cast_fp16)[name = string("op_4759_cast_fp16")];
            fp16 var_4760_to_fp16 = const()[name = string("op_4760_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_57_cast_fp16 = mul(x = var_4759_cast_fp16, y = var_4760_to_fp16)[name = string("scores_pos_57_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_57_cast_fp16 = add(x = scores_content_57_cast_fp16, y = scores_pos_57_cast_fp16)[name = string("logits_57_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_4763_cast_fp16 = softmax(axis = var_4004, x = logits_57_cast_fp16)[name = string("op_4763_cast_fp16")];
            bool var_4765_transpose_x_1 = const()[name = string("op_4765_transpose_x_1"), val = bool(false)];
            bool var_4765_transpose_y_1 = const()[name = string("op_4765_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_4765_cast_fp16 = matmul(transpose_x = var_4765_transpose_x_1, transpose_y = var_4765_transpose_y_1, x = var_4763_cast_fp16, y = v_head_115_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<int32, [1]> var_4766_axes_1 = const()[name = string("op_4766_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_4766_cast_fp16 = squeeze(axes = var_4766_axes_1, x = var_4765_cast_fp16)[name = string("op_4766_cast_fp16")];
            string dense_output_297_pad_type_1 = const()[name = string("dense_output_297_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_297_strides_1 = const()[name = string("dense_output_297_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_297_pad_1 = const()[name = string("dense_output_297_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_297_dilations_1 = const()[name = string("dense_output_297_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_297_groups_1 = const()[name = string("dense_output_297_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106809216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106940352))))[name = string("layers_3_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_297_cast_fp16 = conv(dilations = dense_output_297_dilations_1, groups = dense_output_297_groups_1, pad = dense_output_297_pad_1, pad_type = dense_output_297_pad_type_1, strides = dense_output_297_strides_1, weight = layers_3_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_297_cast_fp16")];
            string sparse_output_297_pad_type_1 = const()[name = string("sparse_output_297_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_297_strides_1 = const()[name = string("sparse_output_297_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_297_pad_1 = const()[name = string("sparse_output_297_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_297_dilations_1 = const()[name = string("sparse_output_297_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_297_groups_1 = const()[name = string("sparse_output_297_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106943616))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106940928))))[name = string("layers_3_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_297_cast_fp16 = conv(dilations = sparse_output_297_dilations_1, groups = sparse_output_297_groups_1, pad = sparse_output_297_pad_1, pad_type = sparse_output_297_pad_type_1, strides = sparse_output_297_strides_1, weight = layers_3_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_297_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4781_cast_fp16 = add(x = dense_output_297_cast_fp16, y = sparse_output_297_cast_fp16)[name = string("op_4781_cast_fp16")];
            tensor<int32, [4]> var_4782 = const()[name = string("op_4782"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4784 = const()[name = string("op_4784"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4783_cast_fp16 = transpose(perm = var_4782, x = var_4781_cast_fp16)[name = string("transpose_837")];
            tensor<fp16, [1, 51, 128]> q_head_59_cast_fp16 = reshape(shape = var_4784, x = var_4783_cast_fp16)[name = string("q_head_59_cast_fp16")];
            string dense_output_299_pad_type_1 = const()[name = string("dense_output_299_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_299_strides_1 = const()[name = string("dense_output_299_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_299_pad_1 = const()[name = string("dense_output_299_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_299_dilations_1 = const()[name = string("dense_output_299_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_299_groups_1 = const()[name = string("dense_output_299_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106960064))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107091200))))[name = string("layers_3_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_299_cast_fp16 = conv(dilations = dense_output_299_dilations_1, groups = dense_output_299_groups_1, pad = dense_output_299_pad_1, pad_type = dense_output_299_pad_type_1, strides = dense_output_299_strides_1, weight = layers_3_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_299_cast_fp16")];
            string sparse_output_299_pad_type_1 = const()[name = string("sparse_output_299_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_299_strides_1 = const()[name = string("sparse_output_299_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_299_pad_1 = const()[name = string("sparse_output_299_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_299_dilations_1 = const()[name = string("sparse_output_299_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_299_groups_1 = const()[name = string("sparse_output_299_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107094464))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107091776))))[name = string("layers_3_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_299_cast_fp16 = conv(dilations = sparse_output_299_dilations_1, groups = sparse_output_299_groups_1, pad = sparse_output_299_pad_1, pad_type = sparse_output_299_pad_type_1, strides = sparse_output_299_strides_1, weight = layers_3_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_299_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4800_cast_fp16 = add(x = dense_output_299_cast_fp16, y = sparse_output_299_cast_fp16)[name = string("op_4800_cast_fp16")];
            tensor<int32, [4]> var_4801 = const()[name = string("op_4801"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4803 = const()[name = string("op_4803"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4802_cast_fp16 = transpose(perm = var_4801, x = var_4800_cast_fp16)[name = string("transpose_836")];
            tensor<fp16, [1, 51, 128]> k_head_117_cast_fp16 = reshape(shape = var_4803, x = var_4802_cast_fp16)[name = string("k_head_117_cast_fp16")];
            string dense_output_301_pad_type_1 = const()[name = string("dense_output_301_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_301_strides_1 = const()[name = string("dense_output_301_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_301_pad_1 = const()[name = string("dense_output_301_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_301_dilations_1 = const()[name = string("dense_output_301_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_301_groups_1 = const()[name = string("dense_output_301_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107110912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107242048))))[name = string("layers_3_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_301_cast_fp16 = conv(dilations = dense_output_301_dilations_1, groups = dense_output_301_groups_1, pad = dense_output_301_pad_1, pad_type = dense_output_301_pad_type_1, strides = dense_output_301_strides_1, weight = layers_3_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_301_cast_fp16")];
            string sparse_output_301_pad_type_1 = const()[name = string("sparse_output_301_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_301_strides_1 = const()[name = string("sparse_output_301_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_301_pad_1 = const()[name = string("sparse_output_301_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_301_dilations_1 = const()[name = string("sparse_output_301_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_301_groups_1 = const()[name = string("sparse_output_301_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107245312))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107242624))))[name = string("layers_3_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_301_cast_fp16 = conv(dilations = sparse_output_301_dilations_1, groups = sparse_output_301_groups_1, pad = sparse_output_301_pad_1, pad_type = sparse_output_301_pad_type_1, strides = sparse_output_301_strides_1, weight = layers_3_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_301_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4819_cast_fp16 = add(x = dense_output_301_cast_fp16, y = sparse_output_301_cast_fp16)[name = string("op_4819_cast_fp16")];
            tensor<int32, [4]> var_4820 = const()[name = string("op_4820"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4822 = const()[name = string("op_4822"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4821_cast_fp16 = transpose(perm = var_4820, x = var_4819_cast_fp16)[name = string("transpose_835")];
            tensor<fp16, [1, 51, 128]> v_head_117_cast_fp16 = reshape(shape = var_4822, x = var_4821_cast_fp16)[name = string("v_head_117_cast_fp16")];
            string dense_output_303_pad_type_1 = const()[name = string("dense_output_303_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_303_strides_1 = const()[name = string("dense_output_303_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_303_pad_1 = const()[name = string("dense_output_303_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_303_dilations_1 = const()[name = string("dense_output_303_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_303_groups_1 = const()[name = string("dense_output_303_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107261760))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107392896))))[name = string("layers_3_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_303_cast_fp16 = conv(dilations = dense_output_303_dilations_1, groups = dense_output_303_groups_1, pad = dense_output_303_pad_1, pad_type = dense_output_303_pad_type_1, strides = dense_output_303_strides_1, weight = layers_3_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_303_cast_fp16")];
            string sparse_output_303_pad_type_1 = const()[name = string("sparse_output_303_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_303_strides_1 = const()[name = string("sparse_output_303_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_303_pad_1 = const()[name = string("sparse_output_303_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_303_dilations_1 = const()[name = string("sparse_output_303_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_303_groups_1 = const()[name = string("sparse_output_303_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107396160))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107393472))))[name = string("layers_3_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_303_cast_fp16 = conv(dilations = sparse_output_303_dilations_1, groups = sparse_output_303_groups_1, pad = sparse_output_303_pad_1, pad_type = sparse_output_303_pad_type_1, strides = sparse_output_303_strides_1, weight = layers_3_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_303_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_4838_cast_fp16 = add(x = dense_output_303_cast_fp16, y = sparse_output_303_cast_fp16)[name = string("op_4838_cast_fp16")];
            tensor<int32, [4]> var_4839 = const()[name = string("op_4839"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4841 = const()[name = string("op_4841"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_4840_cast_fp16 = transpose(perm = var_4839, x = var_4838_cast_fp16)[name = string("transpose_834")];
            tensor<fp16, [1, 101, 128]> p_head_117_cast_fp16 = reshape(shape = var_4841, x = var_4840_cast_fp16)[name = string("p_head_117_cast_fp16")];
            tensor<fp16, [128]> var_4843_to_fp16 = const()[name = string("op_4843_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107412608)))];
            tensor<fp16, [1, 51, 128]> var_4844_cast_fp16 = add(x = q_head_59_cast_fp16, y = var_4843_to_fp16)[name = string("op_4844_cast_fp16")];
            tensor<int32, [1]> q_u_59_axes_1 = const()[name = string("q_u_59_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_59_cast_fp16 = expand_dims(axes = q_u_59_axes_1, x = var_4844_cast_fp16)[name = string("q_u_59_cast_fp16")];
            tensor<fp16, [128]> var_4846_to_fp16 = const()[name = string("op_4846_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107412928)))];
            tensor<fp16, [1, 51, 128]> var_4847_cast_fp16 = add(x = q_head_59_cast_fp16, y = var_4846_to_fp16)[name = string("op_4847_cast_fp16")];
            tensor<int32, [1]> q_v_59_axes_1 = const()[name = string("q_v_59_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_59_cast_fp16 = expand_dims(axes = q_v_59_axes_1, x = var_4847_cast_fp16)[name = string("q_v_59_cast_fp16")];
            tensor<int32, [1]> k_head_119_axes_1 = const()[name = string("k_head_119_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_119_cast_fp16 = expand_dims(axes = k_head_119_axes_1, x = k_head_117_cast_fp16)[name = string("k_head_119_cast_fp16")];
            tensor<int32, [1]> v_head_119_axes_1 = const()[name = string("v_head_119_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_119_cast_fp16 = expand_dims(axes = v_head_119_axes_1, x = v_head_117_cast_fp16)[name = string("v_head_119_cast_fp16")];
            tensor<int32, [1]> p_head_119_axes_1 = const()[name = string("p_head_119_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_119_cast_fp16 = expand_dims(axes = p_head_119_axes_1, x = p_head_117_cast_fp16)[name = string("p_head_119_cast_fp16")];
            bool var_4853_transpose_x_3 = const()[name = string("op_4853_transpose_x_3"), val = bool(false)];
            bool var_4853_transpose_y_3 = const()[name = string("op_4853_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_4853_cast_fp16 = matmul(transpose_x = var_4853_transpose_x_3, transpose_y = var_4853_transpose_y_3, x = q_u_59_cast_fp16, y = k_head_119_cast_fp16)[name = string("op_4853_cast_fp16")];
            fp16 var_4854_to_fp16 = const()[name = string("op_4854_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_59_cast_fp16 = mul(x = var_4853_cast_fp16, y = var_4854_to_fp16)[name = string("scores_content_59_cast_fp16")];
            bool x_313_transpose_x_3 = const()[name = string("x_313_transpose_x_3"), val = bool(false)];
            bool x_313_transpose_y_3 = const()[name = string("x_313_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_313_cast_fp16 = matmul(transpose_x = x_313_transpose_x_3, transpose_y = x_313_transpose_y_3, x = q_v_59_cast_fp16, y = p_head_119_cast_fp16)[name = string("x_313_cast_fp16")];
            tensor<int32, [8]> x_315_pad_1 = const()[name = string("x_315_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_315_mode_1 = const()[name = string("x_315_mode_1"), val = string("constant")];
            fp16 const_1501_to_fp16 = const()[name = string("const_1501_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_315_cast_fp16 = pad(constant_val = const_1501_to_fp16, mode = x_315_mode_1, pad = x_315_pad_1, x = x_313_cast_fp16)[name = string("x_315_cast_fp16")];
            tensor<int32, [4]> var_4868 = const()[name = string("op_4868"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_317_cast_fp16 = reshape(shape = var_4868, x = x_315_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<int32, [4]> var_4872_begin_1 = const()[name = string("op_4872_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4872_end_1 = const()[name = string("op_4872_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_4872_end_mask_1 = const()[name = string("op_4872_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_4872_cast_fp16 = slice_by_index(begin = var_4872_begin_1, end = var_4872_end_1, end_mask = var_4872_end_mask_1, x = x_317_cast_fp16)[name = string("op_4872_cast_fp16")];
            tensor<int32, [4]> var_4874 = const()[name = string("op_4874"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_4875_cast_fp16 = reshape(shape = var_4874, x = var_4872_cast_fp16)[name = string("op_4875_cast_fp16")];
            tensor<int32, [4]> var_4880_begin_1 = const()[name = string("op_4880_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4880_end_1 = const()[name = string("op_4880_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_4880_end_mask_1 = const()[name = string("op_4880_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_4880_cast_fp16 = slice_by_index(begin = var_4880_begin_1, end = var_4880_end_1, end_mask = var_4880_end_mask_1, x = var_4875_cast_fp16)[name = string("op_4880_cast_fp16")];
            fp16 var_4881_to_fp16 = const()[name = string("op_4881_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_59_cast_fp16 = mul(x = var_4880_cast_fp16, y = var_4881_to_fp16)[name = string("scores_pos_59_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_59_cast_fp16 = add(x = scores_content_59_cast_fp16, y = scores_pos_59_cast_fp16)[name = string("logits_59_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_4884_cast_fp16 = softmax(axis = var_4004, x = logits_59_cast_fp16)[name = string("op_4884_cast_fp16")];
            bool var_4886_transpose_x_1 = const()[name = string("op_4886_transpose_x_1"), val = bool(false)];
            bool var_4886_transpose_y_1 = const()[name = string("op_4886_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_4886_cast_fp16 = matmul(transpose_x = var_4886_transpose_x_1, transpose_y = var_4886_transpose_y_1, x = var_4884_cast_fp16, y = v_head_119_cast_fp16)[name = string("op_4886_cast_fp16")];
            tensor<int32, [1]> var_4887_axes_1 = const()[name = string("op_4887_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_4887_cast_fp16 = squeeze(axes = var_4887_axes_1, x = var_4886_cast_fp16)[name = string("op_4887_cast_fp16")];
            string dense_output_305_pad_type_1 = const()[name = string("dense_output_305_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_305_strides_1 = const()[name = string("dense_output_305_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_305_pad_1 = const()[name = string("dense_output_305_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_305_dilations_1 = const()[name = string("dense_output_305_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_305_groups_1 = const()[name = string("dense_output_305_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107413248))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107544384))))[name = string("layers_3_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_305_cast_fp16 = conv(dilations = dense_output_305_dilations_1, groups = dense_output_305_groups_1, pad = dense_output_305_pad_1, pad_type = dense_output_305_pad_type_1, strides = dense_output_305_strides_1, weight = layers_3_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_305_cast_fp16")];
            string sparse_output_305_pad_type_1 = const()[name = string("sparse_output_305_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_305_strides_1 = const()[name = string("sparse_output_305_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_305_pad_1 = const()[name = string("sparse_output_305_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_305_dilations_1 = const()[name = string("sparse_output_305_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_305_groups_1 = const()[name = string("sparse_output_305_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107547648))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107544960))))[name = string("layers_3_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_305_cast_fp16 = conv(dilations = sparse_output_305_dilations_1, groups = sparse_output_305_groups_1, pad = sparse_output_305_pad_1, pad_type = sparse_output_305_pad_type_1, strides = sparse_output_305_strides_1, weight = layers_3_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_305_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4902_cast_fp16 = add(x = dense_output_305_cast_fp16, y = sparse_output_305_cast_fp16)[name = string("op_4902_cast_fp16")];
            tensor<int32, [4]> var_4903 = const()[name = string("op_4903"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4905 = const()[name = string("op_4905"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4904_cast_fp16 = transpose(perm = var_4903, x = var_4902_cast_fp16)[name = string("transpose_833")];
            tensor<fp16, [1, 51, 128]> q_head_61_cast_fp16 = reshape(shape = var_4905, x = var_4904_cast_fp16)[name = string("q_head_61_cast_fp16")];
            string dense_output_307_pad_type_1 = const()[name = string("dense_output_307_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_307_strides_1 = const()[name = string("dense_output_307_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_307_pad_1 = const()[name = string("dense_output_307_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_307_dilations_1 = const()[name = string("dense_output_307_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_307_groups_1 = const()[name = string("dense_output_307_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107564096))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107695232))))[name = string("layers_3_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_307_cast_fp16 = conv(dilations = dense_output_307_dilations_1, groups = dense_output_307_groups_1, pad = dense_output_307_pad_1, pad_type = dense_output_307_pad_type_1, strides = dense_output_307_strides_1, weight = layers_3_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_307_cast_fp16")];
            string sparse_output_307_pad_type_1 = const()[name = string("sparse_output_307_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_307_strides_1 = const()[name = string("sparse_output_307_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_307_pad_1 = const()[name = string("sparse_output_307_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_307_dilations_1 = const()[name = string("sparse_output_307_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_307_groups_1 = const()[name = string("sparse_output_307_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107698496))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107695808))))[name = string("layers_3_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_307_cast_fp16 = conv(dilations = sparse_output_307_dilations_1, groups = sparse_output_307_groups_1, pad = sparse_output_307_pad_1, pad_type = sparse_output_307_pad_type_1, strides = sparse_output_307_strides_1, weight = layers_3_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_307_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4921_cast_fp16 = add(x = dense_output_307_cast_fp16, y = sparse_output_307_cast_fp16)[name = string("op_4921_cast_fp16")];
            tensor<int32, [4]> var_4922 = const()[name = string("op_4922"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4924 = const()[name = string("op_4924"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4923_cast_fp16 = transpose(perm = var_4922, x = var_4921_cast_fp16)[name = string("transpose_832")];
            tensor<fp16, [1, 51, 128]> k_head_121_cast_fp16 = reshape(shape = var_4924, x = var_4923_cast_fp16)[name = string("k_head_121_cast_fp16")];
            string dense_output_309_pad_type_1 = const()[name = string("dense_output_309_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_309_strides_1 = const()[name = string("dense_output_309_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_309_pad_1 = const()[name = string("dense_output_309_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_309_dilations_1 = const()[name = string("dense_output_309_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_309_groups_1 = const()[name = string("dense_output_309_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107714944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107846080))))[name = string("layers_3_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_309_cast_fp16 = conv(dilations = dense_output_309_dilations_1, groups = dense_output_309_groups_1, pad = dense_output_309_pad_1, pad_type = dense_output_309_pad_type_1, strides = dense_output_309_strides_1, weight = layers_3_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_309_cast_fp16")];
            string sparse_output_309_pad_type_1 = const()[name = string("sparse_output_309_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_309_strides_1 = const()[name = string("sparse_output_309_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_309_pad_1 = const()[name = string("sparse_output_309_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_309_dilations_1 = const()[name = string("sparse_output_309_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_309_groups_1 = const()[name = string("sparse_output_309_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107849344))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107846656))))[name = string("layers_3_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_309_cast_fp16 = conv(dilations = sparse_output_309_dilations_1, groups = sparse_output_309_groups_1, pad = sparse_output_309_pad_1, pad_type = sparse_output_309_pad_type_1, strides = sparse_output_309_strides_1, weight = layers_3_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_309_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_4940_cast_fp16 = add(x = dense_output_309_cast_fp16, y = sparse_output_309_cast_fp16)[name = string("op_4940_cast_fp16")];
            tensor<int32, [4]> var_4941 = const()[name = string("op_4941"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4943 = const()[name = string("op_4943"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_4942_cast_fp16 = transpose(perm = var_4941, x = var_4940_cast_fp16)[name = string("transpose_831")];
            tensor<fp16, [1, 51, 128]> v_head_121_cast_fp16 = reshape(shape = var_4943, x = var_4942_cast_fp16)[name = string("v_head_121_cast_fp16")];
            string dense_output_311_pad_type_1 = const()[name = string("dense_output_311_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_311_strides_1 = const()[name = string("dense_output_311_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_311_pad_1 = const()[name = string("dense_output_311_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_311_dilations_1 = const()[name = string("dense_output_311_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_311_groups_1 = const()[name = string("dense_output_311_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107865792))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107996928))))[name = string("layers_3_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_311_cast_fp16 = conv(dilations = dense_output_311_dilations_1, groups = dense_output_311_groups_1, pad = dense_output_311_pad_1, pad_type = dense_output_311_pad_type_1, strides = dense_output_311_strides_1, weight = layers_3_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_311_cast_fp16")];
            string sparse_output_311_pad_type_1 = const()[name = string("sparse_output_311_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_311_strides_1 = const()[name = string("sparse_output_311_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_311_pad_1 = const()[name = string("sparse_output_311_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_311_dilations_1 = const()[name = string("sparse_output_311_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_311_groups_1 = const()[name = string("sparse_output_311_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108000192))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107997504))))[name = string("layers_3_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_311_cast_fp16 = conv(dilations = sparse_output_311_dilations_1, groups = sparse_output_311_groups_1, pad = sparse_output_311_pad_1, pad_type = sparse_output_311_pad_type_1, strides = sparse_output_311_strides_1, weight = layers_3_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_311_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_4959_cast_fp16 = add(x = dense_output_311_cast_fp16, y = sparse_output_311_cast_fp16)[name = string("op_4959_cast_fp16")];
            tensor<int32, [4]> var_4960 = const()[name = string("op_4960"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_4962 = const()[name = string("op_4962"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_4961_cast_fp16 = transpose(perm = var_4960, x = var_4959_cast_fp16)[name = string("transpose_830")];
            tensor<fp16, [1, 101, 128]> p_head_121_cast_fp16 = reshape(shape = var_4962, x = var_4961_cast_fp16)[name = string("p_head_121_cast_fp16")];
            tensor<fp16, [128]> var_4964_to_fp16 = const()[name = string("op_4964_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108016640)))];
            tensor<fp16, [1, 51, 128]> var_4965_cast_fp16 = add(x = q_head_61_cast_fp16, y = var_4964_to_fp16)[name = string("op_4965_cast_fp16")];
            tensor<int32, [1]> q_u_61_axes_1 = const()[name = string("q_u_61_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_61_cast_fp16 = expand_dims(axes = q_u_61_axes_1, x = var_4965_cast_fp16)[name = string("q_u_61_cast_fp16")];
            tensor<fp16, [128]> var_4967_to_fp16 = const()[name = string("op_4967_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108016960)))];
            tensor<fp16, [1, 51, 128]> var_4968_cast_fp16 = add(x = q_head_61_cast_fp16, y = var_4967_to_fp16)[name = string("op_4968_cast_fp16")];
            tensor<int32, [1]> q_v_61_axes_1 = const()[name = string("q_v_61_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_61_cast_fp16 = expand_dims(axes = q_v_61_axes_1, x = var_4968_cast_fp16)[name = string("q_v_61_cast_fp16")];
            tensor<int32, [1]> k_head_123_axes_1 = const()[name = string("k_head_123_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_123_cast_fp16 = expand_dims(axes = k_head_123_axes_1, x = k_head_121_cast_fp16)[name = string("k_head_123_cast_fp16")];
            tensor<int32, [1]> v_head_123_axes_1 = const()[name = string("v_head_123_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_123_cast_fp16 = expand_dims(axes = v_head_123_axes_1, x = v_head_121_cast_fp16)[name = string("v_head_123_cast_fp16")];
            tensor<int32, [1]> p_head_123_axes_1 = const()[name = string("p_head_123_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_123_cast_fp16 = expand_dims(axes = p_head_123_axes_1, x = p_head_121_cast_fp16)[name = string("p_head_123_cast_fp16")];
            bool var_4974_transpose_x_3 = const()[name = string("op_4974_transpose_x_3"), val = bool(false)];
            bool var_4974_transpose_y_3 = const()[name = string("op_4974_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_4974_cast_fp16 = matmul(transpose_x = var_4974_transpose_x_3, transpose_y = var_4974_transpose_y_3, x = q_u_61_cast_fp16, y = k_head_123_cast_fp16)[name = string("op_4974_cast_fp16")];
            fp16 var_4975_to_fp16 = const()[name = string("op_4975_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_61_cast_fp16 = mul(x = var_4974_cast_fp16, y = var_4975_to_fp16)[name = string("scores_content_61_cast_fp16")];
            bool x_321_transpose_x_3 = const()[name = string("x_321_transpose_x_3"), val = bool(false)];
            bool x_321_transpose_y_3 = const()[name = string("x_321_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_321_cast_fp16 = matmul(transpose_x = x_321_transpose_x_3, transpose_y = x_321_transpose_y_3, x = q_v_61_cast_fp16, y = p_head_123_cast_fp16)[name = string("x_321_cast_fp16")];
            tensor<int32, [8]> x_323_pad_1 = const()[name = string("x_323_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_323_mode_1 = const()[name = string("x_323_mode_1"), val = string("constant")];
            fp16 const_1507_to_fp16 = const()[name = string("const_1507_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_323_cast_fp16 = pad(constant_val = const_1507_to_fp16, mode = x_323_mode_1, pad = x_323_pad_1, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")];
            tensor<int32, [4]> var_4989 = const()[name = string("op_4989"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_325_cast_fp16 = reshape(shape = var_4989, x = x_323_cast_fp16)[name = string("x_325_cast_fp16")];
            tensor<int32, [4]> var_4993_begin_1 = const()[name = string("op_4993_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4993_end_1 = const()[name = string("op_4993_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_4993_end_mask_1 = const()[name = string("op_4993_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_4993_cast_fp16 = slice_by_index(begin = var_4993_begin_1, end = var_4993_end_1, end_mask = var_4993_end_mask_1, x = x_325_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<int32, [4]> var_4995 = const()[name = string("op_4995"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_4996_cast_fp16 = reshape(shape = var_4995, x = var_4993_cast_fp16)[name = string("op_4996_cast_fp16")];
            tensor<int32, [4]> var_5001_begin_1 = const()[name = string("op_5001_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5001_end_1 = const()[name = string("op_5001_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_5001_end_mask_1 = const()[name = string("op_5001_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_1, end = var_5001_end_1, end_mask = var_5001_end_mask_1, x = var_4996_cast_fp16)[name = string("op_5001_cast_fp16")];
            fp16 var_5002_to_fp16 = const()[name = string("op_5002_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_61_cast_fp16 = mul(x = var_5001_cast_fp16, y = var_5002_to_fp16)[name = string("scores_pos_61_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_61_cast_fp16 = add(x = scores_content_61_cast_fp16, y = scores_pos_61_cast_fp16)[name = string("logits_61_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_5005_cast_fp16 = softmax(axis = var_4004, x = logits_61_cast_fp16)[name = string("op_5005_cast_fp16")];
            bool var_5007_transpose_x_1 = const()[name = string("op_5007_transpose_x_1"), val = bool(false)];
            bool var_5007_transpose_y_1 = const()[name = string("op_5007_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_5007_cast_fp16 = matmul(transpose_x = var_5007_transpose_x_1, transpose_y = var_5007_transpose_y_1, x = var_5005_cast_fp16, y = v_head_123_cast_fp16)[name = string("op_5007_cast_fp16")];
            tensor<int32, [1]> var_5008_axes_1 = const()[name = string("op_5008_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_5008_cast_fp16 = squeeze(axes = var_5008_axes_1, x = var_5007_cast_fp16)[name = string("op_5008_cast_fp16")];
            string dense_output_313_pad_type_1 = const()[name = string("dense_output_313_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_313_strides_1 = const()[name = string("dense_output_313_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_313_pad_1 = const()[name = string("dense_output_313_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_313_dilations_1 = const()[name = string("dense_output_313_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_313_groups_1 = const()[name = string("dense_output_313_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108017280))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108148416))))[name = string("layers_3_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_313_cast_fp16 = conv(dilations = dense_output_313_dilations_1, groups = dense_output_313_groups_1, pad = dense_output_313_pad_1, pad_type = dense_output_313_pad_type_1, strides = dense_output_313_strides_1, weight = layers_3_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_313_cast_fp16")];
            string sparse_output_313_pad_type_1 = const()[name = string("sparse_output_313_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_313_strides_1 = const()[name = string("sparse_output_313_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_313_pad_1 = const()[name = string("sparse_output_313_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_313_dilations_1 = const()[name = string("sparse_output_313_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_313_groups_1 = const()[name = string("sparse_output_313_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108151680))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108148992))))[name = string("layers_3_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_313_cast_fp16 = conv(dilations = sparse_output_313_dilations_1, groups = sparse_output_313_groups_1, pad = sparse_output_313_pad_1, pad_type = sparse_output_313_pad_type_1, strides = sparse_output_313_strides_1, weight = layers_3_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_313_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5023_cast_fp16 = add(x = dense_output_313_cast_fp16, y = sparse_output_313_cast_fp16)[name = string("op_5023_cast_fp16")];
            tensor<int32, [4]> var_5024 = const()[name = string("op_5024"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5026 = const()[name = string("op_5026"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5025_cast_fp16 = transpose(perm = var_5024, x = var_5023_cast_fp16)[name = string("transpose_829")];
            tensor<fp16, [1, 51, 128]> q_head_63_cast_fp16 = reshape(shape = var_5026, x = var_5025_cast_fp16)[name = string("q_head_63_cast_fp16")];
            string dense_output_315_pad_type_1 = const()[name = string("dense_output_315_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_315_strides_1 = const()[name = string("dense_output_315_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_315_pad_1 = const()[name = string("dense_output_315_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_315_dilations_1 = const()[name = string("dense_output_315_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_315_groups_1 = const()[name = string("dense_output_315_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108168128))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108299264))))[name = string("layers_3_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_315_cast_fp16 = conv(dilations = dense_output_315_dilations_1, groups = dense_output_315_groups_1, pad = dense_output_315_pad_1, pad_type = dense_output_315_pad_type_1, strides = dense_output_315_strides_1, weight = layers_3_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_315_cast_fp16")];
            string sparse_output_315_pad_type_1 = const()[name = string("sparse_output_315_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_315_strides_1 = const()[name = string("sparse_output_315_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_315_pad_1 = const()[name = string("sparse_output_315_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_315_dilations_1 = const()[name = string("sparse_output_315_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_315_groups_1 = const()[name = string("sparse_output_315_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108302528))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108299840))))[name = string("layers_3_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_315_cast_fp16 = conv(dilations = sparse_output_315_dilations_1, groups = sparse_output_315_groups_1, pad = sparse_output_315_pad_1, pad_type = sparse_output_315_pad_type_1, strides = sparse_output_315_strides_1, weight = layers_3_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_315_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5042_cast_fp16 = add(x = dense_output_315_cast_fp16, y = sparse_output_315_cast_fp16)[name = string("op_5042_cast_fp16")];
            tensor<int32, [4]> var_5043 = const()[name = string("op_5043"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5045 = const()[name = string("op_5045"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5044_cast_fp16 = transpose(perm = var_5043, x = var_5042_cast_fp16)[name = string("transpose_828")];
            tensor<fp16, [1, 51, 128]> k_head_125_cast_fp16 = reshape(shape = var_5045, x = var_5044_cast_fp16)[name = string("k_head_125_cast_fp16")];
            string dense_output_317_pad_type_1 = const()[name = string("dense_output_317_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_317_strides_1 = const()[name = string("dense_output_317_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_317_pad_1 = const()[name = string("dense_output_317_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_317_dilations_1 = const()[name = string("dense_output_317_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_317_groups_1 = const()[name = string("dense_output_317_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108318976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108450112))))[name = string("layers_3_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_317_cast_fp16 = conv(dilations = dense_output_317_dilations_1, groups = dense_output_317_groups_1, pad = dense_output_317_pad_1, pad_type = dense_output_317_pad_type_1, strides = dense_output_317_strides_1, weight = layers_3_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = string("dense_output_317_cast_fp16")];
            string sparse_output_317_pad_type_1 = const()[name = string("sparse_output_317_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_317_strides_1 = const()[name = string("sparse_output_317_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_317_pad_1 = const()[name = string("sparse_output_317_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_317_dilations_1 = const()[name = string("sparse_output_317_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_317_groups_1 = const()[name = string("sparse_output_317_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108453376))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108450688))))[name = string("layers_3_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_317_cast_fp16 = conv(dilations = sparse_output_317_dilations_1, groups = sparse_output_317_groups_1, pad = sparse_output_317_pad_1, pad_type = sparse_output_317_pad_type_1, strides = sparse_output_317_strides_1, weight = layers_3_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_167_cast_fp16)[name = string("sparse_output_317_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5061_cast_fp16 = add(x = dense_output_317_cast_fp16, y = sparse_output_317_cast_fp16)[name = string("op_5061_cast_fp16")];
            tensor<int32, [4]> var_5062 = const()[name = string("op_5062"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5064 = const()[name = string("op_5064"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5063_cast_fp16 = transpose(perm = var_5062, x = var_5061_cast_fp16)[name = string("transpose_827")];
            tensor<fp16, [1, 51, 128]> v_head_125_cast_fp16 = reshape(shape = var_5064, x = var_5063_cast_fp16)[name = string("v_head_125_cast_fp16")];
            string dense_output_319_pad_type_1 = const()[name = string("dense_output_319_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_319_strides_1 = const()[name = string("dense_output_319_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_319_pad_1 = const()[name = string("dense_output_319_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_319_dilations_1 = const()[name = string("dense_output_319_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_319_groups_1 = const()[name = string("dense_output_319_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108469824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108600960))))[name = string("layers_3_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_319_cast_fp16 = conv(dilations = dense_output_319_dilations_1, groups = dense_output_319_groups_1, pad = dense_output_319_pad_1, pad_type = dense_output_319_pad_type_1, strides = dense_output_319_strides_1, weight = layers_3_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_319_cast_fp16")];
            string sparse_output_319_pad_type_1 = const()[name = string("sparse_output_319_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_319_strides_1 = const()[name = string("sparse_output_319_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_319_pad_1 = const()[name = string("sparse_output_319_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_319_dilations_1 = const()[name = string("sparse_output_319_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_319_groups_1 = const()[name = string("sparse_output_319_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_3_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108604224))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108601536))))[name = string("layers_3_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_319_cast_fp16 = conv(dilations = sparse_output_319_dilations_1, groups = sparse_output_319_groups_1, pad = sparse_output_319_pad_1, pad_type = sparse_output_319_pad_type_1, strides = sparse_output_319_strides_1, weight = layers_3_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_319_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_5080_cast_fp16 = add(x = dense_output_319_cast_fp16, y = sparse_output_319_cast_fp16)[name = string("op_5080_cast_fp16")];
            tensor<int32, [4]> var_5081 = const()[name = string("op_5081"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5083 = const()[name = string("op_5083"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_5082_cast_fp16 = transpose(perm = var_5081, x = var_5080_cast_fp16)[name = string("transpose_826")];
            tensor<fp16, [1, 101, 128]> p_head_125_cast_fp16 = reshape(shape = var_5083, x = var_5082_cast_fp16)[name = string("p_head_125_cast_fp16")];
            tensor<fp16, [128]> var_5085_to_fp16 = const()[name = string("op_5085_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108620672)))];
            tensor<fp16, [1, 51, 128]> var_5086_cast_fp16 = add(x = q_head_63_cast_fp16, y = var_5085_to_fp16)[name = string("op_5086_cast_fp16")];
            tensor<int32, [1]> q_u_63_axes_1 = const()[name = string("q_u_63_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_63_cast_fp16 = expand_dims(axes = q_u_63_axes_1, x = var_5086_cast_fp16)[name = string("q_u_63_cast_fp16")];
            tensor<fp16, [128]> var_5088_to_fp16 = const()[name = string("op_5088_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108620992)))];
            tensor<fp16, [1, 51, 128]> var_5089_cast_fp16 = add(x = q_head_63_cast_fp16, y = var_5088_to_fp16)[name = string("op_5089_cast_fp16")];
            tensor<int32, [1]> q_v_63_axes_1 = const()[name = string("q_v_63_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_63_cast_fp16 = expand_dims(axes = q_v_63_axes_1, x = var_5089_cast_fp16)[name = string("q_v_63_cast_fp16")];
            tensor<int32, [1]> k_head_127_axes_1 = const()[name = string("k_head_127_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_127_cast_fp16 = expand_dims(axes = k_head_127_axes_1, x = k_head_125_cast_fp16)[name = string("k_head_127_cast_fp16")];
            tensor<int32, [1]> v_head_127_axes_1 = const()[name = string("v_head_127_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_127_cast_fp16 = expand_dims(axes = v_head_127_axes_1, x = v_head_125_cast_fp16)[name = string("v_head_127_cast_fp16")];
            tensor<int32, [1]> p_head_127_axes_1 = const()[name = string("p_head_127_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_127_cast_fp16 = expand_dims(axes = p_head_127_axes_1, x = p_head_125_cast_fp16)[name = string("p_head_127_cast_fp16")];
            bool var_5095_transpose_x_3 = const()[name = string("op_5095_transpose_x_3"), val = bool(false)];
            bool var_5095_transpose_y_3 = const()[name = string("op_5095_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_5095_cast_fp16 = matmul(transpose_x = var_5095_transpose_x_3, transpose_y = var_5095_transpose_y_3, x = q_u_63_cast_fp16, y = k_head_127_cast_fp16)[name = string("op_5095_cast_fp16")];
            fp16 var_5096_to_fp16 = const()[name = string("op_5096_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_63_cast_fp16 = mul(x = var_5095_cast_fp16, y = var_5096_to_fp16)[name = string("scores_content_63_cast_fp16")];
            bool x_329_transpose_x_3 = const()[name = string("x_329_transpose_x_3"), val = bool(false)];
            bool x_329_transpose_y_3 = const()[name = string("x_329_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_329_cast_fp16 = matmul(transpose_x = x_329_transpose_x_3, transpose_y = x_329_transpose_y_3, x = q_v_63_cast_fp16, y = p_head_127_cast_fp16)[name = string("x_329_cast_fp16")];
            tensor<int32, [8]> x_331_pad_1 = const()[name = string("x_331_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_331_mode_1 = const()[name = string("x_331_mode_1"), val = string("constant")];
            fp16 const_1513_to_fp16 = const()[name = string("const_1513_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_331_cast_fp16 = pad(constant_val = const_1513_to_fp16, mode = x_331_mode_1, pad = x_331_pad_1, x = x_329_cast_fp16)[name = string("x_331_cast_fp16")];
            tensor<int32, [4]> var_5110 = const()[name = string("op_5110"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_333_cast_fp16 = reshape(shape = var_5110, x = x_331_cast_fp16)[name = string("x_333_cast_fp16")];
            tensor<int32, [4]> var_5114_begin_1 = const()[name = string("op_5114_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5114_end_1 = const()[name = string("op_5114_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_5114_end_mask_1 = const()[name = string("op_5114_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_5114_cast_fp16 = slice_by_index(begin = var_5114_begin_1, end = var_5114_end_1, end_mask = var_5114_end_mask_1, x = x_333_cast_fp16)[name = string("op_5114_cast_fp16")];
            tensor<int32, [4]> var_5116 = const()[name = string("op_5116"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_5117_cast_fp16 = reshape(shape = var_5116, x = var_5114_cast_fp16)[name = string("op_5117_cast_fp16")];
            tensor<int32, [4]> var_5122_begin_1 = const()[name = string("op_5122_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5122_end_1 = const()[name = string("op_5122_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_5122_end_mask_1 = const()[name = string("op_5122_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_5122_cast_fp16 = slice_by_index(begin = var_5122_begin_1, end = var_5122_end_1, end_mask = var_5122_end_mask_1, x = var_5117_cast_fp16)[name = string("op_5122_cast_fp16")];
            fp16 var_5123_to_fp16 = const()[name = string("op_5123_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_63_cast_fp16 = mul(x = var_5122_cast_fp16, y = var_5123_to_fp16)[name = string("scores_pos_63_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_63_cast_fp16 = add(x = scores_content_63_cast_fp16, y = scores_pos_63_cast_fp16)[name = string("logits_63_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_5126_cast_fp16 = softmax(axis = var_4004, x = logits_63_cast_fp16)[name = string("op_5126_cast_fp16")];
            bool var_5128_transpose_x_1 = const()[name = string("op_5128_transpose_x_1"), val = bool(false)];
            bool var_5128_transpose_y_1 = const()[name = string("op_5128_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_5128_cast_fp16 = matmul(transpose_x = var_5128_transpose_x_1, transpose_y = var_5128_transpose_y_1, x = var_5126_cast_fp16, y = v_head_127_cast_fp16)[name = string("op_5128_cast_fp16")];
            tensor<int32, [1]> o_head_7_axes_1 = const()[name = string("o_head_7_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_7_cast_fp16 = squeeze(axes = o_head_7_axes_1, x = var_5128_cast_fp16)[name = string("o_head_7_cast_fp16")];
            bool out_7_interleave_1 = const()[name = string("out_7_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_7_cast_fp16 = concat(axis = var_4004, interleave = out_7_interleave_1, values = (var_4282_cast_fp16, var_4403_cast_fp16, var_4524_cast_fp16, var_4645_cast_fp16, var_4766_cast_fp16, var_4887_cast_fp16, var_5008_cast_fp16, o_head_7_cast_fp16))[name = string("out_7_cast_fp16")];
            tensor<int32, [3]> var_5132_perm_1 = const()[name = string("op_5132_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_175_axes_1 = const()[name = string("input_175_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_5132_cast_fp16 = transpose(perm = var_5132_perm_1, x = out_7_cast_fp16)[name = string("transpose_825")];
            tensor<fp16, [1, 1024, 51, 1]> input_175_cast_fp16 = expand_dims(axes = input_175_axes_1, x = var_5132_cast_fp16)[name = string("input_175_cast_fp16")];
            string dense_output_321_pad_type_1 = const()[name = string("dense_output_321_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_321_strides_1 = const()[name = string("dense_output_321_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_321_pad_1 = const()[name = string("dense_output_321_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_321_dilations_1 = const()[name = string("dense_output_321_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_321_groups_1 = const()[name = string("dense_output_321_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108621312))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109669952))))[name = string("layers_3_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_321_cast_fp16 = conv(dilations = dense_output_321_dilations_1, groups = dense_output_321_groups_1, pad = dense_output_321_pad_1, pad_type = dense_output_321_pad_type_1, strides = dense_output_321_strides_1, weight = layers_3_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_175_cast_fp16)[name = string("dense_output_321_cast_fp16")];
            string sparse_output_321_pad_type_1 = const()[name = string("sparse_output_321_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_321_strides_1 = const()[name = string("sparse_output_321_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_321_pad_1 = const()[name = string("sparse_output_321_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_321_dilations_1 = const()[name = string("sparse_output_321_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_321_groups_1 = const()[name = string("sparse_output_321_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109691584))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109670528))))[name = string("layers_3_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_321_cast_fp16 = conv(dilations = sparse_output_321_dilations_1, groups = sparse_output_321_groups_1, pad = sparse_output_321_pad_1, pad_type = sparse_output_321_pad_type_1, strides = sparse_output_321_strides_1, weight = layers_3_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_175_cast_fp16)[name = string("sparse_output_321_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_7_cast_fp16 = add(x = dense_output_321_cast_fp16, y = sparse_output_321_cast_fp16)[name = string("out_conv_7_cast_fp16")];
            tensor<int32, [1]> var_5149_axes_1 = const()[name = string("op_5149_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_5149_cast_fp16 = squeeze(axes = var_5149_axes_1, x = out_conv_7_cast_fp16)[name = string("op_5149_cast_fp16")];
            tensor<int32, [3]> var_5150_perm_1 = const()[name = string("op_5150_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_5150_cast_fp16 = transpose(perm = var_5150_perm_1, x = var_5149_cast_fp16)[name = string("transpose_824")];
            tensor<fp16, [1, 51, 1024]> input_177_cast_fp16 = add(x = input_165_cast_fp16, y = var_5150_cast_fp16)[name = string("input_177_cast_fp16")];
            tensor<int32, [1]> x_337_axes_1 = const()[name = string("x_337_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_3_norm_conv_weight_to_fp16 = const()[name = string("layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109822720)))];
            tensor<fp16, [1024]> layers_3_norm_conv_bias_to_fp16 = const()[name = string("layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109824832)))];
            tensor<fp16, [1, 51, 1024]> x_337_cast_fp16 = layer_norm(axes = x_337_axes_1, beta = layers_3_norm_conv_bias_to_fp16, epsilon = var_4019_to_fp16, gamma = layers_3_norm_conv_weight_to_fp16, x = input_177_cast_fp16)[name = string("x_337_cast_fp16")];
            tensor<int32, [3]> var_5160_perm_1 = const()[name = string("op_5160_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_179_axes_1 = const()[name = string("input_179_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_5160_cast_fp16 = transpose(perm = var_5160_perm_1, x = x_337_cast_fp16)[name = string("transpose_823")];
            tensor<fp16, [1, 1024, 51, 1]> input_179_cast_fp16 = expand_dims(axes = input_179_axes_1, x = var_5160_cast_fp16)[name = string("input_179_cast_fp16")];
            string dense_output_323_pad_type_1 = const()[name = string("dense_output_323_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_323_strides_1 = const()[name = string("dense_output_323_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_323_pad_1 = const()[name = string("dense_output_323_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_323_dilations_1 = const()[name = string("dense_output_323_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_323_groups_1 = const()[name = string("dense_output_323_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_3_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109826944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111924160))))[name = string("layers_3_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_323_cast_fp16 = conv(dilations = dense_output_323_dilations_1, groups = dense_output_323_groups_1, pad = dense_output_323_pad_1, pad_type = dense_output_323_pad_type_1, strides = dense_output_323_strides_1, weight = layers_3_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_179_cast_fp16)[name = string("dense_output_323_cast_fp16")];
            string sparse_output_323_pad_type_1 = const()[name = string("sparse_output_323_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_323_strides_1 = const()[name = string("sparse_output_323_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_323_pad_1 = const()[name = string("sparse_output_323_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_323_dilations_1 = const()[name = string("sparse_output_323_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_323_groups_1 = const()[name = string("sparse_output_323_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_3_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111966784))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111924736))))[name = string("layers_3_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_323_cast_fp16 = conv(dilations = sparse_output_323_dilations_1, groups = sparse_output_323_groups_1, pad = sparse_output_323_pad_1, pad_type = sparse_output_323_pad_type_1, strides = sparse_output_323_strides_1, weight = layers_3_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_179_cast_fp16)[name = string("sparse_output_323_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_181_cast_fp16 = add(x = dense_output_323_cast_fp16, y = sparse_output_323_cast_fp16)[name = string("input_181_cast_fp16")];
            int32 input_183_split_num_splits_1 = const()[name = string("input_183_split_num_splits_1"), val = int32(2)];
            int32 input_183_split_axis_1 = const()[name = string("input_183_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_183_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_183_split_cast_fp16_1 = split(axis = input_183_split_axis_1, num_splits = input_183_split_num_splits_1, x = input_181_cast_fp16)[name = string("input_183_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_183_split_1_sigmoid_cast_fp16 = sigmoid(x = input_183_split_cast_fp16_1)[name = string("input_183_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_183_cast_fp16 = mul(x = input_183_split_cast_fp16_0, y = input_183_split_1_sigmoid_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<int32, [8]> input_185_pad_1 = const()[name = string("input_185_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_185_mode_1 = const()[name = string("input_185_mode_1"), val = string("constant")];
            fp16 const_1515_to_fp16 = const()[name = string("const_1515_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_185_cast_fp16 = pad(constant_val = const_1515_to_fp16, mode = input_185_mode_1, pad = input_185_pad_1, x = input_183_cast_fp16)[name = string("input_185_cast_fp16")];
            string dense_output_325_pad_type_1 = const()[name = string("dense_output_325_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_325_strides_1 = const()[name = string("dense_output_325_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_325_pad_1 = const()[name = string("dense_output_325_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_325_dilations_1 = const()[name = string("dense_output_325_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_325_groups_1 = const()[name = string("dense_output_325_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_325_cast_fp16 = conv(dilations = dense_output_325_dilations_1, groups = dense_output_325_groups_1, pad = dense_output_325_pad_1, pad_type = dense_output_325_pad_type_1, strides = dense_output_325_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_185_cast_fp16)[name = string("dense_output_325_cast_fp16")];
            string sparse_output_325_pad_type_1 = const()[name = string("sparse_output_325_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_325_strides_1 = const()[name = string("sparse_output_325_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_325_pad_1 = const()[name = string("sparse_output_325_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_325_dilations_1 = const()[name = string("sparse_output_325_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_325_groups_1 = const()[name = string("sparse_output_325_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_3_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112228992))))[name = string("layers_3_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_325_cast_fp16 = conv(dilations = sparse_output_325_dilations_1, groups = sparse_output_325_groups_1, pad = sparse_output_325_pad_1, pad_type = sparse_output_325_pad_type_1, strides = sparse_output_325_strides_1, weight = layers_3_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_185_cast_fp16)[name = string("sparse_output_325_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_187_cast_fp16 = add(x = dense_output_325_cast_fp16, y = sparse_output_325_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [1024]> layers_3_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_3_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112247488)))];
            tensor<fp16, [1024]> layers_3_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_3_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112249600)))];
            tensor<fp16, [1024]> layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112251712)))];
            tensor<fp16, [1024]> layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112253824)))];
            tensor<fp16, [1, 1024, 51, 1]> input_189_cast_fp16 = batch_norm(beta = layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_4019_to_fp16, gamma = layers_3_conv_batch_norm_weight_to_fp16, mean = layers_3_conv_batch_norm_running_mean_to_fp16, variance = layers_3_conv_batch_norm_running_var_to_fp16, x = input_187_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_191_cast_fp16 = silu(x = input_189_cast_fp16)[name = string("input_191_cast_fp16")];
            string dense_output_327_pad_type_1 = const()[name = string("dense_output_327_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_327_strides_1 = const()[name = string("dense_output_327_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_327_pad_1 = const()[name = string("dense_output_327_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_327_dilations_1 = const()[name = string("dense_output_327_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_327_groups_1 = const()[name = string("dense_output_327_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112255936))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113304576))))[name = string("layers_3_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_327_cast_fp16 = conv(dilations = dense_output_327_dilations_1, groups = dense_output_327_groups_1, pad = dense_output_327_pad_1, pad_type = dense_output_327_pad_type_1, strides = dense_output_327_strides_1, weight = layers_3_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_191_cast_fp16)[name = string("dense_output_327_cast_fp16")];
            string sparse_output_327_pad_type_1 = const()[name = string("sparse_output_327_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_327_strides_1 = const()[name = string("sparse_output_327_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_327_pad_1 = const()[name = string("sparse_output_327_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_327_dilations_1 = const()[name = string("sparse_output_327_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_327_groups_1 = const()[name = string("sparse_output_327_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_3_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113326208))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113305152))))[name = string("layers_3_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_327_cast_fp16 = conv(dilations = sparse_output_327_dilations_1, groups = sparse_output_327_groups_1, pad = sparse_output_327_pad_1, pad_type = sparse_output_327_pad_type_1, strides = sparse_output_327_strides_1, weight = layers_3_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_191_cast_fp16)[name = string("sparse_output_327_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_339_cast_fp16 = add(x = dense_output_327_cast_fp16, y = sparse_output_327_cast_fp16)[name = string("x_339_cast_fp16")];
            tensor<int32, [1]> var_5216_axes_1 = const()[name = string("op_5216_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_5216_cast_fp16 = squeeze(axes = var_5216_axes_1, x = x_339_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<int32, [3]> var_5217_perm_1 = const()[name = string("op_5217_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_5217_cast_fp16 = transpose(perm = var_5217_perm_1, x = var_5216_cast_fp16)[name = string("transpose_822")];
            tensor<fp16, [1, 51, 1024]> input_193_cast_fp16 = add(x = input_177_cast_fp16, y = var_5217_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<int32, [1]> x_341_axes_1 = const()[name = string("x_341_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113457344)))];
            tensor<fp16, [1024]> layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113459456)))];
            tensor<fp16, [1, 51, 1024]> x_341_cast_fp16 = layer_norm(axes = x_341_axes_1, beta = layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_4019_to_fp16, gamma = layers_3_norm_feed_forward2_weight_to_fp16, x = input_193_cast_fp16)[name = string("x_341_cast_fp16")];
            tensor<int32, [4]> var_5227 = const()[name = string("op_5227"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_343_cast_fp16 = reshape(shape = var_5227, x = x_341_cast_fp16)[name = string("x_343_cast_fp16")];
            tensor<int32, [4]> input_195_perm_1 = const()[name = string("input_195_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_329_pad_type_1 = const()[name = string("dense_output_329_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_329_strides_1 = const()[name = string("dense_output_329_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_329_pad_1 = const()[name = string("dense_output_329_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_329_dilations_1 = const()[name = string("dense_output_329_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_329_groups_1 = const()[name = string("dense_output_329_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_3_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113461568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117655936))))[name = string("layers_3_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_195_cast_fp16 = transpose(perm = input_195_perm_1, x = x_343_cast_fp16)[name = string("transpose_821")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_329_cast_fp16 = conv(dilations = dense_output_329_dilations_1, groups = dense_output_329_groups_1, pad = dense_output_329_pad_1, pad_type = dense_output_329_pad_type_1, strides = dense_output_329_strides_1, weight = layers_3_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("dense_output_329_cast_fp16")];
            string sparse_output_329_pad_type_1 = const()[name = string("sparse_output_329_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_329_strides_1 = const()[name = string("sparse_output_329_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_329_pad_1 = const()[name = string("sparse_output_329_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_329_dilations_1 = const()[name = string("sparse_output_329_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_329_groups_1 = const()[name = string("sparse_output_329_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_3_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117740480))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117656512))))[name = string("layers_3_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_329_cast_fp16 = conv(dilations = sparse_output_329_dilations_1, groups = sparse_output_329_groups_1, pad = sparse_output_329_pad_1, pad_type = sparse_output_329_pad_type_1, strides = sparse_output_329_strides_1, weight = layers_3_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_195_cast_fp16)[name = string("sparse_output_329_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_197_cast_fp16 = add(x = dense_output_329_cast_fp16, y = sparse_output_329_cast_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_199_cast_fp16 = silu(x = input_197_cast_fp16)[name = string("input_199_cast_fp16")];
            string dense_output_331_pad_type_1 = const()[name = string("dense_output_331_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_331_strides_1 = const()[name = string("dense_output_331_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_331_pad_1 = const()[name = string("dense_output_331_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_331_dilations_1 = const()[name = string("dense_output_331_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_331_groups_1 = const()[name = string("dense_output_331_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_3_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118264832))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122459200))))[name = string("layers_3_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_331_cast_fp16 = conv(dilations = dense_output_331_dilations_1, groups = dense_output_331_groups_1, pad = dense_output_331_pad_1, pad_type = dense_output_331_pad_type_1, strides = dense_output_331_strides_1, weight = layers_3_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("dense_output_331_cast_fp16")];
            string sparse_output_331_pad_type_1 = const()[name = string("sparse_output_331_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_331_strides_1 = const()[name = string("sparse_output_331_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_331_pad_1 = const()[name = string("sparse_output_331_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_331_dilations_1 = const()[name = string("sparse_output_331_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_331_groups_1 = const()[name = string("sparse_output_331_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_3_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122543744))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122459776))))[name = string("layers_3_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_331_cast_fp16 = conv(dilations = sparse_output_331_dilations_1, groups = sparse_output_331_groups_1, pad = sparse_output_331_pad_1, pad_type = sparse_output_331_pad_type_1, strides = sparse_output_331_strides_1, weight = layers_3_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_199_cast_fp16)[name = string("sparse_output_331_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_345_cast_fp16 = add(x = dense_output_331_cast_fp16, y = sparse_output_331_cast_fp16)[name = string("x_345_cast_fp16")];
            tensor<int32, [4]> x_347_perm_1 = const()[name = string("x_347_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_5262 = const()[name = string("op_5262"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_347_cast_fp16 = transpose(perm = x_347_perm_1, x = x_345_cast_fp16)[name = string("transpose_820")];
            tensor<fp16, [1, 51, 1024]> var_5263_cast_fp16 = reshape(shape = var_5262, x = x_347_cast_fp16)[name = string("op_5263_cast_fp16")];
            fp16 var_5264_to_fp16 = const()[name = string("op_5264_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_5265_cast_fp16 = mul(x = var_5263_cast_fp16, y = var_5264_to_fp16)[name = string("op_5265_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_201_cast_fp16 = add(x = input_193_cast_fp16, y = var_5265_cast_fp16)[name = string("input_201_cast_fp16")];
            tensor<int32, [1]> input_203_axes_1 = const()[name = string("input_203_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_3_norm_out_weight_to_fp16 = const()[name = string("layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123068096)))];
            tensor<fp16, [1024]> layers_3_norm_out_bias_to_fp16 = const()[name = string("layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123070208)))];
            tensor<fp16, [1, 51, 1024]> input_203_cast_fp16 = layer_norm(axes = input_203_axes_1, beta = layers_3_norm_out_bias_to_fp16, epsilon = var_4019_to_fp16, gamma = layers_3_norm_out_weight_to_fp16, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
            int32 var_5273 = const()[name = string("op_5273"), val = int32(-1)];
            tensor<int32, [1]> x_349_axes_1 = const()[name = string("x_349_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123072320)))];
            tensor<fp16, [1024]> layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123074432)))];
            fp16 var_5288_to_fp16 = const()[name = string("op_5288_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_349_cast_fp16 = layer_norm(axes = x_349_axes_1, beta = layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_5288_to_fp16, gamma = layers_4_norm_feed_forward1_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_349_cast_fp16")];
            tensor<int32, [4]> var_5307 = const()[name = string("op_5307"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_351_cast_fp16 = reshape(shape = var_5307, x = x_349_cast_fp16)[name = string("x_351_cast_fp16")];
            tensor<int32, [4]> input_205_perm_1 = const()[name = string("input_205_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_333_pad_type_1 = const()[name = string("dense_output_333_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_333_strides_1 = const()[name = string("dense_output_333_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_333_pad_1 = const()[name = string("dense_output_333_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_333_dilations_1 = const()[name = string("dense_output_333_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_333_groups_1 = const()[name = string("dense_output_333_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_4_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123076544))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127270912))))[name = string("layers_4_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_205_cast_fp16 = transpose(perm = input_205_perm_1, x = x_351_cast_fp16)[name = string("transpose_819")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_333_cast_fp16 = conv(dilations = dense_output_333_dilations_1, groups = dense_output_333_groups_1, pad = dense_output_333_pad_1, pad_type = dense_output_333_pad_type_1, strides = dense_output_333_strides_1, weight = layers_4_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_205_cast_fp16)[name = string("dense_output_333_cast_fp16")];
            string sparse_output_333_pad_type_1 = const()[name = string("sparse_output_333_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_333_strides_1 = const()[name = string("sparse_output_333_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_333_pad_1 = const()[name = string("sparse_output_333_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_333_dilations_1 = const()[name = string("sparse_output_333_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_333_groups_1 = const()[name = string("sparse_output_333_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_4_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127355456))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127271488))))[name = string("layers_4_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_333_cast_fp16 = conv(dilations = sparse_output_333_dilations_1, groups = sparse_output_333_groups_1, pad = sparse_output_333_pad_1, pad_type = sparse_output_333_pad_type_1, strides = sparse_output_333_strides_1, weight = layers_4_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_205_cast_fp16)[name = string("sparse_output_333_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_207_cast_fp16 = add(x = dense_output_333_cast_fp16, y = sparse_output_333_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_209_cast_fp16 = silu(x = input_207_cast_fp16)[name = string("input_209_cast_fp16")];
            string dense_output_335_pad_type_1 = const()[name = string("dense_output_335_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_335_strides_1 = const()[name = string("dense_output_335_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_335_pad_1 = const()[name = string("dense_output_335_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_335_dilations_1 = const()[name = string("dense_output_335_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_335_groups_1 = const()[name = string("dense_output_335_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_4_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127879808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132074176))))[name = string("layers_4_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_335_cast_fp16 = conv(dilations = dense_output_335_dilations_1, groups = dense_output_335_groups_1, pad = dense_output_335_pad_1, pad_type = dense_output_335_pad_type_1, strides = dense_output_335_strides_1, weight = layers_4_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_209_cast_fp16)[name = string("dense_output_335_cast_fp16")];
            string sparse_output_335_pad_type_1 = const()[name = string("sparse_output_335_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_335_strides_1 = const()[name = string("sparse_output_335_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_335_pad_1 = const()[name = string("sparse_output_335_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_335_dilations_1 = const()[name = string("sparse_output_335_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_335_groups_1 = const()[name = string("sparse_output_335_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_4_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132158720))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132074752))))[name = string("layers_4_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_335_cast_fp16 = conv(dilations = sparse_output_335_dilations_1, groups = sparse_output_335_groups_1, pad = sparse_output_335_pad_1, pad_type = sparse_output_335_pad_type_1, strides = sparse_output_335_strides_1, weight = layers_4_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_209_cast_fp16)[name = string("sparse_output_335_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_353_cast_fp16 = add(x = dense_output_335_cast_fp16, y = sparse_output_335_cast_fp16)[name = string("x_353_cast_fp16")];
            tensor<int32, [4]> x_355_perm_1 = const()[name = string("x_355_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_5342 = const()[name = string("op_5342"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_355_cast_fp16 = transpose(perm = x_355_perm_1, x = x_353_cast_fp16)[name = string("transpose_818")];
            tensor<fp16, [1, 51, 1024]> var_5343_cast_fp16 = reshape(shape = var_5342, x = x_355_cast_fp16)[name = string("op_5343_cast_fp16")];
            fp16 var_5344_to_fp16 = const()[name = string("op_5344_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_5345_cast_fp16 = mul(x = var_5343_cast_fp16, y = var_5344_to_fp16)[name = string("op_5345_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_211_cast_fp16 = add(x = input_203_cast_fp16, y = var_5345_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<int32, [1]> q_9_axes_1 = const()[name = string("q_9_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_4_norm_self_att_weight_to_fp16 = const()[name = string("layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132683072)))];
            tensor<fp16, [1024]> layers_4_norm_self_att_bias_to_fp16 = const()[name = string("layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132685184)))];
            tensor<fp16, [1, 51, 1024]> q_9_cast_fp16 = layer_norm(axes = q_9_axes_1, beta = layers_4_norm_self_att_bias_to_fp16, epsilon = var_5288_to_fp16, gamma = layers_4_norm_self_att_weight_to_fp16, x = input_211_cast_fp16)[name = string("q_9_cast_fp16")];
            tensor<int32, [3]> var_5419 = const()[name = string("op_5419"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_213_axes_1 = const()[name = string("input_213_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_5420_cast_fp16 = transpose(perm = var_5419, x = q_9_cast_fp16)[name = string("transpose_817")];
            tensor<fp16, [1, 1024, 51, 1]> input_213_cast_fp16 = expand_dims(axes = input_213_axes_1, x = var_5420_cast_fp16)[name = string("input_213_cast_fp16")];
            string dense_output_337_pad_type_1 = const()[name = string("dense_output_337_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_337_strides_1 = const()[name = string("dense_output_337_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_337_pad_1 = const()[name = string("dense_output_337_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_337_dilations_1 = const()[name = string("dense_output_337_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_337_groups_1 = const()[name = string("dense_output_337_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132687296))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132818432))))[name = string("layers_4_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_337_cast_fp16 = conv(dilations = dense_output_337_dilations_1, groups = dense_output_337_groups_1, pad = dense_output_337_pad_1, pad_type = dense_output_337_pad_type_1, strides = dense_output_337_strides_1, weight = layers_4_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_337_cast_fp16")];
            string sparse_output_337_pad_type_1 = const()[name = string("sparse_output_337_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_337_strides_1 = const()[name = string("sparse_output_337_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_337_pad_1 = const()[name = string("sparse_output_337_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_337_dilations_1 = const()[name = string("sparse_output_337_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_337_groups_1 = const()[name = string("sparse_output_337_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132821696))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132819008))))[name = string("layers_4_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_337_cast_fp16 = conv(dilations = sparse_output_337_dilations_1, groups = sparse_output_337_groups_1, pad = sparse_output_337_pad_1, pad_type = sparse_output_337_pad_type_1, strides = sparse_output_337_strides_1, weight = layers_4_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_337_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5445_cast_fp16 = add(x = dense_output_337_cast_fp16, y = sparse_output_337_cast_fp16)[name = string("op_5445_cast_fp16")];
            tensor<int32, [4]> var_5446 = const()[name = string("op_5446"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5448 = const()[name = string("op_5448"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5447_cast_fp16 = transpose(perm = var_5446, x = var_5445_cast_fp16)[name = string("transpose_816")];
            tensor<fp16, [1, 51, 128]> q_head_65_cast_fp16 = reshape(shape = var_5448, x = var_5447_cast_fp16)[name = string("q_head_65_cast_fp16")];
            string dense_output_339_pad_type_1 = const()[name = string("dense_output_339_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_339_strides_1 = const()[name = string("dense_output_339_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_339_pad_1 = const()[name = string("dense_output_339_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_339_dilations_1 = const()[name = string("dense_output_339_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_339_groups_1 = const()[name = string("dense_output_339_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132838144))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132969280))))[name = string("layers_4_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_339_cast_fp16 = conv(dilations = dense_output_339_dilations_1, groups = dense_output_339_groups_1, pad = dense_output_339_pad_1, pad_type = dense_output_339_pad_type_1, strides = dense_output_339_strides_1, weight = layers_4_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_339_cast_fp16")];
            string sparse_output_339_pad_type_1 = const()[name = string("sparse_output_339_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_339_strides_1 = const()[name = string("sparse_output_339_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_339_pad_1 = const()[name = string("sparse_output_339_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_339_dilations_1 = const()[name = string("sparse_output_339_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_339_groups_1 = const()[name = string("sparse_output_339_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132972544))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132969856))))[name = string("layers_4_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_339_cast_fp16 = conv(dilations = sparse_output_339_dilations_1, groups = sparse_output_339_groups_1, pad = sparse_output_339_pad_1, pad_type = sparse_output_339_pad_type_1, strides = sparse_output_339_strides_1, weight = layers_4_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_339_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5464_cast_fp16 = add(x = dense_output_339_cast_fp16, y = sparse_output_339_cast_fp16)[name = string("op_5464_cast_fp16")];
            tensor<int32, [4]> var_5465 = const()[name = string("op_5465"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5467 = const()[name = string("op_5467"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5466_cast_fp16 = transpose(perm = var_5465, x = var_5464_cast_fp16)[name = string("transpose_815")];
            tensor<fp16, [1, 51, 128]> k_head_129_cast_fp16 = reshape(shape = var_5467, x = var_5466_cast_fp16)[name = string("k_head_129_cast_fp16")];
            string dense_output_341_pad_type_1 = const()[name = string("dense_output_341_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_341_strides_1 = const()[name = string("dense_output_341_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_341_pad_1 = const()[name = string("dense_output_341_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_341_dilations_1 = const()[name = string("dense_output_341_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_341_groups_1 = const()[name = string("dense_output_341_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132988992))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133120128))))[name = string("layers_4_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_341_cast_fp16 = conv(dilations = dense_output_341_dilations_1, groups = dense_output_341_groups_1, pad = dense_output_341_pad_1, pad_type = dense_output_341_pad_type_1, strides = dense_output_341_strides_1, weight = layers_4_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_341_cast_fp16")];
            string sparse_output_341_pad_type_1 = const()[name = string("sparse_output_341_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_341_strides_1 = const()[name = string("sparse_output_341_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_341_pad_1 = const()[name = string("sparse_output_341_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_341_dilations_1 = const()[name = string("sparse_output_341_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_341_groups_1 = const()[name = string("sparse_output_341_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133123392))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133120704))))[name = string("layers_4_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_341_cast_fp16 = conv(dilations = sparse_output_341_dilations_1, groups = sparse_output_341_groups_1, pad = sparse_output_341_pad_1, pad_type = sparse_output_341_pad_type_1, strides = sparse_output_341_strides_1, weight = layers_4_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_341_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5483_cast_fp16 = add(x = dense_output_341_cast_fp16, y = sparse_output_341_cast_fp16)[name = string("op_5483_cast_fp16")];
            tensor<int32, [4]> var_5484 = const()[name = string("op_5484"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5486 = const()[name = string("op_5486"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5485_cast_fp16 = transpose(perm = var_5484, x = var_5483_cast_fp16)[name = string("transpose_814")];
            tensor<fp16, [1, 51, 128]> v_head_129_cast_fp16 = reshape(shape = var_5486, x = var_5485_cast_fp16)[name = string("v_head_129_cast_fp16")];
            string dense_output_343_pad_type_1 = const()[name = string("dense_output_343_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_343_strides_1 = const()[name = string("dense_output_343_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_343_pad_1 = const()[name = string("dense_output_343_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_343_dilations_1 = const()[name = string("dense_output_343_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_343_groups_1 = const()[name = string("dense_output_343_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133139840))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133270976))))[name = string("layers_4_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_343_cast_fp16 = conv(dilations = dense_output_343_dilations_1, groups = dense_output_343_groups_1, pad = dense_output_343_pad_1, pad_type = dense_output_343_pad_type_1, strides = dense_output_343_strides_1, weight = layers_4_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_343_cast_fp16")];
            string sparse_output_343_pad_type_1 = const()[name = string("sparse_output_343_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_343_strides_1 = const()[name = string("sparse_output_343_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_343_pad_1 = const()[name = string("sparse_output_343_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_343_dilations_1 = const()[name = string("sparse_output_343_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_343_groups_1 = const()[name = string("sparse_output_343_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133274240))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133271552))))[name = string("layers_4_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_343_cast_fp16 = conv(dilations = sparse_output_343_dilations_1, groups = sparse_output_343_groups_1, pad = sparse_output_343_pad_1, pad_type = sparse_output_343_pad_type_1, strides = sparse_output_343_strides_1, weight = layers_4_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_343_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_5502_cast_fp16 = add(x = dense_output_343_cast_fp16, y = sparse_output_343_cast_fp16)[name = string("op_5502_cast_fp16")];
            tensor<int32, [4]> var_5503 = const()[name = string("op_5503"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5505 = const()[name = string("op_5505"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_5504_cast_fp16 = transpose(perm = var_5503, x = var_5502_cast_fp16)[name = string("transpose_813")];
            tensor<fp16, [1, 101, 128]> p_head_129_cast_fp16 = reshape(shape = var_5505, x = var_5504_cast_fp16)[name = string("p_head_129_cast_fp16")];
            tensor<fp16, [128]> var_5507_to_fp16 = const()[name = string("op_5507_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133290688)))];
            tensor<fp16, [1, 51, 128]> var_5508_cast_fp16 = add(x = q_head_65_cast_fp16, y = var_5507_to_fp16)[name = string("op_5508_cast_fp16")];
            tensor<int32, [1]> q_u_65_axes_1 = const()[name = string("q_u_65_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_65_cast_fp16 = expand_dims(axes = q_u_65_axes_1, x = var_5508_cast_fp16)[name = string("q_u_65_cast_fp16")];
            tensor<fp16, [128]> var_5510_to_fp16 = const()[name = string("op_5510_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133291008)))];
            tensor<fp16, [1, 51, 128]> var_5511_cast_fp16 = add(x = q_head_65_cast_fp16, y = var_5510_to_fp16)[name = string("op_5511_cast_fp16")];
            tensor<int32, [1]> q_v_65_axes_1 = const()[name = string("q_v_65_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_65_cast_fp16 = expand_dims(axes = q_v_65_axes_1, x = var_5511_cast_fp16)[name = string("q_v_65_cast_fp16")];
            tensor<int32, [1]> k_head_131_axes_1 = const()[name = string("k_head_131_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_131_cast_fp16 = expand_dims(axes = k_head_131_axes_1, x = k_head_129_cast_fp16)[name = string("k_head_131_cast_fp16")];
            tensor<int32, [1]> v_head_131_axes_1 = const()[name = string("v_head_131_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_131_cast_fp16 = expand_dims(axes = v_head_131_axes_1, x = v_head_129_cast_fp16)[name = string("v_head_131_cast_fp16")];
            tensor<int32, [1]> p_head_131_axes_1 = const()[name = string("p_head_131_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_131_cast_fp16 = expand_dims(axes = p_head_131_axes_1, x = p_head_129_cast_fp16)[name = string("p_head_131_cast_fp16")];
            bool var_5517_transpose_x_3 = const()[name = string("op_5517_transpose_x_3"), val = bool(false)];
            bool var_5517_transpose_y_3 = const()[name = string("op_5517_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_5517_cast_fp16 = matmul(transpose_x = var_5517_transpose_x_3, transpose_y = var_5517_transpose_y_3, x = q_u_65_cast_fp16, y = k_head_131_cast_fp16)[name = string("op_5517_cast_fp16")];
            fp16 var_5518_to_fp16 = const()[name = string("op_5518_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_65_cast_fp16 = mul(x = var_5517_cast_fp16, y = var_5518_to_fp16)[name = string("scores_content_65_cast_fp16")];
            bool x_357_transpose_x_3 = const()[name = string("x_357_transpose_x_3"), val = bool(false)];
            bool x_357_transpose_y_3 = const()[name = string("x_357_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_357_cast_fp16 = matmul(transpose_x = x_357_transpose_x_3, transpose_y = x_357_transpose_y_3, x = q_v_65_cast_fp16, y = p_head_131_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<int32, [8]> x_359_pad_1 = const()[name = string("x_359_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_359_mode_1 = const()[name = string("x_359_mode_1"), val = string("constant")];
            fp16 const_1525_to_fp16 = const()[name = string("const_1525_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_359_cast_fp16 = pad(constant_val = const_1525_to_fp16, mode = x_359_mode_1, pad = x_359_pad_1, x = x_357_cast_fp16)[name = string("x_359_cast_fp16")];
            tensor<int32, [4]> var_5532 = const()[name = string("op_5532"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_361_cast_fp16 = reshape(shape = var_5532, x = x_359_cast_fp16)[name = string("x_361_cast_fp16")];
            tensor<int32, [4]> var_5536_begin_1 = const()[name = string("op_5536_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5536_end_1 = const()[name = string("op_5536_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_5536_end_mask_1 = const()[name = string("op_5536_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_5536_cast_fp16 = slice_by_index(begin = var_5536_begin_1, end = var_5536_end_1, end_mask = var_5536_end_mask_1, x = x_361_cast_fp16)[name = string("op_5536_cast_fp16")];
            tensor<int32, [4]> var_5538 = const()[name = string("op_5538"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_5539_cast_fp16 = reshape(shape = var_5538, x = var_5536_cast_fp16)[name = string("op_5539_cast_fp16")];
            tensor<int32, [4]> var_5544_begin_1 = const()[name = string("op_5544_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5544_end_1 = const()[name = string("op_5544_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_5544_end_mask_1 = const()[name = string("op_5544_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_5544_cast_fp16 = slice_by_index(begin = var_5544_begin_1, end = var_5544_end_1, end_mask = var_5544_end_mask_1, x = var_5539_cast_fp16)[name = string("op_5544_cast_fp16")];
            fp16 var_5545_to_fp16 = const()[name = string("op_5545_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_65_cast_fp16 = mul(x = var_5544_cast_fp16, y = var_5545_to_fp16)[name = string("scores_pos_65_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_65_cast_fp16 = add(x = scores_content_65_cast_fp16, y = scores_pos_65_cast_fp16)[name = string("logits_65_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_5548_cast_fp16 = softmax(axis = var_5273, x = logits_65_cast_fp16)[name = string("op_5548_cast_fp16")];
            bool var_5550_transpose_x_1 = const()[name = string("op_5550_transpose_x_1"), val = bool(false)];
            bool var_5550_transpose_y_1 = const()[name = string("op_5550_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_5550_cast_fp16 = matmul(transpose_x = var_5550_transpose_x_1, transpose_y = var_5550_transpose_y_1, x = var_5548_cast_fp16, y = v_head_131_cast_fp16)[name = string("op_5550_cast_fp16")];
            tensor<int32, [1]> var_5551_axes_1 = const()[name = string("op_5551_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_5551_cast_fp16 = squeeze(axes = var_5551_axes_1, x = var_5550_cast_fp16)[name = string("op_5551_cast_fp16")];
            string dense_output_345_pad_type_1 = const()[name = string("dense_output_345_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_345_strides_1 = const()[name = string("dense_output_345_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_345_pad_1 = const()[name = string("dense_output_345_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_345_dilations_1 = const()[name = string("dense_output_345_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_345_groups_1 = const()[name = string("dense_output_345_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133291328))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133422464))))[name = string("layers_4_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_345_cast_fp16 = conv(dilations = dense_output_345_dilations_1, groups = dense_output_345_groups_1, pad = dense_output_345_pad_1, pad_type = dense_output_345_pad_type_1, strides = dense_output_345_strides_1, weight = layers_4_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_345_cast_fp16")];
            string sparse_output_345_pad_type_1 = const()[name = string("sparse_output_345_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_345_strides_1 = const()[name = string("sparse_output_345_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_345_pad_1 = const()[name = string("sparse_output_345_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_345_dilations_1 = const()[name = string("sparse_output_345_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_345_groups_1 = const()[name = string("sparse_output_345_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133425728))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133423040))))[name = string("layers_4_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_345_cast_fp16 = conv(dilations = sparse_output_345_dilations_1, groups = sparse_output_345_groups_1, pad = sparse_output_345_pad_1, pad_type = sparse_output_345_pad_type_1, strides = sparse_output_345_strides_1, weight = layers_4_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_345_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5566_cast_fp16 = add(x = dense_output_345_cast_fp16, y = sparse_output_345_cast_fp16)[name = string("op_5566_cast_fp16")];
            tensor<int32, [4]> var_5567 = const()[name = string("op_5567"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5569 = const()[name = string("op_5569"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5568_cast_fp16 = transpose(perm = var_5567, x = var_5566_cast_fp16)[name = string("transpose_812")];
            tensor<fp16, [1, 51, 128]> q_head_67_cast_fp16 = reshape(shape = var_5569, x = var_5568_cast_fp16)[name = string("q_head_67_cast_fp16")];
            string dense_output_347_pad_type_1 = const()[name = string("dense_output_347_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_347_strides_1 = const()[name = string("dense_output_347_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_347_pad_1 = const()[name = string("dense_output_347_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_347_dilations_1 = const()[name = string("dense_output_347_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_347_groups_1 = const()[name = string("dense_output_347_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133442176))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133573312))))[name = string("layers_4_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_347_cast_fp16 = conv(dilations = dense_output_347_dilations_1, groups = dense_output_347_groups_1, pad = dense_output_347_pad_1, pad_type = dense_output_347_pad_type_1, strides = dense_output_347_strides_1, weight = layers_4_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_347_cast_fp16")];
            string sparse_output_347_pad_type_1 = const()[name = string("sparse_output_347_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_347_strides_1 = const()[name = string("sparse_output_347_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_347_pad_1 = const()[name = string("sparse_output_347_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_347_dilations_1 = const()[name = string("sparse_output_347_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_347_groups_1 = const()[name = string("sparse_output_347_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133576576))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133573888))))[name = string("layers_4_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_347_cast_fp16 = conv(dilations = sparse_output_347_dilations_1, groups = sparse_output_347_groups_1, pad = sparse_output_347_pad_1, pad_type = sparse_output_347_pad_type_1, strides = sparse_output_347_strides_1, weight = layers_4_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_347_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5585_cast_fp16 = add(x = dense_output_347_cast_fp16, y = sparse_output_347_cast_fp16)[name = string("op_5585_cast_fp16")];
            tensor<int32, [4]> var_5586 = const()[name = string("op_5586"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5588 = const()[name = string("op_5588"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5587_cast_fp16 = transpose(perm = var_5586, x = var_5585_cast_fp16)[name = string("transpose_811")];
            tensor<fp16, [1, 51, 128]> k_head_133_cast_fp16 = reshape(shape = var_5588, x = var_5587_cast_fp16)[name = string("k_head_133_cast_fp16")];
            string dense_output_349_pad_type_1 = const()[name = string("dense_output_349_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_349_strides_1 = const()[name = string("dense_output_349_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_349_pad_1 = const()[name = string("dense_output_349_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_349_dilations_1 = const()[name = string("dense_output_349_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_349_groups_1 = const()[name = string("dense_output_349_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133593024))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133724160))))[name = string("layers_4_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_349_cast_fp16 = conv(dilations = dense_output_349_dilations_1, groups = dense_output_349_groups_1, pad = dense_output_349_pad_1, pad_type = dense_output_349_pad_type_1, strides = dense_output_349_strides_1, weight = layers_4_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_349_cast_fp16")];
            string sparse_output_349_pad_type_1 = const()[name = string("sparse_output_349_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_349_strides_1 = const()[name = string("sparse_output_349_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_349_pad_1 = const()[name = string("sparse_output_349_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_349_dilations_1 = const()[name = string("sparse_output_349_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_349_groups_1 = const()[name = string("sparse_output_349_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133727424))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133724736))))[name = string("layers_4_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_349_cast_fp16 = conv(dilations = sparse_output_349_dilations_1, groups = sparse_output_349_groups_1, pad = sparse_output_349_pad_1, pad_type = sparse_output_349_pad_type_1, strides = sparse_output_349_strides_1, weight = layers_4_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_349_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5604_cast_fp16 = add(x = dense_output_349_cast_fp16, y = sparse_output_349_cast_fp16)[name = string("op_5604_cast_fp16")];
            tensor<int32, [4]> var_5605 = const()[name = string("op_5605"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5607 = const()[name = string("op_5607"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5606_cast_fp16 = transpose(perm = var_5605, x = var_5604_cast_fp16)[name = string("transpose_810")];
            tensor<fp16, [1, 51, 128]> v_head_133_cast_fp16 = reshape(shape = var_5607, x = var_5606_cast_fp16)[name = string("v_head_133_cast_fp16")];
            string dense_output_351_pad_type_1 = const()[name = string("dense_output_351_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_351_strides_1 = const()[name = string("dense_output_351_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_351_pad_1 = const()[name = string("dense_output_351_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_351_dilations_1 = const()[name = string("dense_output_351_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_351_groups_1 = const()[name = string("dense_output_351_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133743872))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133875008))))[name = string("layers_4_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_351_cast_fp16 = conv(dilations = dense_output_351_dilations_1, groups = dense_output_351_groups_1, pad = dense_output_351_pad_1, pad_type = dense_output_351_pad_type_1, strides = dense_output_351_strides_1, weight = layers_4_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_351_cast_fp16")];
            string sparse_output_351_pad_type_1 = const()[name = string("sparse_output_351_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_351_strides_1 = const()[name = string("sparse_output_351_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_351_pad_1 = const()[name = string("sparse_output_351_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_351_dilations_1 = const()[name = string("sparse_output_351_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_351_groups_1 = const()[name = string("sparse_output_351_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133878272))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133875584))))[name = string("layers_4_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_351_cast_fp16 = conv(dilations = sparse_output_351_dilations_1, groups = sparse_output_351_groups_1, pad = sparse_output_351_pad_1, pad_type = sparse_output_351_pad_type_1, strides = sparse_output_351_strides_1, weight = layers_4_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_351_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_5623_cast_fp16 = add(x = dense_output_351_cast_fp16, y = sparse_output_351_cast_fp16)[name = string("op_5623_cast_fp16")];
            tensor<int32, [4]> var_5624 = const()[name = string("op_5624"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5626 = const()[name = string("op_5626"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_5625_cast_fp16 = transpose(perm = var_5624, x = var_5623_cast_fp16)[name = string("transpose_809")];
            tensor<fp16, [1, 101, 128]> p_head_133_cast_fp16 = reshape(shape = var_5626, x = var_5625_cast_fp16)[name = string("p_head_133_cast_fp16")];
            tensor<fp16, [128]> var_5628_to_fp16 = const()[name = string("op_5628_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133894720)))];
            tensor<fp16, [1, 51, 128]> var_5629_cast_fp16 = add(x = q_head_67_cast_fp16, y = var_5628_to_fp16)[name = string("op_5629_cast_fp16")];
            tensor<int32, [1]> q_u_67_axes_1 = const()[name = string("q_u_67_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_67_cast_fp16 = expand_dims(axes = q_u_67_axes_1, x = var_5629_cast_fp16)[name = string("q_u_67_cast_fp16")];
            tensor<fp16, [128]> var_5631_to_fp16 = const()[name = string("op_5631_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133895040)))];
            tensor<fp16, [1, 51, 128]> var_5632_cast_fp16 = add(x = q_head_67_cast_fp16, y = var_5631_to_fp16)[name = string("op_5632_cast_fp16")];
            tensor<int32, [1]> q_v_67_axes_1 = const()[name = string("q_v_67_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_67_cast_fp16 = expand_dims(axes = q_v_67_axes_1, x = var_5632_cast_fp16)[name = string("q_v_67_cast_fp16")];
            tensor<int32, [1]> k_head_135_axes_1 = const()[name = string("k_head_135_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_135_cast_fp16 = expand_dims(axes = k_head_135_axes_1, x = k_head_133_cast_fp16)[name = string("k_head_135_cast_fp16")];
            tensor<int32, [1]> v_head_135_axes_1 = const()[name = string("v_head_135_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_135_cast_fp16 = expand_dims(axes = v_head_135_axes_1, x = v_head_133_cast_fp16)[name = string("v_head_135_cast_fp16")];
            tensor<int32, [1]> p_head_135_axes_1 = const()[name = string("p_head_135_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_135_cast_fp16 = expand_dims(axes = p_head_135_axes_1, x = p_head_133_cast_fp16)[name = string("p_head_135_cast_fp16")];
            bool var_5638_transpose_x_3 = const()[name = string("op_5638_transpose_x_3"), val = bool(false)];
            bool var_5638_transpose_y_3 = const()[name = string("op_5638_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_5638_cast_fp16 = matmul(transpose_x = var_5638_transpose_x_3, transpose_y = var_5638_transpose_y_3, x = q_u_67_cast_fp16, y = k_head_135_cast_fp16)[name = string("op_5638_cast_fp16")];
            fp16 var_5639_to_fp16 = const()[name = string("op_5639_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_67_cast_fp16 = mul(x = var_5638_cast_fp16, y = var_5639_to_fp16)[name = string("scores_content_67_cast_fp16")];
            bool x_365_transpose_x_3 = const()[name = string("x_365_transpose_x_3"), val = bool(false)];
            bool x_365_transpose_y_3 = const()[name = string("x_365_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_365_cast_fp16 = matmul(transpose_x = x_365_transpose_x_3, transpose_y = x_365_transpose_y_3, x = q_v_67_cast_fp16, y = p_head_135_cast_fp16)[name = string("x_365_cast_fp16")];
            tensor<int32, [8]> x_367_pad_1 = const()[name = string("x_367_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_367_mode_1 = const()[name = string("x_367_mode_1"), val = string("constant")];
            fp16 const_1531_to_fp16 = const()[name = string("const_1531_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_367_cast_fp16 = pad(constant_val = const_1531_to_fp16, mode = x_367_mode_1, pad = x_367_pad_1, x = x_365_cast_fp16)[name = string("x_367_cast_fp16")];
            tensor<int32, [4]> var_5653 = const()[name = string("op_5653"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_369_cast_fp16 = reshape(shape = var_5653, x = x_367_cast_fp16)[name = string("x_369_cast_fp16")];
            tensor<int32, [4]> var_5657_begin_1 = const()[name = string("op_5657_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5657_end_1 = const()[name = string("op_5657_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_5657_end_mask_1 = const()[name = string("op_5657_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_5657_cast_fp16 = slice_by_index(begin = var_5657_begin_1, end = var_5657_end_1, end_mask = var_5657_end_mask_1, x = x_369_cast_fp16)[name = string("op_5657_cast_fp16")];
            tensor<int32, [4]> var_5659 = const()[name = string("op_5659"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_5660_cast_fp16 = reshape(shape = var_5659, x = var_5657_cast_fp16)[name = string("op_5660_cast_fp16")];
            tensor<int32, [4]> var_5665_begin_1 = const()[name = string("op_5665_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5665_end_1 = const()[name = string("op_5665_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_5665_end_mask_1 = const()[name = string("op_5665_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_5665_cast_fp16 = slice_by_index(begin = var_5665_begin_1, end = var_5665_end_1, end_mask = var_5665_end_mask_1, x = var_5660_cast_fp16)[name = string("op_5665_cast_fp16")];
            fp16 var_5666_to_fp16 = const()[name = string("op_5666_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_67_cast_fp16 = mul(x = var_5665_cast_fp16, y = var_5666_to_fp16)[name = string("scores_pos_67_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_67_cast_fp16 = add(x = scores_content_67_cast_fp16, y = scores_pos_67_cast_fp16)[name = string("logits_67_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_5669_cast_fp16 = softmax(axis = var_5273, x = logits_67_cast_fp16)[name = string("op_5669_cast_fp16")];
            bool var_5671_transpose_x_1 = const()[name = string("op_5671_transpose_x_1"), val = bool(false)];
            bool var_5671_transpose_y_1 = const()[name = string("op_5671_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_5671_cast_fp16 = matmul(transpose_x = var_5671_transpose_x_1, transpose_y = var_5671_transpose_y_1, x = var_5669_cast_fp16, y = v_head_135_cast_fp16)[name = string("op_5671_cast_fp16")];
            tensor<int32, [1]> var_5672_axes_1 = const()[name = string("op_5672_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_5672_cast_fp16 = squeeze(axes = var_5672_axes_1, x = var_5671_cast_fp16)[name = string("op_5672_cast_fp16")];
            string dense_output_353_pad_type_1 = const()[name = string("dense_output_353_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_353_strides_1 = const()[name = string("dense_output_353_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_353_pad_1 = const()[name = string("dense_output_353_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_353_dilations_1 = const()[name = string("dense_output_353_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_353_groups_1 = const()[name = string("dense_output_353_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133895360))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134026496))))[name = string("layers_4_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_353_cast_fp16 = conv(dilations = dense_output_353_dilations_1, groups = dense_output_353_groups_1, pad = dense_output_353_pad_1, pad_type = dense_output_353_pad_type_1, strides = dense_output_353_strides_1, weight = layers_4_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_353_cast_fp16")];
            string sparse_output_353_pad_type_1 = const()[name = string("sparse_output_353_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_353_strides_1 = const()[name = string("sparse_output_353_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_353_pad_1 = const()[name = string("sparse_output_353_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_353_dilations_1 = const()[name = string("sparse_output_353_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_353_groups_1 = const()[name = string("sparse_output_353_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134029760))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134027072))))[name = string("layers_4_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_353_cast_fp16 = conv(dilations = sparse_output_353_dilations_1, groups = sparse_output_353_groups_1, pad = sparse_output_353_pad_1, pad_type = sparse_output_353_pad_type_1, strides = sparse_output_353_strides_1, weight = layers_4_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_353_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5687_cast_fp16 = add(x = dense_output_353_cast_fp16, y = sparse_output_353_cast_fp16)[name = string("op_5687_cast_fp16")];
            tensor<int32, [4]> var_5688 = const()[name = string("op_5688"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5690 = const()[name = string("op_5690"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5689_cast_fp16 = transpose(perm = var_5688, x = var_5687_cast_fp16)[name = string("transpose_808")];
            tensor<fp16, [1, 51, 128]> q_head_69_cast_fp16 = reshape(shape = var_5690, x = var_5689_cast_fp16)[name = string("q_head_69_cast_fp16")];
            string dense_output_355_pad_type_1 = const()[name = string("dense_output_355_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_355_strides_1 = const()[name = string("dense_output_355_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_355_pad_1 = const()[name = string("dense_output_355_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_355_dilations_1 = const()[name = string("dense_output_355_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_355_groups_1 = const()[name = string("dense_output_355_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134046208))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134177344))))[name = string("layers_4_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_355_cast_fp16 = conv(dilations = dense_output_355_dilations_1, groups = dense_output_355_groups_1, pad = dense_output_355_pad_1, pad_type = dense_output_355_pad_type_1, strides = dense_output_355_strides_1, weight = layers_4_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_355_cast_fp16")];
            string sparse_output_355_pad_type_1 = const()[name = string("sparse_output_355_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_355_strides_1 = const()[name = string("sparse_output_355_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_355_pad_1 = const()[name = string("sparse_output_355_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_355_dilations_1 = const()[name = string("sparse_output_355_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_355_groups_1 = const()[name = string("sparse_output_355_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134180608))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134177920))))[name = string("layers_4_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_355_cast_fp16 = conv(dilations = sparse_output_355_dilations_1, groups = sparse_output_355_groups_1, pad = sparse_output_355_pad_1, pad_type = sparse_output_355_pad_type_1, strides = sparse_output_355_strides_1, weight = layers_4_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_355_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5706_cast_fp16 = add(x = dense_output_355_cast_fp16, y = sparse_output_355_cast_fp16)[name = string("op_5706_cast_fp16")];
            tensor<int32, [4]> var_5707 = const()[name = string("op_5707"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5709 = const()[name = string("op_5709"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5708_cast_fp16 = transpose(perm = var_5707, x = var_5706_cast_fp16)[name = string("transpose_807")];
            tensor<fp16, [1, 51, 128]> k_head_137_cast_fp16 = reshape(shape = var_5709, x = var_5708_cast_fp16)[name = string("k_head_137_cast_fp16")];
            string dense_output_357_pad_type_1 = const()[name = string("dense_output_357_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_357_strides_1 = const()[name = string("dense_output_357_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_357_pad_1 = const()[name = string("dense_output_357_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_357_dilations_1 = const()[name = string("dense_output_357_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_357_groups_1 = const()[name = string("dense_output_357_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134197056))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134328192))))[name = string("layers_4_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_357_cast_fp16 = conv(dilations = dense_output_357_dilations_1, groups = dense_output_357_groups_1, pad = dense_output_357_pad_1, pad_type = dense_output_357_pad_type_1, strides = dense_output_357_strides_1, weight = layers_4_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_357_cast_fp16")];
            string sparse_output_357_pad_type_1 = const()[name = string("sparse_output_357_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_357_strides_1 = const()[name = string("sparse_output_357_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_357_pad_1 = const()[name = string("sparse_output_357_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_357_dilations_1 = const()[name = string("sparse_output_357_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_357_groups_1 = const()[name = string("sparse_output_357_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134331456))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134328768))))[name = string("layers_4_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_357_cast_fp16 = conv(dilations = sparse_output_357_dilations_1, groups = sparse_output_357_groups_1, pad = sparse_output_357_pad_1, pad_type = sparse_output_357_pad_type_1, strides = sparse_output_357_strides_1, weight = layers_4_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_357_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5725_cast_fp16 = add(x = dense_output_357_cast_fp16, y = sparse_output_357_cast_fp16)[name = string("op_5725_cast_fp16")];
            tensor<int32, [4]> var_5726 = const()[name = string("op_5726"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5728 = const()[name = string("op_5728"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5727_cast_fp16 = transpose(perm = var_5726, x = var_5725_cast_fp16)[name = string("transpose_806")];
            tensor<fp16, [1, 51, 128]> v_head_137_cast_fp16 = reshape(shape = var_5728, x = var_5727_cast_fp16)[name = string("v_head_137_cast_fp16")];
            string dense_output_359_pad_type_1 = const()[name = string("dense_output_359_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_359_strides_1 = const()[name = string("dense_output_359_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_359_pad_1 = const()[name = string("dense_output_359_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_359_dilations_1 = const()[name = string("dense_output_359_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_359_groups_1 = const()[name = string("dense_output_359_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134347904))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134479040))))[name = string("layers_4_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_359_cast_fp16 = conv(dilations = dense_output_359_dilations_1, groups = dense_output_359_groups_1, pad = dense_output_359_pad_1, pad_type = dense_output_359_pad_type_1, strides = dense_output_359_strides_1, weight = layers_4_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_359_cast_fp16")];
            string sparse_output_359_pad_type_1 = const()[name = string("sparse_output_359_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_359_strides_1 = const()[name = string("sparse_output_359_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_359_pad_1 = const()[name = string("sparse_output_359_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_359_dilations_1 = const()[name = string("sparse_output_359_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_359_groups_1 = const()[name = string("sparse_output_359_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134482304))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134479616))))[name = string("layers_4_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_359_cast_fp16 = conv(dilations = sparse_output_359_dilations_1, groups = sparse_output_359_groups_1, pad = sparse_output_359_pad_1, pad_type = sparse_output_359_pad_type_1, strides = sparse_output_359_strides_1, weight = layers_4_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_359_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_5744_cast_fp16 = add(x = dense_output_359_cast_fp16, y = sparse_output_359_cast_fp16)[name = string("op_5744_cast_fp16")];
            tensor<int32, [4]> var_5745 = const()[name = string("op_5745"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5747 = const()[name = string("op_5747"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_5746_cast_fp16 = transpose(perm = var_5745, x = var_5744_cast_fp16)[name = string("transpose_805")];
            tensor<fp16, [1, 101, 128]> p_head_137_cast_fp16 = reshape(shape = var_5747, x = var_5746_cast_fp16)[name = string("p_head_137_cast_fp16")];
            tensor<fp16, [128]> var_5749_to_fp16 = const()[name = string("op_5749_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134498752)))];
            tensor<fp16, [1, 51, 128]> var_5750_cast_fp16 = add(x = q_head_69_cast_fp16, y = var_5749_to_fp16)[name = string("op_5750_cast_fp16")];
            tensor<int32, [1]> q_u_69_axes_1 = const()[name = string("q_u_69_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_69_cast_fp16 = expand_dims(axes = q_u_69_axes_1, x = var_5750_cast_fp16)[name = string("q_u_69_cast_fp16")];
            tensor<fp16, [128]> var_5752_to_fp16 = const()[name = string("op_5752_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134499072)))];
            tensor<fp16, [1, 51, 128]> var_5753_cast_fp16 = add(x = q_head_69_cast_fp16, y = var_5752_to_fp16)[name = string("op_5753_cast_fp16")];
            tensor<int32, [1]> q_v_69_axes_1 = const()[name = string("q_v_69_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_69_cast_fp16 = expand_dims(axes = q_v_69_axes_1, x = var_5753_cast_fp16)[name = string("q_v_69_cast_fp16")];
            tensor<int32, [1]> k_head_139_axes_1 = const()[name = string("k_head_139_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_139_cast_fp16 = expand_dims(axes = k_head_139_axes_1, x = k_head_137_cast_fp16)[name = string("k_head_139_cast_fp16")];
            tensor<int32, [1]> v_head_139_axes_1 = const()[name = string("v_head_139_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_139_cast_fp16 = expand_dims(axes = v_head_139_axes_1, x = v_head_137_cast_fp16)[name = string("v_head_139_cast_fp16")];
            tensor<int32, [1]> p_head_139_axes_1 = const()[name = string("p_head_139_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_139_cast_fp16 = expand_dims(axes = p_head_139_axes_1, x = p_head_137_cast_fp16)[name = string("p_head_139_cast_fp16")];
            bool var_5759_transpose_x_3 = const()[name = string("op_5759_transpose_x_3"), val = bool(false)];
            bool var_5759_transpose_y_3 = const()[name = string("op_5759_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_5759_cast_fp16 = matmul(transpose_x = var_5759_transpose_x_3, transpose_y = var_5759_transpose_y_3, x = q_u_69_cast_fp16, y = k_head_139_cast_fp16)[name = string("op_5759_cast_fp16")];
            fp16 var_5760_to_fp16 = const()[name = string("op_5760_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_69_cast_fp16 = mul(x = var_5759_cast_fp16, y = var_5760_to_fp16)[name = string("scores_content_69_cast_fp16")];
            bool x_373_transpose_x_3 = const()[name = string("x_373_transpose_x_3"), val = bool(false)];
            bool x_373_transpose_y_3 = const()[name = string("x_373_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_373_cast_fp16 = matmul(transpose_x = x_373_transpose_x_3, transpose_y = x_373_transpose_y_3, x = q_v_69_cast_fp16, y = p_head_139_cast_fp16)[name = string("x_373_cast_fp16")];
            tensor<int32, [8]> x_375_pad_1 = const()[name = string("x_375_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_375_mode_1 = const()[name = string("x_375_mode_1"), val = string("constant")];
            fp16 const_1537_to_fp16 = const()[name = string("const_1537_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_375_cast_fp16 = pad(constant_val = const_1537_to_fp16, mode = x_375_mode_1, pad = x_375_pad_1, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")];
            tensor<int32, [4]> var_5774 = const()[name = string("op_5774"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_377_cast_fp16 = reshape(shape = var_5774, x = x_375_cast_fp16)[name = string("x_377_cast_fp16")];
            tensor<int32, [4]> var_5778_begin_1 = const()[name = string("op_5778_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5778_end_1 = const()[name = string("op_5778_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_5778_end_mask_1 = const()[name = string("op_5778_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_5778_cast_fp16 = slice_by_index(begin = var_5778_begin_1, end = var_5778_end_1, end_mask = var_5778_end_mask_1, x = x_377_cast_fp16)[name = string("op_5778_cast_fp16")];
            tensor<int32, [4]> var_5780 = const()[name = string("op_5780"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_5781_cast_fp16 = reshape(shape = var_5780, x = var_5778_cast_fp16)[name = string("op_5781_cast_fp16")];
            tensor<int32, [4]> var_5786_begin_1 = const()[name = string("op_5786_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5786_end_1 = const()[name = string("op_5786_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_5786_end_mask_1 = const()[name = string("op_5786_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_5786_cast_fp16 = slice_by_index(begin = var_5786_begin_1, end = var_5786_end_1, end_mask = var_5786_end_mask_1, x = var_5781_cast_fp16)[name = string("op_5786_cast_fp16")];
            fp16 var_5787_to_fp16 = const()[name = string("op_5787_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_69_cast_fp16 = mul(x = var_5786_cast_fp16, y = var_5787_to_fp16)[name = string("scores_pos_69_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_69_cast_fp16 = add(x = scores_content_69_cast_fp16, y = scores_pos_69_cast_fp16)[name = string("logits_69_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_5790_cast_fp16 = softmax(axis = var_5273, x = logits_69_cast_fp16)[name = string("op_5790_cast_fp16")];
            bool var_5792_transpose_x_1 = const()[name = string("op_5792_transpose_x_1"), val = bool(false)];
            bool var_5792_transpose_y_1 = const()[name = string("op_5792_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_5792_cast_fp16 = matmul(transpose_x = var_5792_transpose_x_1, transpose_y = var_5792_transpose_y_1, x = var_5790_cast_fp16, y = v_head_139_cast_fp16)[name = string("op_5792_cast_fp16")];
            tensor<int32, [1]> var_5793_axes_1 = const()[name = string("op_5793_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_5793_cast_fp16 = squeeze(axes = var_5793_axes_1, x = var_5792_cast_fp16)[name = string("op_5793_cast_fp16")];
            string dense_output_361_pad_type_1 = const()[name = string("dense_output_361_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_361_strides_1 = const()[name = string("dense_output_361_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_361_pad_1 = const()[name = string("dense_output_361_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_361_dilations_1 = const()[name = string("dense_output_361_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_361_groups_1 = const()[name = string("dense_output_361_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134499392))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134630528))))[name = string("layers_4_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_361_cast_fp16 = conv(dilations = dense_output_361_dilations_1, groups = dense_output_361_groups_1, pad = dense_output_361_pad_1, pad_type = dense_output_361_pad_type_1, strides = dense_output_361_strides_1, weight = layers_4_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_361_cast_fp16")];
            string sparse_output_361_pad_type_1 = const()[name = string("sparse_output_361_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_361_strides_1 = const()[name = string("sparse_output_361_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_361_pad_1 = const()[name = string("sparse_output_361_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_361_dilations_1 = const()[name = string("sparse_output_361_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_361_groups_1 = const()[name = string("sparse_output_361_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134633792))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134631104))))[name = string("layers_4_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_361_cast_fp16 = conv(dilations = sparse_output_361_dilations_1, groups = sparse_output_361_groups_1, pad = sparse_output_361_pad_1, pad_type = sparse_output_361_pad_type_1, strides = sparse_output_361_strides_1, weight = layers_4_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_361_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5808_cast_fp16 = add(x = dense_output_361_cast_fp16, y = sparse_output_361_cast_fp16)[name = string("op_5808_cast_fp16")];
            tensor<int32, [4]> var_5809 = const()[name = string("op_5809"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5811 = const()[name = string("op_5811"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5810_cast_fp16 = transpose(perm = var_5809, x = var_5808_cast_fp16)[name = string("transpose_804")];
            tensor<fp16, [1, 51, 128]> q_head_71_cast_fp16 = reshape(shape = var_5811, x = var_5810_cast_fp16)[name = string("q_head_71_cast_fp16")];
            string dense_output_363_pad_type_1 = const()[name = string("dense_output_363_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_363_strides_1 = const()[name = string("dense_output_363_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_363_pad_1 = const()[name = string("dense_output_363_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_363_dilations_1 = const()[name = string("dense_output_363_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_363_groups_1 = const()[name = string("dense_output_363_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134650240))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134781376))))[name = string("layers_4_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_363_cast_fp16 = conv(dilations = dense_output_363_dilations_1, groups = dense_output_363_groups_1, pad = dense_output_363_pad_1, pad_type = dense_output_363_pad_type_1, strides = dense_output_363_strides_1, weight = layers_4_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_363_cast_fp16")];
            string sparse_output_363_pad_type_1 = const()[name = string("sparse_output_363_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_363_strides_1 = const()[name = string("sparse_output_363_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_363_pad_1 = const()[name = string("sparse_output_363_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_363_dilations_1 = const()[name = string("sparse_output_363_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_363_groups_1 = const()[name = string("sparse_output_363_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134784640))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134781952))))[name = string("layers_4_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_363_cast_fp16 = conv(dilations = sparse_output_363_dilations_1, groups = sparse_output_363_groups_1, pad = sparse_output_363_pad_1, pad_type = sparse_output_363_pad_type_1, strides = sparse_output_363_strides_1, weight = layers_4_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_363_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5827_cast_fp16 = add(x = dense_output_363_cast_fp16, y = sparse_output_363_cast_fp16)[name = string("op_5827_cast_fp16")];
            tensor<int32, [4]> var_5828 = const()[name = string("op_5828"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5830 = const()[name = string("op_5830"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5829_cast_fp16 = transpose(perm = var_5828, x = var_5827_cast_fp16)[name = string("transpose_803")];
            tensor<fp16, [1, 51, 128]> k_head_141_cast_fp16 = reshape(shape = var_5830, x = var_5829_cast_fp16)[name = string("k_head_141_cast_fp16")];
            string dense_output_365_pad_type_1 = const()[name = string("dense_output_365_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_365_strides_1 = const()[name = string("dense_output_365_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_365_pad_1 = const()[name = string("dense_output_365_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_365_dilations_1 = const()[name = string("dense_output_365_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_365_groups_1 = const()[name = string("dense_output_365_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134801088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134932224))))[name = string("layers_4_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_365_cast_fp16 = conv(dilations = dense_output_365_dilations_1, groups = dense_output_365_groups_1, pad = dense_output_365_pad_1, pad_type = dense_output_365_pad_type_1, strides = dense_output_365_strides_1, weight = layers_4_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_365_cast_fp16")];
            string sparse_output_365_pad_type_1 = const()[name = string("sparse_output_365_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_365_strides_1 = const()[name = string("sparse_output_365_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_365_pad_1 = const()[name = string("sparse_output_365_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_365_dilations_1 = const()[name = string("sparse_output_365_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_365_groups_1 = const()[name = string("sparse_output_365_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134935488))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134932800))))[name = string("layers_4_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_365_cast_fp16 = conv(dilations = sparse_output_365_dilations_1, groups = sparse_output_365_groups_1, pad = sparse_output_365_pad_1, pad_type = sparse_output_365_pad_type_1, strides = sparse_output_365_strides_1, weight = layers_4_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_365_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5846_cast_fp16 = add(x = dense_output_365_cast_fp16, y = sparse_output_365_cast_fp16)[name = string("op_5846_cast_fp16")];
            tensor<int32, [4]> var_5847 = const()[name = string("op_5847"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5849 = const()[name = string("op_5849"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5848_cast_fp16 = transpose(perm = var_5847, x = var_5846_cast_fp16)[name = string("transpose_802")];
            tensor<fp16, [1, 51, 128]> v_head_141_cast_fp16 = reshape(shape = var_5849, x = var_5848_cast_fp16)[name = string("v_head_141_cast_fp16")];
            string dense_output_367_pad_type_1 = const()[name = string("dense_output_367_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_367_strides_1 = const()[name = string("dense_output_367_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_367_pad_1 = const()[name = string("dense_output_367_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_367_dilations_1 = const()[name = string("dense_output_367_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_367_groups_1 = const()[name = string("dense_output_367_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134951936))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135083072))))[name = string("layers_4_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_367_cast_fp16 = conv(dilations = dense_output_367_dilations_1, groups = dense_output_367_groups_1, pad = dense_output_367_pad_1, pad_type = dense_output_367_pad_type_1, strides = dense_output_367_strides_1, weight = layers_4_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_367_cast_fp16")];
            string sparse_output_367_pad_type_1 = const()[name = string("sparse_output_367_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_367_strides_1 = const()[name = string("sparse_output_367_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_367_pad_1 = const()[name = string("sparse_output_367_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_367_dilations_1 = const()[name = string("sparse_output_367_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_367_groups_1 = const()[name = string("sparse_output_367_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135086336))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135083648))))[name = string("layers_4_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_367_cast_fp16 = conv(dilations = sparse_output_367_dilations_1, groups = sparse_output_367_groups_1, pad = sparse_output_367_pad_1, pad_type = sparse_output_367_pad_type_1, strides = sparse_output_367_strides_1, weight = layers_4_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_367_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_5865_cast_fp16 = add(x = dense_output_367_cast_fp16, y = sparse_output_367_cast_fp16)[name = string("op_5865_cast_fp16")];
            tensor<int32, [4]> var_5866 = const()[name = string("op_5866"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5868 = const()[name = string("op_5868"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_5867_cast_fp16 = transpose(perm = var_5866, x = var_5865_cast_fp16)[name = string("transpose_801")];
            tensor<fp16, [1, 101, 128]> p_head_141_cast_fp16 = reshape(shape = var_5868, x = var_5867_cast_fp16)[name = string("p_head_141_cast_fp16")];
            tensor<fp16, [128]> var_5870_to_fp16 = const()[name = string("op_5870_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135102784)))];
            tensor<fp16, [1, 51, 128]> var_5871_cast_fp16 = add(x = q_head_71_cast_fp16, y = var_5870_to_fp16)[name = string("op_5871_cast_fp16")];
            tensor<int32, [1]> q_u_71_axes_1 = const()[name = string("q_u_71_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_71_cast_fp16 = expand_dims(axes = q_u_71_axes_1, x = var_5871_cast_fp16)[name = string("q_u_71_cast_fp16")];
            tensor<fp16, [128]> var_5873_to_fp16 = const()[name = string("op_5873_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135103104)))];
            tensor<fp16, [1, 51, 128]> var_5874_cast_fp16 = add(x = q_head_71_cast_fp16, y = var_5873_to_fp16)[name = string("op_5874_cast_fp16")];
            tensor<int32, [1]> q_v_71_axes_1 = const()[name = string("q_v_71_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_71_cast_fp16 = expand_dims(axes = q_v_71_axes_1, x = var_5874_cast_fp16)[name = string("q_v_71_cast_fp16")];
            tensor<int32, [1]> k_head_143_axes_1 = const()[name = string("k_head_143_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_143_cast_fp16 = expand_dims(axes = k_head_143_axes_1, x = k_head_141_cast_fp16)[name = string("k_head_143_cast_fp16")];
            tensor<int32, [1]> v_head_143_axes_1 = const()[name = string("v_head_143_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_143_cast_fp16 = expand_dims(axes = v_head_143_axes_1, x = v_head_141_cast_fp16)[name = string("v_head_143_cast_fp16")];
            tensor<int32, [1]> p_head_143_axes_1 = const()[name = string("p_head_143_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_143_cast_fp16 = expand_dims(axes = p_head_143_axes_1, x = p_head_141_cast_fp16)[name = string("p_head_143_cast_fp16")];
            bool var_5880_transpose_x_3 = const()[name = string("op_5880_transpose_x_3"), val = bool(false)];
            bool var_5880_transpose_y_3 = const()[name = string("op_5880_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_5880_cast_fp16 = matmul(transpose_x = var_5880_transpose_x_3, transpose_y = var_5880_transpose_y_3, x = q_u_71_cast_fp16, y = k_head_143_cast_fp16)[name = string("op_5880_cast_fp16")];
            fp16 var_5881_to_fp16 = const()[name = string("op_5881_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_71_cast_fp16 = mul(x = var_5880_cast_fp16, y = var_5881_to_fp16)[name = string("scores_content_71_cast_fp16")];
            bool x_381_transpose_x_3 = const()[name = string("x_381_transpose_x_3"), val = bool(false)];
            bool x_381_transpose_y_3 = const()[name = string("x_381_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_381_cast_fp16 = matmul(transpose_x = x_381_transpose_x_3, transpose_y = x_381_transpose_y_3, x = q_v_71_cast_fp16, y = p_head_143_cast_fp16)[name = string("x_381_cast_fp16")];
            tensor<int32, [8]> x_383_pad_1 = const()[name = string("x_383_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_383_mode_1 = const()[name = string("x_383_mode_1"), val = string("constant")];
            fp16 const_1543_to_fp16 = const()[name = string("const_1543_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_383_cast_fp16 = pad(constant_val = const_1543_to_fp16, mode = x_383_mode_1, pad = x_383_pad_1, x = x_381_cast_fp16)[name = string("x_383_cast_fp16")];
            tensor<int32, [4]> var_5895 = const()[name = string("op_5895"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_385_cast_fp16 = reshape(shape = var_5895, x = x_383_cast_fp16)[name = string("x_385_cast_fp16")];
            tensor<int32, [4]> var_5899_begin_1 = const()[name = string("op_5899_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5899_end_1 = const()[name = string("op_5899_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_5899_end_mask_1 = const()[name = string("op_5899_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_5899_cast_fp16 = slice_by_index(begin = var_5899_begin_1, end = var_5899_end_1, end_mask = var_5899_end_mask_1, x = x_385_cast_fp16)[name = string("op_5899_cast_fp16")];
            tensor<int32, [4]> var_5901 = const()[name = string("op_5901"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_5902_cast_fp16 = reshape(shape = var_5901, x = var_5899_cast_fp16)[name = string("op_5902_cast_fp16")];
            tensor<int32, [4]> var_5907_begin_1 = const()[name = string("op_5907_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5907_end_1 = const()[name = string("op_5907_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_5907_end_mask_1 = const()[name = string("op_5907_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_5907_cast_fp16 = slice_by_index(begin = var_5907_begin_1, end = var_5907_end_1, end_mask = var_5907_end_mask_1, x = var_5902_cast_fp16)[name = string("op_5907_cast_fp16")];
            fp16 var_5908_to_fp16 = const()[name = string("op_5908_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_71_cast_fp16 = mul(x = var_5907_cast_fp16, y = var_5908_to_fp16)[name = string("scores_pos_71_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_71_cast_fp16 = add(x = scores_content_71_cast_fp16, y = scores_pos_71_cast_fp16)[name = string("logits_71_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_5911_cast_fp16 = softmax(axis = var_5273, x = logits_71_cast_fp16)[name = string("op_5911_cast_fp16")];
            bool var_5913_transpose_x_1 = const()[name = string("op_5913_transpose_x_1"), val = bool(false)];
            bool var_5913_transpose_y_1 = const()[name = string("op_5913_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_5913_cast_fp16 = matmul(transpose_x = var_5913_transpose_x_1, transpose_y = var_5913_transpose_y_1, x = var_5911_cast_fp16, y = v_head_143_cast_fp16)[name = string("op_5913_cast_fp16")];
            tensor<int32, [1]> var_5914_axes_1 = const()[name = string("op_5914_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_5914_cast_fp16 = squeeze(axes = var_5914_axes_1, x = var_5913_cast_fp16)[name = string("op_5914_cast_fp16")];
            string dense_output_369_pad_type_1 = const()[name = string("dense_output_369_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_369_strides_1 = const()[name = string("dense_output_369_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_369_pad_1 = const()[name = string("dense_output_369_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_369_dilations_1 = const()[name = string("dense_output_369_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_369_groups_1 = const()[name = string("dense_output_369_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135103424))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135234560))))[name = string("layers_4_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_369_cast_fp16 = conv(dilations = dense_output_369_dilations_1, groups = dense_output_369_groups_1, pad = dense_output_369_pad_1, pad_type = dense_output_369_pad_type_1, strides = dense_output_369_strides_1, weight = layers_4_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_369_cast_fp16")];
            string sparse_output_369_pad_type_1 = const()[name = string("sparse_output_369_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_369_strides_1 = const()[name = string("sparse_output_369_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_369_pad_1 = const()[name = string("sparse_output_369_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_369_dilations_1 = const()[name = string("sparse_output_369_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_369_groups_1 = const()[name = string("sparse_output_369_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135237824))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135235136))))[name = string("layers_4_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_369_cast_fp16 = conv(dilations = sparse_output_369_dilations_1, groups = sparse_output_369_groups_1, pad = sparse_output_369_pad_1, pad_type = sparse_output_369_pad_type_1, strides = sparse_output_369_strides_1, weight = layers_4_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_369_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5929_cast_fp16 = add(x = dense_output_369_cast_fp16, y = sparse_output_369_cast_fp16)[name = string("op_5929_cast_fp16")];
            tensor<int32, [4]> var_5930 = const()[name = string("op_5930"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5932 = const()[name = string("op_5932"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5931_cast_fp16 = transpose(perm = var_5930, x = var_5929_cast_fp16)[name = string("transpose_800")];
            tensor<fp16, [1, 51, 128]> q_head_73_cast_fp16 = reshape(shape = var_5932, x = var_5931_cast_fp16)[name = string("q_head_73_cast_fp16")];
            string dense_output_371_pad_type_1 = const()[name = string("dense_output_371_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_371_strides_1 = const()[name = string("dense_output_371_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_371_pad_1 = const()[name = string("dense_output_371_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_371_dilations_1 = const()[name = string("dense_output_371_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_371_groups_1 = const()[name = string("dense_output_371_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135254272))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135385408))))[name = string("layers_4_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_371_cast_fp16 = conv(dilations = dense_output_371_dilations_1, groups = dense_output_371_groups_1, pad = dense_output_371_pad_1, pad_type = dense_output_371_pad_type_1, strides = dense_output_371_strides_1, weight = layers_4_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_371_cast_fp16")];
            string sparse_output_371_pad_type_1 = const()[name = string("sparse_output_371_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_371_strides_1 = const()[name = string("sparse_output_371_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_371_pad_1 = const()[name = string("sparse_output_371_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_371_dilations_1 = const()[name = string("sparse_output_371_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_371_groups_1 = const()[name = string("sparse_output_371_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135388672))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135385984))))[name = string("layers_4_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_371_cast_fp16 = conv(dilations = sparse_output_371_dilations_1, groups = sparse_output_371_groups_1, pad = sparse_output_371_pad_1, pad_type = sparse_output_371_pad_type_1, strides = sparse_output_371_strides_1, weight = layers_4_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_371_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5948_cast_fp16 = add(x = dense_output_371_cast_fp16, y = sparse_output_371_cast_fp16)[name = string("op_5948_cast_fp16")];
            tensor<int32, [4]> var_5949 = const()[name = string("op_5949"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5951 = const()[name = string("op_5951"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5950_cast_fp16 = transpose(perm = var_5949, x = var_5948_cast_fp16)[name = string("transpose_799")];
            tensor<fp16, [1, 51, 128]> k_head_145_cast_fp16 = reshape(shape = var_5951, x = var_5950_cast_fp16)[name = string("k_head_145_cast_fp16")];
            string dense_output_373_pad_type_1 = const()[name = string("dense_output_373_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_373_strides_1 = const()[name = string("dense_output_373_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_373_pad_1 = const()[name = string("dense_output_373_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_373_dilations_1 = const()[name = string("dense_output_373_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_373_groups_1 = const()[name = string("dense_output_373_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135405120))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135536256))))[name = string("layers_4_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_373_cast_fp16 = conv(dilations = dense_output_373_dilations_1, groups = dense_output_373_groups_1, pad = dense_output_373_pad_1, pad_type = dense_output_373_pad_type_1, strides = dense_output_373_strides_1, weight = layers_4_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_373_cast_fp16")];
            string sparse_output_373_pad_type_1 = const()[name = string("sparse_output_373_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_373_strides_1 = const()[name = string("sparse_output_373_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_373_pad_1 = const()[name = string("sparse_output_373_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_373_dilations_1 = const()[name = string("sparse_output_373_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_373_groups_1 = const()[name = string("sparse_output_373_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135539520))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135536832))))[name = string("layers_4_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_373_cast_fp16 = conv(dilations = sparse_output_373_dilations_1, groups = sparse_output_373_groups_1, pad = sparse_output_373_pad_1, pad_type = sparse_output_373_pad_type_1, strides = sparse_output_373_strides_1, weight = layers_4_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_373_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_5967_cast_fp16 = add(x = dense_output_373_cast_fp16, y = sparse_output_373_cast_fp16)[name = string("op_5967_cast_fp16")];
            tensor<int32, [4]> var_5968 = const()[name = string("op_5968"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5970 = const()[name = string("op_5970"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_5969_cast_fp16 = transpose(perm = var_5968, x = var_5967_cast_fp16)[name = string("transpose_798")];
            tensor<fp16, [1, 51, 128]> v_head_145_cast_fp16 = reshape(shape = var_5970, x = var_5969_cast_fp16)[name = string("v_head_145_cast_fp16")];
            string dense_output_375_pad_type_1 = const()[name = string("dense_output_375_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_375_strides_1 = const()[name = string("dense_output_375_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_375_pad_1 = const()[name = string("dense_output_375_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_375_dilations_1 = const()[name = string("dense_output_375_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_375_groups_1 = const()[name = string("dense_output_375_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135555968))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135687104))))[name = string("layers_4_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_375_cast_fp16 = conv(dilations = dense_output_375_dilations_1, groups = dense_output_375_groups_1, pad = dense_output_375_pad_1, pad_type = dense_output_375_pad_type_1, strides = dense_output_375_strides_1, weight = layers_4_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_375_cast_fp16")];
            string sparse_output_375_pad_type_1 = const()[name = string("sparse_output_375_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_375_strides_1 = const()[name = string("sparse_output_375_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_375_pad_1 = const()[name = string("sparse_output_375_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_375_dilations_1 = const()[name = string("sparse_output_375_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_375_groups_1 = const()[name = string("sparse_output_375_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135690368))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135687680))))[name = string("layers_4_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_375_cast_fp16 = conv(dilations = sparse_output_375_dilations_1, groups = sparse_output_375_groups_1, pad = sparse_output_375_pad_1, pad_type = sparse_output_375_pad_type_1, strides = sparse_output_375_strides_1, weight = layers_4_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_375_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_5986_cast_fp16 = add(x = dense_output_375_cast_fp16, y = sparse_output_375_cast_fp16)[name = string("op_5986_cast_fp16")];
            tensor<int32, [4]> var_5987 = const()[name = string("op_5987"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_5989 = const()[name = string("op_5989"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_5988_cast_fp16 = transpose(perm = var_5987, x = var_5986_cast_fp16)[name = string("transpose_797")];
            tensor<fp16, [1, 101, 128]> p_head_145_cast_fp16 = reshape(shape = var_5989, x = var_5988_cast_fp16)[name = string("p_head_145_cast_fp16")];
            tensor<fp16, [128]> var_5991_to_fp16 = const()[name = string("op_5991_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135706816)))];
            tensor<fp16, [1, 51, 128]> var_5992_cast_fp16 = add(x = q_head_73_cast_fp16, y = var_5991_to_fp16)[name = string("op_5992_cast_fp16")];
            tensor<int32, [1]> q_u_73_axes_1 = const()[name = string("q_u_73_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_73_cast_fp16 = expand_dims(axes = q_u_73_axes_1, x = var_5992_cast_fp16)[name = string("q_u_73_cast_fp16")];
            tensor<fp16, [128]> var_5994_to_fp16 = const()[name = string("op_5994_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135707136)))];
            tensor<fp16, [1, 51, 128]> var_5995_cast_fp16 = add(x = q_head_73_cast_fp16, y = var_5994_to_fp16)[name = string("op_5995_cast_fp16")];
            tensor<int32, [1]> q_v_73_axes_1 = const()[name = string("q_v_73_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_73_cast_fp16 = expand_dims(axes = q_v_73_axes_1, x = var_5995_cast_fp16)[name = string("q_v_73_cast_fp16")];
            tensor<int32, [1]> k_head_147_axes_1 = const()[name = string("k_head_147_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_147_cast_fp16 = expand_dims(axes = k_head_147_axes_1, x = k_head_145_cast_fp16)[name = string("k_head_147_cast_fp16")];
            tensor<int32, [1]> v_head_147_axes_1 = const()[name = string("v_head_147_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_147_cast_fp16 = expand_dims(axes = v_head_147_axes_1, x = v_head_145_cast_fp16)[name = string("v_head_147_cast_fp16")];
            tensor<int32, [1]> p_head_147_axes_1 = const()[name = string("p_head_147_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_147_cast_fp16 = expand_dims(axes = p_head_147_axes_1, x = p_head_145_cast_fp16)[name = string("p_head_147_cast_fp16")];
            bool var_6001_transpose_x_3 = const()[name = string("op_6001_transpose_x_3"), val = bool(false)];
            bool var_6001_transpose_y_3 = const()[name = string("op_6001_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_6001_cast_fp16 = matmul(transpose_x = var_6001_transpose_x_3, transpose_y = var_6001_transpose_y_3, x = q_u_73_cast_fp16, y = k_head_147_cast_fp16)[name = string("op_6001_cast_fp16")];
            fp16 var_6002_to_fp16 = const()[name = string("op_6002_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_73_cast_fp16 = mul(x = var_6001_cast_fp16, y = var_6002_to_fp16)[name = string("scores_content_73_cast_fp16")];
            bool x_389_transpose_x_3 = const()[name = string("x_389_transpose_x_3"), val = bool(false)];
            bool x_389_transpose_y_3 = const()[name = string("x_389_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_389_cast_fp16 = matmul(transpose_x = x_389_transpose_x_3, transpose_y = x_389_transpose_y_3, x = q_v_73_cast_fp16, y = p_head_147_cast_fp16)[name = string("x_389_cast_fp16")];
            tensor<int32, [8]> x_391_pad_1 = const()[name = string("x_391_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_391_mode_1 = const()[name = string("x_391_mode_1"), val = string("constant")];
            fp16 const_1549_to_fp16 = const()[name = string("const_1549_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_391_cast_fp16 = pad(constant_val = const_1549_to_fp16, mode = x_391_mode_1, pad = x_391_pad_1, x = x_389_cast_fp16)[name = string("x_391_cast_fp16")];
            tensor<int32, [4]> var_6016 = const()[name = string("op_6016"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_393_cast_fp16 = reshape(shape = var_6016, x = x_391_cast_fp16)[name = string("x_393_cast_fp16")];
            tensor<int32, [4]> var_6020_begin_1 = const()[name = string("op_6020_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6020_end_1 = const()[name = string("op_6020_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_6020_end_mask_1 = const()[name = string("op_6020_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_6020_cast_fp16 = slice_by_index(begin = var_6020_begin_1, end = var_6020_end_1, end_mask = var_6020_end_mask_1, x = x_393_cast_fp16)[name = string("op_6020_cast_fp16")];
            tensor<int32, [4]> var_6022 = const()[name = string("op_6022"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_6023_cast_fp16 = reshape(shape = var_6022, x = var_6020_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<int32, [4]> var_6028_begin_1 = const()[name = string("op_6028_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6028_end_1 = const()[name = string("op_6028_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_6028_end_mask_1 = const()[name = string("op_6028_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_6028_cast_fp16 = slice_by_index(begin = var_6028_begin_1, end = var_6028_end_1, end_mask = var_6028_end_mask_1, x = var_6023_cast_fp16)[name = string("op_6028_cast_fp16")];
            fp16 var_6029_to_fp16 = const()[name = string("op_6029_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_73_cast_fp16 = mul(x = var_6028_cast_fp16, y = var_6029_to_fp16)[name = string("scores_pos_73_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_73_cast_fp16 = add(x = scores_content_73_cast_fp16, y = scores_pos_73_cast_fp16)[name = string("logits_73_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_6032_cast_fp16 = softmax(axis = var_5273, x = logits_73_cast_fp16)[name = string("op_6032_cast_fp16")];
            bool var_6034_transpose_x_1 = const()[name = string("op_6034_transpose_x_1"), val = bool(false)];
            bool var_6034_transpose_y_1 = const()[name = string("op_6034_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_6034_cast_fp16 = matmul(transpose_x = var_6034_transpose_x_1, transpose_y = var_6034_transpose_y_1, x = var_6032_cast_fp16, y = v_head_147_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [1]> var_6035_axes_1 = const()[name = string("op_6035_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_6035_cast_fp16 = squeeze(axes = var_6035_axes_1, x = var_6034_cast_fp16)[name = string("op_6035_cast_fp16")];
            string dense_output_377_pad_type_1 = const()[name = string("dense_output_377_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_377_strides_1 = const()[name = string("dense_output_377_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_377_pad_1 = const()[name = string("dense_output_377_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_377_dilations_1 = const()[name = string("dense_output_377_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_377_groups_1 = const()[name = string("dense_output_377_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135707456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135838592))))[name = string("layers_4_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_377_cast_fp16 = conv(dilations = dense_output_377_dilations_1, groups = dense_output_377_groups_1, pad = dense_output_377_pad_1, pad_type = dense_output_377_pad_type_1, strides = dense_output_377_strides_1, weight = layers_4_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_377_cast_fp16")];
            string sparse_output_377_pad_type_1 = const()[name = string("sparse_output_377_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_377_strides_1 = const()[name = string("sparse_output_377_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_377_pad_1 = const()[name = string("sparse_output_377_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_377_dilations_1 = const()[name = string("sparse_output_377_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_377_groups_1 = const()[name = string("sparse_output_377_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135841856))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135839168))))[name = string("layers_4_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_377_cast_fp16 = conv(dilations = sparse_output_377_dilations_1, groups = sparse_output_377_groups_1, pad = sparse_output_377_pad_1, pad_type = sparse_output_377_pad_type_1, strides = sparse_output_377_strides_1, weight = layers_4_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_377_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6050_cast_fp16 = add(x = dense_output_377_cast_fp16, y = sparse_output_377_cast_fp16)[name = string("op_6050_cast_fp16")];
            tensor<int32, [4]> var_6051 = const()[name = string("op_6051"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6053 = const()[name = string("op_6053"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6052_cast_fp16 = transpose(perm = var_6051, x = var_6050_cast_fp16)[name = string("transpose_796")];
            tensor<fp16, [1, 51, 128]> q_head_75_cast_fp16 = reshape(shape = var_6053, x = var_6052_cast_fp16)[name = string("q_head_75_cast_fp16")];
            string dense_output_379_pad_type_1 = const()[name = string("dense_output_379_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_379_strides_1 = const()[name = string("dense_output_379_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_379_pad_1 = const()[name = string("dense_output_379_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_379_dilations_1 = const()[name = string("dense_output_379_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_379_groups_1 = const()[name = string("dense_output_379_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135858304))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135989440))))[name = string("layers_4_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_379_cast_fp16 = conv(dilations = dense_output_379_dilations_1, groups = dense_output_379_groups_1, pad = dense_output_379_pad_1, pad_type = dense_output_379_pad_type_1, strides = dense_output_379_strides_1, weight = layers_4_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_379_cast_fp16")];
            string sparse_output_379_pad_type_1 = const()[name = string("sparse_output_379_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_379_strides_1 = const()[name = string("sparse_output_379_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_379_pad_1 = const()[name = string("sparse_output_379_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_379_dilations_1 = const()[name = string("sparse_output_379_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_379_groups_1 = const()[name = string("sparse_output_379_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135992704))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135990016))))[name = string("layers_4_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_379_cast_fp16 = conv(dilations = sparse_output_379_dilations_1, groups = sparse_output_379_groups_1, pad = sparse_output_379_pad_1, pad_type = sparse_output_379_pad_type_1, strides = sparse_output_379_strides_1, weight = layers_4_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_379_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6069_cast_fp16 = add(x = dense_output_379_cast_fp16, y = sparse_output_379_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<int32, [4]> var_6070 = const()[name = string("op_6070"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6072 = const()[name = string("op_6072"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6071_cast_fp16 = transpose(perm = var_6070, x = var_6069_cast_fp16)[name = string("transpose_795")];
            tensor<fp16, [1, 51, 128]> k_head_149_cast_fp16 = reshape(shape = var_6072, x = var_6071_cast_fp16)[name = string("k_head_149_cast_fp16")];
            string dense_output_381_pad_type_1 = const()[name = string("dense_output_381_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_381_strides_1 = const()[name = string("dense_output_381_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_381_pad_1 = const()[name = string("dense_output_381_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_381_dilations_1 = const()[name = string("dense_output_381_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_381_groups_1 = const()[name = string("dense_output_381_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136009152))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136140288))))[name = string("layers_4_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_381_cast_fp16 = conv(dilations = dense_output_381_dilations_1, groups = dense_output_381_groups_1, pad = dense_output_381_pad_1, pad_type = dense_output_381_pad_type_1, strides = dense_output_381_strides_1, weight = layers_4_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_381_cast_fp16")];
            string sparse_output_381_pad_type_1 = const()[name = string("sparse_output_381_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_381_strides_1 = const()[name = string("sparse_output_381_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_381_pad_1 = const()[name = string("sparse_output_381_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_381_dilations_1 = const()[name = string("sparse_output_381_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_381_groups_1 = const()[name = string("sparse_output_381_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136143552))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136140864))))[name = string("layers_4_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_381_cast_fp16 = conv(dilations = sparse_output_381_dilations_1, groups = sparse_output_381_groups_1, pad = sparse_output_381_pad_1, pad_type = sparse_output_381_pad_type_1, strides = sparse_output_381_strides_1, weight = layers_4_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_381_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6088_cast_fp16 = add(x = dense_output_381_cast_fp16, y = sparse_output_381_cast_fp16)[name = string("op_6088_cast_fp16")];
            tensor<int32, [4]> var_6089 = const()[name = string("op_6089"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6091 = const()[name = string("op_6091"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6090_cast_fp16 = transpose(perm = var_6089, x = var_6088_cast_fp16)[name = string("transpose_794")];
            tensor<fp16, [1, 51, 128]> v_head_149_cast_fp16 = reshape(shape = var_6091, x = var_6090_cast_fp16)[name = string("v_head_149_cast_fp16")];
            string dense_output_383_pad_type_1 = const()[name = string("dense_output_383_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_383_strides_1 = const()[name = string("dense_output_383_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_383_pad_1 = const()[name = string("dense_output_383_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_383_dilations_1 = const()[name = string("dense_output_383_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_383_groups_1 = const()[name = string("dense_output_383_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136160000))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136291136))))[name = string("layers_4_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_383_cast_fp16 = conv(dilations = dense_output_383_dilations_1, groups = dense_output_383_groups_1, pad = dense_output_383_pad_1, pad_type = dense_output_383_pad_type_1, strides = dense_output_383_strides_1, weight = layers_4_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_383_cast_fp16")];
            string sparse_output_383_pad_type_1 = const()[name = string("sparse_output_383_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_383_strides_1 = const()[name = string("sparse_output_383_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_383_pad_1 = const()[name = string("sparse_output_383_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_383_dilations_1 = const()[name = string("sparse_output_383_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_383_groups_1 = const()[name = string("sparse_output_383_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136294400))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136291712))))[name = string("layers_4_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_383_cast_fp16 = conv(dilations = sparse_output_383_dilations_1, groups = sparse_output_383_groups_1, pad = sparse_output_383_pad_1, pad_type = sparse_output_383_pad_type_1, strides = sparse_output_383_strides_1, weight = layers_4_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_383_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_6107_cast_fp16 = add(x = dense_output_383_cast_fp16, y = sparse_output_383_cast_fp16)[name = string("op_6107_cast_fp16")];
            tensor<int32, [4]> var_6108 = const()[name = string("op_6108"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6110 = const()[name = string("op_6110"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_6109_cast_fp16 = transpose(perm = var_6108, x = var_6107_cast_fp16)[name = string("transpose_793")];
            tensor<fp16, [1, 101, 128]> p_head_149_cast_fp16 = reshape(shape = var_6110, x = var_6109_cast_fp16)[name = string("p_head_149_cast_fp16")];
            tensor<fp16, [128]> var_6112_to_fp16 = const()[name = string("op_6112_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136310848)))];
            tensor<fp16, [1, 51, 128]> var_6113_cast_fp16 = add(x = q_head_75_cast_fp16, y = var_6112_to_fp16)[name = string("op_6113_cast_fp16")];
            tensor<int32, [1]> q_u_75_axes_1 = const()[name = string("q_u_75_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_75_cast_fp16 = expand_dims(axes = q_u_75_axes_1, x = var_6113_cast_fp16)[name = string("q_u_75_cast_fp16")];
            tensor<fp16, [128]> var_6115_to_fp16 = const()[name = string("op_6115_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136311168)))];
            tensor<fp16, [1, 51, 128]> var_6116_cast_fp16 = add(x = q_head_75_cast_fp16, y = var_6115_to_fp16)[name = string("op_6116_cast_fp16")];
            tensor<int32, [1]> q_v_75_axes_1 = const()[name = string("q_v_75_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_75_cast_fp16 = expand_dims(axes = q_v_75_axes_1, x = var_6116_cast_fp16)[name = string("q_v_75_cast_fp16")];
            tensor<int32, [1]> k_head_151_axes_1 = const()[name = string("k_head_151_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_151_cast_fp16 = expand_dims(axes = k_head_151_axes_1, x = k_head_149_cast_fp16)[name = string("k_head_151_cast_fp16")];
            tensor<int32, [1]> v_head_151_axes_1 = const()[name = string("v_head_151_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_151_cast_fp16 = expand_dims(axes = v_head_151_axes_1, x = v_head_149_cast_fp16)[name = string("v_head_151_cast_fp16")];
            tensor<int32, [1]> p_head_151_axes_1 = const()[name = string("p_head_151_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_151_cast_fp16 = expand_dims(axes = p_head_151_axes_1, x = p_head_149_cast_fp16)[name = string("p_head_151_cast_fp16")];
            bool var_6122_transpose_x_3 = const()[name = string("op_6122_transpose_x_3"), val = bool(false)];
            bool var_6122_transpose_y_3 = const()[name = string("op_6122_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_6122_cast_fp16 = matmul(transpose_x = var_6122_transpose_x_3, transpose_y = var_6122_transpose_y_3, x = q_u_75_cast_fp16, y = k_head_151_cast_fp16)[name = string("op_6122_cast_fp16")];
            fp16 var_6123_to_fp16 = const()[name = string("op_6123_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_75_cast_fp16 = mul(x = var_6122_cast_fp16, y = var_6123_to_fp16)[name = string("scores_content_75_cast_fp16")];
            bool x_397_transpose_x_3 = const()[name = string("x_397_transpose_x_3"), val = bool(false)];
            bool x_397_transpose_y_3 = const()[name = string("x_397_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_3, transpose_y = x_397_transpose_y_3, x = q_v_75_cast_fp16, y = p_head_151_cast_fp16)[name = string("x_397_cast_fp16")];
            tensor<int32, [8]> x_399_pad_1 = const()[name = string("x_399_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_399_mode_1 = const()[name = string("x_399_mode_1"), val = string("constant")];
            fp16 const_1555_to_fp16 = const()[name = string("const_1555_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_399_cast_fp16 = pad(constant_val = const_1555_to_fp16, mode = x_399_mode_1, pad = x_399_pad_1, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")];
            tensor<int32, [4]> var_6137 = const()[name = string("op_6137"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_401_cast_fp16 = reshape(shape = var_6137, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")];
            tensor<int32, [4]> var_6141_begin_1 = const()[name = string("op_6141_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6141_end_1 = const()[name = string("op_6141_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_6141_end_mask_1 = const()[name = string("op_6141_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_6141_cast_fp16 = slice_by_index(begin = var_6141_begin_1, end = var_6141_end_1, end_mask = var_6141_end_mask_1, x = x_401_cast_fp16)[name = string("op_6141_cast_fp16")];
            tensor<int32, [4]> var_6143 = const()[name = string("op_6143"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_6144_cast_fp16 = reshape(shape = var_6143, x = var_6141_cast_fp16)[name = string("op_6144_cast_fp16")];
            tensor<int32, [4]> var_6149_begin_1 = const()[name = string("op_6149_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6149_end_1 = const()[name = string("op_6149_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_6149_end_mask_1 = const()[name = string("op_6149_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_6149_cast_fp16 = slice_by_index(begin = var_6149_begin_1, end = var_6149_end_1, end_mask = var_6149_end_mask_1, x = var_6144_cast_fp16)[name = string("op_6149_cast_fp16")];
            fp16 var_6150_to_fp16 = const()[name = string("op_6150_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_75_cast_fp16 = mul(x = var_6149_cast_fp16, y = var_6150_to_fp16)[name = string("scores_pos_75_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_75_cast_fp16 = add(x = scores_content_75_cast_fp16, y = scores_pos_75_cast_fp16)[name = string("logits_75_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_6153_cast_fp16 = softmax(axis = var_5273, x = logits_75_cast_fp16)[name = string("op_6153_cast_fp16")];
            bool var_6155_transpose_x_1 = const()[name = string("op_6155_transpose_x_1"), val = bool(false)];
            bool var_6155_transpose_y_1 = const()[name = string("op_6155_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_6155_cast_fp16 = matmul(transpose_x = var_6155_transpose_x_1, transpose_y = var_6155_transpose_y_1, x = var_6153_cast_fp16, y = v_head_151_cast_fp16)[name = string("op_6155_cast_fp16")];
            tensor<int32, [1]> var_6156_axes_1 = const()[name = string("op_6156_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_6156_cast_fp16 = squeeze(axes = var_6156_axes_1, x = var_6155_cast_fp16)[name = string("op_6156_cast_fp16")];
            string dense_output_385_pad_type_1 = const()[name = string("dense_output_385_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_385_strides_1 = const()[name = string("dense_output_385_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_385_pad_1 = const()[name = string("dense_output_385_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_385_dilations_1 = const()[name = string("dense_output_385_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_385_groups_1 = const()[name = string("dense_output_385_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136311488))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136442624))))[name = string("layers_4_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_385_cast_fp16 = conv(dilations = dense_output_385_dilations_1, groups = dense_output_385_groups_1, pad = dense_output_385_pad_1, pad_type = dense_output_385_pad_type_1, strides = dense_output_385_strides_1, weight = layers_4_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_385_cast_fp16")];
            string sparse_output_385_pad_type_1 = const()[name = string("sparse_output_385_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_385_strides_1 = const()[name = string("sparse_output_385_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_385_pad_1 = const()[name = string("sparse_output_385_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_385_dilations_1 = const()[name = string("sparse_output_385_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_385_groups_1 = const()[name = string("sparse_output_385_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136445888))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136443200))))[name = string("layers_4_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_385_cast_fp16 = conv(dilations = sparse_output_385_dilations_1, groups = sparse_output_385_groups_1, pad = sparse_output_385_pad_1, pad_type = sparse_output_385_pad_type_1, strides = sparse_output_385_strides_1, weight = layers_4_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_385_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6171_cast_fp16 = add(x = dense_output_385_cast_fp16, y = sparse_output_385_cast_fp16)[name = string("op_6171_cast_fp16")];
            tensor<int32, [4]> var_6172 = const()[name = string("op_6172"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6174 = const()[name = string("op_6174"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6173_cast_fp16 = transpose(perm = var_6172, x = var_6171_cast_fp16)[name = string("transpose_792")];
            tensor<fp16, [1, 51, 128]> q_head_77_cast_fp16 = reshape(shape = var_6174, x = var_6173_cast_fp16)[name = string("q_head_77_cast_fp16")];
            string dense_output_387_pad_type_1 = const()[name = string("dense_output_387_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_387_strides_1 = const()[name = string("dense_output_387_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_387_pad_1 = const()[name = string("dense_output_387_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_387_dilations_1 = const()[name = string("dense_output_387_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_387_groups_1 = const()[name = string("dense_output_387_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136462336))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136593472))))[name = string("layers_4_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_387_cast_fp16 = conv(dilations = dense_output_387_dilations_1, groups = dense_output_387_groups_1, pad = dense_output_387_pad_1, pad_type = dense_output_387_pad_type_1, strides = dense_output_387_strides_1, weight = layers_4_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_387_cast_fp16")];
            string sparse_output_387_pad_type_1 = const()[name = string("sparse_output_387_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_387_strides_1 = const()[name = string("sparse_output_387_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_387_pad_1 = const()[name = string("sparse_output_387_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_387_dilations_1 = const()[name = string("sparse_output_387_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_387_groups_1 = const()[name = string("sparse_output_387_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136596736))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136594048))))[name = string("layers_4_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_387_cast_fp16 = conv(dilations = sparse_output_387_dilations_1, groups = sparse_output_387_groups_1, pad = sparse_output_387_pad_1, pad_type = sparse_output_387_pad_type_1, strides = sparse_output_387_strides_1, weight = layers_4_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_387_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6190_cast_fp16 = add(x = dense_output_387_cast_fp16, y = sparse_output_387_cast_fp16)[name = string("op_6190_cast_fp16")];
            tensor<int32, [4]> var_6191 = const()[name = string("op_6191"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6193 = const()[name = string("op_6193"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6192_cast_fp16 = transpose(perm = var_6191, x = var_6190_cast_fp16)[name = string("transpose_791")];
            tensor<fp16, [1, 51, 128]> k_head_153_cast_fp16 = reshape(shape = var_6193, x = var_6192_cast_fp16)[name = string("k_head_153_cast_fp16")];
            string dense_output_389_pad_type_1 = const()[name = string("dense_output_389_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_389_strides_1 = const()[name = string("dense_output_389_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_389_pad_1 = const()[name = string("dense_output_389_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_389_dilations_1 = const()[name = string("dense_output_389_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_389_groups_1 = const()[name = string("dense_output_389_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136613184))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136744320))))[name = string("layers_4_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_389_cast_fp16 = conv(dilations = dense_output_389_dilations_1, groups = dense_output_389_groups_1, pad = dense_output_389_pad_1, pad_type = dense_output_389_pad_type_1, strides = dense_output_389_strides_1, weight = layers_4_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_389_cast_fp16")];
            string sparse_output_389_pad_type_1 = const()[name = string("sparse_output_389_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_389_strides_1 = const()[name = string("sparse_output_389_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_389_pad_1 = const()[name = string("sparse_output_389_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_389_dilations_1 = const()[name = string("sparse_output_389_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_389_groups_1 = const()[name = string("sparse_output_389_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136747584))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136744896))))[name = string("layers_4_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_389_cast_fp16 = conv(dilations = sparse_output_389_dilations_1, groups = sparse_output_389_groups_1, pad = sparse_output_389_pad_1, pad_type = sparse_output_389_pad_type_1, strides = sparse_output_389_strides_1, weight = layers_4_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_389_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6209_cast_fp16 = add(x = dense_output_389_cast_fp16, y = sparse_output_389_cast_fp16)[name = string("op_6209_cast_fp16")];
            tensor<int32, [4]> var_6210 = const()[name = string("op_6210"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6212 = const()[name = string("op_6212"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6211_cast_fp16 = transpose(perm = var_6210, x = var_6209_cast_fp16)[name = string("transpose_790")];
            tensor<fp16, [1, 51, 128]> v_head_153_cast_fp16 = reshape(shape = var_6212, x = var_6211_cast_fp16)[name = string("v_head_153_cast_fp16")];
            string dense_output_391_pad_type_1 = const()[name = string("dense_output_391_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_391_strides_1 = const()[name = string("dense_output_391_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_391_pad_1 = const()[name = string("dense_output_391_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_391_dilations_1 = const()[name = string("dense_output_391_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_391_groups_1 = const()[name = string("dense_output_391_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136764032))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136895168))))[name = string("layers_4_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_391_cast_fp16 = conv(dilations = dense_output_391_dilations_1, groups = dense_output_391_groups_1, pad = dense_output_391_pad_1, pad_type = dense_output_391_pad_type_1, strides = dense_output_391_strides_1, weight = layers_4_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_391_cast_fp16")];
            string sparse_output_391_pad_type_1 = const()[name = string("sparse_output_391_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_391_strides_1 = const()[name = string("sparse_output_391_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_391_pad_1 = const()[name = string("sparse_output_391_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_391_dilations_1 = const()[name = string("sparse_output_391_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_391_groups_1 = const()[name = string("sparse_output_391_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136898432))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136895744))))[name = string("layers_4_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_391_cast_fp16 = conv(dilations = sparse_output_391_dilations_1, groups = sparse_output_391_groups_1, pad = sparse_output_391_pad_1, pad_type = sparse_output_391_pad_type_1, strides = sparse_output_391_strides_1, weight = layers_4_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_391_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_6228_cast_fp16 = add(x = dense_output_391_cast_fp16, y = sparse_output_391_cast_fp16)[name = string("op_6228_cast_fp16")];
            tensor<int32, [4]> var_6229 = const()[name = string("op_6229"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6231 = const()[name = string("op_6231"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_6230_cast_fp16 = transpose(perm = var_6229, x = var_6228_cast_fp16)[name = string("transpose_789")];
            tensor<fp16, [1, 101, 128]> p_head_153_cast_fp16 = reshape(shape = var_6231, x = var_6230_cast_fp16)[name = string("p_head_153_cast_fp16")];
            tensor<fp16, [128]> var_6233_to_fp16 = const()[name = string("op_6233_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136914880)))];
            tensor<fp16, [1, 51, 128]> var_6234_cast_fp16 = add(x = q_head_77_cast_fp16, y = var_6233_to_fp16)[name = string("op_6234_cast_fp16")];
            tensor<int32, [1]> q_u_77_axes_1 = const()[name = string("q_u_77_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_77_cast_fp16 = expand_dims(axes = q_u_77_axes_1, x = var_6234_cast_fp16)[name = string("q_u_77_cast_fp16")];
            tensor<fp16, [128]> var_6236_to_fp16 = const()[name = string("op_6236_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136915200)))];
            tensor<fp16, [1, 51, 128]> var_6237_cast_fp16 = add(x = q_head_77_cast_fp16, y = var_6236_to_fp16)[name = string("op_6237_cast_fp16")];
            tensor<int32, [1]> q_v_77_axes_1 = const()[name = string("q_v_77_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_77_cast_fp16 = expand_dims(axes = q_v_77_axes_1, x = var_6237_cast_fp16)[name = string("q_v_77_cast_fp16")];
            tensor<int32, [1]> k_head_155_axes_1 = const()[name = string("k_head_155_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_155_cast_fp16 = expand_dims(axes = k_head_155_axes_1, x = k_head_153_cast_fp16)[name = string("k_head_155_cast_fp16")];
            tensor<int32, [1]> v_head_155_axes_1 = const()[name = string("v_head_155_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_155_cast_fp16 = expand_dims(axes = v_head_155_axes_1, x = v_head_153_cast_fp16)[name = string("v_head_155_cast_fp16")];
            tensor<int32, [1]> p_head_155_axes_1 = const()[name = string("p_head_155_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_155_cast_fp16 = expand_dims(axes = p_head_155_axes_1, x = p_head_153_cast_fp16)[name = string("p_head_155_cast_fp16")];
            bool var_6243_transpose_x_3 = const()[name = string("op_6243_transpose_x_3"), val = bool(false)];
            bool var_6243_transpose_y_3 = const()[name = string("op_6243_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_6243_cast_fp16 = matmul(transpose_x = var_6243_transpose_x_3, transpose_y = var_6243_transpose_y_3, x = q_u_77_cast_fp16, y = k_head_155_cast_fp16)[name = string("op_6243_cast_fp16")];
            fp16 var_6244_to_fp16 = const()[name = string("op_6244_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_77_cast_fp16 = mul(x = var_6243_cast_fp16, y = var_6244_to_fp16)[name = string("scores_content_77_cast_fp16")];
            bool x_405_transpose_x_3 = const()[name = string("x_405_transpose_x_3"), val = bool(false)];
            bool x_405_transpose_y_3 = const()[name = string("x_405_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_405_cast_fp16 = matmul(transpose_x = x_405_transpose_x_3, transpose_y = x_405_transpose_y_3, x = q_v_77_cast_fp16, y = p_head_155_cast_fp16)[name = string("x_405_cast_fp16")];
            tensor<int32, [8]> x_407_pad_1 = const()[name = string("x_407_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_407_mode_1 = const()[name = string("x_407_mode_1"), val = string("constant")];
            fp16 const_1561_to_fp16 = const()[name = string("const_1561_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_407_cast_fp16 = pad(constant_val = const_1561_to_fp16, mode = x_407_mode_1, pad = x_407_pad_1, x = x_405_cast_fp16)[name = string("x_407_cast_fp16")];
            tensor<int32, [4]> var_6258 = const()[name = string("op_6258"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_409_cast_fp16 = reshape(shape = var_6258, x = x_407_cast_fp16)[name = string("x_409_cast_fp16")];
            tensor<int32, [4]> var_6262_begin_1 = const()[name = string("op_6262_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6262_end_1 = const()[name = string("op_6262_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_6262_end_mask_1 = const()[name = string("op_6262_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_6262_cast_fp16 = slice_by_index(begin = var_6262_begin_1, end = var_6262_end_1, end_mask = var_6262_end_mask_1, x = x_409_cast_fp16)[name = string("op_6262_cast_fp16")];
            tensor<int32, [4]> var_6264 = const()[name = string("op_6264"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_6265_cast_fp16 = reshape(shape = var_6264, x = var_6262_cast_fp16)[name = string("op_6265_cast_fp16")];
            tensor<int32, [4]> var_6270_begin_1 = const()[name = string("op_6270_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6270_end_1 = const()[name = string("op_6270_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_6270_end_mask_1 = const()[name = string("op_6270_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_6270_cast_fp16 = slice_by_index(begin = var_6270_begin_1, end = var_6270_end_1, end_mask = var_6270_end_mask_1, x = var_6265_cast_fp16)[name = string("op_6270_cast_fp16")];
            fp16 var_6271_to_fp16 = const()[name = string("op_6271_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_77_cast_fp16 = mul(x = var_6270_cast_fp16, y = var_6271_to_fp16)[name = string("scores_pos_77_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_77_cast_fp16 = add(x = scores_content_77_cast_fp16, y = scores_pos_77_cast_fp16)[name = string("logits_77_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_6274_cast_fp16 = softmax(axis = var_5273, x = logits_77_cast_fp16)[name = string("op_6274_cast_fp16")];
            bool var_6276_transpose_x_1 = const()[name = string("op_6276_transpose_x_1"), val = bool(false)];
            bool var_6276_transpose_y_1 = const()[name = string("op_6276_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_6276_cast_fp16 = matmul(transpose_x = var_6276_transpose_x_1, transpose_y = var_6276_transpose_y_1, x = var_6274_cast_fp16, y = v_head_155_cast_fp16)[name = string("op_6276_cast_fp16")];
            tensor<int32, [1]> var_6277_axes_1 = const()[name = string("op_6277_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_6277_cast_fp16 = squeeze(axes = var_6277_axes_1, x = var_6276_cast_fp16)[name = string("op_6277_cast_fp16")];
            string dense_output_393_pad_type_1 = const()[name = string("dense_output_393_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_393_strides_1 = const()[name = string("dense_output_393_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_393_pad_1 = const()[name = string("dense_output_393_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_393_dilations_1 = const()[name = string("dense_output_393_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_393_groups_1 = const()[name = string("dense_output_393_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136915520))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137046656))))[name = string("layers_4_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_393_cast_fp16 = conv(dilations = dense_output_393_dilations_1, groups = dense_output_393_groups_1, pad = dense_output_393_pad_1, pad_type = dense_output_393_pad_type_1, strides = dense_output_393_strides_1, weight = layers_4_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_393_cast_fp16")];
            string sparse_output_393_pad_type_1 = const()[name = string("sparse_output_393_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_393_strides_1 = const()[name = string("sparse_output_393_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_393_pad_1 = const()[name = string("sparse_output_393_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_393_dilations_1 = const()[name = string("sparse_output_393_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_393_groups_1 = const()[name = string("sparse_output_393_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137049920))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137047232))))[name = string("layers_4_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_393_cast_fp16 = conv(dilations = sparse_output_393_dilations_1, groups = sparse_output_393_groups_1, pad = sparse_output_393_pad_1, pad_type = sparse_output_393_pad_type_1, strides = sparse_output_393_strides_1, weight = layers_4_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_393_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6292_cast_fp16 = add(x = dense_output_393_cast_fp16, y = sparse_output_393_cast_fp16)[name = string("op_6292_cast_fp16")];
            tensor<int32, [4]> var_6293 = const()[name = string("op_6293"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6295 = const()[name = string("op_6295"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6294_cast_fp16 = transpose(perm = var_6293, x = var_6292_cast_fp16)[name = string("transpose_788")];
            tensor<fp16, [1, 51, 128]> q_head_79_cast_fp16 = reshape(shape = var_6295, x = var_6294_cast_fp16)[name = string("q_head_79_cast_fp16")];
            string dense_output_395_pad_type_1 = const()[name = string("dense_output_395_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_395_strides_1 = const()[name = string("dense_output_395_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_395_pad_1 = const()[name = string("dense_output_395_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_395_dilations_1 = const()[name = string("dense_output_395_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_395_groups_1 = const()[name = string("dense_output_395_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137066368))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137197504))))[name = string("layers_4_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_395_cast_fp16 = conv(dilations = dense_output_395_dilations_1, groups = dense_output_395_groups_1, pad = dense_output_395_pad_1, pad_type = dense_output_395_pad_type_1, strides = dense_output_395_strides_1, weight = layers_4_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_395_cast_fp16")];
            string sparse_output_395_pad_type_1 = const()[name = string("sparse_output_395_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_395_strides_1 = const()[name = string("sparse_output_395_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_395_pad_1 = const()[name = string("sparse_output_395_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_395_dilations_1 = const()[name = string("sparse_output_395_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_395_groups_1 = const()[name = string("sparse_output_395_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137200768))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137198080))))[name = string("layers_4_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_395_cast_fp16 = conv(dilations = sparse_output_395_dilations_1, groups = sparse_output_395_groups_1, pad = sparse_output_395_pad_1, pad_type = sparse_output_395_pad_type_1, strides = sparse_output_395_strides_1, weight = layers_4_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_395_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6311_cast_fp16 = add(x = dense_output_395_cast_fp16, y = sparse_output_395_cast_fp16)[name = string("op_6311_cast_fp16")];
            tensor<int32, [4]> var_6312 = const()[name = string("op_6312"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6314 = const()[name = string("op_6314"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6313_cast_fp16 = transpose(perm = var_6312, x = var_6311_cast_fp16)[name = string("transpose_787")];
            tensor<fp16, [1, 51, 128]> k_head_157_cast_fp16 = reshape(shape = var_6314, x = var_6313_cast_fp16)[name = string("k_head_157_cast_fp16")];
            string dense_output_397_pad_type_1 = const()[name = string("dense_output_397_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_397_strides_1 = const()[name = string("dense_output_397_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_397_pad_1 = const()[name = string("dense_output_397_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_397_dilations_1 = const()[name = string("dense_output_397_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_397_groups_1 = const()[name = string("dense_output_397_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137217216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137348352))))[name = string("layers_4_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_397_cast_fp16 = conv(dilations = dense_output_397_dilations_1, groups = dense_output_397_groups_1, pad = dense_output_397_pad_1, pad_type = dense_output_397_pad_type_1, strides = dense_output_397_strides_1, weight = layers_4_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_213_cast_fp16)[name = string("dense_output_397_cast_fp16")];
            string sparse_output_397_pad_type_1 = const()[name = string("sparse_output_397_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_397_strides_1 = const()[name = string("sparse_output_397_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_397_pad_1 = const()[name = string("sparse_output_397_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_397_dilations_1 = const()[name = string("sparse_output_397_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_397_groups_1 = const()[name = string("sparse_output_397_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137351616))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137348928))))[name = string("layers_4_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_397_cast_fp16 = conv(dilations = sparse_output_397_dilations_1, groups = sparse_output_397_groups_1, pad = sparse_output_397_pad_1, pad_type = sparse_output_397_pad_type_1, strides = sparse_output_397_strides_1, weight = layers_4_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_213_cast_fp16)[name = string("sparse_output_397_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6330_cast_fp16 = add(x = dense_output_397_cast_fp16, y = sparse_output_397_cast_fp16)[name = string("op_6330_cast_fp16")];
            tensor<int32, [4]> var_6331 = const()[name = string("op_6331"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6333 = const()[name = string("op_6333"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6332_cast_fp16 = transpose(perm = var_6331, x = var_6330_cast_fp16)[name = string("transpose_786")];
            tensor<fp16, [1, 51, 128]> v_head_157_cast_fp16 = reshape(shape = var_6333, x = var_6332_cast_fp16)[name = string("v_head_157_cast_fp16")];
            string dense_output_399_pad_type_1 = const()[name = string("dense_output_399_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_399_strides_1 = const()[name = string("dense_output_399_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_399_pad_1 = const()[name = string("dense_output_399_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_399_dilations_1 = const()[name = string("dense_output_399_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_399_groups_1 = const()[name = string("dense_output_399_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137368064))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137499200))))[name = string("layers_4_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_399_cast_fp16 = conv(dilations = dense_output_399_dilations_1, groups = dense_output_399_groups_1, pad = dense_output_399_pad_1, pad_type = dense_output_399_pad_type_1, strides = dense_output_399_strides_1, weight = layers_4_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_399_cast_fp16")];
            string sparse_output_399_pad_type_1 = const()[name = string("sparse_output_399_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_399_strides_1 = const()[name = string("sparse_output_399_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_399_pad_1 = const()[name = string("sparse_output_399_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_399_dilations_1 = const()[name = string("sparse_output_399_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_399_groups_1 = const()[name = string("sparse_output_399_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_4_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137502464))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137499776))))[name = string("layers_4_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_399_cast_fp16 = conv(dilations = sparse_output_399_dilations_1, groups = sparse_output_399_groups_1, pad = sparse_output_399_pad_1, pad_type = sparse_output_399_pad_type_1, strides = sparse_output_399_strides_1, weight = layers_4_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_399_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_6349_cast_fp16 = add(x = dense_output_399_cast_fp16, y = sparse_output_399_cast_fp16)[name = string("op_6349_cast_fp16")];
            tensor<int32, [4]> var_6350 = const()[name = string("op_6350"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6352 = const()[name = string("op_6352"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_6351_cast_fp16 = transpose(perm = var_6350, x = var_6349_cast_fp16)[name = string("transpose_785")];
            tensor<fp16, [1, 101, 128]> p_head_157_cast_fp16 = reshape(shape = var_6352, x = var_6351_cast_fp16)[name = string("p_head_157_cast_fp16")];
            tensor<fp16, [128]> var_6354_to_fp16 = const()[name = string("op_6354_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137518912)))];
            tensor<fp16, [1, 51, 128]> var_6355_cast_fp16 = add(x = q_head_79_cast_fp16, y = var_6354_to_fp16)[name = string("op_6355_cast_fp16")];
            tensor<int32, [1]> q_u_79_axes_1 = const()[name = string("q_u_79_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_79_cast_fp16 = expand_dims(axes = q_u_79_axes_1, x = var_6355_cast_fp16)[name = string("q_u_79_cast_fp16")];
            tensor<fp16, [128]> var_6357_to_fp16 = const()[name = string("op_6357_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137519232)))];
            tensor<fp16, [1, 51, 128]> var_6358_cast_fp16 = add(x = q_head_79_cast_fp16, y = var_6357_to_fp16)[name = string("op_6358_cast_fp16")];
            tensor<int32, [1]> q_v_79_axes_1 = const()[name = string("q_v_79_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_79_cast_fp16 = expand_dims(axes = q_v_79_axes_1, x = var_6358_cast_fp16)[name = string("q_v_79_cast_fp16")];
            tensor<int32, [1]> k_head_159_axes_1 = const()[name = string("k_head_159_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_159_cast_fp16 = expand_dims(axes = k_head_159_axes_1, x = k_head_157_cast_fp16)[name = string("k_head_159_cast_fp16")];
            tensor<int32, [1]> v_head_159_axes_1 = const()[name = string("v_head_159_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_159_cast_fp16 = expand_dims(axes = v_head_159_axes_1, x = v_head_157_cast_fp16)[name = string("v_head_159_cast_fp16")];
            tensor<int32, [1]> p_head_159_axes_1 = const()[name = string("p_head_159_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_159_cast_fp16 = expand_dims(axes = p_head_159_axes_1, x = p_head_157_cast_fp16)[name = string("p_head_159_cast_fp16")];
            bool var_6364_transpose_x_3 = const()[name = string("op_6364_transpose_x_3"), val = bool(false)];
            bool var_6364_transpose_y_3 = const()[name = string("op_6364_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_6364_cast_fp16 = matmul(transpose_x = var_6364_transpose_x_3, transpose_y = var_6364_transpose_y_3, x = q_u_79_cast_fp16, y = k_head_159_cast_fp16)[name = string("op_6364_cast_fp16")];
            fp16 var_6365_to_fp16 = const()[name = string("op_6365_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_79_cast_fp16 = mul(x = var_6364_cast_fp16, y = var_6365_to_fp16)[name = string("scores_content_79_cast_fp16")];
            bool x_413_transpose_x_3 = const()[name = string("x_413_transpose_x_3"), val = bool(false)];
            bool x_413_transpose_y_3 = const()[name = string("x_413_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_413_cast_fp16 = matmul(transpose_x = x_413_transpose_x_3, transpose_y = x_413_transpose_y_3, x = q_v_79_cast_fp16, y = p_head_159_cast_fp16)[name = string("x_413_cast_fp16")];
            tensor<int32, [8]> x_415_pad_1 = const()[name = string("x_415_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_415_mode_1 = const()[name = string("x_415_mode_1"), val = string("constant")];
            fp16 const_1567_to_fp16 = const()[name = string("const_1567_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_415_cast_fp16 = pad(constant_val = const_1567_to_fp16, mode = x_415_mode_1, pad = x_415_pad_1, x = x_413_cast_fp16)[name = string("x_415_cast_fp16")];
            tensor<int32, [4]> var_6379 = const()[name = string("op_6379"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_417_cast_fp16 = reshape(shape = var_6379, x = x_415_cast_fp16)[name = string("x_417_cast_fp16")];
            tensor<int32, [4]> var_6383_begin_1 = const()[name = string("op_6383_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6383_end_1 = const()[name = string("op_6383_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_6383_end_mask_1 = const()[name = string("op_6383_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_6383_cast_fp16 = slice_by_index(begin = var_6383_begin_1, end = var_6383_end_1, end_mask = var_6383_end_mask_1, x = x_417_cast_fp16)[name = string("op_6383_cast_fp16")];
            tensor<int32, [4]> var_6385 = const()[name = string("op_6385"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_6386_cast_fp16 = reshape(shape = var_6385, x = var_6383_cast_fp16)[name = string("op_6386_cast_fp16")];
            tensor<int32, [4]> var_6391_begin_1 = const()[name = string("op_6391_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6391_end_1 = const()[name = string("op_6391_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_6391_end_mask_1 = const()[name = string("op_6391_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_6391_cast_fp16 = slice_by_index(begin = var_6391_begin_1, end = var_6391_end_1, end_mask = var_6391_end_mask_1, x = var_6386_cast_fp16)[name = string("op_6391_cast_fp16")];
            fp16 var_6392_to_fp16 = const()[name = string("op_6392_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_79_cast_fp16 = mul(x = var_6391_cast_fp16, y = var_6392_to_fp16)[name = string("scores_pos_79_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_79_cast_fp16 = add(x = scores_content_79_cast_fp16, y = scores_pos_79_cast_fp16)[name = string("logits_79_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_6395_cast_fp16 = softmax(axis = var_5273, x = logits_79_cast_fp16)[name = string("op_6395_cast_fp16")];
            bool var_6397_transpose_x_1 = const()[name = string("op_6397_transpose_x_1"), val = bool(false)];
            bool var_6397_transpose_y_1 = const()[name = string("op_6397_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_6397_cast_fp16 = matmul(transpose_x = var_6397_transpose_x_1, transpose_y = var_6397_transpose_y_1, x = var_6395_cast_fp16, y = v_head_159_cast_fp16)[name = string("op_6397_cast_fp16")];
            tensor<int32, [1]> o_head_9_axes_1 = const()[name = string("o_head_9_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_9_cast_fp16 = squeeze(axes = o_head_9_axes_1, x = var_6397_cast_fp16)[name = string("o_head_9_cast_fp16")];
            bool out_9_interleave_1 = const()[name = string("out_9_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_9_cast_fp16 = concat(axis = var_5273, interleave = out_9_interleave_1, values = (var_5551_cast_fp16, var_5672_cast_fp16, var_5793_cast_fp16, var_5914_cast_fp16, var_6035_cast_fp16, var_6156_cast_fp16, var_6277_cast_fp16, o_head_9_cast_fp16))[name = string("out_9_cast_fp16")];
            tensor<int32, [3]> var_6401_perm_1 = const()[name = string("op_6401_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_221_axes_1 = const()[name = string("input_221_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_6401_cast_fp16 = transpose(perm = var_6401_perm_1, x = out_9_cast_fp16)[name = string("transpose_784")];
            tensor<fp16, [1, 1024, 51, 1]> input_221_cast_fp16 = expand_dims(axes = input_221_axes_1, x = var_6401_cast_fp16)[name = string("input_221_cast_fp16")];
            string dense_output_401_pad_type_1 = const()[name = string("dense_output_401_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_401_strides_1 = const()[name = string("dense_output_401_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_401_pad_1 = const()[name = string("dense_output_401_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_401_dilations_1 = const()[name = string("dense_output_401_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_401_groups_1 = const()[name = string("dense_output_401_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137519552))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138568192))))[name = string("layers_4_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_401_cast_fp16 = conv(dilations = dense_output_401_dilations_1, groups = dense_output_401_groups_1, pad = dense_output_401_pad_1, pad_type = dense_output_401_pad_type_1, strides = dense_output_401_strides_1, weight = layers_4_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("dense_output_401_cast_fp16")];
            string sparse_output_401_pad_type_1 = const()[name = string("sparse_output_401_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_401_strides_1 = const()[name = string("sparse_output_401_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_401_pad_1 = const()[name = string("sparse_output_401_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_401_dilations_1 = const()[name = string("sparse_output_401_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_401_groups_1 = const()[name = string("sparse_output_401_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138589824))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138568768))))[name = string("layers_4_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_401_cast_fp16 = conv(dilations = sparse_output_401_dilations_1, groups = sparse_output_401_groups_1, pad = sparse_output_401_pad_1, pad_type = sparse_output_401_pad_type_1, strides = sparse_output_401_strides_1, weight = layers_4_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_221_cast_fp16)[name = string("sparse_output_401_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_9_cast_fp16 = add(x = dense_output_401_cast_fp16, y = sparse_output_401_cast_fp16)[name = string("out_conv_9_cast_fp16")];
            tensor<int32, [1]> var_6418_axes_1 = const()[name = string("op_6418_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_6418_cast_fp16 = squeeze(axes = var_6418_axes_1, x = out_conv_9_cast_fp16)[name = string("op_6418_cast_fp16")];
            tensor<int32, [3]> var_6419_perm_1 = const()[name = string("op_6419_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_6419_cast_fp16 = transpose(perm = var_6419_perm_1, x = var_6418_cast_fp16)[name = string("transpose_783")];
            tensor<fp16, [1, 51, 1024]> input_223_cast_fp16 = add(x = input_211_cast_fp16, y = var_6419_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<int32, [1]> x_421_axes_1 = const()[name = string("x_421_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_4_norm_conv_weight_to_fp16 = const()[name = string("layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138720960)))];
            tensor<fp16, [1024]> layers_4_norm_conv_bias_to_fp16 = const()[name = string("layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138723072)))];
            tensor<fp16, [1, 51, 1024]> x_421_cast_fp16 = layer_norm(axes = x_421_axes_1, beta = layers_4_norm_conv_bias_to_fp16, epsilon = var_5288_to_fp16, gamma = layers_4_norm_conv_weight_to_fp16, x = input_223_cast_fp16)[name = string("x_421_cast_fp16")];
            tensor<int32, [3]> var_6429_perm_1 = const()[name = string("op_6429_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_225_axes_1 = const()[name = string("input_225_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_6429_cast_fp16 = transpose(perm = var_6429_perm_1, x = x_421_cast_fp16)[name = string("transpose_782")];
            tensor<fp16, [1, 1024, 51, 1]> input_225_cast_fp16 = expand_dims(axes = input_225_axes_1, x = var_6429_cast_fp16)[name = string("input_225_cast_fp16")];
            string dense_output_403_pad_type_1 = const()[name = string("dense_output_403_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_403_strides_1 = const()[name = string("dense_output_403_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_403_pad_1 = const()[name = string("dense_output_403_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_403_dilations_1 = const()[name = string("dense_output_403_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_403_groups_1 = const()[name = string("dense_output_403_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_4_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138725184))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140822400))))[name = string("layers_4_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_403_cast_fp16 = conv(dilations = dense_output_403_dilations_1, groups = dense_output_403_groups_1, pad = dense_output_403_pad_1, pad_type = dense_output_403_pad_type_1, strides = dense_output_403_strides_1, weight = layers_4_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("dense_output_403_cast_fp16")];
            string sparse_output_403_pad_type_1 = const()[name = string("sparse_output_403_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_403_strides_1 = const()[name = string("sparse_output_403_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_403_pad_1 = const()[name = string("sparse_output_403_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_403_dilations_1 = const()[name = string("sparse_output_403_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_403_groups_1 = const()[name = string("sparse_output_403_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_4_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140865024))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140822976))))[name = string("layers_4_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_403_cast_fp16 = conv(dilations = sparse_output_403_dilations_1, groups = sparse_output_403_groups_1, pad = sparse_output_403_pad_1, pad_type = sparse_output_403_pad_type_1, strides = sparse_output_403_strides_1, weight = layers_4_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_225_cast_fp16)[name = string("sparse_output_403_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_227_cast_fp16 = add(x = dense_output_403_cast_fp16, y = sparse_output_403_cast_fp16)[name = string("input_227_cast_fp16")];
            int32 input_229_split_num_splits_1 = const()[name = string("input_229_split_num_splits_1"), val = int32(2)];
            int32 input_229_split_axis_1 = const()[name = string("input_229_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_229_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_229_split_cast_fp16_1 = split(axis = input_229_split_axis_1, num_splits = input_229_split_num_splits_1, x = input_227_cast_fp16)[name = string("input_229_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_229_split_1_sigmoid_cast_fp16 = sigmoid(x = input_229_split_cast_fp16_1)[name = string("input_229_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_229_cast_fp16 = mul(x = input_229_split_cast_fp16_0, y = input_229_split_1_sigmoid_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<int32, [8]> input_231_pad_1 = const()[name = string("input_231_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_231_mode_1 = const()[name = string("input_231_mode_1"), val = string("constant")];
            fp16 const_1569_to_fp16 = const()[name = string("const_1569_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_231_cast_fp16 = pad(constant_val = const_1569_to_fp16, mode = input_231_mode_1, pad = input_231_pad_1, x = input_229_cast_fp16)[name = string("input_231_cast_fp16")];
            string dense_output_405_pad_type_1 = const()[name = string("dense_output_405_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_405_strides_1 = const()[name = string("dense_output_405_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_405_pad_1 = const()[name = string("dense_output_405_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_405_dilations_1 = const()[name = string("dense_output_405_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_405_groups_1 = const()[name = string("dense_output_405_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_405_cast_fp16 = conv(dilations = dense_output_405_dilations_1, groups = dense_output_405_groups_1, pad = dense_output_405_pad_1, pad_type = dense_output_405_pad_type_1, strides = dense_output_405_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_231_cast_fp16)[name = string("dense_output_405_cast_fp16")];
            string sparse_output_405_pad_type_1 = const()[name = string("sparse_output_405_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_405_strides_1 = const()[name = string("sparse_output_405_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_405_pad_1 = const()[name = string("sparse_output_405_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_405_dilations_1 = const()[name = string("sparse_output_405_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_405_groups_1 = const()[name = string("sparse_output_405_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_4_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141127232))))[name = string("layers_4_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_405_cast_fp16 = conv(dilations = sparse_output_405_dilations_1, groups = sparse_output_405_groups_1, pad = sparse_output_405_pad_1, pad_type = sparse_output_405_pad_type_1, strides = sparse_output_405_strides_1, weight = layers_4_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_231_cast_fp16)[name = string("sparse_output_405_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_233_cast_fp16 = add(x = dense_output_405_cast_fp16, y = sparse_output_405_cast_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [1024]> layers_4_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_4_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141145728)))];
            tensor<fp16, [1024]> layers_4_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_4_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141147840)))];
            tensor<fp16, [1024]> layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141149952)))];
            tensor<fp16, [1024]> layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141152064)))];
            tensor<fp16, [1, 1024, 51, 1]> input_235_cast_fp16 = batch_norm(beta = layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_5288_to_fp16, gamma = layers_4_conv_batch_norm_weight_to_fp16, mean = layers_4_conv_batch_norm_running_mean_to_fp16, variance = layers_4_conv_batch_norm_running_var_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_237_cast_fp16 = silu(x = input_235_cast_fp16)[name = string("input_237_cast_fp16")];
            string dense_output_407_pad_type_1 = const()[name = string("dense_output_407_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_407_strides_1 = const()[name = string("dense_output_407_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_407_pad_1 = const()[name = string("dense_output_407_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_407_dilations_1 = const()[name = string("dense_output_407_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_407_groups_1 = const()[name = string("dense_output_407_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141154176))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142202816))))[name = string("layers_4_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_407_cast_fp16 = conv(dilations = dense_output_407_dilations_1, groups = dense_output_407_groups_1, pad = dense_output_407_pad_1, pad_type = dense_output_407_pad_type_1, strides = dense_output_407_strides_1, weight = layers_4_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_237_cast_fp16)[name = string("dense_output_407_cast_fp16")];
            string sparse_output_407_pad_type_1 = const()[name = string("sparse_output_407_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_407_strides_1 = const()[name = string("sparse_output_407_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_407_pad_1 = const()[name = string("sparse_output_407_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_407_dilations_1 = const()[name = string("sparse_output_407_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_407_groups_1 = const()[name = string("sparse_output_407_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_4_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142224448))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142203392))))[name = string("layers_4_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_407_cast_fp16 = conv(dilations = sparse_output_407_dilations_1, groups = sparse_output_407_groups_1, pad = sparse_output_407_pad_1, pad_type = sparse_output_407_pad_type_1, strides = sparse_output_407_strides_1, weight = layers_4_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_237_cast_fp16)[name = string("sparse_output_407_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_423_cast_fp16 = add(x = dense_output_407_cast_fp16, y = sparse_output_407_cast_fp16)[name = string("x_423_cast_fp16")];
            tensor<int32, [1]> var_6485_axes_1 = const()[name = string("op_6485_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_6485_cast_fp16 = squeeze(axes = var_6485_axes_1, x = x_423_cast_fp16)[name = string("op_6485_cast_fp16")];
            tensor<int32, [3]> var_6486_perm_1 = const()[name = string("op_6486_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_6486_cast_fp16 = transpose(perm = var_6486_perm_1, x = var_6485_cast_fp16)[name = string("transpose_781")];
            tensor<fp16, [1, 51, 1024]> input_239_cast_fp16 = add(x = input_223_cast_fp16, y = var_6486_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<int32, [1]> x_425_axes_1 = const()[name = string("x_425_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142355584)))];
            tensor<fp16, [1024]> layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142357696)))];
            tensor<fp16, [1, 51, 1024]> x_425_cast_fp16 = layer_norm(axes = x_425_axes_1, beta = layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_5288_to_fp16, gamma = layers_4_norm_feed_forward2_weight_to_fp16, x = input_239_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<int32, [4]> var_6496 = const()[name = string("op_6496"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_427_cast_fp16 = reshape(shape = var_6496, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")];
            tensor<int32, [4]> input_241_perm_1 = const()[name = string("input_241_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_409_pad_type_1 = const()[name = string("dense_output_409_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_409_strides_1 = const()[name = string("dense_output_409_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_409_pad_1 = const()[name = string("dense_output_409_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_409_dilations_1 = const()[name = string("dense_output_409_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_409_groups_1 = const()[name = string("dense_output_409_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_4_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142359808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146554176))))[name = string("layers_4_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_241_cast_fp16 = transpose(perm = input_241_perm_1, x = x_427_cast_fp16)[name = string("transpose_780")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_409_cast_fp16 = conv(dilations = dense_output_409_dilations_1, groups = dense_output_409_groups_1, pad = dense_output_409_pad_1, pad_type = dense_output_409_pad_type_1, strides = dense_output_409_strides_1, weight = layers_4_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_241_cast_fp16)[name = string("dense_output_409_cast_fp16")];
            string sparse_output_409_pad_type_1 = const()[name = string("sparse_output_409_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_409_strides_1 = const()[name = string("sparse_output_409_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_409_pad_1 = const()[name = string("sparse_output_409_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_409_dilations_1 = const()[name = string("sparse_output_409_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_409_groups_1 = const()[name = string("sparse_output_409_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_4_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146638720))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146554752))))[name = string("layers_4_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_409_cast_fp16 = conv(dilations = sparse_output_409_dilations_1, groups = sparse_output_409_groups_1, pad = sparse_output_409_pad_1, pad_type = sparse_output_409_pad_type_1, strides = sparse_output_409_strides_1, weight = layers_4_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_241_cast_fp16)[name = string("sparse_output_409_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_243_cast_fp16 = add(x = dense_output_409_cast_fp16, y = sparse_output_409_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_245_cast_fp16 = silu(x = input_243_cast_fp16)[name = string("input_245_cast_fp16")];
            string dense_output_411_pad_type_1 = const()[name = string("dense_output_411_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_411_strides_1 = const()[name = string("dense_output_411_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_411_pad_1 = const()[name = string("dense_output_411_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_411_dilations_1 = const()[name = string("dense_output_411_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_411_groups_1 = const()[name = string("dense_output_411_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_4_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147163072))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151357440))))[name = string("layers_4_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_411_cast_fp16 = conv(dilations = dense_output_411_dilations_1, groups = dense_output_411_groups_1, pad = dense_output_411_pad_1, pad_type = dense_output_411_pad_type_1, strides = dense_output_411_strides_1, weight = layers_4_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_245_cast_fp16)[name = string("dense_output_411_cast_fp16")];
            string sparse_output_411_pad_type_1 = const()[name = string("sparse_output_411_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_411_strides_1 = const()[name = string("sparse_output_411_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_411_pad_1 = const()[name = string("sparse_output_411_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_411_dilations_1 = const()[name = string("sparse_output_411_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_411_groups_1 = const()[name = string("sparse_output_411_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_4_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151441984))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151358016))))[name = string("layers_4_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_411_cast_fp16 = conv(dilations = sparse_output_411_dilations_1, groups = sparse_output_411_groups_1, pad = sparse_output_411_pad_1, pad_type = sparse_output_411_pad_type_1, strides = sparse_output_411_strides_1, weight = layers_4_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_245_cast_fp16)[name = string("sparse_output_411_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_429_cast_fp16 = add(x = dense_output_411_cast_fp16, y = sparse_output_411_cast_fp16)[name = string("x_429_cast_fp16")];
            tensor<int32, [4]> x_431_perm_1 = const()[name = string("x_431_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_6531 = const()[name = string("op_6531"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_431_cast_fp16 = transpose(perm = x_431_perm_1, x = x_429_cast_fp16)[name = string("transpose_779")];
            tensor<fp16, [1, 51, 1024]> var_6532_cast_fp16 = reshape(shape = var_6531, x = x_431_cast_fp16)[name = string("op_6532_cast_fp16")];
            fp16 var_6533_to_fp16 = const()[name = string("op_6533_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_6534_cast_fp16 = mul(x = var_6532_cast_fp16, y = var_6533_to_fp16)[name = string("op_6534_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_247_cast_fp16 = add(x = input_239_cast_fp16, y = var_6534_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<int32, [1]> input_249_axes_1 = const()[name = string("input_249_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_4_norm_out_weight_to_fp16 = const()[name = string("layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151966336)))];
            tensor<fp16, [1024]> layers_4_norm_out_bias_to_fp16 = const()[name = string("layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151968448)))];
            tensor<fp16, [1, 51, 1024]> input_249_cast_fp16 = layer_norm(axes = input_249_axes_1, beta = layers_4_norm_out_bias_to_fp16, epsilon = var_5288_to_fp16, gamma = layers_4_norm_out_weight_to_fp16, x = input_247_cast_fp16)[name = string("input_249_cast_fp16")];
            int32 var_6542 = const()[name = string("op_6542"), val = int32(-1)];
            tensor<int32, [1]> x_433_axes_1 = const()[name = string("x_433_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151970560)))];
            tensor<fp16, [1024]> layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151972672)))];
            fp16 var_6557_to_fp16 = const()[name = string("op_6557_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_1, beta = layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_6557_to_fp16, gamma = layers_5_norm_feed_forward1_weight_to_fp16, x = input_249_cast_fp16)[name = string("x_433_cast_fp16")];
            tensor<int32, [4]> var_6576 = const()[name = string("op_6576"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_435_cast_fp16 = reshape(shape = var_6576, x = x_433_cast_fp16)[name = string("x_435_cast_fp16")];
            tensor<int32, [4]> input_251_perm_1 = const()[name = string("input_251_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_413_pad_type_1 = const()[name = string("dense_output_413_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_413_strides_1 = const()[name = string("dense_output_413_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_413_pad_1 = const()[name = string("dense_output_413_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_413_dilations_1 = const()[name = string("dense_output_413_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_413_groups_1 = const()[name = string("dense_output_413_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_5_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151974784))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156169152))))[name = string("layers_5_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_251_cast_fp16 = transpose(perm = input_251_perm_1, x = x_435_cast_fp16)[name = string("transpose_778")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_413_cast_fp16 = conv(dilations = dense_output_413_dilations_1, groups = dense_output_413_groups_1, pad = dense_output_413_pad_1, pad_type = dense_output_413_pad_type_1, strides = dense_output_413_strides_1, weight = layers_5_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("dense_output_413_cast_fp16")];
            string sparse_output_413_pad_type_1 = const()[name = string("sparse_output_413_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_413_strides_1 = const()[name = string("sparse_output_413_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_413_pad_1 = const()[name = string("sparse_output_413_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_413_dilations_1 = const()[name = string("sparse_output_413_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_413_groups_1 = const()[name = string("sparse_output_413_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_5_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156253696))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156169728))))[name = string("layers_5_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_413_cast_fp16 = conv(dilations = sparse_output_413_dilations_1, groups = sparse_output_413_groups_1, pad = sparse_output_413_pad_1, pad_type = sparse_output_413_pad_type_1, strides = sparse_output_413_strides_1, weight = layers_5_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_251_cast_fp16)[name = string("sparse_output_413_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_253_cast_fp16 = add(x = dense_output_413_cast_fp16, y = sparse_output_413_cast_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_255_cast_fp16 = silu(x = input_253_cast_fp16)[name = string("input_255_cast_fp16")];
            string dense_output_415_pad_type_1 = const()[name = string("dense_output_415_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_415_strides_1 = const()[name = string("dense_output_415_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_415_pad_1 = const()[name = string("dense_output_415_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_415_dilations_1 = const()[name = string("dense_output_415_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_415_groups_1 = const()[name = string("dense_output_415_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_5_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156778048))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160972416))))[name = string("layers_5_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_415_cast_fp16 = conv(dilations = dense_output_415_dilations_1, groups = dense_output_415_groups_1, pad = dense_output_415_pad_1, pad_type = dense_output_415_pad_type_1, strides = dense_output_415_strides_1, weight = layers_5_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_255_cast_fp16)[name = string("dense_output_415_cast_fp16")];
            string sparse_output_415_pad_type_1 = const()[name = string("sparse_output_415_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_415_strides_1 = const()[name = string("sparse_output_415_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_415_pad_1 = const()[name = string("sparse_output_415_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_415_dilations_1 = const()[name = string("sparse_output_415_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_415_groups_1 = const()[name = string("sparse_output_415_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_5_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161056960))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160972992))))[name = string("layers_5_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_415_cast_fp16 = conv(dilations = sparse_output_415_dilations_1, groups = sparse_output_415_groups_1, pad = sparse_output_415_pad_1, pad_type = sparse_output_415_pad_type_1, strides = sparse_output_415_strides_1, weight = layers_5_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_255_cast_fp16)[name = string("sparse_output_415_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_437_cast_fp16 = add(x = dense_output_415_cast_fp16, y = sparse_output_415_cast_fp16)[name = string("x_437_cast_fp16")];
            tensor<int32, [4]> x_439_perm_1 = const()[name = string("x_439_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_6611 = const()[name = string("op_6611"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_439_cast_fp16 = transpose(perm = x_439_perm_1, x = x_437_cast_fp16)[name = string("transpose_777")];
            tensor<fp16, [1, 51, 1024]> var_6612_cast_fp16 = reshape(shape = var_6611, x = x_439_cast_fp16)[name = string("op_6612_cast_fp16")];
            fp16 var_6613_to_fp16 = const()[name = string("op_6613_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_6614_cast_fp16 = mul(x = var_6612_cast_fp16, y = var_6613_to_fp16)[name = string("op_6614_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_257_cast_fp16 = add(x = input_249_cast_fp16, y = var_6614_cast_fp16)[name = string("input_257_cast_fp16")];
            tensor<int32, [1]> q_11_axes_1 = const()[name = string("q_11_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_5_norm_self_att_weight_to_fp16 = const()[name = string("layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161581312)))];
            tensor<fp16, [1024]> layers_5_norm_self_att_bias_to_fp16 = const()[name = string("layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161583424)))];
            tensor<fp16, [1, 51, 1024]> q_11_cast_fp16 = layer_norm(axes = q_11_axes_1, beta = layers_5_norm_self_att_bias_to_fp16, epsilon = var_6557_to_fp16, gamma = layers_5_norm_self_att_weight_to_fp16, x = input_257_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<int32, [3]> var_6688 = const()[name = string("op_6688"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_259_axes_1 = const()[name = string("input_259_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_6689_cast_fp16 = transpose(perm = var_6688, x = q_11_cast_fp16)[name = string("transpose_776")];
            tensor<fp16, [1, 1024, 51, 1]> input_259_cast_fp16 = expand_dims(axes = input_259_axes_1, x = var_6689_cast_fp16)[name = string("input_259_cast_fp16")];
            string dense_output_417_pad_type_1 = const()[name = string("dense_output_417_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_417_strides_1 = const()[name = string("dense_output_417_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_417_pad_1 = const()[name = string("dense_output_417_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_417_dilations_1 = const()[name = string("dense_output_417_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_417_groups_1 = const()[name = string("dense_output_417_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161585536))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161716672))))[name = string("layers_5_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_417_cast_fp16 = conv(dilations = dense_output_417_dilations_1, groups = dense_output_417_groups_1, pad = dense_output_417_pad_1, pad_type = dense_output_417_pad_type_1, strides = dense_output_417_strides_1, weight = layers_5_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_417_cast_fp16")];
            string sparse_output_417_pad_type_1 = const()[name = string("sparse_output_417_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_417_strides_1 = const()[name = string("sparse_output_417_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_417_pad_1 = const()[name = string("sparse_output_417_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_417_dilations_1 = const()[name = string("sparse_output_417_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_417_groups_1 = const()[name = string("sparse_output_417_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161719936))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161717248))))[name = string("layers_5_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_417_cast_fp16 = conv(dilations = sparse_output_417_dilations_1, groups = sparse_output_417_groups_1, pad = sparse_output_417_pad_1, pad_type = sparse_output_417_pad_type_1, strides = sparse_output_417_strides_1, weight = layers_5_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_417_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6714_cast_fp16 = add(x = dense_output_417_cast_fp16, y = sparse_output_417_cast_fp16)[name = string("op_6714_cast_fp16")];
            tensor<int32, [4]> var_6715 = const()[name = string("op_6715"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6717 = const()[name = string("op_6717"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6716_cast_fp16 = transpose(perm = var_6715, x = var_6714_cast_fp16)[name = string("transpose_775")];
            tensor<fp16, [1, 51, 128]> q_head_81_cast_fp16 = reshape(shape = var_6717, x = var_6716_cast_fp16)[name = string("q_head_81_cast_fp16")];
            string dense_output_419_pad_type_1 = const()[name = string("dense_output_419_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_419_strides_1 = const()[name = string("dense_output_419_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_419_pad_1 = const()[name = string("dense_output_419_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_419_dilations_1 = const()[name = string("dense_output_419_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_419_groups_1 = const()[name = string("dense_output_419_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161736384))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161867520))))[name = string("layers_5_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_419_cast_fp16 = conv(dilations = dense_output_419_dilations_1, groups = dense_output_419_groups_1, pad = dense_output_419_pad_1, pad_type = dense_output_419_pad_type_1, strides = dense_output_419_strides_1, weight = layers_5_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_419_cast_fp16")];
            string sparse_output_419_pad_type_1 = const()[name = string("sparse_output_419_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_419_strides_1 = const()[name = string("sparse_output_419_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_419_pad_1 = const()[name = string("sparse_output_419_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_419_dilations_1 = const()[name = string("sparse_output_419_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_419_groups_1 = const()[name = string("sparse_output_419_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161870784))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161868096))))[name = string("layers_5_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_419_cast_fp16 = conv(dilations = sparse_output_419_dilations_1, groups = sparse_output_419_groups_1, pad = sparse_output_419_pad_1, pad_type = sparse_output_419_pad_type_1, strides = sparse_output_419_strides_1, weight = layers_5_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_419_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6733_cast_fp16 = add(x = dense_output_419_cast_fp16, y = sparse_output_419_cast_fp16)[name = string("op_6733_cast_fp16")];
            tensor<int32, [4]> var_6734 = const()[name = string("op_6734"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6736 = const()[name = string("op_6736"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6735_cast_fp16 = transpose(perm = var_6734, x = var_6733_cast_fp16)[name = string("transpose_774")];
            tensor<fp16, [1, 51, 128]> k_head_161_cast_fp16 = reshape(shape = var_6736, x = var_6735_cast_fp16)[name = string("k_head_161_cast_fp16")];
            string dense_output_421_pad_type_1 = const()[name = string("dense_output_421_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_421_strides_1 = const()[name = string("dense_output_421_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_421_pad_1 = const()[name = string("dense_output_421_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_421_dilations_1 = const()[name = string("dense_output_421_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_421_groups_1 = const()[name = string("dense_output_421_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161887232))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162018368))))[name = string("layers_5_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_421_cast_fp16 = conv(dilations = dense_output_421_dilations_1, groups = dense_output_421_groups_1, pad = dense_output_421_pad_1, pad_type = dense_output_421_pad_type_1, strides = dense_output_421_strides_1, weight = layers_5_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_421_cast_fp16")];
            string sparse_output_421_pad_type_1 = const()[name = string("sparse_output_421_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_421_strides_1 = const()[name = string("sparse_output_421_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_421_pad_1 = const()[name = string("sparse_output_421_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_421_dilations_1 = const()[name = string("sparse_output_421_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_421_groups_1 = const()[name = string("sparse_output_421_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162021632))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162018944))))[name = string("layers_5_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_421_cast_fp16 = conv(dilations = sparse_output_421_dilations_1, groups = sparse_output_421_groups_1, pad = sparse_output_421_pad_1, pad_type = sparse_output_421_pad_type_1, strides = sparse_output_421_strides_1, weight = layers_5_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_421_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6752_cast_fp16 = add(x = dense_output_421_cast_fp16, y = sparse_output_421_cast_fp16)[name = string("op_6752_cast_fp16")];
            tensor<int32, [4]> var_6753 = const()[name = string("op_6753"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6755 = const()[name = string("op_6755"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6754_cast_fp16 = transpose(perm = var_6753, x = var_6752_cast_fp16)[name = string("transpose_773")];
            tensor<fp16, [1, 51, 128]> v_head_161_cast_fp16 = reshape(shape = var_6755, x = var_6754_cast_fp16)[name = string("v_head_161_cast_fp16")];
            string dense_output_423_pad_type_1 = const()[name = string("dense_output_423_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_423_strides_1 = const()[name = string("dense_output_423_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_423_pad_1 = const()[name = string("dense_output_423_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_423_dilations_1 = const()[name = string("dense_output_423_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_423_groups_1 = const()[name = string("dense_output_423_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162038080))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162169216))))[name = string("layers_5_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_423_cast_fp16 = conv(dilations = dense_output_423_dilations_1, groups = dense_output_423_groups_1, pad = dense_output_423_pad_1, pad_type = dense_output_423_pad_type_1, strides = dense_output_423_strides_1, weight = layers_5_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_423_cast_fp16")];
            string sparse_output_423_pad_type_1 = const()[name = string("sparse_output_423_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_423_strides_1 = const()[name = string("sparse_output_423_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_423_pad_1 = const()[name = string("sparse_output_423_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_423_dilations_1 = const()[name = string("sparse_output_423_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_423_groups_1 = const()[name = string("sparse_output_423_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162172480))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162169792))))[name = string("layers_5_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_423_cast_fp16 = conv(dilations = sparse_output_423_dilations_1, groups = sparse_output_423_groups_1, pad = sparse_output_423_pad_1, pad_type = sparse_output_423_pad_type_1, strides = sparse_output_423_strides_1, weight = layers_5_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_423_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_6771_cast_fp16 = add(x = dense_output_423_cast_fp16, y = sparse_output_423_cast_fp16)[name = string("op_6771_cast_fp16")];
            tensor<int32, [4]> var_6772 = const()[name = string("op_6772"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6774 = const()[name = string("op_6774"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_6773_cast_fp16 = transpose(perm = var_6772, x = var_6771_cast_fp16)[name = string("transpose_772")];
            tensor<fp16, [1, 101, 128]> p_head_161_cast_fp16 = reshape(shape = var_6774, x = var_6773_cast_fp16)[name = string("p_head_161_cast_fp16")];
            tensor<fp16, [128]> var_6776_to_fp16 = const()[name = string("op_6776_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162188928)))];
            tensor<fp16, [1, 51, 128]> var_6777_cast_fp16 = add(x = q_head_81_cast_fp16, y = var_6776_to_fp16)[name = string("op_6777_cast_fp16")];
            tensor<int32, [1]> q_u_81_axes_1 = const()[name = string("q_u_81_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_81_cast_fp16 = expand_dims(axes = q_u_81_axes_1, x = var_6777_cast_fp16)[name = string("q_u_81_cast_fp16")];
            tensor<fp16, [128]> var_6779_to_fp16 = const()[name = string("op_6779_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162189248)))];
            tensor<fp16, [1, 51, 128]> var_6780_cast_fp16 = add(x = q_head_81_cast_fp16, y = var_6779_to_fp16)[name = string("op_6780_cast_fp16")];
            tensor<int32, [1]> q_v_81_axes_1 = const()[name = string("q_v_81_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_81_cast_fp16 = expand_dims(axes = q_v_81_axes_1, x = var_6780_cast_fp16)[name = string("q_v_81_cast_fp16")];
            tensor<int32, [1]> k_head_163_axes_1 = const()[name = string("k_head_163_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_163_cast_fp16 = expand_dims(axes = k_head_163_axes_1, x = k_head_161_cast_fp16)[name = string("k_head_163_cast_fp16")];
            tensor<int32, [1]> v_head_163_axes_1 = const()[name = string("v_head_163_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_163_cast_fp16 = expand_dims(axes = v_head_163_axes_1, x = v_head_161_cast_fp16)[name = string("v_head_163_cast_fp16")];
            tensor<int32, [1]> p_head_163_axes_1 = const()[name = string("p_head_163_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_163_cast_fp16 = expand_dims(axes = p_head_163_axes_1, x = p_head_161_cast_fp16)[name = string("p_head_163_cast_fp16")];
            bool var_6786_transpose_x_3 = const()[name = string("op_6786_transpose_x_3"), val = bool(false)];
            bool var_6786_transpose_y_3 = const()[name = string("op_6786_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_6786_cast_fp16 = matmul(transpose_x = var_6786_transpose_x_3, transpose_y = var_6786_transpose_y_3, x = q_u_81_cast_fp16, y = k_head_163_cast_fp16)[name = string("op_6786_cast_fp16")];
            fp16 var_6787_to_fp16 = const()[name = string("op_6787_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_81_cast_fp16 = mul(x = var_6786_cast_fp16, y = var_6787_to_fp16)[name = string("scores_content_81_cast_fp16")];
            bool x_441_transpose_x_3 = const()[name = string("x_441_transpose_x_3"), val = bool(false)];
            bool x_441_transpose_y_3 = const()[name = string("x_441_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_441_cast_fp16 = matmul(transpose_x = x_441_transpose_x_3, transpose_y = x_441_transpose_y_3, x = q_v_81_cast_fp16, y = p_head_163_cast_fp16)[name = string("x_441_cast_fp16")];
            tensor<int32, [8]> x_443_pad_1 = const()[name = string("x_443_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_443_mode_1 = const()[name = string("x_443_mode_1"), val = string("constant")];
            fp16 const_1579_to_fp16 = const()[name = string("const_1579_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_443_cast_fp16 = pad(constant_val = const_1579_to_fp16, mode = x_443_mode_1, pad = x_443_pad_1, x = x_441_cast_fp16)[name = string("x_443_cast_fp16")];
            tensor<int32, [4]> var_6801 = const()[name = string("op_6801"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_445_cast_fp16 = reshape(shape = var_6801, x = x_443_cast_fp16)[name = string("x_445_cast_fp16")];
            tensor<int32, [4]> var_6805_begin_1 = const()[name = string("op_6805_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6805_end_1 = const()[name = string("op_6805_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_6805_end_mask_1 = const()[name = string("op_6805_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_6805_cast_fp16 = slice_by_index(begin = var_6805_begin_1, end = var_6805_end_1, end_mask = var_6805_end_mask_1, x = x_445_cast_fp16)[name = string("op_6805_cast_fp16")];
            tensor<int32, [4]> var_6807 = const()[name = string("op_6807"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_6808_cast_fp16 = reshape(shape = var_6807, x = var_6805_cast_fp16)[name = string("op_6808_cast_fp16")];
            tensor<int32, [4]> var_6813_begin_1 = const()[name = string("op_6813_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6813_end_1 = const()[name = string("op_6813_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_6813_end_mask_1 = const()[name = string("op_6813_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_6813_cast_fp16 = slice_by_index(begin = var_6813_begin_1, end = var_6813_end_1, end_mask = var_6813_end_mask_1, x = var_6808_cast_fp16)[name = string("op_6813_cast_fp16")];
            fp16 var_6814_to_fp16 = const()[name = string("op_6814_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_81_cast_fp16 = mul(x = var_6813_cast_fp16, y = var_6814_to_fp16)[name = string("scores_pos_81_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_81_cast_fp16 = add(x = scores_content_81_cast_fp16, y = scores_pos_81_cast_fp16)[name = string("logits_81_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_6817_cast_fp16 = softmax(axis = var_6542, x = logits_81_cast_fp16)[name = string("op_6817_cast_fp16")];
            bool var_6819_transpose_x_1 = const()[name = string("op_6819_transpose_x_1"), val = bool(false)];
            bool var_6819_transpose_y_1 = const()[name = string("op_6819_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_6819_cast_fp16 = matmul(transpose_x = var_6819_transpose_x_1, transpose_y = var_6819_transpose_y_1, x = var_6817_cast_fp16, y = v_head_163_cast_fp16)[name = string("op_6819_cast_fp16")];
            tensor<int32, [1]> var_6820_axes_1 = const()[name = string("op_6820_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_6820_cast_fp16 = squeeze(axes = var_6820_axes_1, x = var_6819_cast_fp16)[name = string("op_6820_cast_fp16")];
            string dense_output_425_pad_type_1 = const()[name = string("dense_output_425_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_425_strides_1 = const()[name = string("dense_output_425_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_425_pad_1 = const()[name = string("dense_output_425_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_425_dilations_1 = const()[name = string("dense_output_425_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_425_groups_1 = const()[name = string("dense_output_425_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162189568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162320704))))[name = string("layers_5_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_425_cast_fp16 = conv(dilations = dense_output_425_dilations_1, groups = dense_output_425_groups_1, pad = dense_output_425_pad_1, pad_type = dense_output_425_pad_type_1, strides = dense_output_425_strides_1, weight = layers_5_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_425_cast_fp16")];
            string sparse_output_425_pad_type_1 = const()[name = string("sparse_output_425_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_425_strides_1 = const()[name = string("sparse_output_425_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_425_pad_1 = const()[name = string("sparse_output_425_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_425_dilations_1 = const()[name = string("sparse_output_425_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_425_groups_1 = const()[name = string("sparse_output_425_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162323968))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162321280))))[name = string("layers_5_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_425_cast_fp16 = conv(dilations = sparse_output_425_dilations_1, groups = sparse_output_425_groups_1, pad = sparse_output_425_pad_1, pad_type = sparse_output_425_pad_type_1, strides = sparse_output_425_strides_1, weight = layers_5_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_425_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6835_cast_fp16 = add(x = dense_output_425_cast_fp16, y = sparse_output_425_cast_fp16)[name = string("op_6835_cast_fp16")];
            tensor<int32, [4]> var_6836 = const()[name = string("op_6836"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6838 = const()[name = string("op_6838"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6837_cast_fp16 = transpose(perm = var_6836, x = var_6835_cast_fp16)[name = string("transpose_771")];
            tensor<fp16, [1, 51, 128]> q_head_83_cast_fp16 = reshape(shape = var_6838, x = var_6837_cast_fp16)[name = string("q_head_83_cast_fp16")];
            string dense_output_427_pad_type_1 = const()[name = string("dense_output_427_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_427_strides_1 = const()[name = string("dense_output_427_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_427_pad_1 = const()[name = string("dense_output_427_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_427_dilations_1 = const()[name = string("dense_output_427_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_427_groups_1 = const()[name = string("dense_output_427_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162340416))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162471552))))[name = string("layers_5_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_427_cast_fp16 = conv(dilations = dense_output_427_dilations_1, groups = dense_output_427_groups_1, pad = dense_output_427_pad_1, pad_type = dense_output_427_pad_type_1, strides = dense_output_427_strides_1, weight = layers_5_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_427_cast_fp16")];
            string sparse_output_427_pad_type_1 = const()[name = string("sparse_output_427_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_427_strides_1 = const()[name = string("sparse_output_427_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_427_pad_1 = const()[name = string("sparse_output_427_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_427_dilations_1 = const()[name = string("sparse_output_427_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_427_groups_1 = const()[name = string("sparse_output_427_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162474816))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162472128))))[name = string("layers_5_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_427_cast_fp16 = conv(dilations = sparse_output_427_dilations_1, groups = sparse_output_427_groups_1, pad = sparse_output_427_pad_1, pad_type = sparse_output_427_pad_type_1, strides = sparse_output_427_strides_1, weight = layers_5_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_427_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6854_cast_fp16 = add(x = dense_output_427_cast_fp16, y = sparse_output_427_cast_fp16)[name = string("op_6854_cast_fp16")];
            tensor<int32, [4]> var_6855 = const()[name = string("op_6855"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6857 = const()[name = string("op_6857"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6856_cast_fp16 = transpose(perm = var_6855, x = var_6854_cast_fp16)[name = string("transpose_770")];
            tensor<fp16, [1, 51, 128]> k_head_165_cast_fp16 = reshape(shape = var_6857, x = var_6856_cast_fp16)[name = string("k_head_165_cast_fp16")];
            string dense_output_429_pad_type_1 = const()[name = string("dense_output_429_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_429_strides_1 = const()[name = string("dense_output_429_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_429_pad_1 = const()[name = string("dense_output_429_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_429_dilations_1 = const()[name = string("dense_output_429_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_429_groups_1 = const()[name = string("dense_output_429_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162491264))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162622400))))[name = string("layers_5_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_429_cast_fp16 = conv(dilations = dense_output_429_dilations_1, groups = dense_output_429_groups_1, pad = dense_output_429_pad_1, pad_type = dense_output_429_pad_type_1, strides = dense_output_429_strides_1, weight = layers_5_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_429_cast_fp16")];
            string sparse_output_429_pad_type_1 = const()[name = string("sparse_output_429_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_429_strides_1 = const()[name = string("sparse_output_429_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_429_pad_1 = const()[name = string("sparse_output_429_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_429_dilations_1 = const()[name = string("sparse_output_429_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_429_groups_1 = const()[name = string("sparse_output_429_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162625664))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162622976))))[name = string("layers_5_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_429_cast_fp16 = conv(dilations = sparse_output_429_dilations_1, groups = sparse_output_429_groups_1, pad = sparse_output_429_pad_1, pad_type = sparse_output_429_pad_type_1, strides = sparse_output_429_strides_1, weight = layers_5_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_429_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6873_cast_fp16 = add(x = dense_output_429_cast_fp16, y = sparse_output_429_cast_fp16)[name = string("op_6873_cast_fp16")];
            tensor<int32, [4]> var_6874 = const()[name = string("op_6874"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6876 = const()[name = string("op_6876"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6875_cast_fp16 = transpose(perm = var_6874, x = var_6873_cast_fp16)[name = string("transpose_769")];
            tensor<fp16, [1, 51, 128]> v_head_165_cast_fp16 = reshape(shape = var_6876, x = var_6875_cast_fp16)[name = string("v_head_165_cast_fp16")];
            string dense_output_431_pad_type_1 = const()[name = string("dense_output_431_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_431_strides_1 = const()[name = string("dense_output_431_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_431_pad_1 = const()[name = string("dense_output_431_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_431_dilations_1 = const()[name = string("dense_output_431_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_431_groups_1 = const()[name = string("dense_output_431_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162642112))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162773248))))[name = string("layers_5_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_431_cast_fp16 = conv(dilations = dense_output_431_dilations_1, groups = dense_output_431_groups_1, pad = dense_output_431_pad_1, pad_type = dense_output_431_pad_type_1, strides = dense_output_431_strides_1, weight = layers_5_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_431_cast_fp16")];
            string sparse_output_431_pad_type_1 = const()[name = string("sparse_output_431_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_431_strides_1 = const()[name = string("sparse_output_431_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_431_pad_1 = const()[name = string("sparse_output_431_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_431_dilations_1 = const()[name = string("sparse_output_431_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_431_groups_1 = const()[name = string("sparse_output_431_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162776512))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162773824))))[name = string("layers_5_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_431_cast_fp16 = conv(dilations = sparse_output_431_dilations_1, groups = sparse_output_431_groups_1, pad = sparse_output_431_pad_1, pad_type = sparse_output_431_pad_type_1, strides = sparse_output_431_strides_1, weight = layers_5_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_431_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_6892_cast_fp16 = add(x = dense_output_431_cast_fp16, y = sparse_output_431_cast_fp16)[name = string("op_6892_cast_fp16")];
            tensor<int32, [4]> var_6893 = const()[name = string("op_6893"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6895 = const()[name = string("op_6895"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_6894_cast_fp16 = transpose(perm = var_6893, x = var_6892_cast_fp16)[name = string("transpose_768")];
            tensor<fp16, [1, 101, 128]> p_head_165_cast_fp16 = reshape(shape = var_6895, x = var_6894_cast_fp16)[name = string("p_head_165_cast_fp16")];
            tensor<fp16, [128]> var_6897_to_fp16 = const()[name = string("op_6897_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162792960)))];
            tensor<fp16, [1, 51, 128]> var_6898_cast_fp16 = add(x = q_head_83_cast_fp16, y = var_6897_to_fp16)[name = string("op_6898_cast_fp16")];
            tensor<int32, [1]> q_u_83_axes_1 = const()[name = string("q_u_83_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_83_cast_fp16 = expand_dims(axes = q_u_83_axes_1, x = var_6898_cast_fp16)[name = string("q_u_83_cast_fp16")];
            tensor<fp16, [128]> var_6900_to_fp16 = const()[name = string("op_6900_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162793280)))];
            tensor<fp16, [1, 51, 128]> var_6901_cast_fp16 = add(x = q_head_83_cast_fp16, y = var_6900_to_fp16)[name = string("op_6901_cast_fp16")];
            tensor<int32, [1]> q_v_83_axes_1 = const()[name = string("q_v_83_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_83_cast_fp16 = expand_dims(axes = q_v_83_axes_1, x = var_6901_cast_fp16)[name = string("q_v_83_cast_fp16")];
            tensor<int32, [1]> k_head_167_axes_1 = const()[name = string("k_head_167_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_167_cast_fp16 = expand_dims(axes = k_head_167_axes_1, x = k_head_165_cast_fp16)[name = string("k_head_167_cast_fp16")];
            tensor<int32, [1]> v_head_167_axes_1 = const()[name = string("v_head_167_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_167_cast_fp16 = expand_dims(axes = v_head_167_axes_1, x = v_head_165_cast_fp16)[name = string("v_head_167_cast_fp16")];
            tensor<int32, [1]> p_head_167_axes_1 = const()[name = string("p_head_167_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_167_cast_fp16 = expand_dims(axes = p_head_167_axes_1, x = p_head_165_cast_fp16)[name = string("p_head_167_cast_fp16")];
            bool var_6907_transpose_x_3 = const()[name = string("op_6907_transpose_x_3"), val = bool(false)];
            bool var_6907_transpose_y_3 = const()[name = string("op_6907_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_6907_cast_fp16 = matmul(transpose_x = var_6907_transpose_x_3, transpose_y = var_6907_transpose_y_3, x = q_u_83_cast_fp16, y = k_head_167_cast_fp16)[name = string("op_6907_cast_fp16")];
            fp16 var_6908_to_fp16 = const()[name = string("op_6908_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_83_cast_fp16 = mul(x = var_6907_cast_fp16, y = var_6908_to_fp16)[name = string("scores_content_83_cast_fp16")];
            bool x_449_transpose_x_3 = const()[name = string("x_449_transpose_x_3"), val = bool(false)];
            bool x_449_transpose_y_3 = const()[name = string("x_449_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_3, transpose_y = x_449_transpose_y_3, x = q_v_83_cast_fp16, y = p_head_167_cast_fp16)[name = string("x_449_cast_fp16")];
            tensor<int32, [8]> x_451_pad_1 = const()[name = string("x_451_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_451_mode_1 = const()[name = string("x_451_mode_1"), val = string("constant")];
            fp16 const_1585_to_fp16 = const()[name = string("const_1585_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_451_cast_fp16 = pad(constant_val = const_1585_to_fp16, mode = x_451_mode_1, pad = x_451_pad_1, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")];
            tensor<int32, [4]> var_6922 = const()[name = string("op_6922"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_453_cast_fp16 = reshape(shape = var_6922, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")];
            tensor<int32, [4]> var_6926_begin_1 = const()[name = string("op_6926_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_6926_end_1 = const()[name = string("op_6926_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_6926_end_mask_1 = const()[name = string("op_6926_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_6926_cast_fp16 = slice_by_index(begin = var_6926_begin_1, end = var_6926_end_1, end_mask = var_6926_end_mask_1, x = x_453_cast_fp16)[name = string("op_6926_cast_fp16")];
            tensor<int32, [4]> var_6928 = const()[name = string("op_6928"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_6929_cast_fp16 = reshape(shape = var_6928, x = var_6926_cast_fp16)[name = string("op_6929_cast_fp16")];
            tensor<int32, [4]> var_6934_begin_1 = const()[name = string("op_6934_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6934_end_1 = const()[name = string("op_6934_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_6934_end_mask_1 = const()[name = string("op_6934_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_6934_cast_fp16 = slice_by_index(begin = var_6934_begin_1, end = var_6934_end_1, end_mask = var_6934_end_mask_1, x = var_6929_cast_fp16)[name = string("op_6934_cast_fp16")];
            fp16 var_6935_to_fp16 = const()[name = string("op_6935_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_83_cast_fp16 = mul(x = var_6934_cast_fp16, y = var_6935_to_fp16)[name = string("scores_pos_83_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_83_cast_fp16 = add(x = scores_content_83_cast_fp16, y = scores_pos_83_cast_fp16)[name = string("logits_83_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_6938_cast_fp16 = softmax(axis = var_6542, x = logits_83_cast_fp16)[name = string("op_6938_cast_fp16")];
            bool var_6940_transpose_x_1 = const()[name = string("op_6940_transpose_x_1"), val = bool(false)];
            bool var_6940_transpose_y_1 = const()[name = string("op_6940_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_6940_cast_fp16 = matmul(transpose_x = var_6940_transpose_x_1, transpose_y = var_6940_transpose_y_1, x = var_6938_cast_fp16, y = v_head_167_cast_fp16)[name = string("op_6940_cast_fp16")];
            tensor<int32, [1]> var_6941_axes_1 = const()[name = string("op_6941_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_6941_cast_fp16 = squeeze(axes = var_6941_axes_1, x = var_6940_cast_fp16)[name = string("op_6941_cast_fp16")];
            string dense_output_433_pad_type_1 = const()[name = string("dense_output_433_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_433_strides_1 = const()[name = string("dense_output_433_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_433_pad_1 = const()[name = string("dense_output_433_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_433_dilations_1 = const()[name = string("dense_output_433_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_433_groups_1 = const()[name = string("dense_output_433_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162793600))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162924736))))[name = string("layers_5_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_433_cast_fp16 = conv(dilations = dense_output_433_dilations_1, groups = dense_output_433_groups_1, pad = dense_output_433_pad_1, pad_type = dense_output_433_pad_type_1, strides = dense_output_433_strides_1, weight = layers_5_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_433_cast_fp16")];
            string sparse_output_433_pad_type_1 = const()[name = string("sparse_output_433_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_433_strides_1 = const()[name = string("sparse_output_433_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_433_pad_1 = const()[name = string("sparse_output_433_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_433_dilations_1 = const()[name = string("sparse_output_433_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_433_groups_1 = const()[name = string("sparse_output_433_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162928000))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162925312))))[name = string("layers_5_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_433_cast_fp16 = conv(dilations = sparse_output_433_dilations_1, groups = sparse_output_433_groups_1, pad = sparse_output_433_pad_1, pad_type = sparse_output_433_pad_type_1, strides = sparse_output_433_strides_1, weight = layers_5_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_433_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6956_cast_fp16 = add(x = dense_output_433_cast_fp16, y = sparse_output_433_cast_fp16)[name = string("op_6956_cast_fp16")];
            tensor<int32, [4]> var_6957 = const()[name = string("op_6957"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6959 = const()[name = string("op_6959"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6958_cast_fp16 = transpose(perm = var_6957, x = var_6956_cast_fp16)[name = string("transpose_767")];
            tensor<fp16, [1, 51, 128]> q_head_85_cast_fp16 = reshape(shape = var_6959, x = var_6958_cast_fp16)[name = string("q_head_85_cast_fp16")];
            string dense_output_435_pad_type_1 = const()[name = string("dense_output_435_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_435_strides_1 = const()[name = string("dense_output_435_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_435_pad_1 = const()[name = string("dense_output_435_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_435_dilations_1 = const()[name = string("dense_output_435_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_435_groups_1 = const()[name = string("dense_output_435_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162944448))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163075584))))[name = string("layers_5_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_435_cast_fp16 = conv(dilations = dense_output_435_dilations_1, groups = dense_output_435_groups_1, pad = dense_output_435_pad_1, pad_type = dense_output_435_pad_type_1, strides = dense_output_435_strides_1, weight = layers_5_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_435_cast_fp16")];
            string sparse_output_435_pad_type_1 = const()[name = string("sparse_output_435_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_435_strides_1 = const()[name = string("sparse_output_435_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_435_pad_1 = const()[name = string("sparse_output_435_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_435_dilations_1 = const()[name = string("sparse_output_435_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_435_groups_1 = const()[name = string("sparse_output_435_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163078848))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163076160))))[name = string("layers_5_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_435_cast_fp16 = conv(dilations = sparse_output_435_dilations_1, groups = sparse_output_435_groups_1, pad = sparse_output_435_pad_1, pad_type = sparse_output_435_pad_type_1, strides = sparse_output_435_strides_1, weight = layers_5_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_435_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6975_cast_fp16 = add(x = dense_output_435_cast_fp16, y = sparse_output_435_cast_fp16)[name = string("op_6975_cast_fp16")];
            tensor<int32, [4]> var_6976 = const()[name = string("op_6976"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6978 = const()[name = string("op_6978"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6977_cast_fp16 = transpose(perm = var_6976, x = var_6975_cast_fp16)[name = string("transpose_766")];
            tensor<fp16, [1, 51, 128]> k_head_169_cast_fp16 = reshape(shape = var_6978, x = var_6977_cast_fp16)[name = string("k_head_169_cast_fp16")];
            string dense_output_437_pad_type_1 = const()[name = string("dense_output_437_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_437_strides_1 = const()[name = string("dense_output_437_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_437_pad_1 = const()[name = string("dense_output_437_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_437_dilations_1 = const()[name = string("dense_output_437_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_437_groups_1 = const()[name = string("dense_output_437_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163095296))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163226432))))[name = string("layers_5_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_437_cast_fp16 = conv(dilations = dense_output_437_dilations_1, groups = dense_output_437_groups_1, pad = dense_output_437_pad_1, pad_type = dense_output_437_pad_type_1, strides = dense_output_437_strides_1, weight = layers_5_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_437_cast_fp16")];
            string sparse_output_437_pad_type_1 = const()[name = string("sparse_output_437_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_437_strides_1 = const()[name = string("sparse_output_437_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_437_pad_1 = const()[name = string("sparse_output_437_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_437_dilations_1 = const()[name = string("sparse_output_437_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_437_groups_1 = const()[name = string("sparse_output_437_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163229696))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163227008))))[name = string("layers_5_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_437_cast_fp16 = conv(dilations = sparse_output_437_dilations_1, groups = sparse_output_437_groups_1, pad = sparse_output_437_pad_1, pad_type = sparse_output_437_pad_type_1, strides = sparse_output_437_strides_1, weight = layers_5_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_437_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_6994_cast_fp16 = add(x = dense_output_437_cast_fp16, y = sparse_output_437_cast_fp16)[name = string("op_6994_cast_fp16")];
            tensor<int32, [4]> var_6995 = const()[name = string("op_6995"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_6997 = const()[name = string("op_6997"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_6996_cast_fp16 = transpose(perm = var_6995, x = var_6994_cast_fp16)[name = string("transpose_765")];
            tensor<fp16, [1, 51, 128]> v_head_169_cast_fp16 = reshape(shape = var_6997, x = var_6996_cast_fp16)[name = string("v_head_169_cast_fp16")];
            string dense_output_439_pad_type_1 = const()[name = string("dense_output_439_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_439_strides_1 = const()[name = string("dense_output_439_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_439_pad_1 = const()[name = string("dense_output_439_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_439_dilations_1 = const()[name = string("dense_output_439_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_439_groups_1 = const()[name = string("dense_output_439_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163246144))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163377280))))[name = string("layers_5_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_439_cast_fp16 = conv(dilations = dense_output_439_dilations_1, groups = dense_output_439_groups_1, pad = dense_output_439_pad_1, pad_type = dense_output_439_pad_type_1, strides = dense_output_439_strides_1, weight = layers_5_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_439_cast_fp16")];
            string sparse_output_439_pad_type_1 = const()[name = string("sparse_output_439_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_439_strides_1 = const()[name = string("sparse_output_439_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_439_pad_1 = const()[name = string("sparse_output_439_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_439_dilations_1 = const()[name = string("sparse_output_439_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_439_groups_1 = const()[name = string("sparse_output_439_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163380544))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163377856))))[name = string("layers_5_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_439_cast_fp16 = conv(dilations = sparse_output_439_dilations_1, groups = sparse_output_439_groups_1, pad = sparse_output_439_pad_1, pad_type = sparse_output_439_pad_type_1, strides = sparse_output_439_strides_1, weight = layers_5_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_439_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_7013_cast_fp16 = add(x = dense_output_439_cast_fp16, y = sparse_output_439_cast_fp16)[name = string("op_7013_cast_fp16")];
            tensor<int32, [4]> var_7014 = const()[name = string("op_7014"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7016 = const()[name = string("op_7016"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_7015_cast_fp16 = transpose(perm = var_7014, x = var_7013_cast_fp16)[name = string("transpose_764")];
            tensor<fp16, [1, 101, 128]> p_head_169_cast_fp16 = reshape(shape = var_7016, x = var_7015_cast_fp16)[name = string("p_head_169_cast_fp16")];
            tensor<fp16, [128]> var_7018_to_fp16 = const()[name = string("op_7018_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163396992)))];
            tensor<fp16, [1, 51, 128]> var_7019_cast_fp16 = add(x = q_head_85_cast_fp16, y = var_7018_to_fp16)[name = string("op_7019_cast_fp16")];
            tensor<int32, [1]> q_u_85_axes_1 = const()[name = string("q_u_85_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_85_cast_fp16 = expand_dims(axes = q_u_85_axes_1, x = var_7019_cast_fp16)[name = string("q_u_85_cast_fp16")];
            tensor<fp16, [128]> var_7021_to_fp16 = const()[name = string("op_7021_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163397312)))];
            tensor<fp16, [1, 51, 128]> var_7022_cast_fp16 = add(x = q_head_85_cast_fp16, y = var_7021_to_fp16)[name = string("op_7022_cast_fp16")];
            tensor<int32, [1]> q_v_85_axes_1 = const()[name = string("q_v_85_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_85_cast_fp16 = expand_dims(axes = q_v_85_axes_1, x = var_7022_cast_fp16)[name = string("q_v_85_cast_fp16")];
            tensor<int32, [1]> k_head_171_axes_1 = const()[name = string("k_head_171_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_171_cast_fp16 = expand_dims(axes = k_head_171_axes_1, x = k_head_169_cast_fp16)[name = string("k_head_171_cast_fp16")];
            tensor<int32, [1]> v_head_171_axes_1 = const()[name = string("v_head_171_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_171_cast_fp16 = expand_dims(axes = v_head_171_axes_1, x = v_head_169_cast_fp16)[name = string("v_head_171_cast_fp16")];
            tensor<int32, [1]> p_head_171_axes_1 = const()[name = string("p_head_171_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_171_cast_fp16 = expand_dims(axes = p_head_171_axes_1, x = p_head_169_cast_fp16)[name = string("p_head_171_cast_fp16")];
            bool var_7028_transpose_x_3 = const()[name = string("op_7028_transpose_x_3"), val = bool(false)];
            bool var_7028_transpose_y_3 = const()[name = string("op_7028_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_7028_cast_fp16 = matmul(transpose_x = var_7028_transpose_x_3, transpose_y = var_7028_transpose_y_3, x = q_u_85_cast_fp16, y = k_head_171_cast_fp16)[name = string("op_7028_cast_fp16")];
            fp16 var_7029_to_fp16 = const()[name = string("op_7029_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_85_cast_fp16 = mul(x = var_7028_cast_fp16, y = var_7029_to_fp16)[name = string("scores_content_85_cast_fp16")];
            bool x_457_transpose_x_3 = const()[name = string("x_457_transpose_x_3"), val = bool(false)];
            bool x_457_transpose_y_3 = const()[name = string("x_457_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_457_cast_fp16 = matmul(transpose_x = x_457_transpose_x_3, transpose_y = x_457_transpose_y_3, x = q_v_85_cast_fp16, y = p_head_171_cast_fp16)[name = string("x_457_cast_fp16")];
            tensor<int32, [8]> x_459_pad_1 = const()[name = string("x_459_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_459_mode_1 = const()[name = string("x_459_mode_1"), val = string("constant")];
            fp16 const_1591_to_fp16 = const()[name = string("const_1591_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_459_cast_fp16 = pad(constant_val = const_1591_to_fp16, mode = x_459_mode_1, pad = x_459_pad_1, x = x_457_cast_fp16)[name = string("x_459_cast_fp16")];
            tensor<int32, [4]> var_7043 = const()[name = string("op_7043"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_461_cast_fp16 = reshape(shape = var_7043, x = x_459_cast_fp16)[name = string("x_461_cast_fp16")];
            tensor<int32, [4]> var_7047_begin_1 = const()[name = string("op_7047_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7047_end_1 = const()[name = string("op_7047_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_7047_end_mask_1 = const()[name = string("op_7047_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_7047_cast_fp16 = slice_by_index(begin = var_7047_begin_1, end = var_7047_end_1, end_mask = var_7047_end_mask_1, x = x_461_cast_fp16)[name = string("op_7047_cast_fp16")];
            tensor<int32, [4]> var_7049 = const()[name = string("op_7049"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_7050_cast_fp16 = reshape(shape = var_7049, x = var_7047_cast_fp16)[name = string("op_7050_cast_fp16")];
            tensor<int32, [4]> var_7055_begin_1 = const()[name = string("op_7055_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7055_end_1 = const()[name = string("op_7055_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_7055_end_mask_1 = const()[name = string("op_7055_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_7055_cast_fp16 = slice_by_index(begin = var_7055_begin_1, end = var_7055_end_1, end_mask = var_7055_end_mask_1, x = var_7050_cast_fp16)[name = string("op_7055_cast_fp16")];
            fp16 var_7056_to_fp16 = const()[name = string("op_7056_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_85_cast_fp16 = mul(x = var_7055_cast_fp16, y = var_7056_to_fp16)[name = string("scores_pos_85_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_85_cast_fp16 = add(x = scores_content_85_cast_fp16, y = scores_pos_85_cast_fp16)[name = string("logits_85_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_7059_cast_fp16 = softmax(axis = var_6542, x = logits_85_cast_fp16)[name = string("op_7059_cast_fp16")];
            bool var_7061_transpose_x_1 = const()[name = string("op_7061_transpose_x_1"), val = bool(false)];
            bool var_7061_transpose_y_1 = const()[name = string("op_7061_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_7061_cast_fp16 = matmul(transpose_x = var_7061_transpose_x_1, transpose_y = var_7061_transpose_y_1, x = var_7059_cast_fp16, y = v_head_171_cast_fp16)[name = string("op_7061_cast_fp16")];
            tensor<int32, [1]> var_7062_axes_1 = const()[name = string("op_7062_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_7062_cast_fp16 = squeeze(axes = var_7062_axes_1, x = var_7061_cast_fp16)[name = string("op_7062_cast_fp16")];
            string dense_output_441_pad_type_1 = const()[name = string("dense_output_441_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_441_strides_1 = const()[name = string("dense_output_441_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_441_pad_1 = const()[name = string("dense_output_441_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_441_dilations_1 = const()[name = string("dense_output_441_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_441_groups_1 = const()[name = string("dense_output_441_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163397632))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163528768))))[name = string("layers_5_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_441_cast_fp16 = conv(dilations = dense_output_441_dilations_1, groups = dense_output_441_groups_1, pad = dense_output_441_pad_1, pad_type = dense_output_441_pad_type_1, strides = dense_output_441_strides_1, weight = layers_5_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_441_cast_fp16")];
            string sparse_output_441_pad_type_1 = const()[name = string("sparse_output_441_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_441_strides_1 = const()[name = string("sparse_output_441_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_441_pad_1 = const()[name = string("sparse_output_441_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_441_dilations_1 = const()[name = string("sparse_output_441_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_441_groups_1 = const()[name = string("sparse_output_441_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163532032))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163529344))))[name = string("layers_5_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_441_cast_fp16 = conv(dilations = sparse_output_441_dilations_1, groups = sparse_output_441_groups_1, pad = sparse_output_441_pad_1, pad_type = sparse_output_441_pad_type_1, strides = sparse_output_441_strides_1, weight = layers_5_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_441_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7077_cast_fp16 = add(x = dense_output_441_cast_fp16, y = sparse_output_441_cast_fp16)[name = string("op_7077_cast_fp16")];
            tensor<int32, [4]> var_7078 = const()[name = string("op_7078"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7080 = const()[name = string("op_7080"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7079_cast_fp16 = transpose(perm = var_7078, x = var_7077_cast_fp16)[name = string("transpose_763")];
            tensor<fp16, [1, 51, 128]> q_head_87_cast_fp16 = reshape(shape = var_7080, x = var_7079_cast_fp16)[name = string("q_head_87_cast_fp16")];
            string dense_output_443_pad_type_1 = const()[name = string("dense_output_443_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_443_strides_1 = const()[name = string("dense_output_443_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_443_pad_1 = const()[name = string("dense_output_443_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_443_dilations_1 = const()[name = string("dense_output_443_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_443_groups_1 = const()[name = string("dense_output_443_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163548480))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163679616))))[name = string("layers_5_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_443_cast_fp16 = conv(dilations = dense_output_443_dilations_1, groups = dense_output_443_groups_1, pad = dense_output_443_pad_1, pad_type = dense_output_443_pad_type_1, strides = dense_output_443_strides_1, weight = layers_5_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_443_cast_fp16")];
            string sparse_output_443_pad_type_1 = const()[name = string("sparse_output_443_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_443_strides_1 = const()[name = string("sparse_output_443_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_443_pad_1 = const()[name = string("sparse_output_443_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_443_dilations_1 = const()[name = string("sparse_output_443_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_443_groups_1 = const()[name = string("sparse_output_443_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163682880))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163680192))))[name = string("layers_5_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_443_cast_fp16 = conv(dilations = sparse_output_443_dilations_1, groups = sparse_output_443_groups_1, pad = sparse_output_443_pad_1, pad_type = sparse_output_443_pad_type_1, strides = sparse_output_443_strides_1, weight = layers_5_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_443_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7096_cast_fp16 = add(x = dense_output_443_cast_fp16, y = sparse_output_443_cast_fp16)[name = string("op_7096_cast_fp16")];
            tensor<int32, [4]> var_7097 = const()[name = string("op_7097"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7099 = const()[name = string("op_7099"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7098_cast_fp16 = transpose(perm = var_7097, x = var_7096_cast_fp16)[name = string("transpose_762")];
            tensor<fp16, [1, 51, 128]> k_head_173_cast_fp16 = reshape(shape = var_7099, x = var_7098_cast_fp16)[name = string("k_head_173_cast_fp16")];
            string dense_output_445_pad_type_1 = const()[name = string("dense_output_445_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_445_strides_1 = const()[name = string("dense_output_445_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_445_pad_1 = const()[name = string("dense_output_445_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_445_dilations_1 = const()[name = string("dense_output_445_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_445_groups_1 = const()[name = string("dense_output_445_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163699328))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163830464))))[name = string("layers_5_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_445_cast_fp16 = conv(dilations = dense_output_445_dilations_1, groups = dense_output_445_groups_1, pad = dense_output_445_pad_1, pad_type = dense_output_445_pad_type_1, strides = dense_output_445_strides_1, weight = layers_5_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_445_cast_fp16")];
            string sparse_output_445_pad_type_1 = const()[name = string("sparse_output_445_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_445_strides_1 = const()[name = string("sparse_output_445_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_445_pad_1 = const()[name = string("sparse_output_445_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_445_dilations_1 = const()[name = string("sparse_output_445_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_445_groups_1 = const()[name = string("sparse_output_445_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163833728))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163831040))))[name = string("layers_5_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_445_cast_fp16 = conv(dilations = sparse_output_445_dilations_1, groups = sparse_output_445_groups_1, pad = sparse_output_445_pad_1, pad_type = sparse_output_445_pad_type_1, strides = sparse_output_445_strides_1, weight = layers_5_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_445_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7115_cast_fp16 = add(x = dense_output_445_cast_fp16, y = sparse_output_445_cast_fp16)[name = string("op_7115_cast_fp16")];
            tensor<int32, [4]> var_7116 = const()[name = string("op_7116"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7118 = const()[name = string("op_7118"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7117_cast_fp16 = transpose(perm = var_7116, x = var_7115_cast_fp16)[name = string("transpose_761")];
            tensor<fp16, [1, 51, 128]> v_head_173_cast_fp16 = reshape(shape = var_7118, x = var_7117_cast_fp16)[name = string("v_head_173_cast_fp16")];
            string dense_output_447_pad_type_1 = const()[name = string("dense_output_447_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_447_strides_1 = const()[name = string("dense_output_447_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_447_pad_1 = const()[name = string("dense_output_447_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_447_dilations_1 = const()[name = string("dense_output_447_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_447_groups_1 = const()[name = string("dense_output_447_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163850176))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163981312))))[name = string("layers_5_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_447_cast_fp16 = conv(dilations = dense_output_447_dilations_1, groups = dense_output_447_groups_1, pad = dense_output_447_pad_1, pad_type = dense_output_447_pad_type_1, strides = dense_output_447_strides_1, weight = layers_5_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_447_cast_fp16")];
            string sparse_output_447_pad_type_1 = const()[name = string("sparse_output_447_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_447_strides_1 = const()[name = string("sparse_output_447_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_447_pad_1 = const()[name = string("sparse_output_447_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_447_dilations_1 = const()[name = string("sparse_output_447_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_447_groups_1 = const()[name = string("sparse_output_447_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163984576))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163981888))))[name = string("layers_5_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_447_cast_fp16 = conv(dilations = sparse_output_447_dilations_1, groups = sparse_output_447_groups_1, pad = sparse_output_447_pad_1, pad_type = sparse_output_447_pad_type_1, strides = sparse_output_447_strides_1, weight = layers_5_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_447_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_7134_cast_fp16 = add(x = dense_output_447_cast_fp16, y = sparse_output_447_cast_fp16)[name = string("op_7134_cast_fp16")];
            tensor<int32, [4]> var_7135 = const()[name = string("op_7135"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7137 = const()[name = string("op_7137"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_7136_cast_fp16 = transpose(perm = var_7135, x = var_7134_cast_fp16)[name = string("transpose_760")];
            tensor<fp16, [1, 101, 128]> p_head_173_cast_fp16 = reshape(shape = var_7137, x = var_7136_cast_fp16)[name = string("p_head_173_cast_fp16")];
            tensor<fp16, [128]> var_7139_to_fp16 = const()[name = string("op_7139_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164001024)))];
            tensor<fp16, [1, 51, 128]> var_7140_cast_fp16 = add(x = q_head_87_cast_fp16, y = var_7139_to_fp16)[name = string("op_7140_cast_fp16")];
            tensor<int32, [1]> q_u_87_axes_1 = const()[name = string("q_u_87_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_87_cast_fp16 = expand_dims(axes = q_u_87_axes_1, x = var_7140_cast_fp16)[name = string("q_u_87_cast_fp16")];
            tensor<fp16, [128]> var_7142_to_fp16 = const()[name = string("op_7142_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164001344)))];
            tensor<fp16, [1, 51, 128]> var_7143_cast_fp16 = add(x = q_head_87_cast_fp16, y = var_7142_to_fp16)[name = string("op_7143_cast_fp16")];
            tensor<int32, [1]> q_v_87_axes_1 = const()[name = string("q_v_87_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_87_cast_fp16 = expand_dims(axes = q_v_87_axes_1, x = var_7143_cast_fp16)[name = string("q_v_87_cast_fp16")];
            tensor<int32, [1]> k_head_175_axes_1 = const()[name = string("k_head_175_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_175_cast_fp16 = expand_dims(axes = k_head_175_axes_1, x = k_head_173_cast_fp16)[name = string("k_head_175_cast_fp16")];
            tensor<int32, [1]> v_head_175_axes_1 = const()[name = string("v_head_175_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_175_cast_fp16 = expand_dims(axes = v_head_175_axes_1, x = v_head_173_cast_fp16)[name = string("v_head_175_cast_fp16")];
            tensor<int32, [1]> p_head_175_axes_1 = const()[name = string("p_head_175_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_175_cast_fp16 = expand_dims(axes = p_head_175_axes_1, x = p_head_173_cast_fp16)[name = string("p_head_175_cast_fp16")];
            bool var_7149_transpose_x_3 = const()[name = string("op_7149_transpose_x_3"), val = bool(false)];
            bool var_7149_transpose_y_3 = const()[name = string("op_7149_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_7149_cast_fp16 = matmul(transpose_x = var_7149_transpose_x_3, transpose_y = var_7149_transpose_y_3, x = q_u_87_cast_fp16, y = k_head_175_cast_fp16)[name = string("op_7149_cast_fp16")];
            fp16 var_7150_to_fp16 = const()[name = string("op_7150_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_87_cast_fp16 = mul(x = var_7149_cast_fp16, y = var_7150_to_fp16)[name = string("scores_content_87_cast_fp16")];
            bool x_465_transpose_x_3 = const()[name = string("x_465_transpose_x_3"), val = bool(false)];
            bool x_465_transpose_y_3 = const()[name = string("x_465_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_465_cast_fp16 = matmul(transpose_x = x_465_transpose_x_3, transpose_y = x_465_transpose_y_3, x = q_v_87_cast_fp16, y = p_head_175_cast_fp16)[name = string("x_465_cast_fp16")];
            tensor<int32, [8]> x_467_pad_1 = const()[name = string("x_467_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_467_mode_1 = const()[name = string("x_467_mode_1"), val = string("constant")];
            fp16 const_1597_to_fp16 = const()[name = string("const_1597_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_467_cast_fp16 = pad(constant_val = const_1597_to_fp16, mode = x_467_mode_1, pad = x_467_pad_1, x = x_465_cast_fp16)[name = string("x_467_cast_fp16")];
            tensor<int32, [4]> var_7164 = const()[name = string("op_7164"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_469_cast_fp16 = reshape(shape = var_7164, x = x_467_cast_fp16)[name = string("x_469_cast_fp16")];
            tensor<int32, [4]> var_7168_begin_1 = const()[name = string("op_7168_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7168_end_1 = const()[name = string("op_7168_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_7168_end_mask_1 = const()[name = string("op_7168_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_7168_cast_fp16 = slice_by_index(begin = var_7168_begin_1, end = var_7168_end_1, end_mask = var_7168_end_mask_1, x = x_469_cast_fp16)[name = string("op_7168_cast_fp16")];
            tensor<int32, [4]> var_7170 = const()[name = string("op_7170"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_7171_cast_fp16 = reshape(shape = var_7170, x = var_7168_cast_fp16)[name = string("op_7171_cast_fp16")];
            tensor<int32, [4]> var_7176_begin_1 = const()[name = string("op_7176_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7176_end_1 = const()[name = string("op_7176_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_7176_end_mask_1 = const()[name = string("op_7176_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_7176_cast_fp16 = slice_by_index(begin = var_7176_begin_1, end = var_7176_end_1, end_mask = var_7176_end_mask_1, x = var_7171_cast_fp16)[name = string("op_7176_cast_fp16")];
            fp16 var_7177_to_fp16 = const()[name = string("op_7177_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_87_cast_fp16 = mul(x = var_7176_cast_fp16, y = var_7177_to_fp16)[name = string("scores_pos_87_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_87_cast_fp16 = add(x = scores_content_87_cast_fp16, y = scores_pos_87_cast_fp16)[name = string("logits_87_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_7180_cast_fp16 = softmax(axis = var_6542, x = logits_87_cast_fp16)[name = string("op_7180_cast_fp16")];
            bool var_7182_transpose_x_1 = const()[name = string("op_7182_transpose_x_1"), val = bool(false)];
            bool var_7182_transpose_y_1 = const()[name = string("op_7182_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_7182_cast_fp16 = matmul(transpose_x = var_7182_transpose_x_1, transpose_y = var_7182_transpose_y_1, x = var_7180_cast_fp16, y = v_head_175_cast_fp16)[name = string("op_7182_cast_fp16")];
            tensor<int32, [1]> var_7183_axes_1 = const()[name = string("op_7183_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_7183_cast_fp16 = squeeze(axes = var_7183_axes_1, x = var_7182_cast_fp16)[name = string("op_7183_cast_fp16")];
            string dense_output_449_pad_type_1 = const()[name = string("dense_output_449_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_449_strides_1 = const()[name = string("dense_output_449_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_449_pad_1 = const()[name = string("dense_output_449_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_449_dilations_1 = const()[name = string("dense_output_449_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_449_groups_1 = const()[name = string("dense_output_449_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164001664))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164132800))))[name = string("layers_5_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_449_cast_fp16 = conv(dilations = dense_output_449_dilations_1, groups = dense_output_449_groups_1, pad = dense_output_449_pad_1, pad_type = dense_output_449_pad_type_1, strides = dense_output_449_strides_1, weight = layers_5_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_449_cast_fp16")];
            string sparse_output_449_pad_type_1 = const()[name = string("sparse_output_449_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_449_strides_1 = const()[name = string("sparse_output_449_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_449_pad_1 = const()[name = string("sparse_output_449_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_449_dilations_1 = const()[name = string("sparse_output_449_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_449_groups_1 = const()[name = string("sparse_output_449_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164136064))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164133376))))[name = string("layers_5_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_449_cast_fp16 = conv(dilations = sparse_output_449_dilations_1, groups = sparse_output_449_groups_1, pad = sparse_output_449_pad_1, pad_type = sparse_output_449_pad_type_1, strides = sparse_output_449_strides_1, weight = layers_5_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_449_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7198_cast_fp16 = add(x = dense_output_449_cast_fp16, y = sparse_output_449_cast_fp16)[name = string("op_7198_cast_fp16")];
            tensor<int32, [4]> var_7199 = const()[name = string("op_7199"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7201 = const()[name = string("op_7201"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7200_cast_fp16 = transpose(perm = var_7199, x = var_7198_cast_fp16)[name = string("transpose_759")];
            tensor<fp16, [1, 51, 128]> q_head_89_cast_fp16 = reshape(shape = var_7201, x = var_7200_cast_fp16)[name = string("q_head_89_cast_fp16")];
            string dense_output_451_pad_type_1 = const()[name = string("dense_output_451_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_451_strides_1 = const()[name = string("dense_output_451_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_451_pad_1 = const()[name = string("dense_output_451_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_451_dilations_1 = const()[name = string("dense_output_451_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_451_groups_1 = const()[name = string("dense_output_451_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164152512))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164283648))))[name = string("layers_5_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_451_cast_fp16 = conv(dilations = dense_output_451_dilations_1, groups = dense_output_451_groups_1, pad = dense_output_451_pad_1, pad_type = dense_output_451_pad_type_1, strides = dense_output_451_strides_1, weight = layers_5_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_451_cast_fp16")];
            string sparse_output_451_pad_type_1 = const()[name = string("sparse_output_451_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_451_strides_1 = const()[name = string("sparse_output_451_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_451_pad_1 = const()[name = string("sparse_output_451_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_451_dilations_1 = const()[name = string("sparse_output_451_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_451_groups_1 = const()[name = string("sparse_output_451_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164286912))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164284224))))[name = string("layers_5_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_451_cast_fp16 = conv(dilations = sparse_output_451_dilations_1, groups = sparse_output_451_groups_1, pad = sparse_output_451_pad_1, pad_type = sparse_output_451_pad_type_1, strides = sparse_output_451_strides_1, weight = layers_5_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_451_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7217_cast_fp16 = add(x = dense_output_451_cast_fp16, y = sparse_output_451_cast_fp16)[name = string("op_7217_cast_fp16")];
            tensor<int32, [4]> var_7218 = const()[name = string("op_7218"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7220 = const()[name = string("op_7220"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7219_cast_fp16 = transpose(perm = var_7218, x = var_7217_cast_fp16)[name = string("transpose_758")];
            tensor<fp16, [1, 51, 128]> k_head_177_cast_fp16 = reshape(shape = var_7220, x = var_7219_cast_fp16)[name = string("k_head_177_cast_fp16")];
            string dense_output_453_pad_type_1 = const()[name = string("dense_output_453_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_453_strides_1 = const()[name = string("dense_output_453_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_453_pad_1 = const()[name = string("dense_output_453_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_453_dilations_1 = const()[name = string("dense_output_453_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_453_groups_1 = const()[name = string("dense_output_453_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164303360))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164434496))))[name = string("layers_5_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_453_cast_fp16 = conv(dilations = dense_output_453_dilations_1, groups = dense_output_453_groups_1, pad = dense_output_453_pad_1, pad_type = dense_output_453_pad_type_1, strides = dense_output_453_strides_1, weight = layers_5_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_453_cast_fp16")];
            string sparse_output_453_pad_type_1 = const()[name = string("sparse_output_453_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_453_strides_1 = const()[name = string("sparse_output_453_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_453_pad_1 = const()[name = string("sparse_output_453_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_453_dilations_1 = const()[name = string("sparse_output_453_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_453_groups_1 = const()[name = string("sparse_output_453_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164437760))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164435072))))[name = string("layers_5_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_453_cast_fp16 = conv(dilations = sparse_output_453_dilations_1, groups = sparse_output_453_groups_1, pad = sparse_output_453_pad_1, pad_type = sparse_output_453_pad_type_1, strides = sparse_output_453_strides_1, weight = layers_5_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_453_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7236_cast_fp16 = add(x = dense_output_453_cast_fp16, y = sparse_output_453_cast_fp16)[name = string("op_7236_cast_fp16")];
            tensor<int32, [4]> var_7237 = const()[name = string("op_7237"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7239 = const()[name = string("op_7239"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7238_cast_fp16 = transpose(perm = var_7237, x = var_7236_cast_fp16)[name = string("transpose_757")];
            tensor<fp16, [1, 51, 128]> v_head_177_cast_fp16 = reshape(shape = var_7239, x = var_7238_cast_fp16)[name = string("v_head_177_cast_fp16")];
            string dense_output_455_pad_type_1 = const()[name = string("dense_output_455_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_455_strides_1 = const()[name = string("dense_output_455_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_455_pad_1 = const()[name = string("dense_output_455_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_455_dilations_1 = const()[name = string("dense_output_455_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_455_groups_1 = const()[name = string("dense_output_455_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164454208))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164585344))))[name = string("layers_5_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_455_cast_fp16 = conv(dilations = dense_output_455_dilations_1, groups = dense_output_455_groups_1, pad = dense_output_455_pad_1, pad_type = dense_output_455_pad_type_1, strides = dense_output_455_strides_1, weight = layers_5_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_455_cast_fp16")];
            string sparse_output_455_pad_type_1 = const()[name = string("sparse_output_455_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_455_strides_1 = const()[name = string("sparse_output_455_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_455_pad_1 = const()[name = string("sparse_output_455_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_455_dilations_1 = const()[name = string("sparse_output_455_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_455_groups_1 = const()[name = string("sparse_output_455_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164588608))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164585920))))[name = string("layers_5_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_455_cast_fp16 = conv(dilations = sparse_output_455_dilations_1, groups = sparse_output_455_groups_1, pad = sparse_output_455_pad_1, pad_type = sparse_output_455_pad_type_1, strides = sparse_output_455_strides_1, weight = layers_5_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_455_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_7255_cast_fp16 = add(x = dense_output_455_cast_fp16, y = sparse_output_455_cast_fp16)[name = string("op_7255_cast_fp16")];
            tensor<int32, [4]> var_7256 = const()[name = string("op_7256"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7258 = const()[name = string("op_7258"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_7257_cast_fp16 = transpose(perm = var_7256, x = var_7255_cast_fp16)[name = string("transpose_756")];
            tensor<fp16, [1, 101, 128]> p_head_177_cast_fp16 = reshape(shape = var_7258, x = var_7257_cast_fp16)[name = string("p_head_177_cast_fp16")];
            tensor<fp16, [128]> var_7260_to_fp16 = const()[name = string("op_7260_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164605056)))];
            tensor<fp16, [1, 51, 128]> var_7261_cast_fp16 = add(x = q_head_89_cast_fp16, y = var_7260_to_fp16)[name = string("op_7261_cast_fp16")];
            tensor<int32, [1]> q_u_89_axes_1 = const()[name = string("q_u_89_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_89_cast_fp16 = expand_dims(axes = q_u_89_axes_1, x = var_7261_cast_fp16)[name = string("q_u_89_cast_fp16")];
            tensor<fp16, [128]> var_7263_to_fp16 = const()[name = string("op_7263_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164605376)))];
            tensor<fp16, [1, 51, 128]> var_7264_cast_fp16 = add(x = q_head_89_cast_fp16, y = var_7263_to_fp16)[name = string("op_7264_cast_fp16")];
            tensor<int32, [1]> q_v_89_axes_1 = const()[name = string("q_v_89_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_89_cast_fp16 = expand_dims(axes = q_v_89_axes_1, x = var_7264_cast_fp16)[name = string("q_v_89_cast_fp16")];
            tensor<int32, [1]> k_head_179_axes_1 = const()[name = string("k_head_179_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_179_cast_fp16 = expand_dims(axes = k_head_179_axes_1, x = k_head_177_cast_fp16)[name = string("k_head_179_cast_fp16")];
            tensor<int32, [1]> v_head_179_axes_1 = const()[name = string("v_head_179_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_179_cast_fp16 = expand_dims(axes = v_head_179_axes_1, x = v_head_177_cast_fp16)[name = string("v_head_179_cast_fp16")];
            tensor<int32, [1]> p_head_179_axes_1 = const()[name = string("p_head_179_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_179_cast_fp16 = expand_dims(axes = p_head_179_axes_1, x = p_head_177_cast_fp16)[name = string("p_head_179_cast_fp16")];
            bool var_7270_transpose_x_3 = const()[name = string("op_7270_transpose_x_3"), val = bool(false)];
            bool var_7270_transpose_y_3 = const()[name = string("op_7270_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_7270_cast_fp16 = matmul(transpose_x = var_7270_transpose_x_3, transpose_y = var_7270_transpose_y_3, x = q_u_89_cast_fp16, y = k_head_179_cast_fp16)[name = string("op_7270_cast_fp16")];
            fp16 var_7271_to_fp16 = const()[name = string("op_7271_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_89_cast_fp16 = mul(x = var_7270_cast_fp16, y = var_7271_to_fp16)[name = string("scores_content_89_cast_fp16")];
            bool x_473_transpose_x_3 = const()[name = string("x_473_transpose_x_3"), val = bool(false)];
            bool x_473_transpose_y_3 = const()[name = string("x_473_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_473_cast_fp16 = matmul(transpose_x = x_473_transpose_x_3, transpose_y = x_473_transpose_y_3, x = q_v_89_cast_fp16, y = p_head_179_cast_fp16)[name = string("x_473_cast_fp16")];
            tensor<int32, [8]> x_475_pad_1 = const()[name = string("x_475_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_475_mode_1 = const()[name = string("x_475_mode_1"), val = string("constant")];
            fp16 const_1603_to_fp16 = const()[name = string("const_1603_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_475_cast_fp16 = pad(constant_val = const_1603_to_fp16, mode = x_475_mode_1, pad = x_475_pad_1, x = x_473_cast_fp16)[name = string("x_475_cast_fp16")];
            tensor<int32, [4]> var_7285 = const()[name = string("op_7285"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_477_cast_fp16 = reshape(shape = var_7285, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")];
            tensor<int32, [4]> var_7289_begin_1 = const()[name = string("op_7289_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7289_end_1 = const()[name = string("op_7289_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_7289_end_mask_1 = const()[name = string("op_7289_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_7289_cast_fp16 = slice_by_index(begin = var_7289_begin_1, end = var_7289_end_1, end_mask = var_7289_end_mask_1, x = x_477_cast_fp16)[name = string("op_7289_cast_fp16")];
            tensor<int32, [4]> var_7291 = const()[name = string("op_7291"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_7292_cast_fp16 = reshape(shape = var_7291, x = var_7289_cast_fp16)[name = string("op_7292_cast_fp16")];
            tensor<int32, [4]> var_7297_begin_1 = const()[name = string("op_7297_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7297_end_1 = const()[name = string("op_7297_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_7297_end_mask_1 = const()[name = string("op_7297_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_7297_cast_fp16 = slice_by_index(begin = var_7297_begin_1, end = var_7297_end_1, end_mask = var_7297_end_mask_1, x = var_7292_cast_fp16)[name = string("op_7297_cast_fp16")];
            fp16 var_7298_to_fp16 = const()[name = string("op_7298_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_89_cast_fp16 = mul(x = var_7297_cast_fp16, y = var_7298_to_fp16)[name = string("scores_pos_89_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_89_cast_fp16 = add(x = scores_content_89_cast_fp16, y = scores_pos_89_cast_fp16)[name = string("logits_89_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_7301_cast_fp16 = softmax(axis = var_6542, x = logits_89_cast_fp16)[name = string("op_7301_cast_fp16")];
            bool var_7303_transpose_x_1 = const()[name = string("op_7303_transpose_x_1"), val = bool(false)];
            bool var_7303_transpose_y_1 = const()[name = string("op_7303_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_7303_cast_fp16 = matmul(transpose_x = var_7303_transpose_x_1, transpose_y = var_7303_transpose_y_1, x = var_7301_cast_fp16, y = v_head_179_cast_fp16)[name = string("op_7303_cast_fp16")];
            tensor<int32, [1]> var_7304_axes_1 = const()[name = string("op_7304_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_7304_cast_fp16 = squeeze(axes = var_7304_axes_1, x = var_7303_cast_fp16)[name = string("op_7304_cast_fp16")];
            string dense_output_457_pad_type_1 = const()[name = string("dense_output_457_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_457_strides_1 = const()[name = string("dense_output_457_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_457_pad_1 = const()[name = string("dense_output_457_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_457_dilations_1 = const()[name = string("dense_output_457_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_457_groups_1 = const()[name = string("dense_output_457_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164605696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164736832))))[name = string("layers_5_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_457_cast_fp16 = conv(dilations = dense_output_457_dilations_1, groups = dense_output_457_groups_1, pad = dense_output_457_pad_1, pad_type = dense_output_457_pad_type_1, strides = dense_output_457_strides_1, weight = layers_5_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_457_cast_fp16")];
            string sparse_output_457_pad_type_1 = const()[name = string("sparse_output_457_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_457_strides_1 = const()[name = string("sparse_output_457_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_457_pad_1 = const()[name = string("sparse_output_457_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_457_dilations_1 = const()[name = string("sparse_output_457_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_457_groups_1 = const()[name = string("sparse_output_457_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164740096))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164737408))))[name = string("layers_5_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_457_cast_fp16 = conv(dilations = sparse_output_457_dilations_1, groups = sparse_output_457_groups_1, pad = sparse_output_457_pad_1, pad_type = sparse_output_457_pad_type_1, strides = sparse_output_457_strides_1, weight = layers_5_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_457_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7319_cast_fp16 = add(x = dense_output_457_cast_fp16, y = sparse_output_457_cast_fp16)[name = string("op_7319_cast_fp16")];
            tensor<int32, [4]> var_7320 = const()[name = string("op_7320"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7322 = const()[name = string("op_7322"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7321_cast_fp16 = transpose(perm = var_7320, x = var_7319_cast_fp16)[name = string("transpose_755")];
            tensor<fp16, [1, 51, 128]> q_head_91_cast_fp16 = reshape(shape = var_7322, x = var_7321_cast_fp16)[name = string("q_head_91_cast_fp16")];
            string dense_output_459_pad_type_1 = const()[name = string("dense_output_459_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_459_strides_1 = const()[name = string("dense_output_459_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_459_pad_1 = const()[name = string("dense_output_459_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_459_dilations_1 = const()[name = string("dense_output_459_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_459_groups_1 = const()[name = string("dense_output_459_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164756544))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164887680))))[name = string("layers_5_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_459_cast_fp16 = conv(dilations = dense_output_459_dilations_1, groups = dense_output_459_groups_1, pad = dense_output_459_pad_1, pad_type = dense_output_459_pad_type_1, strides = dense_output_459_strides_1, weight = layers_5_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_459_cast_fp16")];
            string sparse_output_459_pad_type_1 = const()[name = string("sparse_output_459_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_459_strides_1 = const()[name = string("sparse_output_459_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_459_pad_1 = const()[name = string("sparse_output_459_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_459_dilations_1 = const()[name = string("sparse_output_459_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_459_groups_1 = const()[name = string("sparse_output_459_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164890944))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164888256))))[name = string("layers_5_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_459_cast_fp16 = conv(dilations = sparse_output_459_dilations_1, groups = sparse_output_459_groups_1, pad = sparse_output_459_pad_1, pad_type = sparse_output_459_pad_type_1, strides = sparse_output_459_strides_1, weight = layers_5_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_459_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7338_cast_fp16 = add(x = dense_output_459_cast_fp16, y = sparse_output_459_cast_fp16)[name = string("op_7338_cast_fp16")];
            tensor<int32, [4]> var_7339 = const()[name = string("op_7339"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7341 = const()[name = string("op_7341"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7340_cast_fp16 = transpose(perm = var_7339, x = var_7338_cast_fp16)[name = string("transpose_754")];
            tensor<fp16, [1, 51, 128]> k_head_181_cast_fp16 = reshape(shape = var_7341, x = var_7340_cast_fp16)[name = string("k_head_181_cast_fp16")];
            string dense_output_461_pad_type_1 = const()[name = string("dense_output_461_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_461_strides_1 = const()[name = string("dense_output_461_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_461_pad_1 = const()[name = string("dense_output_461_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_461_dilations_1 = const()[name = string("dense_output_461_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_461_groups_1 = const()[name = string("dense_output_461_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164907392))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165038528))))[name = string("layers_5_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_461_cast_fp16 = conv(dilations = dense_output_461_dilations_1, groups = dense_output_461_groups_1, pad = dense_output_461_pad_1, pad_type = dense_output_461_pad_type_1, strides = dense_output_461_strides_1, weight = layers_5_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_461_cast_fp16")];
            string sparse_output_461_pad_type_1 = const()[name = string("sparse_output_461_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_461_strides_1 = const()[name = string("sparse_output_461_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_461_pad_1 = const()[name = string("sparse_output_461_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_461_dilations_1 = const()[name = string("sparse_output_461_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_461_groups_1 = const()[name = string("sparse_output_461_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165041792))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165039104))))[name = string("layers_5_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_461_cast_fp16 = conv(dilations = sparse_output_461_dilations_1, groups = sparse_output_461_groups_1, pad = sparse_output_461_pad_1, pad_type = sparse_output_461_pad_type_1, strides = sparse_output_461_strides_1, weight = layers_5_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_461_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7357_cast_fp16 = add(x = dense_output_461_cast_fp16, y = sparse_output_461_cast_fp16)[name = string("op_7357_cast_fp16")];
            tensor<int32, [4]> var_7358 = const()[name = string("op_7358"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7360 = const()[name = string("op_7360"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7359_cast_fp16 = transpose(perm = var_7358, x = var_7357_cast_fp16)[name = string("transpose_753")];
            tensor<fp16, [1, 51, 128]> v_head_181_cast_fp16 = reshape(shape = var_7360, x = var_7359_cast_fp16)[name = string("v_head_181_cast_fp16")];
            string dense_output_463_pad_type_1 = const()[name = string("dense_output_463_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_463_strides_1 = const()[name = string("dense_output_463_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_463_pad_1 = const()[name = string("dense_output_463_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_463_dilations_1 = const()[name = string("dense_output_463_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_463_groups_1 = const()[name = string("dense_output_463_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165058240))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165189376))))[name = string("layers_5_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_463_cast_fp16 = conv(dilations = dense_output_463_dilations_1, groups = dense_output_463_groups_1, pad = dense_output_463_pad_1, pad_type = dense_output_463_pad_type_1, strides = dense_output_463_strides_1, weight = layers_5_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_463_cast_fp16")];
            string sparse_output_463_pad_type_1 = const()[name = string("sparse_output_463_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_463_strides_1 = const()[name = string("sparse_output_463_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_463_pad_1 = const()[name = string("sparse_output_463_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_463_dilations_1 = const()[name = string("sparse_output_463_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_463_groups_1 = const()[name = string("sparse_output_463_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165192640))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165189952))))[name = string("layers_5_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_463_cast_fp16 = conv(dilations = sparse_output_463_dilations_1, groups = sparse_output_463_groups_1, pad = sparse_output_463_pad_1, pad_type = sparse_output_463_pad_type_1, strides = sparse_output_463_strides_1, weight = layers_5_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_463_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_7376_cast_fp16 = add(x = dense_output_463_cast_fp16, y = sparse_output_463_cast_fp16)[name = string("op_7376_cast_fp16")];
            tensor<int32, [4]> var_7377 = const()[name = string("op_7377"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7379 = const()[name = string("op_7379"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_7378_cast_fp16 = transpose(perm = var_7377, x = var_7376_cast_fp16)[name = string("transpose_752")];
            tensor<fp16, [1, 101, 128]> p_head_181_cast_fp16 = reshape(shape = var_7379, x = var_7378_cast_fp16)[name = string("p_head_181_cast_fp16")];
            tensor<fp16, [128]> var_7381_to_fp16 = const()[name = string("op_7381_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165209088)))];
            tensor<fp16, [1, 51, 128]> var_7382_cast_fp16 = add(x = q_head_91_cast_fp16, y = var_7381_to_fp16)[name = string("op_7382_cast_fp16")];
            tensor<int32, [1]> q_u_91_axes_1 = const()[name = string("q_u_91_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_91_cast_fp16 = expand_dims(axes = q_u_91_axes_1, x = var_7382_cast_fp16)[name = string("q_u_91_cast_fp16")];
            tensor<fp16, [128]> var_7384_to_fp16 = const()[name = string("op_7384_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165209408)))];
            tensor<fp16, [1, 51, 128]> var_7385_cast_fp16 = add(x = q_head_91_cast_fp16, y = var_7384_to_fp16)[name = string("op_7385_cast_fp16")];
            tensor<int32, [1]> q_v_91_axes_1 = const()[name = string("q_v_91_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_91_cast_fp16 = expand_dims(axes = q_v_91_axes_1, x = var_7385_cast_fp16)[name = string("q_v_91_cast_fp16")];
            tensor<int32, [1]> k_head_183_axes_1 = const()[name = string("k_head_183_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_183_cast_fp16 = expand_dims(axes = k_head_183_axes_1, x = k_head_181_cast_fp16)[name = string("k_head_183_cast_fp16")];
            tensor<int32, [1]> v_head_183_axes_1 = const()[name = string("v_head_183_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_183_cast_fp16 = expand_dims(axes = v_head_183_axes_1, x = v_head_181_cast_fp16)[name = string("v_head_183_cast_fp16")];
            tensor<int32, [1]> p_head_183_axes_1 = const()[name = string("p_head_183_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_183_cast_fp16 = expand_dims(axes = p_head_183_axes_1, x = p_head_181_cast_fp16)[name = string("p_head_183_cast_fp16")];
            bool var_7391_transpose_x_3 = const()[name = string("op_7391_transpose_x_3"), val = bool(false)];
            bool var_7391_transpose_y_3 = const()[name = string("op_7391_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_7391_cast_fp16 = matmul(transpose_x = var_7391_transpose_x_3, transpose_y = var_7391_transpose_y_3, x = q_u_91_cast_fp16, y = k_head_183_cast_fp16)[name = string("op_7391_cast_fp16")];
            fp16 var_7392_to_fp16 = const()[name = string("op_7392_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_91_cast_fp16 = mul(x = var_7391_cast_fp16, y = var_7392_to_fp16)[name = string("scores_content_91_cast_fp16")];
            bool x_481_transpose_x_3 = const()[name = string("x_481_transpose_x_3"), val = bool(false)];
            bool x_481_transpose_y_3 = const()[name = string("x_481_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_3, transpose_y = x_481_transpose_y_3, x = q_v_91_cast_fp16, y = p_head_183_cast_fp16)[name = string("x_481_cast_fp16")];
            tensor<int32, [8]> x_483_pad_1 = const()[name = string("x_483_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_483_mode_1 = const()[name = string("x_483_mode_1"), val = string("constant")];
            fp16 const_1609_to_fp16 = const()[name = string("const_1609_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_483_cast_fp16 = pad(constant_val = const_1609_to_fp16, mode = x_483_mode_1, pad = x_483_pad_1, x = x_481_cast_fp16)[name = string("x_483_cast_fp16")];
            tensor<int32, [4]> var_7406 = const()[name = string("op_7406"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_485_cast_fp16 = reshape(shape = var_7406, x = x_483_cast_fp16)[name = string("x_485_cast_fp16")];
            tensor<int32, [4]> var_7410_begin_1 = const()[name = string("op_7410_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7410_end_1 = const()[name = string("op_7410_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_7410_end_mask_1 = const()[name = string("op_7410_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_7410_cast_fp16 = slice_by_index(begin = var_7410_begin_1, end = var_7410_end_1, end_mask = var_7410_end_mask_1, x = x_485_cast_fp16)[name = string("op_7410_cast_fp16")];
            tensor<int32, [4]> var_7412 = const()[name = string("op_7412"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_7413_cast_fp16 = reshape(shape = var_7412, x = var_7410_cast_fp16)[name = string("op_7413_cast_fp16")];
            tensor<int32, [4]> var_7418_begin_1 = const()[name = string("op_7418_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7418_end_1 = const()[name = string("op_7418_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_7418_end_mask_1 = const()[name = string("op_7418_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_7418_cast_fp16 = slice_by_index(begin = var_7418_begin_1, end = var_7418_end_1, end_mask = var_7418_end_mask_1, x = var_7413_cast_fp16)[name = string("op_7418_cast_fp16")];
            fp16 var_7419_to_fp16 = const()[name = string("op_7419_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_91_cast_fp16 = mul(x = var_7418_cast_fp16, y = var_7419_to_fp16)[name = string("scores_pos_91_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_91_cast_fp16 = add(x = scores_content_91_cast_fp16, y = scores_pos_91_cast_fp16)[name = string("logits_91_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_7422_cast_fp16 = softmax(axis = var_6542, x = logits_91_cast_fp16)[name = string("op_7422_cast_fp16")];
            bool var_7424_transpose_x_1 = const()[name = string("op_7424_transpose_x_1"), val = bool(false)];
            bool var_7424_transpose_y_1 = const()[name = string("op_7424_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_7424_cast_fp16 = matmul(transpose_x = var_7424_transpose_x_1, transpose_y = var_7424_transpose_y_1, x = var_7422_cast_fp16, y = v_head_183_cast_fp16)[name = string("op_7424_cast_fp16")];
            tensor<int32, [1]> var_7425_axes_1 = const()[name = string("op_7425_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_7425_cast_fp16 = squeeze(axes = var_7425_axes_1, x = var_7424_cast_fp16)[name = string("op_7425_cast_fp16")];
            string dense_output_465_pad_type_1 = const()[name = string("dense_output_465_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_465_strides_1 = const()[name = string("dense_output_465_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_465_pad_1 = const()[name = string("dense_output_465_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_465_dilations_1 = const()[name = string("dense_output_465_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_465_groups_1 = const()[name = string("dense_output_465_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165209728))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165340864))))[name = string("layers_5_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_465_cast_fp16 = conv(dilations = dense_output_465_dilations_1, groups = dense_output_465_groups_1, pad = dense_output_465_pad_1, pad_type = dense_output_465_pad_type_1, strides = dense_output_465_strides_1, weight = layers_5_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_465_cast_fp16")];
            string sparse_output_465_pad_type_1 = const()[name = string("sparse_output_465_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_465_strides_1 = const()[name = string("sparse_output_465_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_465_pad_1 = const()[name = string("sparse_output_465_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_465_dilations_1 = const()[name = string("sparse_output_465_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_465_groups_1 = const()[name = string("sparse_output_465_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165344128))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165341440))))[name = string("layers_5_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_465_cast_fp16 = conv(dilations = sparse_output_465_dilations_1, groups = sparse_output_465_groups_1, pad = sparse_output_465_pad_1, pad_type = sparse_output_465_pad_type_1, strides = sparse_output_465_strides_1, weight = layers_5_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_465_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7440_cast_fp16 = add(x = dense_output_465_cast_fp16, y = sparse_output_465_cast_fp16)[name = string("op_7440_cast_fp16")];
            tensor<int32, [4]> var_7441 = const()[name = string("op_7441"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7443 = const()[name = string("op_7443"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7442_cast_fp16 = transpose(perm = var_7441, x = var_7440_cast_fp16)[name = string("transpose_751")];
            tensor<fp16, [1, 51, 128]> q_head_93_cast_fp16 = reshape(shape = var_7443, x = var_7442_cast_fp16)[name = string("q_head_93_cast_fp16")];
            string dense_output_467_pad_type_1 = const()[name = string("dense_output_467_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_467_strides_1 = const()[name = string("dense_output_467_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_467_pad_1 = const()[name = string("dense_output_467_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_467_dilations_1 = const()[name = string("dense_output_467_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_467_groups_1 = const()[name = string("dense_output_467_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165360576))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165491712))))[name = string("layers_5_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_467_cast_fp16 = conv(dilations = dense_output_467_dilations_1, groups = dense_output_467_groups_1, pad = dense_output_467_pad_1, pad_type = dense_output_467_pad_type_1, strides = dense_output_467_strides_1, weight = layers_5_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_467_cast_fp16")];
            string sparse_output_467_pad_type_1 = const()[name = string("sparse_output_467_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_467_strides_1 = const()[name = string("sparse_output_467_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_467_pad_1 = const()[name = string("sparse_output_467_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_467_dilations_1 = const()[name = string("sparse_output_467_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_467_groups_1 = const()[name = string("sparse_output_467_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165494976))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165492288))))[name = string("layers_5_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_467_cast_fp16 = conv(dilations = sparse_output_467_dilations_1, groups = sparse_output_467_groups_1, pad = sparse_output_467_pad_1, pad_type = sparse_output_467_pad_type_1, strides = sparse_output_467_strides_1, weight = layers_5_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_467_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7459_cast_fp16 = add(x = dense_output_467_cast_fp16, y = sparse_output_467_cast_fp16)[name = string("op_7459_cast_fp16")];
            tensor<int32, [4]> var_7460 = const()[name = string("op_7460"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7462 = const()[name = string("op_7462"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7461_cast_fp16 = transpose(perm = var_7460, x = var_7459_cast_fp16)[name = string("transpose_750")];
            tensor<fp16, [1, 51, 128]> k_head_185_cast_fp16 = reshape(shape = var_7462, x = var_7461_cast_fp16)[name = string("k_head_185_cast_fp16")];
            string dense_output_469_pad_type_1 = const()[name = string("dense_output_469_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_469_strides_1 = const()[name = string("dense_output_469_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_469_pad_1 = const()[name = string("dense_output_469_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_469_dilations_1 = const()[name = string("dense_output_469_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_469_groups_1 = const()[name = string("dense_output_469_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165511424))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165642560))))[name = string("layers_5_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_469_cast_fp16 = conv(dilations = dense_output_469_dilations_1, groups = dense_output_469_groups_1, pad = dense_output_469_pad_1, pad_type = dense_output_469_pad_type_1, strides = dense_output_469_strides_1, weight = layers_5_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_469_cast_fp16")];
            string sparse_output_469_pad_type_1 = const()[name = string("sparse_output_469_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_469_strides_1 = const()[name = string("sparse_output_469_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_469_pad_1 = const()[name = string("sparse_output_469_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_469_dilations_1 = const()[name = string("sparse_output_469_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_469_groups_1 = const()[name = string("sparse_output_469_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165645824))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165643136))))[name = string("layers_5_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_469_cast_fp16 = conv(dilations = sparse_output_469_dilations_1, groups = sparse_output_469_groups_1, pad = sparse_output_469_pad_1, pad_type = sparse_output_469_pad_type_1, strides = sparse_output_469_strides_1, weight = layers_5_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_469_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7478_cast_fp16 = add(x = dense_output_469_cast_fp16, y = sparse_output_469_cast_fp16)[name = string("op_7478_cast_fp16")];
            tensor<int32, [4]> var_7479 = const()[name = string("op_7479"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7481 = const()[name = string("op_7481"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7480_cast_fp16 = transpose(perm = var_7479, x = var_7478_cast_fp16)[name = string("transpose_749")];
            tensor<fp16, [1, 51, 128]> v_head_185_cast_fp16 = reshape(shape = var_7481, x = var_7480_cast_fp16)[name = string("v_head_185_cast_fp16")];
            string dense_output_471_pad_type_1 = const()[name = string("dense_output_471_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_471_strides_1 = const()[name = string("dense_output_471_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_471_pad_1 = const()[name = string("dense_output_471_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_471_dilations_1 = const()[name = string("dense_output_471_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_471_groups_1 = const()[name = string("dense_output_471_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165662272))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165793408))))[name = string("layers_5_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_471_cast_fp16 = conv(dilations = dense_output_471_dilations_1, groups = dense_output_471_groups_1, pad = dense_output_471_pad_1, pad_type = dense_output_471_pad_type_1, strides = dense_output_471_strides_1, weight = layers_5_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_471_cast_fp16")];
            string sparse_output_471_pad_type_1 = const()[name = string("sparse_output_471_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_471_strides_1 = const()[name = string("sparse_output_471_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_471_pad_1 = const()[name = string("sparse_output_471_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_471_dilations_1 = const()[name = string("sparse_output_471_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_471_groups_1 = const()[name = string("sparse_output_471_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165796672))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165793984))))[name = string("layers_5_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_471_cast_fp16 = conv(dilations = sparse_output_471_dilations_1, groups = sparse_output_471_groups_1, pad = sparse_output_471_pad_1, pad_type = sparse_output_471_pad_type_1, strides = sparse_output_471_strides_1, weight = layers_5_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_471_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_7497_cast_fp16 = add(x = dense_output_471_cast_fp16, y = sparse_output_471_cast_fp16)[name = string("op_7497_cast_fp16")];
            tensor<int32, [4]> var_7498 = const()[name = string("op_7498"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7500 = const()[name = string("op_7500"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_7499_cast_fp16 = transpose(perm = var_7498, x = var_7497_cast_fp16)[name = string("transpose_748")];
            tensor<fp16, [1, 101, 128]> p_head_185_cast_fp16 = reshape(shape = var_7500, x = var_7499_cast_fp16)[name = string("p_head_185_cast_fp16")];
            tensor<fp16, [128]> var_7502_to_fp16 = const()[name = string("op_7502_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165813120)))];
            tensor<fp16, [1, 51, 128]> var_7503_cast_fp16 = add(x = q_head_93_cast_fp16, y = var_7502_to_fp16)[name = string("op_7503_cast_fp16")];
            tensor<int32, [1]> q_u_93_axes_1 = const()[name = string("q_u_93_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_93_cast_fp16 = expand_dims(axes = q_u_93_axes_1, x = var_7503_cast_fp16)[name = string("q_u_93_cast_fp16")];
            tensor<fp16, [128]> var_7505_to_fp16 = const()[name = string("op_7505_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165813440)))];
            tensor<fp16, [1, 51, 128]> var_7506_cast_fp16 = add(x = q_head_93_cast_fp16, y = var_7505_to_fp16)[name = string("op_7506_cast_fp16")];
            tensor<int32, [1]> q_v_93_axes_1 = const()[name = string("q_v_93_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_93_cast_fp16 = expand_dims(axes = q_v_93_axes_1, x = var_7506_cast_fp16)[name = string("q_v_93_cast_fp16")];
            tensor<int32, [1]> k_head_187_axes_1 = const()[name = string("k_head_187_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_187_cast_fp16 = expand_dims(axes = k_head_187_axes_1, x = k_head_185_cast_fp16)[name = string("k_head_187_cast_fp16")];
            tensor<int32, [1]> v_head_187_axes_1 = const()[name = string("v_head_187_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_187_cast_fp16 = expand_dims(axes = v_head_187_axes_1, x = v_head_185_cast_fp16)[name = string("v_head_187_cast_fp16")];
            tensor<int32, [1]> p_head_187_axes_1 = const()[name = string("p_head_187_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_187_cast_fp16 = expand_dims(axes = p_head_187_axes_1, x = p_head_185_cast_fp16)[name = string("p_head_187_cast_fp16")];
            bool var_7512_transpose_x_3 = const()[name = string("op_7512_transpose_x_3"), val = bool(false)];
            bool var_7512_transpose_y_3 = const()[name = string("op_7512_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_7512_cast_fp16 = matmul(transpose_x = var_7512_transpose_x_3, transpose_y = var_7512_transpose_y_3, x = q_u_93_cast_fp16, y = k_head_187_cast_fp16)[name = string("op_7512_cast_fp16")];
            fp16 var_7513_to_fp16 = const()[name = string("op_7513_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_93_cast_fp16 = mul(x = var_7512_cast_fp16, y = var_7513_to_fp16)[name = string("scores_content_93_cast_fp16")];
            bool x_489_transpose_x_3 = const()[name = string("x_489_transpose_x_3"), val = bool(false)];
            bool x_489_transpose_y_3 = const()[name = string("x_489_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_489_cast_fp16 = matmul(transpose_x = x_489_transpose_x_3, transpose_y = x_489_transpose_y_3, x = q_v_93_cast_fp16, y = p_head_187_cast_fp16)[name = string("x_489_cast_fp16")];
            tensor<int32, [8]> x_491_pad_1 = const()[name = string("x_491_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_491_mode_1 = const()[name = string("x_491_mode_1"), val = string("constant")];
            fp16 const_1615_to_fp16 = const()[name = string("const_1615_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_491_cast_fp16 = pad(constant_val = const_1615_to_fp16, mode = x_491_mode_1, pad = x_491_pad_1, x = x_489_cast_fp16)[name = string("x_491_cast_fp16")];
            tensor<int32, [4]> var_7527 = const()[name = string("op_7527"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_493_cast_fp16 = reshape(shape = var_7527, x = x_491_cast_fp16)[name = string("x_493_cast_fp16")];
            tensor<int32, [4]> var_7531_begin_1 = const()[name = string("op_7531_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7531_end_1 = const()[name = string("op_7531_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_7531_end_mask_1 = const()[name = string("op_7531_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_7531_cast_fp16 = slice_by_index(begin = var_7531_begin_1, end = var_7531_end_1, end_mask = var_7531_end_mask_1, x = x_493_cast_fp16)[name = string("op_7531_cast_fp16")];
            tensor<int32, [4]> var_7533 = const()[name = string("op_7533"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_7534_cast_fp16 = reshape(shape = var_7533, x = var_7531_cast_fp16)[name = string("op_7534_cast_fp16")];
            tensor<int32, [4]> var_7539_begin_1 = const()[name = string("op_7539_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7539_end_1 = const()[name = string("op_7539_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_7539_end_mask_1 = const()[name = string("op_7539_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_7539_cast_fp16 = slice_by_index(begin = var_7539_begin_1, end = var_7539_end_1, end_mask = var_7539_end_mask_1, x = var_7534_cast_fp16)[name = string("op_7539_cast_fp16")];
            fp16 var_7540_to_fp16 = const()[name = string("op_7540_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_93_cast_fp16 = mul(x = var_7539_cast_fp16, y = var_7540_to_fp16)[name = string("scores_pos_93_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_93_cast_fp16 = add(x = scores_content_93_cast_fp16, y = scores_pos_93_cast_fp16)[name = string("logits_93_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_7543_cast_fp16 = softmax(axis = var_6542, x = logits_93_cast_fp16)[name = string("op_7543_cast_fp16")];
            bool var_7545_transpose_x_1 = const()[name = string("op_7545_transpose_x_1"), val = bool(false)];
            bool var_7545_transpose_y_1 = const()[name = string("op_7545_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_7545_cast_fp16 = matmul(transpose_x = var_7545_transpose_x_1, transpose_y = var_7545_transpose_y_1, x = var_7543_cast_fp16, y = v_head_187_cast_fp16)[name = string("op_7545_cast_fp16")];
            tensor<int32, [1]> var_7546_axes_1 = const()[name = string("op_7546_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_7546_cast_fp16 = squeeze(axes = var_7546_axes_1, x = var_7545_cast_fp16)[name = string("op_7546_cast_fp16")];
            string dense_output_473_pad_type_1 = const()[name = string("dense_output_473_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_473_strides_1 = const()[name = string("dense_output_473_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_473_pad_1 = const()[name = string("dense_output_473_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_473_dilations_1 = const()[name = string("dense_output_473_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_473_groups_1 = const()[name = string("dense_output_473_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165813760))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165944896))))[name = string("layers_5_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_473_cast_fp16 = conv(dilations = dense_output_473_dilations_1, groups = dense_output_473_groups_1, pad = dense_output_473_pad_1, pad_type = dense_output_473_pad_type_1, strides = dense_output_473_strides_1, weight = layers_5_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_473_cast_fp16")];
            string sparse_output_473_pad_type_1 = const()[name = string("sparse_output_473_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_473_strides_1 = const()[name = string("sparse_output_473_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_473_pad_1 = const()[name = string("sparse_output_473_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_473_dilations_1 = const()[name = string("sparse_output_473_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_473_groups_1 = const()[name = string("sparse_output_473_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165948160))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165945472))))[name = string("layers_5_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_473_cast_fp16 = conv(dilations = sparse_output_473_dilations_1, groups = sparse_output_473_groups_1, pad = sparse_output_473_pad_1, pad_type = sparse_output_473_pad_type_1, strides = sparse_output_473_strides_1, weight = layers_5_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_473_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7561_cast_fp16 = add(x = dense_output_473_cast_fp16, y = sparse_output_473_cast_fp16)[name = string("op_7561_cast_fp16")];
            tensor<int32, [4]> var_7562 = const()[name = string("op_7562"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7564 = const()[name = string("op_7564"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7563_cast_fp16 = transpose(perm = var_7562, x = var_7561_cast_fp16)[name = string("transpose_747")];
            tensor<fp16, [1, 51, 128]> q_head_95_cast_fp16 = reshape(shape = var_7564, x = var_7563_cast_fp16)[name = string("q_head_95_cast_fp16")];
            string dense_output_475_pad_type_1 = const()[name = string("dense_output_475_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_475_strides_1 = const()[name = string("dense_output_475_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_475_pad_1 = const()[name = string("dense_output_475_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_475_dilations_1 = const()[name = string("dense_output_475_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_475_groups_1 = const()[name = string("dense_output_475_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165964608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166095744))))[name = string("layers_5_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_475_cast_fp16 = conv(dilations = dense_output_475_dilations_1, groups = dense_output_475_groups_1, pad = dense_output_475_pad_1, pad_type = dense_output_475_pad_type_1, strides = dense_output_475_strides_1, weight = layers_5_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_475_cast_fp16")];
            string sparse_output_475_pad_type_1 = const()[name = string("sparse_output_475_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_475_strides_1 = const()[name = string("sparse_output_475_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_475_pad_1 = const()[name = string("sparse_output_475_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_475_dilations_1 = const()[name = string("sparse_output_475_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_475_groups_1 = const()[name = string("sparse_output_475_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166099008))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166096320))))[name = string("layers_5_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_475_cast_fp16 = conv(dilations = sparse_output_475_dilations_1, groups = sparse_output_475_groups_1, pad = sparse_output_475_pad_1, pad_type = sparse_output_475_pad_type_1, strides = sparse_output_475_strides_1, weight = layers_5_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_475_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7580_cast_fp16 = add(x = dense_output_475_cast_fp16, y = sparse_output_475_cast_fp16)[name = string("op_7580_cast_fp16")];
            tensor<int32, [4]> var_7581 = const()[name = string("op_7581"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7583 = const()[name = string("op_7583"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7582_cast_fp16 = transpose(perm = var_7581, x = var_7580_cast_fp16)[name = string("transpose_746")];
            tensor<fp16, [1, 51, 128]> k_head_189_cast_fp16 = reshape(shape = var_7583, x = var_7582_cast_fp16)[name = string("k_head_189_cast_fp16")];
            string dense_output_477_pad_type_1 = const()[name = string("dense_output_477_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_477_strides_1 = const()[name = string("dense_output_477_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_477_pad_1 = const()[name = string("dense_output_477_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_477_dilations_1 = const()[name = string("dense_output_477_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_477_groups_1 = const()[name = string("dense_output_477_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166115456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166246592))))[name = string("layers_5_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_477_cast_fp16 = conv(dilations = dense_output_477_dilations_1, groups = dense_output_477_groups_1, pad = dense_output_477_pad_1, pad_type = dense_output_477_pad_type_1, strides = dense_output_477_strides_1, weight = layers_5_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = string("dense_output_477_cast_fp16")];
            string sparse_output_477_pad_type_1 = const()[name = string("sparse_output_477_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_477_strides_1 = const()[name = string("sparse_output_477_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_477_pad_1 = const()[name = string("sparse_output_477_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_477_dilations_1 = const()[name = string("sparse_output_477_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_477_groups_1 = const()[name = string("sparse_output_477_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166249856))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166247168))))[name = string("layers_5_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_477_cast_fp16 = conv(dilations = sparse_output_477_dilations_1, groups = sparse_output_477_groups_1, pad = sparse_output_477_pad_1, pad_type = sparse_output_477_pad_type_1, strides = sparse_output_477_strides_1, weight = layers_5_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_259_cast_fp16)[name = string("sparse_output_477_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7599_cast_fp16 = add(x = dense_output_477_cast_fp16, y = sparse_output_477_cast_fp16)[name = string("op_7599_cast_fp16")];
            tensor<int32, [4]> var_7600 = const()[name = string("op_7600"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7602 = const()[name = string("op_7602"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7601_cast_fp16 = transpose(perm = var_7600, x = var_7599_cast_fp16)[name = string("transpose_745")];
            tensor<fp16, [1, 51, 128]> v_head_189_cast_fp16 = reshape(shape = var_7602, x = var_7601_cast_fp16)[name = string("v_head_189_cast_fp16")];
            string dense_output_479_pad_type_1 = const()[name = string("dense_output_479_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_479_strides_1 = const()[name = string("dense_output_479_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_479_pad_1 = const()[name = string("dense_output_479_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_479_dilations_1 = const()[name = string("dense_output_479_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_479_groups_1 = const()[name = string("dense_output_479_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166266304))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166397440))))[name = string("layers_5_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_479_cast_fp16 = conv(dilations = dense_output_479_dilations_1, groups = dense_output_479_groups_1, pad = dense_output_479_pad_1, pad_type = dense_output_479_pad_type_1, strides = dense_output_479_strides_1, weight = layers_5_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_479_cast_fp16")];
            string sparse_output_479_pad_type_1 = const()[name = string("sparse_output_479_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_479_strides_1 = const()[name = string("sparse_output_479_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_479_pad_1 = const()[name = string("sparse_output_479_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_479_dilations_1 = const()[name = string("sparse_output_479_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_479_groups_1 = const()[name = string("sparse_output_479_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_5_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166400704))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166398016))))[name = string("layers_5_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_479_cast_fp16 = conv(dilations = sparse_output_479_dilations_1, groups = sparse_output_479_groups_1, pad = sparse_output_479_pad_1, pad_type = sparse_output_479_pad_type_1, strides = sparse_output_479_strides_1, weight = layers_5_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_479_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_7618_cast_fp16 = add(x = dense_output_479_cast_fp16, y = sparse_output_479_cast_fp16)[name = string("op_7618_cast_fp16")];
            tensor<int32, [4]> var_7619 = const()[name = string("op_7619"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7621 = const()[name = string("op_7621"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_7620_cast_fp16 = transpose(perm = var_7619, x = var_7618_cast_fp16)[name = string("transpose_744")];
            tensor<fp16, [1, 101, 128]> p_head_189_cast_fp16 = reshape(shape = var_7621, x = var_7620_cast_fp16)[name = string("p_head_189_cast_fp16")];
            tensor<fp16, [128]> var_7623_to_fp16 = const()[name = string("op_7623_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166417152)))];
            tensor<fp16, [1, 51, 128]> var_7624_cast_fp16 = add(x = q_head_95_cast_fp16, y = var_7623_to_fp16)[name = string("op_7624_cast_fp16")];
            tensor<int32, [1]> q_u_95_axes_1 = const()[name = string("q_u_95_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_95_cast_fp16 = expand_dims(axes = q_u_95_axes_1, x = var_7624_cast_fp16)[name = string("q_u_95_cast_fp16")];
            tensor<fp16, [128]> var_7626_to_fp16 = const()[name = string("op_7626_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166417472)))];
            tensor<fp16, [1, 51, 128]> var_7627_cast_fp16 = add(x = q_head_95_cast_fp16, y = var_7626_to_fp16)[name = string("op_7627_cast_fp16")];
            tensor<int32, [1]> q_v_95_axes_1 = const()[name = string("q_v_95_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_95_cast_fp16 = expand_dims(axes = q_v_95_axes_1, x = var_7627_cast_fp16)[name = string("q_v_95_cast_fp16")];
            tensor<int32, [1]> k_head_191_axes_1 = const()[name = string("k_head_191_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_191_cast_fp16 = expand_dims(axes = k_head_191_axes_1, x = k_head_189_cast_fp16)[name = string("k_head_191_cast_fp16")];
            tensor<int32, [1]> v_head_191_axes_1 = const()[name = string("v_head_191_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_191_cast_fp16 = expand_dims(axes = v_head_191_axes_1, x = v_head_189_cast_fp16)[name = string("v_head_191_cast_fp16")];
            tensor<int32, [1]> p_head_191_axes_1 = const()[name = string("p_head_191_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_191_cast_fp16 = expand_dims(axes = p_head_191_axes_1, x = p_head_189_cast_fp16)[name = string("p_head_191_cast_fp16")];
            bool var_7633_transpose_x_3 = const()[name = string("op_7633_transpose_x_3"), val = bool(false)];
            bool var_7633_transpose_y_3 = const()[name = string("op_7633_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_7633_cast_fp16 = matmul(transpose_x = var_7633_transpose_x_3, transpose_y = var_7633_transpose_y_3, x = q_u_95_cast_fp16, y = k_head_191_cast_fp16)[name = string("op_7633_cast_fp16")];
            fp16 var_7634_to_fp16 = const()[name = string("op_7634_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_95_cast_fp16 = mul(x = var_7633_cast_fp16, y = var_7634_to_fp16)[name = string("scores_content_95_cast_fp16")];
            bool x_497_transpose_x_3 = const()[name = string("x_497_transpose_x_3"), val = bool(false)];
            bool x_497_transpose_y_3 = const()[name = string("x_497_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_497_cast_fp16 = matmul(transpose_x = x_497_transpose_x_3, transpose_y = x_497_transpose_y_3, x = q_v_95_cast_fp16, y = p_head_191_cast_fp16)[name = string("x_497_cast_fp16")];
            tensor<int32, [8]> x_499_pad_1 = const()[name = string("x_499_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_499_mode_1 = const()[name = string("x_499_mode_1"), val = string("constant")];
            fp16 const_1621_to_fp16 = const()[name = string("const_1621_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_499_cast_fp16 = pad(constant_val = const_1621_to_fp16, mode = x_499_mode_1, pad = x_499_pad_1, x = x_497_cast_fp16)[name = string("x_499_cast_fp16")];
            tensor<int32, [4]> var_7648 = const()[name = string("op_7648"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_501_cast_fp16 = reshape(shape = var_7648, x = x_499_cast_fp16)[name = string("x_501_cast_fp16")];
            tensor<int32, [4]> var_7652_begin_1 = const()[name = string("op_7652_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_7652_end_1 = const()[name = string("op_7652_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_7652_end_mask_1 = const()[name = string("op_7652_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_7652_cast_fp16 = slice_by_index(begin = var_7652_begin_1, end = var_7652_end_1, end_mask = var_7652_end_mask_1, x = x_501_cast_fp16)[name = string("op_7652_cast_fp16")];
            tensor<int32, [4]> var_7654 = const()[name = string("op_7654"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_7655_cast_fp16 = reshape(shape = var_7654, x = var_7652_cast_fp16)[name = string("op_7655_cast_fp16")];
            tensor<int32, [4]> var_7660_begin_1 = const()[name = string("op_7660_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7660_end_1 = const()[name = string("op_7660_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_7660_end_mask_1 = const()[name = string("op_7660_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_7660_cast_fp16 = slice_by_index(begin = var_7660_begin_1, end = var_7660_end_1, end_mask = var_7660_end_mask_1, x = var_7655_cast_fp16)[name = string("op_7660_cast_fp16")];
            fp16 var_7661_to_fp16 = const()[name = string("op_7661_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_95_cast_fp16 = mul(x = var_7660_cast_fp16, y = var_7661_to_fp16)[name = string("scores_pos_95_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_95_cast_fp16 = add(x = scores_content_95_cast_fp16, y = scores_pos_95_cast_fp16)[name = string("logits_95_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_7664_cast_fp16 = softmax(axis = var_6542, x = logits_95_cast_fp16)[name = string("op_7664_cast_fp16")];
            bool var_7666_transpose_x_1 = const()[name = string("op_7666_transpose_x_1"), val = bool(false)];
            bool var_7666_transpose_y_1 = const()[name = string("op_7666_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_7666_cast_fp16 = matmul(transpose_x = var_7666_transpose_x_1, transpose_y = var_7666_transpose_y_1, x = var_7664_cast_fp16, y = v_head_191_cast_fp16)[name = string("op_7666_cast_fp16")];
            tensor<int32, [1]> o_head_11_axes_1 = const()[name = string("o_head_11_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_11_cast_fp16 = squeeze(axes = o_head_11_axes_1, x = var_7666_cast_fp16)[name = string("o_head_11_cast_fp16")];
            bool out_11_interleave_1 = const()[name = string("out_11_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_11_cast_fp16 = concat(axis = var_6542, interleave = out_11_interleave_1, values = (var_6820_cast_fp16, var_6941_cast_fp16, var_7062_cast_fp16, var_7183_cast_fp16, var_7304_cast_fp16, var_7425_cast_fp16, var_7546_cast_fp16, o_head_11_cast_fp16))[name = string("out_11_cast_fp16")];
            tensor<int32, [3]> var_7670_perm_1 = const()[name = string("op_7670_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_267_axes_1 = const()[name = string("input_267_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_7670_cast_fp16 = transpose(perm = var_7670_perm_1, x = out_11_cast_fp16)[name = string("transpose_743")];
            tensor<fp16, [1, 1024, 51, 1]> input_267_cast_fp16 = expand_dims(axes = input_267_axes_1, x = var_7670_cast_fp16)[name = string("input_267_cast_fp16")];
            string dense_output_481_pad_type_1 = const()[name = string("dense_output_481_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_481_strides_1 = const()[name = string("dense_output_481_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_481_pad_1 = const()[name = string("dense_output_481_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_481_dilations_1 = const()[name = string("dense_output_481_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_481_groups_1 = const()[name = string("dense_output_481_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166417792))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167466432))))[name = string("layers_5_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_481_cast_fp16 = conv(dilations = dense_output_481_dilations_1, groups = dense_output_481_groups_1, pad = dense_output_481_pad_1, pad_type = dense_output_481_pad_type_1, strides = dense_output_481_strides_1, weight = layers_5_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = string("dense_output_481_cast_fp16")];
            string sparse_output_481_pad_type_1 = const()[name = string("sparse_output_481_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_481_strides_1 = const()[name = string("sparse_output_481_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_481_pad_1 = const()[name = string("sparse_output_481_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_481_dilations_1 = const()[name = string("sparse_output_481_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_481_groups_1 = const()[name = string("sparse_output_481_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167488064))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167467008))))[name = string("layers_5_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_481_cast_fp16 = conv(dilations = sparse_output_481_dilations_1, groups = sparse_output_481_groups_1, pad = sparse_output_481_pad_1, pad_type = sparse_output_481_pad_type_1, strides = sparse_output_481_strides_1, weight = layers_5_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_267_cast_fp16)[name = string("sparse_output_481_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_11_cast_fp16 = add(x = dense_output_481_cast_fp16, y = sparse_output_481_cast_fp16)[name = string("out_conv_11_cast_fp16")];
            tensor<int32, [1]> var_7687_axes_1 = const()[name = string("op_7687_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_7687_cast_fp16 = squeeze(axes = var_7687_axes_1, x = out_conv_11_cast_fp16)[name = string("op_7687_cast_fp16")];
            tensor<int32, [3]> var_7688_perm_1 = const()[name = string("op_7688_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_7688_cast_fp16 = transpose(perm = var_7688_perm_1, x = var_7687_cast_fp16)[name = string("transpose_742")];
            tensor<fp16, [1, 51, 1024]> input_269_cast_fp16 = add(x = input_257_cast_fp16, y = var_7688_cast_fp16)[name = string("input_269_cast_fp16")];
            tensor<int32, [1]> x_505_axes_1 = const()[name = string("x_505_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_5_norm_conv_weight_to_fp16 = const()[name = string("layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167619200)))];
            tensor<fp16, [1024]> layers_5_norm_conv_bias_to_fp16 = const()[name = string("layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167621312)))];
            tensor<fp16, [1, 51, 1024]> x_505_cast_fp16 = layer_norm(axes = x_505_axes_1, beta = layers_5_norm_conv_bias_to_fp16, epsilon = var_6557_to_fp16, gamma = layers_5_norm_conv_weight_to_fp16, x = input_269_cast_fp16)[name = string("x_505_cast_fp16")];
            tensor<int32, [3]> var_7698_perm_1 = const()[name = string("op_7698_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_271_axes_1 = const()[name = string("input_271_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_7698_cast_fp16 = transpose(perm = var_7698_perm_1, x = x_505_cast_fp16)[name = string("transpose_741")];
            tensor<fp16, [1, 1024, 51, 1]> input_271_cast_fp16 = expand_dims(axes = input_271_axes_1, x = var_7698_cast_fp16)[name = string("input_271_cast_fp16")];
            string dense_output_483_pad_type_1 = const()[name = string("dense_output_483_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_483_strides_1 = const()[name = string("dense_output_483_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_483_pad_1 = const()[name = string("dense_output_483_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_483_dilations_1 = const()[name = string("dense_output_483_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_483_groups_1 = const()[name = string("dense_output_483_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_5_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167623424))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169720640))))[name = string("layers_5_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_483_cast_fp16 = conv(dilations = dense_output_483_dilations_1, groups = dense_output_483_groups_1, pad = dense_output_483_pad_1, pad_type = dense_output_483_pad_type_1, strides = dense_output_483_strides_1, weight = layers_5_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_271_cast_fp16)[name = string("dense_output_483_cast_fp16")];
            string sparse_output_483_pad_type_1 = const()[name = string("sparse_output_483_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_483_strides_1 = const()[name = string("sparse_output_483_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_483_pad_1 = const()[name = string("sparse_output_483_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_483_dilations_1 = const()[name = string("sparse_output_483_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_483_groups_1 = const()[name = string("sparse_output_483_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_5_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169763264))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169721216))))[name = string("layers_5_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_483_cast_fp16 = conv(dilations = sparse_output_483_dilations_1, groups = sparse_output_483_groups_1, pad = sparse_output_483_pad_1, pad_type = sparse_output_483_pad_type_1, strides = sparse_output_483_strides_1, weight = layers_5_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_271_cast_fp16)[name = string("sparse_output_483_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_273_cast_fp16 = add(x = dense_output_483_cast_fp16, y = sparse_output_483_cast_fp16)[name = string("input_273_cast_fp16")];
            int32 input_275_split_num_splits_1 = const()[name = string("input_275_split_num_splits_1"), val = int32(2)];
            int32 input_275_split_axis_1 = const()[name = string("input_275_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_275_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_275_split_cast_fp16_1 = split(axis = input_275_split_axis_1, num_splits = input_275_split_num_splits_1, x = input_273_cast_fp16)[name = string("input_275_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_275_split_1_sigmoid_cast_fp16 = sigmoid(x = input_275_split_cast_fp16_1)[name = string("input_275_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_275_cast_fp16 = mul(x = input_275_split_cast_fp16_0, y = input_275_split_1_sigmoid_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<int32, [8]> input_277_pad_1 = const()[name = string("input_277_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_277_mode_1 = const()[name = string("input_277_mode_1"), val = string("constant")];
            fp16 const_1623_to_fp16 = const()[name = string("const_1623_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_277_cast_fp16 = pad(constant_val = const_1623_to_fp16, mode = input_277_mode_1, pad = input_277_pad_1, x = input_275_cast_fp16)[name = string("input_277_cast_fp16")];
            string dense_output_485_pad_type_1 = const()[name = string("dense_output_485_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_485_strides_1 = const()[name = string("dense_output_485_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_485_pad_1 = const()[name = string("dense_output_485_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_485_dilations_1 = const()[name = string("dense_output_485_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_485_groups_1 = const()[name = string("dense_output_485_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_485_cast_fp16 = conv(dilations = dense_output_485_dilations_1, groups = dense_output_485_groups_1, pad = dense_output_485_pad_1, pad_type = dense_output_485_pad_type_1, strides = dense_output_485_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_277_cast_fp16)[name = string("dense_output_485_cast_fp16")];
            string sparse_output_485_pad_type_1 = const()[name = string("sparse_output_485_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_485_strides_1 = const()[name = string("sparse_output_485_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_485_pad_1 = const()[name = string("sparse_output_485_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_485_dilations_1 = const()[name = string("sparse_output_485_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_485_groups_1 = const()[name = string("sparse_output_485_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_5_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170025472))))[name = string("layers_5_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_485_cast_fp16 = conv(dilations = sparse_output_485_dilations_1, groups = sparse_output_485_groups_1, pad = sparse_output_485_pad_1, pad_type = sparse_output_485_pad_type_1, strides = sparse_output_485_strides_1, weight = layers_5_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_277_cast_fp16)[name = string("sparse_output_485_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_279_cast_fp16 = add(x = dense_output_485_cast_fp16, y = sparse_output_485_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [1024]> layers_5_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_5_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170043968)))];
            tensor<fp16, [1024]> layers_5_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_5_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170046080)))];
            tensor<fp16, [1024]> layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170048192)))];
            tensor<fp16, [1024]> layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170050304)))];
            tensor<fp16, [1, 1024, 51, 1]> input_281_cast_fp16 = batch_norm(beta = layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_6557_to_fp16, gamma = layers_5_conv_batch_norm_weight_to_fp16, mean = layers_5_conv_batch_norm_running_mean_to_fp16, variance = layers_5_conv_batch_norm_running_var_to_fp16, x = input_279_cast_fp16)[name = string("input_281_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_283_cast_fp16 = silu(x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
            string dense_output_487_pad_type_1 = const()[name = string("dense_output_487_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_487_strides_1 = const()[name = string("dense_output_487_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_487_pad_1 = const()[name = string("dense_output_487_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_487_dilations_1 = const()[name = string("dense_output_487_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_487_groups_1 = const()[name = string("dense_output_487_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170052416))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171101056))))[name = string("layers_5_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_487_cast_fp16 = conv(dilations = dense_output_487_dilations_1, groups = dense_output_487_groups_1, pad = dense_output_487_pad_1, pad_type = dense_output_487_pad_type_1, strides = dense_output_487_strides_1, weight = layers_5_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_283_cast_fp16)[name = string("dense_output_487_cast_fp16")];
            string sparse_output_487_pad_type_1 = const()[name = string("sparse_output_487_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_487_strides_1 = const()[name = string("sparse_output_487_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_487_pad_1 = const()[name = string("sparse_output_487_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_487_dilations_1 = const()[name = string("sparse_output_487_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_487_groups_1 = const()[name = string("sparse_output_487_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_5_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171122688))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171101632))))[name = string("layers_5_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_487_cast_fp16 = conv(dilations = sparse_output_487_dilations_1, groups = sparse_output_487_groups_1, pad = sparse_output_487_pad_1, pad_type = sparse_output_487_pad_type_1, strides = sparse_output_487_strides_1, weight = layers_5_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_283_cast_fp16)[name = string("sparse_output_487_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_507_cast_fp16 = add(x = dense_output_487_cast_fp16, y = sparse_output_487_cast_fp16)[name = string("x_507_cast_fp16")];
            tensor<int32, [1]> var_7754_axes_1 = const()[name = string("op_7754_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_7754_cast_fp16 = squeeze(axes = var_7754_axes_1, x = x_507_cast_fp16)[name = string("op_7754_cast_fp16")];
            tensor<int32, [3]> var_7755_perm_1 = const()[name = string("op_7755_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_7755_cast_fp16 = transpose(perm = var_7755_perm_1, x = var_7754_cast_fp16)[name = string("transpose_740")];
            tensor<fp16, [1, 51, 1024]> input_285_cast_fp16 = add(x = input_269_cast_fp16, y = var_7755_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<int32, [1]> x_509_axes_1 = const()[name = string("x_509_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171253824)))];
            tensor<fp16, [1024]> layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171255936)))];
            tensor<fp16, [1, 51, 1024]> x_509_cast_fp16 = layer_norm(axes = x_509_axes_1, beta = layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_6557_to_fp16, gamma = layers_5_norm_feed_forward2_weight_to_fp16, x = input_285_cast_fp16)[name = string("x_509_cast_fp16")];
            tensor<int32, [4]> var_7765 = const()[name = string("op_7765"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_511_cast_fp16 = reshape(shape = var_7765, x = x_509_cast_fp16)[name = string("x_511_cast_fp16")];
            tensor<int32, [4]> input_287_perm_1 = const()[name = string("input_287_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_489_pad_type_1 = const()[name = string("dense_output_489_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_489_strides_1 = const()[name = string("dense_output_489_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_489_pad_1 = const()[name = string("dense_output_489_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_489_dilations_1 = const()[name = string("dense_output_489_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_489_groups_1 = const()[name = string("dense_output_489_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_5_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171258048))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175452416))))[name = string("layers_5_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_287_cast_fp16 = transpose(perm = input_287_perm_1, x = x_511_cast_fp16)[name = string("transpose_739")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_489_cast_fp16 = conv(dilations = dense_output_489_dilations_1, groups = dense_output_489_groups_1, pad = dense_output_489_pad_1, pad_type = dense_output_489_pad_type_1, strides = dense_output_489_strides_1, weight = layers_5_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("dense_output_489_cast_fp16")];
            string sparse_output_489_pad_type_1 = const()[name = string("sparse_output_489_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_489_strides_1 = const()[name = string("sparse_output_489_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_489_pad_1 = const()[name = string("sparse_output_489_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_489_dilations_1 = const()[name = string("sparse_output_489_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_489_groups_1 = const()[name = string("sparse_output_489_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_5_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175536960))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175452992))))[name = string("layers_5_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_489_cast_fp16 = conv(dilations = sparse_output_489_dilations_1, groups = sparse_output_489_groups_1, pad = sparse_output_489_pad_1, pad_type = sparse_output_489_pad_type_1, strides = sparse_output_489_strides_1, weight = layers_5_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_287_cast_fp16)[name = string("sparse_output_489_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_289_cast_fp16 = add(x = dense_output_489_cast_fp16, y = sparse_output_489_cast_fp16)[name = string("input_289_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_291_cast_fp16 = silu(x = input_289_cast_fp16)[name = string("input_291_cast_fp16")];
            string dense_output_491_pad_type_1 = const()[name = string("dense_output_491_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_491_strides_1 = const()[name = string("dense_output_491_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_491_pad_1 = const()[name = string("dense_output_491_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_491_dilations_1 = const()[name = string("dense_output_491_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_491_groups_1 = const()[name = string("dense_output_491_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_5_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176061312))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180255680))))[name = string("layers_5_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_491_cast_fp16 = conv(dilations = dense_output_491_dilations_1, groups = dense_output_491_groups_1, pad = dense_output_491_pad_1, pad_type = dense_output_491_pad_type_1, strides = dense_output_491_strides_1, weight = layers_5_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("dense_output_491_cast_fp16")];
            string sparse_output_491_pad_type_1 = const()[name = string("sparse_output_491_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_491_strides_1 = const()[name = string("sparse_output_491_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_491_pad_1 = const()[name = string("sparse_output_491_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_491_dilations_1 = const()[name = string("sparse_output_491_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_491_groups_1 = const()[name = string("sparse_output_491_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_5_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180340224))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180256256))))[name = string("layers_5_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_491_cast_fp16 = conv(dilations = sparse_output_491_dilations_1, groups = sparse_output_491_groups_1, pad = sparse_output_491_pad_1, pad_type = sparse_output_491_pad_type_1, strides = sparse_output_491_strides_1, weight = layers_5_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_291_cast_fp16)[name = string("sparse_output_491_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_513_cast_fp16 = add(x = dense_output_491_cast_fp16, y = sparse_output_491_cast_fp16)[name = string("x_513_cast_fp16")];
            tensor<int32, [4]> x_515_perm_1 = const()[name = string("x_515_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_7800 = const()[name = string("op_7800"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_515_cast_fp16 = transpose(perm = x_515_perm_1, x = x_513_cast_fp16)[name = string("transpose_738")];
            tensor<fp16, [1, 51, 1024]> var_7801_cast_fp16 = reshape(shape = var_7800, x = x_515_cast_fp16)[name = string("op_7801_cast_fp16")];
            fp16 var_7802_to_fp16 = const()[name = string("op_7802_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_7803_cast_fp16 = mul(x = var_7801_cast_fp16, y = var_7802_to_fp16)[name = string("op_7803_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_293_cast_fp16 = add(x = input_285_cast_fp16, y = var_7803_cast_fp16)[name = string("input_293_cast_fp16")];
            tensor<int32, [1]> input_295_axes_1 = const()[name = string("input_295_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_5_norm_out_weight_to_fp16 = const()[name = string("layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180864576)))];
            tensor<fp16, [1024]> layers_5_norm_out_bias_to_fp16 = const()[name = string("layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180866688)))];
            tensor<fp16, [1, 51, 1024]> input_295_cast_fp16 = layer_norm(axes = input_295_axes_1, beta = layers_5_norm_out_bias_to_fp16, epsilon = var_6557_to_fp16, gamma = layers_5_norm_out_weight_to_fp16, x = input_293_cast_fp16)[name = string("input_295_cast_fp16")];
            int32 var_7811 = const()[name = string("op_7811"), val = int32(-1)];
            tensor<int32, [1]> x_517_axes_1 = const()[name = string("x_517_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180868800)))];
            tensor<fp16, [1024]> layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180870912)))];
            fp16 var_7826_to_fp16 = const()[name = string("op_7826_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_517_cast_fp16 = layer_norm(axes = x_517_axes_1, beta = layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_7826_to_fp16, gamma = layers_6_norm_feed_forward1_weight_to_fp16, x = input_295_cast_fp16)[name = string("x_517_cast_fp16")];
            tensor<int32, [4]> var_7845 = const()[name = string("op_7845"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_519_cast_fp16 = reshape(shape = var_7845, x = x_517_cast_fp16)[name = string("x_519_cast_fp16")];
            tensor<int32, [4]> input_297_perm_1 = const()[name = string("input_297_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_493_pad_type_1 = const()[name = string("dense_output_493_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_493_strides_1 = const()[name = string("dense_output_493_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_493_pad_1 = const()[name = string("dense_output_493_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_493_dilations_1 = const()[name = string("dense_output_493_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_493_groups_1 = const()[name = string("dense_output_493_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_6_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180873024))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185067392))))[name = string("layers_6_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_297_cast_fp16 = transpose(perm = input_297_perm_1, x = x_519_cast_fp16)[name = string("transpose_737")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_493_cast_fp16 = conv(dilations = dense_output_493_dilations_1, groups = dense_output_493_groups_1, pad = dense_output_493_pad_1, pad_type = dense_output_493_pad_type_1, strides = dense_output_493_strides_1, weight = layers_6_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_297_cast_fp16)[name = string("dense_output_493_cast_fp16")];
            string sparse_output_493_pad_type_1 = const()[name = string("sparse_output_493_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_493_strides_1 = const()[name = string("sparse_output_493_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_493_pad_1 = const()[name = string("sparse_output_493_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_493_dilations_1 = const()[name = string("sparse_output_493_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_493_groups_1 = const()[name = string("sparse_output_493_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_6_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185151936))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185067968))))[name = string("layers_6_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_493_cast_fp16 = conv(dilations = sparse_output_493_dilations_1, groups = sparse_output_493_groups_1, pad = sparse_output_493_pad_1, pad_type = sparse_output_493_pad_type_1, strides = sparse_output_493_strides_1, weight = layers_6_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_297_cast_fp16)[name = string("sparse_output_493_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_299_cast_fp16 = add(x = dense_output_493_cast_fp16, y = sparse_output_493_cast_fp16)[name = string("input_299_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_301_cast_fp16 = silu(x = input_299_cast_fp16)[name = string("input_301_cast_fp16")];
            string dense_output_495_pad_type_1 = const()[name = string("dense_output_495_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_495_strides_1 = const()[name = string("dense_output_495_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_495_pad_1 = const()[name = string("dense_output_495_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_495_dilations_1 = const()[name = string("dense_output_495_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_495_groups_1 = const()[name = string("dense_output_495_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_6_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185676288))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189870656))))[name = string("layers_6_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_495_cast_fp16 = conv(dilations = dense_output_495_dilations_1, groups = dense_output_495_groups_1, pad = dense_output_495_pad_1, pad_type = dense_output_495_pad_type_1, strides = dense_output_495_strides_1, weight = layers_6_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_301_cast_fp16)[name = string("dense_output_495_cast_fp16")];
            string sparse_output_495_pad_type_1 = const()[name = string("sparse_output_495_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_495_strides_1 = const()[name = string("sparse_output_495_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_495_pad_1 = const()[name = string("sparse_output_495_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_495_dilations_1 = const()[name = string("sparse_output_495_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_495_groups_1 = const()[name = string("sparse_output_495_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_6_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189955200))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189871232))))[name = string("layers_6_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_495_cast_fp16 = conv(dilations = sparse_output_495_dilations_1, groups = sparse_output_495_groups_1, pad = sparse_output_495_pad_1, pad_type = sparse_output_495_pad_type_1, strides = sparse_output_495_strides_1, weight = layers_6_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_301_cast_fp16)[name = string("sparse_output_495_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_521_cast_fp16 = add(x = dense_output_495_cast_fp16, y = sparse_output_495_cast_fp16)[name = string("x_521_cast_fp16")];
            tensor<int32, [4]> x_523_perm_1 = const()[name = string("x_523_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_7880 = const()[name = string("op_7880"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_523_cast_fp16 = transpose(perm = x_523_perm_1, x = x_521_cast_fp16)[name = string("transpose_736")];
            tensor<fp16, [1, 51, 1024]> var_7881_cast_fp16 = reshape(shape = var_7880, x = x_523_cast_fp16)[name = string("op_7881_cast_fp16")];
            fp16 var_7882_to_fp16 = const()[name = string("op_7882_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_7883_cast_fp16 = mul(x = var_7881_cast_fp16, y = var_7882_to_fp16)[name = string("op_7883_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_303_cast_fp16 = add(x = input_295_cast_fp16, y = var_7883_cast_fp16)[name = string("input_303_cast_fp16")];
            tensor<int32, [1]> q_13_axes_1 = const()[name = string("q_13_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_6_norm_self_att_weight_to_fp16 = const()[name = string("layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190479552)))];
            tensor<fp16, [1024]> layers_6_norm_self_att_bias_to_fp16 = const()[name = string("layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190481664)))];
            tensor<fp16, [1, 51, 1024]> q_13_cast_fp16 = layer_norm(axes = q_13_axes_1, beta = layers_6_norm_self_att_bias_to_fp16, epsilon = var_7826_to_fp16, gamma = layers_6_norm_self_att_weight_to_fp16, x = input_303_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<int32, [3]> var_7957 = const()[name = string("op_7957"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_305_axes_1 = const()[name = string("input_305_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_7958_cast_fp16 = transpose(perm = var_7957, x = q_13_cast_fp16)[name = string("transpose_735")];
            tensor<fp16, [1, 1024, 51, 1]> input_305_cast_fp16 = expand_dims(axes = input_305_axes_1, x = var_7958_cast_fp16)[name = string("input_305_cast_fp16")];
            string dense_output_497_pad_type_1 = const()[name = string("dense_output_497_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_497_strides_1 = const()[name = string("dense_output_497_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_497_pad_1 = const()[name = string("dense_output_497_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_497_dilations_1 = const()[name = string("dense_output_497_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_497_groups_1 = const()[name = string("dense_output_497_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190483776))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190614912))))[name = string("layers_6_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_497_cast_fp16 = conv(dilations = dense_output_497_dilations_1, groups = dense_output_497_groups_1, pad = dense_output_497_pad_1, pad_type = dense_output_497_pad_type_1, strides = dense_output_497_strides_1, weight = layers_6_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_497_cast_fp16")];
            string sparse_output_497_pad_type_1 = const()[name = string("sparse_output_497_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_497_strides_1 = const()[name = string("sparse_output_497_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_497_pad_1 = const()[name = string("sparse_output_497_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_497_dilations_1 = const()[name = string("sparse_output_497_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_497_groups_1 = const()[name = string("sparse_output_497_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190618176))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190615488))))[name = string("layers_6_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_497_cast_fp16 = conv(dilations = sparse_output_497_dilations_1, groups = sparse_output_497_groups_1, pad = sparse_output_497_pad_1, pad_type = sparse_output_497_pad_type_1, strides = sparse_output_497_strides_1, weight = layers_6_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_497_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_7983_cast_fp16 = add(x = dense_output_497_cast_fp16, y = sparse_output_497_cast_fp16)[name = string("op_7983_cast_fp16")];
            tensor<int32, [4]> var_7984 = const()[name = string("op_7984"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_7986 = const()[name = string("op_7986"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_7985_cast_fp16 = transpose(perm = var_7984, x = var_7983_cast_fp16)[name = string("transpose_734")];
            tensor<fp16, [1, 51, 128]> q_head_97_cast_fp16 = reshape(shape = var_7986, x = var_7985_cast_fp16)[name = string("q_head_97_cast_fp16")];
            string dense_output_499_pad_type_1 = const()[name = string("dense_output_499_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_499_strides_1 = const()[name = string("dense_output_499_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_499_pad_1 = const()[name = string("dense_output_499_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_499_dilations_1 = const()[name = string("dense_output_499_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_499_groups_1 = const()[name = string("dense_output_499_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190634624))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190765760))))[name = string("layers_6_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_499_cast_fp16 = conv(dilations = dense_output_499_dilations_1, groups = dense_output_499_groups_1, pad = dense_output_499_pad_1, pad_type = dense_output_499_pad_type_1, strides = dense_output_499_strides_1, weight = layers_6_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_499_cast_fp16")];
            string sparse_output_499_pad_type_1 = const()[name = string("sparse_output_499_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_499_strides_1 = const()[name = string("sparse_output_499_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_499_pad_1 = const()[name = string("sparse_output_499_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_499_dilations_1 = const()[name = string("sparse_output_499_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_499_groups_1 = const()[name = string("sparse_output_499_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190769024))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190766336))))[name = string("layers_6_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_499_cast_fp16 = conv(dilations = sparse_output_499_dilations_1, groups = sparse_output_499_groups_1, pad = sparse_output_499_pad_1, pad_type = sparse_output_499_pad_type_1, strides = sparse_output_499_strides_1, weight = layers_6_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_499_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8002_cast_fp16 = add(x = dense_output_499_cast_fp16, y = sparse_output_499_cast_fp16)[name = string("op_8002_cast_fp16")];
            tensor<int32, [4]> var_8003 = const()[name = string("op_8003"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8005 = const()[name = string("op_8005"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8004_cast_fp16 = transpose(perm = var_8003, x = var_8002_cast_fp16)[name = string("transpose_733")];
            tensor<fp16, [1, 51, 128]> k_head_193_cast_fp16 = reshape(shape = var_8005, x = var_8004_cast_fp16)[name = string("k_head_193_cast_fp16")];
            string dense_output_501_pad_type_1 = const()[name = string("dense_output_501_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_501_strides_1 = const()[name = string("dense_output_501_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_501_pad_1 = const()[name = string("dense_output_501_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_501_dilations_1 = const()[name = string("dense_output_501_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_501_groups_1 = const()[name = string("dense_output_501_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190785472))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190916608))))[name = string("layers_6_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_501_cast_fp16 = conv(dilations = dense_output_501_dilations_1, groups = dense_output_501_groups_1, pad = dense_output_501_pad_1, pad_type = dense_output_501_pad_type_1, strides = dense_output_501_strides_1, weight = layers_6_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_501_cast_fp16")];
            string sparse_output_501_pad_type_1 = const()[name = string("sparse_output_501_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_501_strides_1 = const()[name = string("sparse_output_501_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_501_pad_1 = const()[name = string("sparse_output_501_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_501_dilations_1 = const()[name = string("sparse_output_501_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_501_groups_1 = const()[name = string("sparse_output_501_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190919872))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190917184))))[name = string("layers_6_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_501_cast_fp16 = conv(dilations = sparse_output_501_dilations_1, groups = sparse_output_501_groups_1, pad = sparse_output_501_pad_1, pad_type = sparse_output_501_pad_type_1, strides = sparse_output_501_strides_1, weight = layers_6_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_501_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8021_cast_fp16 = add(x = dense_output_501_cast_fp16, y = sparse_output_501_cast_fp16)[name = string("op_8021_cast_fp16")];
            tensor<int32, [4]> var_8022 = const()[name = string("op_8022"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8024 = const()[name = string("op_8024"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8023_cast_fp16 = transpose(perm = var_8022, x = var_8021_cast_fp16)[name = string("transpose_732")];
            tensor<fp16, [1, 51, 128]> v_head_193_cast_fp16 = reshape(shape = var_8024, x = var_8023_cast_fp16)[name = string("v_head_193_cast_fp16")];
            string dense_output_503_pad_type_1 = const()[name = string("dense_output_503_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_503_strides_1 = const()[name = string("dense_output_503_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_503_pad_1 = const()[name = string("dense_output_503_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_503_dilations_1 = const()[name = string("dense_output_503_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_503_groups_1 = const()[name = string("dense_output_503_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190936320))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191067456))))[name = string("layers_6_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_503_cast_fp16 = conv(dilations = dense_output_503_dilations_1, groups = dense_output_503_groups_1, pad = dense_output_503_pad_1, pad_type = dense_output_503_pad_type_1, strides = dense_output_503_strides_1, weight = layers_6_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_503_cast_fp16")];
            string sparse_output_503_pad_type_1 = const()[name = string("sparse_output_503_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_503_strides_1 = const()[name = string("sparse_output_503_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_503_pad_1 = const()[name = string("sparse_output_503_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_503_dilations_1 = const()[name = string("sparse_output_503_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_503_groups_1 = const()[name = string("sparse_output_503_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191070720))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191068032))))[name = string("layers_6_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_503_cast_fp16 = conv(dilations = sparse_output_503_dilations_1, groups = sparse_output_503_groups_1, pad = sparse_output_503_pad_1, pad_type = sparse_output_503_pad_type_1, strides = sparse_output_503_strides_1, weight = layers_6_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_503_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_8040_cast_fp16 = add(x = dense_output_503_cast_fp16, y = sparse_output_503_cast_fp16)[name = string("op_8040_cast_fp16")];
            tensor<int32, [4]> var_8041 = const()[name = string("op_8041"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8043 = const()[name = string("op_8043"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_8042_cast_fp16 = transpose(perm = var_8041, x = var_8040_cast_fp16)[name = string("transpose_731")];
            tensor<fp16, [1, 101, 128]> p_head_193_cast_fp16 = reshape(shape = var_8043, x = var_8042_cast_fp16)[name = string("p_head_193_cast_fp16")];
            tensor<fp16, [128]> var_8045_to_fp16 = const()[name = string("op_8045_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191087168)))];
            tensor<fp16, [1, 51, 128]> var_8046_cast_fp16 = add(x = q_head_97_cast_fp16, y = var_8045_to_fp16)[name = string("op_8046_cast_fp16")];
            tensor<int32, [1]> q_u_97_axes_1 = const()[name = string("q_u_97_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_97_cast_fp16 = expand_dims(axes = q_u_97_axes_1, x = var_8046_cast_fp16)[name = string("q_u_97_cast_fp16")];
            tensor<fp16, [128]> var_8048_to_fp16 = const()[name = string("op_8048_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191087488)))];
            tensor<fp16, [1, 51, 128]> var_8049_cast_fp16 = add(x = q_head_97_cast_fp16, y = var_8048_to_fp16)[name = string("op_8049_cast_fp16")];
            tensor<int32, [1]> q_v_97_axes_1 = const()[name = string("q_v_97_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_97_cast_fp16 = expand_dims(axes = q_v_97_axes_1, x = var_8049_cast_fp16)[name = string("q_v_97_cast_fp16")];
            tensor<int32, [1]> k_head_195_axes_1 = const()[name = string("k_head_195_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_195_cast_fp16 = expand_dims(axes = k_head_195_axes_1, x = k_head_193_cast_fp16)[name = string("k_head_195_cast_fp16")];
            tensor<int32, [1]> v_head_195_axes_1 = const()[name = string("v_head_195_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_195_cast_fp16 = expand_dims(axes = v_head_195_axes_1, x = v_head_193_cast_fp16)[name = string("v_head_195_cast_fp16")];
            tensor<int32, [1]> p_head_195_axes_1 = const()[name = string("p_head_195_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_195_cast_fp16 = expand_dims(axes = p_head_195_axes_1, x = p_head_193_cast_fp16)[name = string("p_head_195_cast_fp16")];
            bool var_8055_transpose_x_3 = const()[name = string("op_8055_transpose_x_3"), val = bool(false)];
            bool var_8055_transpose_y_3 = const()[name = string("op_8055_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_8055_cast_fp16 = matmul(transpose_x = var_8055_transpose_x_3, transpose_y = var_8055_transpose_y_3, x = q_u_97_cast_fp16, y = k_head_195_cast_fp16)[name = string("op_8055_cast_fp16")];
            fp16 var_8056_to_fp16 = const()[name = string("op_8056_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_97_cast_fp16 = mul(x = var_8055_cast_fp16, y = var_8056_to_fp16)[name = string("scores_content_97_cast_fp16")];
            bool x_525_transpose_x_3 = const()[name = string("x_525_transpose_x_3"), val = bool(false)];
            bool x_525_transpose_y_3 = const()[name = string("x_525_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_525_cast_fp16 = matmul(transpose_x = x_525_transpose_x_3, transpose_y = x_525_transpose_y_3, x = q_v_97_cast_fp16, y = p_head_195_cast_fp16)[name = string("x_525_cast_fp16")];
            tensor<int32, [8]> x_527_pad_1 = const()[name = string("x_527_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_527_mode_1 = const()[name = string("x_527_mode_1"), val = string("constant")];
            fp16 const_1633_to_fp16 = const()[name = string("const_1633_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_527_cast_fp16 = pad(constant_val = const_1633_to_fp16, mode = x_527_mode_1, pad = x_527_pad_1, x = x_525_cast_fp16)[name = string("x_527_cast_fp16")];
            tensor<int32, [4]> var_8070 = const()[name = string("op_8070"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_529_cast_fp16 = reshape(shape = var_8070, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")];
            tensor<int32, [4]> var_8074_begin_1 = const()[name = string("op_8074_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8074_end_1 = const()[name = string("op_8074_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_8074_end_mask_1 = const()[name = string("op_8074_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_8074_cast_fp16 = slice_by_index(begin = var_8074_begin_1, end = var_8074_end_1, end_mask = var_8074_end_mask_1, x = x_529_cast_fp16)[name = string("op_8074_cast_fp16")];
            tensor<int32, [4]> var_8076 = const()[name = string("op_8076"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_8077_cast_fp16 = reshape(shape = var_8076, x = var_8074_cast_fp16)[name = string("op_8077_cast_fp16")];
            tensor<int32, [4]> var_8082_begin_1 = const()[name = string("op_8082_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8082_end_1 = const()[name = string("op_8082_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_8082_end_mask_1 = const()[name = string("op_8082_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_8082_cast_fp16 = slice_by_index(begin = var_8082_begin_1, end = var_8082_end_1, end_mask = var_8082_end_mask_1, x = var_8077_cast_fp16)[name = string("op_8082_cast_fp16")];
            fp16 var_8083_to_fp16 = const()[name = string("op_8083_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_97_cast_fp16 = mul(x = var_8082_cast_fp16, y = var_8083_to_fp16)[name = string("scores_pos_97_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_97_cast_fp16 = add(x = scores_content_97_cast_fp16, y = scores_pos_97_cast_fp16)[name = string("logits_97_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_8086_cast_fp16 = softmax(axis = var_7811, x = logits_97_cast_fp16)[name = string("op_8086_cast_fp16")];
            bool var_8088_transpose_x_1 = const()[name = string("op_8088_transpose_x_1"), val = bool(false)];
            bool var_8088_transpose_y_1 = const()[name = string("op_8088_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_8088_cast_fp16 = matmul(transpose_x = var_8088_transpose_x_1, transpose_y = var_8088_transpose_y_1, x = var_8086_cast_fp16, y = v_head_195_cast_fp16)[name = string("op_8088_cast_fp16")];
            tensor<int32, [1]> var_8089_axes_1 = const()[name = string("op_8089_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_8089_cast_fp16 = squeeze(axes = var_8089_axes_1, x = var_8088_cast_fp16)[name = string("op_8089_cast_fp16")];
            string dense_output_505_pad_type_1 = const()[name = string("dense_output_505_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_505_strides_1 = const()[name = string("dense_output_505_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_505_pad_1 = const()[name = string("dense_output_505_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_505_dilations_1 = const()[name = string("dense_output_505_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_505_groups_1 = const()[name = string("dense_output_505_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191087808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191218944))))[name = string("layers_6_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_505_cast_fp16 = conv(dilations = dense_output_505_dilations_1, groups = dense_output_505_groups_1, pad = dense_output_505_pad_1, pad_type = dense_output_505_pad_type_1, strides = dense_output_505_strides_1, weight = layers_6_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_505_cast_fp16")];
            string sparse_output_505_pad_type_1 = const()[name = string("sparse_output_505_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_505_strides_1 = const()[name = string("sparse_output_505_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_505_pad_1 = const()[name = string("sparse_output_505_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_505_dilations_1 = const()[name = string("sparse_output_505_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_505_groups_1 = const()[name = string("sparse_output_505_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191222208))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191219520))))[name = string("layers_6_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_505_cast_fp16 = conv(dilations = sparse_output_505_dilations_1, groups = sparse_output_505_groups_1, pad = sparse_output_505_pad_1, pad_type = sparse_output_505_pad_type_1, strides = sparse_output_505_strides_1, weight = layers_6_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_505_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8104_cast_fp16 = add(x = dense_output_505_cast_fp16, y = sparse_output_505_cast_fp16)[name = string("op_8104_cast_fp16")];
            tensor<int32, [4]> var_8105 = const()[name = string("op_8105"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8107 = const()[name = string("op_8107"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8106_cast_fp16 = transpose(perm = var_8105, x = var_8104_cast_fp16)[name = string("transpose_730")];
            tensor<fp16, [1, 51, 128]> q_head_99_cast_fp16 = reshape(shape = var_8107, x = var_8106_cast_fp16)[name = string("q_head_99_cast_fp16")];
            string dense_output_507_pad_type_1 = const()[name = string("dense_output_507_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_507_strides_1 = const()[name = string("dense_output_507_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_507_pad_1 = const()[name = string("dense_output_507_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_507_dilations_1 = const()[name = string("dense_output_507_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_507_groups_1 = const()[name = string("dense_output_507_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191238656))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191369792))))[name = string("layers_6_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_507_cast_fp16 = conv(dilations = dense_output_507_dilations_1, groups = dense_output_507_groups_1, pad = dense_output_507_pad_1, pad_type = dense_output_507_pad_type_1, strides = dense_output_507_strides_1, weight = layers_6_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_507_cast_fp16")];
            string sparse_output_507_pad_type_1 = const()[name = string("sparse_output_507_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_507_strides_1 = const()[name = string("sparse_output_507_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_507_pad_1 = const()[name = string("sparse_output_507_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_507_dilations_1 = const()[name = string("sparse_output_507_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_507_groups_1 = const()[name = string("sparse_output_507_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191373056))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191370368))))[name = string("layers_6_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_507_cast_fp16 = conv(dilations = sparse_output_507_dilations_1, groups = sparse_output_507_groups_1, pad = sparse_output_507_pad_1, pad_type = sparse_output_507_pad_type_1, strides = sparse_output_507_strides_1, weight = layers_6_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_507_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8123_cast_fp16 = add(x = dense_output_507_cast_fp16, y = sparse_output_507_cast_fp16)[name = string("op_8123_cast_fp16")];
            tensor<int32, [4]> var_8124 = const()[name = string("op_8124"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8126 = const()[name = string("op_8126"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8125_cast_fp16 = transpose(perm = var_8124, x = var_8123_cast_fp16)[name = string("transpose_729")];
            tensor<fp16, [1, 51, 128]> k_head_197_cast_fp16 = reshape(shape = var_8126, x = var_8125_cast_fp16)[name = string("k_head_197_cast_fp16")];
            string dense_output_509_pad_type_1 = const()[name = string("dense_output_509_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_509_strides_1 = const()[name = string("dense_output_509_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_509_pad_1 = const()[name = string("dense_output_509_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_509_dilations_1 = const()[name = string("dense_output_509_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_509_groups_1 = const()[name = string("dense_output_509_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191389504))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191520640))))[name = string("layers_6_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_509_cast_fp16 = conv(dilations = dense_output_509_dilations_1, groups = dense_output_509_groups_1, pad = dense_output_509_pad_1, pad_type = dense_output_509_pad_type_1, strides = dense_output_509_strides_1, weight = layers_6_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_509_cast_fp16")];
            string sparse_output_509_pad_type_1 = const()[name = string("sparse_output_509_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_509_strides_1 = const()[name = string("sparse_output_509_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_509_pad_1 = const()[name = string("sparse_output_509_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_509_dilations_1 = const()[name = string("sparse_output_509_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_509_groups_1 = const()[name = string("sparse_output_509_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191523904))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191521216))))[name = string("layers_6_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_509_cast_fp16 = conv(dilations = sparse_output_509_dilations_1, groups = sparse_output_509_groups_1, pad = sparse_output_509_pad_1, pad_type = sparse_output_509_pad_type_1, strides = sparse_output_509_strides_1, weight = layers_6_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_509_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8142_cast_fp16 = add(x = dense_output_509_cast_fp16, y = sparse_output_509_cast_fp16)[name = string("op_8142_cast_fp16")];
            tensor<int32, [4]> var_8143 = const()[name = string("op_8143"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8145 = const()[name = string("op_8145"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8144_cast_fp16 = transpose(perm = var_8143, x = var_8142_cast_fp16)[name = string("transpose_728")];
            tensor<fp16, [1, 51, 128]> v_head_197_cast_fp16 = reshape(shape = var_8145, x = var_8144_cast_fp16)[name = string("v_head_197_cast_fp16")];
            string dense_output_511_pad_type_1 = const()[name = string("dense_output_511_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_511_strides_1 = const()[name = string("dense_output_511_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_511_pad_1 = const()[name = string("dense_output_511_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_511_dilations_1 = const()[name = string("dense_output_511_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_511_groups_1 = const()[name = string("dense_output_511_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191540352))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191671488))))[name = string("layers_6_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_511_cast_fp16 = conv(dilations = dense_output_511_dilations_1, groups = dense_output_511_groups_1, pad = dense_output_511_pad_1, pad_type = dense_output_511_pad_type_1, strides = dense_output_511_strides_1, weight = layers_6_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_511_cast_fp16")];
            string sparse_output_511_pad_type_1 = const()[name = string("sparse_output_511_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_511_strides_1 = const()[name = string("sparse_output_511_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_511_pad_1 = const()[name = string("sparse_output_511_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_511_dilations_1 = const()[name = string("sparse_output_511_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_511_groups_1 = const()[name = string("sparse_output_511_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191674752))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191672064))))[name = string("layers_6_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_511_cast_fp16 = conv(dilations = sparse_output_511_dilations_1, groups = sparse_output_511_groups_1, pad = sparse_output_511_pad_1, pad_type = sparse_output_511_pad_type_1, strides = sparse_output_511_strides_1, weight = layers_6_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_511_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_8161_cast_fp16 = add(x = dense_output_511_cast_fp16, y = sparse_output_511_cast_fp16)[name = string("op_8161_cast_fp16")];
            tensor<int32, [4]> var_8162 = const()[name = string("op_8162"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8164 = const()[name = string("op_8164"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_8163_cast_fp16 = transpose(perm = var_8162, x = var_8161_cast_fp16)[name = string("transpose_727")];
            tensor<fp16, [1, 101, 128]> p_head_197_cast_fp16 = reshape(shape = var_8164, x = var_8163_cast_fp16)[name = string("p_head_197_cast_fp16")];
            tensor<fp16, [128]> var_8166_to_fp16 = const()[name = string("op_8166_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191691200)))];
            tensor<fp16, [1, 51, 128]> var_8167_cast_fp16 = add(x = q_head_99_cast_fp16, y = var_8166_to_fp16)[name = string("op_8167_cast_fp16")];
            tensor<int32, [1]> q_u_99_axes_1 = const()[name = string("q_u_99_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_99_cast_fp16 = expand_dims(axes = q_u_99_axes_1, x = var_8167_cast_fp16)[name = string("q_u_99_cast_fp16")];
            tensor<fp16, [128]> var_8169_to_fp16 = const()[name = string("op_8169_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191691520)))];
            tensor<fp16, [1, 51, 128]> var_8170_cast_fp16 = add(x = q_head_99_cast_fp16, y = var_8169_to_fp16)[name = string("op_8170_cast_fp16")];
            tensor<int32, [1]> q_v_99_axes_1 = const()[name = string("q_v_99_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_99_cast_fp16 = expand_dims(axes = q_v_99_axes_1, x = var_8170_cast_fp16)[name = string("q_v_99_cast_fp16")];
            tensor<int32, [1]> k_head_199_axes_1 = const()[name = string("k_head_199_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_199_cast_fp16 = expand_dims(axes = k_head_199_axes_1, x = k_head_197_cast_fp16)[name = string("k_head_199_cast_fp16")];
            tensor<int32, [1]> v_head_199_axes_1 = const()[name = string("v_head_199_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_199_cast_fp16 = expand_dims(axes = v_head_199_axes_1, x = v_head_197_cast_fp16)[name = string("v_head_199_cast_fp16")];
            tensor<int32, [1]> p_head_199_axes_1 = const()[name = string("p_head_199_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_199_cast_fp16 = expand_dims(axes = p_head_199_axes_1, x = p_head_197_cast_fp16)[name = string("p_head_199_cast_fp16")];
            bool var_8176_transpose_x_3 = const()[name = string("op_8176_transpose_x_3"), val = bool(false)];
            bool var_8176_transpose_y_3 = const()[name = string("op_8176_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_8176_cast_fp16 = matmul(transpose_x = var_8176_transpose_x_3, transpose_y = var_8176_transpose_y_3, x = q_u_99_cast_fp16, y = k_head_199_cast_fp16)[name = string("op_8176_cast_fp16")];
            fp16 var_8177_to_fp16 = const()[name = string("op_8177_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_99_cast_fp16 = mul(x = var_8176_cast_fp16, y = var_8177_to_fp16)[name = string("scores_content_99_cast_fp16")];
            bool x_533_transpose_x_3 = const()[name = string("x_533_transpose_x_3"), val = bool(false)];
            bool x_533_transpose_y_3 = const()[name = string("x_533_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_3, transpose_y = x_533_transpose_y_3, x = q_v_99_cast_fp16, y = p_head_199_cast_fp16)[name = string("x_533_cast_fp16")];
            tensor<int32, [8]> x_535_pad_1 = const()[name = string("x_535_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_535_mode_1 = const()[name = string("x_535_mode_1"), val = string("constant")];
            fp16 const_1639_to_fp16 = const()[name = string("const_1639_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_535_cast_fp16 = pad(constant_val = const_1639_to_fp16, mode = x_535_mode_1, pad = x_535_pad_1, x = x_533_cast_fp16)[name = string("x_535_cast_fp16")];
            tensor<int32, [4]> var_8191 = const()[name = string("op_8191"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_537_cast_fp16 = reshape(shape = var_8191, x = x_535_cast_fp16)[name = string("x_537_cast_fp16")];
            tensor<int32, [4]> var_8195_begin_1 = const()[name = string("op_8195_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8195_end_1 = const()[name = string("op_8195_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_8195_end_mask_1 = const()[name = string("op_8195_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_8195_cast_fp16 = slice_by_index(begin = var_8195_begin_1, end = var_8195_end_1, end_mask = var_8195_end_mask_1, x = x_537_cast_fp16)[name = string("op_8195_cast_fp16")];
            tensor<int32, [4]> var_8197 = const()[name = string("op_8197"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_8198_cast_fp16 = reshape(shape = var_8197, x = var_8195_cast_fp16)[name = string("op_8198_cast_fp16")];
            tensor<int32, [4]> var_8203_begin_1 = const()[name = string("op_8203_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8203_end_1 = const()[name = string("op_8203_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_8203_end_mask_1 = const()[name = string("op_8203_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_8203_cast_fp16 = slice_by_index(begin = var_8203_begin_1, end = var_8203_end_1, end_mask = var_8203_end_mask_1, x = var_8198_cast_fp16)[name = string("op_8203_cast_fp16")];
            fp16 var_8204_to_fp16 = const()[name = string("op_8204_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_99_cast_fp16 = mul(x = var_8203_cast_fp16, y = var_8204_to_fp16)[name = string("scores_pos_99_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_99_cast_fp16 = add(x = scores_content_99_cast_fp16, y = scores_pos_99_cast_fp16)[name = string("logits_99_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_8207_cast_fp16 = softmax(axis = var_7811, x = logits_99_cast_fp16)[name = string("op_8207_cast_fp16")];
            bool var_8209_transpose_x_1 = const()[name = string("op_8209_transpose_x_1"), val = bool(false)];
            bool var_8209_transpose_y_1 = const()[name = string("op_8209_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_8209_cast_fp16 = matmul(transpose_x = var_8209_transpose_x_1, transpose_y = var_8209_transpose_y_1, x = var_8207_cast_fp16, y = v_head_199_cast_fp16)[name = string("op_8209_cast_fp16")];
            tensor<int32, [1]> var_8210_axes_1 = const()[name = string("op_8210_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_8210_cast_fp16 = squeeze(axes = var_8210_axes_1, x = var_8209_cast_fp16)[name = string("op_8210_cast_fp16")];
            string dense_output_513_pad_type_1 = const()[name = string("dense_output_513_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_513_strides_1 = const()[name = string("dense_output_513_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_513_pad_1 = const()[name = string("dense_output_513_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_513_dilations_1 = const()[name = string("dense_output_513_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_513_groups_1 = const()[name = string("dense_output_513_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191691840))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191822976))))[name = string("layers_6_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_513_cast_fp16 = conv(dilations = dense_output_513_dilations_1, groups = dense_output_513_groups_1, pad = dense_output_513_pad_1, pad_type = dense_output_513_pad_type_1, strides = dense_output_513_strides_1, weight = layers_6_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_513_cast_fp16")];
            string sparse_output_513_pad_type_1 = const()[name = string("sparse_output_513_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_513_strides_1 = const()[name = string("sparse_output_513_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_513_pad_1 = const()[name = string("sparse_output_513_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_513_dilations_1 = const()[name = string("sparse_output_513_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_513_groups_1 = const()[name = string("sparse_output_513_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191826240))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191823552))))[name = string("layers_6_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_513_cast_fp16 = conv(dilations = sparse_output_513_dilations_1, groups = sparse_output_513_groups_1, pad = sparse_output_513_pad_1, pad_type = sparse_output_513_pad_type_1, strides = sparse_output_513_strides_1, weight = layers_6_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_513_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8225_cast_fp16 = add(x = dense_output_513_cast_fp16, y = sparse_output_513_cast_fp16)[name = string("op_8225_cast_fp16")];
            tensor<int32, [4]> var_8226 = const()[name = string("op_8226"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8228 = const()[name = string("op_8228"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8227_cast_fp16 = transpose(perm = var_8226, x = var_8225_cast_fp16)[name = string("transpose_726")];
            tensor<fp16, [1, 51, 128]> q_head_101_cast_fp16 = reshape(shape = var_8228, x = var_8227_cast_fp16)[name = string("q_head_101_cast_fp16")];
            string dense_output_515_pad_type_1 = const()[name = string("dense_output_515_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_515_strides_1 = const()[name = string("dense_output_515_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_515_pad_1 = const()[name = string("dense_output_515_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_515_dilations_1 = const()[name = string("dense_output_515_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_515_groups_1 = const()[name = string("dense_output_515_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191842688))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191973824))))[name = string("layers_6_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_515_cast_fp16 = conv(dilations = dense_output_515_dilations_1, groups = dense_output_515_groups_1, pad = dense_output_515_pad_1, pad_type = dense_output_515_pad_type_1, strides = dense_output_515_strides_1, weight = layers_6_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_515_cast_fp16")];
            string sparse_output_515_pad_type_1 = const()[name = string("sparse_output_515_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_515_strides_1 = const()[name = string("sparse_output_515_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_515_pad_1 = const()[name = string("sparse_output_515_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_515_dilations_1 = const()[name = string("sparse_output_515_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_515_groups_1 = const()[name = string("sparse_output_515_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191977088))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191974400))))[name = string("layers_6_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_515_cast_fp16 = conv(dilations = sparse_output_515_dilations_1, groups = sparse_output_515_groups_1, pad = sparse_output_515_pad_1, pad_type = sparse_output_515_pad_type_1, strides = sparse_output_515_strides_1, weight = layers_6_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_515_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8244_cast_fp16 = add(x = dense_output_515_cast_fp16, y = sparse_output_515_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<int32, [4]> var_8245 = const()[name = string("op_8245"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8247 = const()[name = string("op_8247"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8246_cast_fp16 = transpose(perm = var_8245, x = var_8244_cast_fp16)[name = string("transpose_725")];
            tensor<fp16, [1, 51, 128]> k_head_201_cast_fp16 = reshape(shape = var_8247, x = var_8246_cast_fp16)[name = string("k_head_201_cast_fp16")];
            string dense_output_517_pad_type_1 = const()[name = string("dense_output_517_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_517_strides_1 = const()[name = string("dense_output_517_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_517_pad_1 = const()[name = string("dense_output_517_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_517_dilations_1 = const()[name = string("dense_output_517_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_517_groups_1 = const()[name = string("dense_output_517_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191993536))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192124672))))[name = string("layers_6_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_517_cast_fp16 = conv(dilations = dense_output_517_dilations_1, groups = dense_output_517_groups_1, pad = dense_output_517_pad_1, pad_type = dense_output_517_pad_type_1, strides = dense_output_517_strides_1, weight = layers_6_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_517_cast_fp16")];
            string sparse_output_517_pad_type_1 = const()[name = string("sparse_output_517_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_517_strides_1 = const()[name = string("sparse_output_517_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_517_pad_1 = const()[name = string("sparse_output_517_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_517_dilations_1 = const()[name = string("sparse_output_517_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_517_groups_1 = const()[name = string("sparse_output_517_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192127936))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192125248))))[name = string("layers_6_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_517_cast_fp16 = conv(dilations = sparse_output_517_dilations_1, groups = sparse_output_517_groups_1, pad = sparse_output_517_pad_1, pad_type = sparse_output_517_pad_type_1, strides = sparse_output_517_strides_1, weight = layers_6_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_517_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8263_cast_fp16 = add(x = dense_output_517_cast_fp16, y = sparse_output_517_cast_fp16)[name = string("op_8263_cast_fp16")];
            tensor<int32, [4]> var_8264 = const()[name = string("op_8264"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8266 = const()[name = string("op_8266"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8265_cast_fp16 = transpose(perm = var_8264, x = var_8263_cast_fp16)[name = string("transpose_724")];
            tensor<fp16, [1, 51, 128]> v_head_201_cast_fp16 = reshape(shape = var_8266, x = var_8265_cast_fp16)[name = string("v_head_201_cast_fp16")];
            string dense_output_519_pad_type_1 = const()[name = string("dense_output_519_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_519_strides_1 = const()[name = string("dense_output_519_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_519_pad_1 = const()[name = string("dense_output_519_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_519_dilations_1 = const()[name = string("dense_output_519_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_519_groups_1 = const()[name = string("dense_output_519_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192144384))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192275520))))[name = string("layers_6_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_519_cast_fp16 = conv(dilations = dense_output_519_dilations_1, groups = dense_output_519_groups_1, pad = dense_output_519_pad_1, pad_type = dense_output_519_pad_type_1, strides = dense_output_519_strides_1, weight = layers_6_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_519_cast_fp16")];
            string sparse_output_519_pad_type_1 = const()[name = string("sparse_output_519_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_519_strides_1 = const()[name = string("sparse_output_519_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_519_pad_1 = const()[name = string("sparse_output_519_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_519_dilations_1 = const()[name = string("sparse_output_519_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_519_groups_1 = const()[name = string("sparse_output_519_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192278784))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192276096))))[name = string("layers_6_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_519_cast_fp16 = conv(dilations = sparse_output_519_dilations_1, groups = sparse_output_519_groups_1, pad = sparse_output_519_pad_1, pad_type = sparse_output_519_pad_type_1, strides = sparse_output_519_strides_1, weight = layers_6_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_519_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_8282_cast_fp16 = add(x = dense_output_519_cast_fp16, y = sparse_output_519_cast_fp16)[name = string("op_8282_cast_fp16")];
            tensor<int32, [4]> var_8283 = const()[name = string("op_8283"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8285 = const()[name = string("op_8285"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_8284_cast_fp16 = transpose(perm = var_8283, x = var_8282_cast_fp16)[name = string("transpose_723")];
            tensor<fp16, [1, 101, 128]> p_head_201_cast_fp16 = reshape(shape = var_8285, x = var_8284_cast_fp16)[name = string("p_head_201_cast_fp16")];
            tensor<fp16, [128]> var_8287_to_fp16 = const()[name = string("op_8287_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192295232)))];
            tensor<fp16, [1, 51, 128]> var_8288_cast_fp16 = add(x = q_head_101_cast_fp16, y = var_8287_to_fp16)[name = string("op_8288_cast_fp16")];
            tensor<int32, [1]> q_u_101_axes_1 = const()[name = string("q_u_101_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_101_cast_fp16 = expand_dims(axes = q_u_101_axes_1, x = var_8288_cast_fp16)[name = string("q_u_101_cast_fp16")];
            tensor<fp16, [128]> var_8290_to_fp16 = const()[name = string("op_8290_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192295552)))];
            tensor<fp16, [1, 51, 128]> var_8291_cast_fp16 = add(x = q_head_101_cast_fp16, y = var_8290_to_fp16)[name = string("op_8291_cast_fp16")];
            tensor<int32, [1]> q_v_101_axes_1 = const()[name = string("q_v_101_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_101_cast_fp16 = expand_dims(axes = q_v_101_axes_1, x = var_8291_cast_fp16)[name = string("q_v_101_cast_fp16")];
            tensor<int32, [1]> k_head_203_axes_1 = const()[name = string("k_head_203_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_203_cast_fp16 = expand_dims(axes = k_head_203_axes_1, x = k_head_201_cast_fp16)[name = string("k_head_203_cast_fp16")];
            tensor<int32, [1]> v_head_203_axes_1 = const()[name = string("v_head_203_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_203_cast_fp16 = expand_dims(axes = v_head_203_axes_1, x = v_head_201_cast_fp16)[name = string("v_head_203_cast_fp16")];
            tensor<int32, [1]> p_head_203_axes_1 = const()[name = string("p_head_203_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_203_cast_fp16 = expand_dims(axes = p_head_203_axes_1, x = p_head_201_cast_fp16)[name = string("p_head_203_cast_fp16")];
            bool var_8297_transpose_x_3 = const()[name = string("op_8297_transpose_x_3"), val = bool(false)];
            bool var_8297_transpose_y_3 = const()[name = string("op_8297_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_8297_cast_fp16 = matmul(transpose_x = var_8297_transpose_x_3, transpose_y = var_8297_transpose_y_3, x = q_u_101_cast_fp16, y = k_head_203_cast_fp16)[name = string("op_8297_cast_fp16")];
            fp16 var_8298_to_fp16 = const()[name = string("op_8298_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_101_cast_fp16 = mul(x = var_8297_cast_fp16, y = var_8298_to_fp16)[name = string("scores_content_101_cast_fp16")];
            bool x_541_transpose_x_3 = const()[name = string("x_541_transpose_x_3"), val = bool(false)];
            bool x_541_transpose_y_3 = const()[name = string("x_541_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_541_cast_fp16 = matmul(transpose_x = x_541_transpose_x_3, transpose_y = x_541_transpose_y_3, x = q_v_101_cast_fp16, y = p_head_203_cast_fp16)[name = string("x_541_cast_fp16")];
            tensor<int32, [8]> x_543_pad_1 = const()[name = string("x_543_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_543_mode_1 = const()[name = string("x_543_mode_1"), val = string("constant")];
            fp16 const_1645_to_fp16 = const()[name = string("const_1645_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_543_cast_fp16 = pad(constant_val = const_1645_to_fp16, mode = x_543_mode_1, pad = x_543_pad_1, x = x_541_cast_fp16)[name = string("x_543_cast_fp16")];
            tensor<int32, [4]> var_8312 = const()[name = string("op_8312"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_545_cast_fp16 = reshape(shape = var_8312, x = x_543_cast_fp16)[name = string("x_545_cast_fp16")];
            tensor<int32, [4]> var_8316_begin_1 = const()[name = string("op_8316_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8316_end_1 = const()[name = string("op_8316_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_8316_end_mask_1 = const()[name = string("op_8316_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_8316_cast_fp16 = slice_by_index(begin = var_8316_begin_1, end = var_8316_end_1, end_mask = var_8316_end_mask_1, x = x_545_cast_fp16)[name = string("op_8316_cast_fp16")];
            tensor<int32, [4]> var_8318 = const()[name = string("op_8318"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_8319_cast_fp16 = reshape(shape = var_8318, x = var_8316_cast_fp16)[name = string("op_8319_cast_fp16")];
            tensor<int32, [4]> var_8324_begin_1 = const()[name = string("op_8324_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8324_end_1 = const()[name = string("op_8324_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_8324_end_mask_1 = const()[name = string("op_8324_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_8324_cast_fp16 = slice_by_index(begin = var_8324_begin_1, end = var_8324_end_1, end_mask = var_8324_end_mask_1, x = var_8319_cast_fp16)[name = string("op_8324_cast_fp16")];
            fp16 var_8325_to_fp16 = const()[name = string("op_8325_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_101_cast_fp16 = mul(x = var_8324_cast_fp16, y = var_8325_to_fp16)[name = string("scores_pos_101_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_101_cast_fp16 = add(x = scores_content_101_cast_fp16, y = scores_pos_101_cast_fp16)[name = string("logits_101_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_8328_cast_fp16 = softmax(axis = var_7811, x = logits_101_cast_fp16)[name = string("op_8328_cast_fp16")];
            bool var_8330_transpose_x_1 = const()[name = string("op_8330_transpose_x_1"), val = bool(false)];
            bool var_8330_transpose_y_1 = const()[name = string("op_8330_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_8330_cast_fp16 = matmul(transpose_x = var_8330_transpose_x_1, transpose_y = var_8330_transpose_y_1, x = var_8328_cast_fp16, y = v_head_203_cast_fp16)[name = string("op_8330_cast_fp16")];
            tensor<int32, [1]> var_8331_axes_1 = const()[name = string("op_8331_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_8331_cast_fp16 = squeeze(axes = var_8331_axes_1, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            string dense_output_521_pad_type_1 = const()[name = string("dense_output_521_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_521_strides_1 = const()[name = string("dense_output_521_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_521_pad_1 = const()[name = string("dense_output_521_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_521_dilations_1 = const()[name = string("dense_output_521_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_521_groups_1 = const()[name = string("dense_output_521_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192295872))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192427008))))[name = string("layers_6_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_521_cast_fp16 = conv(dilations = dense_output_521_dilations_1, groups = dense_output_521_groups_1, pad = dense_output_521_pad_1, pad_type = dense_output_521_pad_type_1, strides = dense_output_521_strides_1, weight = layers_6_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_521_cast_fp16")];
            string sparse_output_521_pad_type_1 = const()[name = string("sparse_output_521_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_521_strides_1 = const()[name = string("sparse_output_521_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_521_pad_1 = const()[name = string("sparse_output_521_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_521_dilations_1 = const()[name = string("sparse_output_521_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_521_groups_1 = const()[name = string("sparse_output_521_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192430272))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192427584))))[name = string("layers_6_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_521_cast_fp16 = conv(dilations = sparse_output_521_dilations_1, groups = sparse_output_521_groups_1, pad = sparse_output_521_pad_1, pad_type = sparse_output_521_pad_type_1, strides = sparse_output_521_strides_1, weight = layers_6_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_521_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8346_cast_fp16 = add(x = dense_output_521_cast_fp16, y = sparse_output_521_cast_fp16)[name = string("op_8346_cast_fp16")];
            tensor<int32, [4]> var_8347 = const()[name = string("op_8347"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8349 = const()[name = string("op_8349"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8348_cast_fp16 = transpose(perm = var_8347, x = var_8346_cast_fp16)[name = string("transpose_722")];
            tensor<fp16, [1, 51, 128]> q_head_103_cast_fp16 = reshape(shape = var_8349, x = var_8348_cast_fp16)[name = string("q_head_103_cast_fp16")];
            string dense_output_523_pad_type_1 = const()[name = string("dense_output_523_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_523_strides_1 = const()[name = string("dense_output_523_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_523_pad_1 = const()[name = string("dense_output_523_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_523_dilations_1 = const()[name = string("dense_output_523_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_523_groups_1 = const()[name = string("dense_output_523_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192446720))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192577856))))[name = string("layers_6_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_523_cast_fp16 = conv(dilations = dense_output_523_dilations_1, groups = dense_output_523_groups_1, pad = dense_output_523_pad_1, pad_type = dense_output_523_pad_type_1, strides = dense_output_523_strides_1, weight = layers_6_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_523_cast_fp16")];
            string sparse_output_523_pad_type_1 = const()[name = string("sparse_output_523_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_523_strides_1 = const()[name = string("sparse_output_523_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_523_pad_1 = const()[name = string("sparse_output_523_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_523_dilations_1 = const()[name = string("sparse_output_523_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_523_groups_1 = const()[name = string("sparse_output_523_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192581120))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192578432))))[name = string("layers_6_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_523_cast_fp16 = conv(dilations = sparse_output_523_dilations_1, groups = sparse_output_523_groups_1, pad = sparse_output_523_pad_1, pad_type = sparse_output_523_pad_type_1, strides = sparse_output_523_strides_1, weight = layers_6_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_523_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8365_cast_fp16 = add(x = dense_output_523_cast_fp16, y = sparse_output_523_cast_fp16)[name = string("op_8365_cast_fp16")];
            tensor<int32, [4]> var_8366 = const()[name = string("op_8366"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8368 = const()[name = string("op_8368"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8367_cast_fp16 = transpose(perm = var_8366, x = var_8365_cast_fp16)[name = string("transpose_721")];
            tensor<fp16, [1, 51, 128]> k_head_205_cast_fp16 = reshape(shape = var_8368, x = var_8367_cast_fp16)[name = string("k_head_205_cast_fp16")];
            string dense_output_525_pad_type_1 = const()[name = string("dense_output_525_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_525_strides_1 = const()[name = string("dense_output_525_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_525_pad_1 = const()[name = string("dense_output_525_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_525_dilations_1 = const()[name = string("dense_output_525_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_525_groups_1 = const()[name = string("dense_output_525_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192597568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192728704))))[name = string("layers_6_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_525_cast_fp16 = conv(dilations = dense_output_525_dilations_1, groups = dense_output_525_groups_1, pad = dense_output_525_pad_1, pad_type = dense_output_525_pad_type_1, strides = dense_output_525_strides_1, weight = layers_6_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_525_cast_fp16")];
            string sparse_output_525_pad_type_1 = const()[name = string("sparse_output_525_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_525_strides_1 = const()[name = string("sparse_output_525_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_525_pad_1 = const()[name = string("sparse_output_525_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_525_dilations_1 = const()[name = string("sparse_output_525_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_525_groups_1 = const()[name = string("sparse_output_525_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192731968))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192729280))))[name = string("layers_6_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_525_cast_fp16 = conv(dilations = sparse_output_525_dilations_1, groups = sparse_output_525_groups_1, pad = sparse_output_525_pad_1, pad_type = sparse_output_525_pad_type_1, strides = sparse_output_525_strides_1, weight = layers_6_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_525_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8384_cast_fp16 = add(x = dense_output_525_cast_fp16, y = sparse_output_525_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<int32, [4]> var_8385 = const()[name = string("op_8385"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8387 = const()[name = string("op_8387"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8386_cast_fp16 = transpose(perm = var_8385, x = var_8384_cast_fp16)[name = string("transpose_720")];
            tensor<fp16, [1, 51, 128]> v_head_205_cast_fp16 = reshape(shape = var_8387, x = var_8386_cast_fp16)[name = string("v_head_205_cast_fp16")];
            string dense_output_527_pad_type_1 = const()[name = string("dense_output_527_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_527_strides_1 = const()[name = string("dense_output_527_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_527_pad_1 = const()[name = string("dense_output_527_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_527_dilations_1 = const()[name = string("dense_output_527_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_527_groups_1 = const()[name = string("dense_output_527_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192748416))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192879552))))[name = string("layers_6_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_527_cast_fp16 = conv(dilations = dense_output_527_dilations_1, groups = dense_output_527_groups_1, pad = dense_output_527_pad_1, pad_type = dense_output_527_pad_type_1, strides = dense_output_527_strides_1, weight = layers_6_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_527_cast_fp16")];
            string sparse_output_527_pad_type_1 = const()[name = string("sparse_output_527_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_527_strides_1 = const()[name = string("sparse_output_527_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_527_pad_1 = const()[name = string("sparse_output_527_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_527_dilations_1 = const()[name = string("sparse_output_527_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_527_groups_1 = const()[name = string("sparse_output_527_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192882816))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192880128))))[name = string("layers_6_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_527_cast_fp16 = conv(dilations = sparse_output_527_dilations_1, groups = sparse_output_527_groups_1, pad = sparse_output_527_pad_1, pad_type = sparse_output_527_pad_type_1, strides = sparse_output_527_strides_1, weight = layers_6_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_527_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_8403_cast_fp16 = add(x = dense_output_527_cast_fp16, y = sparse_output_527_cast_fp16)[name = string("op_8403_cast_fp16")];
            tensor<int32, [4]> var_8404 = const()[name = string("op_8404"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8406 = const()[name = string("op_8406"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_8405_cast_fp16 = transpose(perm = var_8404, x = var_8403_cast_fp16)[name = string("transpose_719")];
            tensor<fp16, [1, 101, 128]> p_head_205_cast_fp16 = reshape(shape = var_8406, x = var_8405_cast_fp16)[name = string("p_head_205_cast_fp16")];
            tensor<fp16, [128]> var_8408_to_fp16 = const()[name = string("op_8408_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192899264)))];
            tensor<fp16, [1, 51, 128]> var_8409_cast_fp16 = add(x = q_head_103_cast_fp16, y = var_8408_to_fp16)[name = string("op_8409_cast_fp16")];
            tensor<int32, [1]> q_u_103_axes_1 = const()[name = string("q_u_103_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_103_cast_fp16 = expand_dims(axes = q_u_103_axes_1, x = var_8409_cast_fp16)[name = string("q_u_103_cast_fp16")];
            tensor<fp16, [128]> var_8411_to_fp16 = const()[name = string("op_8411_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192899584)))];
            tensor<fp16, [1, 51, 128]> var_8412_cast_fp16 = add(x = q_head_103_cast_fp16, y = var_8411_to_fp16)[name = string("op_8412_cast_fp16")];
            tensor<int32, [1]> q_v_103_axes_1 = const()[name = string("q_v_103_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_103_cast_fp16 = expand_dims(axes = q_v_103_axes_1, x = var_8412_cast_fp16)[name = string("q_v_103_cast_fp16")];
            tensor<int32, [1]> k_head_207_axes_1 = const()[name = string("k_head_207_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_207_cast_fp16 = expand_dims(axes = k_head_207_axes_1, x = k_head_205_cast_fp16)[name = string("k_head_207_cast_fp16")];
            tensor<int32, [1]> v_head_207_axes_1 = const()[name = string("v_head_207_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_207_cast_fp16 = expand_dims(axes = v_head_207_axes_1, x = v_head_205_cast_fp16)[name = string("v_head_207_cast_fp16")];
            tensor<int32, [1]> p_head_207_axes_1 = const()[name = string("p_head_207_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_207_cast_fp16 = expand_dims(axes = p_head_207_axes_1, x = p_head_205_cast_fp16)[name = string("p_head_207_cast_fp16")];
            bool var_8418_transpose_x_3 = const()[name = string("op_8418_transpose_x_3"), val = bool(false)];
            bool var_8418_transpose_y_3 = const()[name = string("op_8418_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_8418_cast_fp16 = matmul(transpose_x = var_8418_transpose_x_3, transpose_y = var_8418_transpose_y_3, x = q_u_103_cast_fp16, y = k_head_207_cast_fp16)[name = string("op_8418_cast_fp16")];
            fp16 var_8419_to_fp16 = const()[name = string("op_8419_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_103_cast_fp16 = mul(x = var_8418_cast_fp16, y = var_8419_to_fp16)[name = string("scores_content_103_cast_fp16")];
            bool x_549_transpose_x_3 = const()[name = string("x_549_transpose_x_3"), val = bool(false)];
            bool x_549_transpose_y_3 = const()[name = string("x_549_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_549_cast_fp16 = matmul(transpose_x = x_549_transpose_x_3, transpose_y = x_549_transpose_y_3, x = q_v_103_cast_fp16, y = p_head_207_cast_fp16)[name = string("x_549_cast_fp16")];
            tensor<int32, [8]> x_551_pad_1 = const()[name = string("x_551_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_551_mode_1 = const()[name = string("x_551_mode_1"), val = string("constant")];
            fp16 const_1651_to_fp16 = const()[name = string("const_1651_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_551_cast_fp16 = pad(constant_val = const_1651_to_fp16, mode = x_551_mode_1, pad = x_551_pad_1, x = x_549_cast_fp16)[name = string("x_551_cast_fp16")];
            tensor<int32, [4]> var_8433 = const()[name = string("op_8433"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_553_cast_fp16 = reshape(shape = var_8433, x = x_551_cast_fp16)[name = string("x_553_cast_fp16")];
            tensor<int32, [4]> var_8437_begin_1 = const()[name = string("op_8437_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8437_end_1 = const()[name = string("op_8437_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_8437_end_mask_1 = const()[name = string("op_8437_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_8437_cast_fp16 = slice_by_index(begin = var_8437_begin_1, end = var_8437_end_1, end_mask = var_8437_end_mask_1, x = x_553_cast_fp16)[name = string("op_8437_cast_fp16")];
            tensor<int32, [4]> var_8439 = const()[name = string("op_8439"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_8440_cast_fp16 = reshape(shape = var_8439, x = var_8437_cast_fp16)[name = string("op_8440_cast_fp16")];
            tensor<int32, [4]> var_8445_begin_1 = const()[name = string("op_8445_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8445_end_1 = const()[name = string("op_8445_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_8445_end_mask_1 = const()[name = string("op_8445_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_8445_cast_fp16 = slice_by_index(begin = var_8445_begin_1, end = var_8445_end_1, end_mask = var_8445_end_mask_1, x = var_8440_cast_fp16)[name = string("op_8445_cast_fp16")];
            fp16 var_8446_to_fp16 = const()[name = string("op_8446_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_103_cast_fp16 = mul(x = var_8445_cast_fp16, y = var_8446_to_fp16)[name = string("scores_pos_103_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_103_cast_fp16 = add(x = scores_content_103_cast_fp16, y = scores_pos_103_cast_fp16)[name = string("logits_103_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_8449_cast_fp16 = softmax(axis = var_7811, x = logits_103_cast_fp16)[name = string("op_8449_cast_fp16")];
            bool var_8451_transpose_x_1 = const()[name = string("op_8451_transpose_x_1"), val = bool(false)];
            bool var_8451_transpose_y_1 = const()[name = string("op_8451_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_8451_cast_fp16 = matmul(transpose_x = var_8451_transpose_x_1, transpose_y = var_8451_transpose_y_1, x = var_8449_cast_fp16, y = v_head_207_cast_fp16)[name = string("op_8451_cast_fp16")];
            tensor<int32, [1]> var_8452_axes_1 = const()[name = string("op_8452_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_8452_cast_fp16 = squeeze(axes = var_8452_axes_1, x = var_8451_cast_fp16)[name = string("op_8452_cast_fp16")];
            string dense_output_529_pad_type_1 = const()[name = string("dense_output_529_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_529_strides_1 = const()[name = string("dense_output_529_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_529_pad_1 = const()[name = string("dense_output_529_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_529_dilations_1 = const()[name = string("dense_output_529_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_529_groups_1 = const()[name = string("dense_output_529_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192899904))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193031040))))[name = string("layers_6_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_529_cast_fp16 = conv(dilations = dense_output_529_dilations_1, groups = dense_output_529_groups_1, pad = dense_output_529_pad_1, pad_type = dense_output_529_pad_type_1, strides = dense_output_529_strides_1, weight = layers_6_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_529_cast_fp16")];
            string sparse_output_529_pad_type_1 = const()[name = string("sparse_output_529_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_529_strides_1 = const()[name = string("sparse_output_529_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_529_pad_1 = const()[name = string("sparse_output_529_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_529_dilations_1 = const()[name = string("sparse_output_529_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_529_groups_1 = const()[name = string("sparse_output_529_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193034304))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193031616))))[name = string("layers_6_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_529_cast_fp16 = conv(dilations = sparse_output_529_dilations_1, groups = sparse_output_529_groups_1, pad = sparse_output_529_pad_1, pad_type = sparse_output_529_pad_type_1, strides = sparse_output_529_strides_1, weight = layers_6_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_529_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8467_cast_fp16 = add(x = dense_output_529_cast_fp16, y = sparse_output_529_cast_fp16)[name = string("op_8467_cast_fp16")];
            tensor<int32, [4]> var_8468 = const()[name = string("op_8468"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8470 = const()[name = string("op_8470"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8469_cast_fp16 = transpose(perm = var_8468, x = var_8467_cast_fp16)[name = string("transpose_718")];
            tensor<fp16, [1, 51, 128]> q_head_105_cast_fp16 = reshape(shape = var_8470, x = var_8469_cast_fp16)[name = string("q_head_105_cast_fp16")];
            string dense_output_531_pad_type_1 = const()[name = string("dense_output_531_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_531_strides_1 = const()[name = string("dense_output_531_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_531_pad_1 = const()[name = string("dense_output_531_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_531_dilations_1 = const()[name = string("dense_output_531_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_531_groups_1 = const()[name = string("dense_output_531_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193050752))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193181888))))[name = string("layers_6_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_531_cast_fp16 = conv(dilations = dense_output_531_dilations_1, groups = dense_output_531_groups_1, pad = dense_output_531_pad_1, pad_type = dense_output_531_pad_type_1, strides = dense_output_531_strides_1, weight = layers_6_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_531_cast_fp16")];
            string sparse_output_531_pad_type_1 = const()[name = string("sparse_output_531_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_531_strides_1 = const()[name = string("sparse_output_531_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_531_pad_1 = const()[name = string("sparse_output_531_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_531_dilations_1 = const()[name = string("sparse_output_531_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_531_groups_1 = const()[name = string("sparse_output_531_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193185152))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193182464))))[name = string("layers_6_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_531_cast_fp16 = conv(dilations = sparse_output_531_dilations_1, groups = sparse_output_531_groups_1, pad = sparse_output_531_pad_1, pad_type = sparse_output_531_pad_type_1, strides = sparse_output_531_strides_1, weight = layers_6_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_531_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8486_cast_fp16 = add(x = dense_output_531_cast_fp16, y = sparse_output_531_cast_fp16)[name = string("op_8486_cast_fp16")];
            tensor<int32, [4]> var_8487 = const()[name = string("op_8487"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8489 = const()[name = string("op_8489"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8488_cast_fp16 = transpose(perm = var_8487, x = var_8486_cast_fp16)[name = string("transpose_717")];
            tensor<fp16, [1, 51, 128]> k_head_209_cast_fp16 = reshape(shape = var_8489, x = var_8488_cast_fp16)[name = string("k_head_209_cast_fp16")];
            string dense_output_533_pad_type_1 = const()[name = string("dense_output_533_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_533_strides_1 = const()[name = string("dense_output_533_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_533_pad_1 = const()[name = string("dense_output_533_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_533_dilations_1 = const()[name = string("dense_output_533_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_533_groups_1 = const()[name = string("dense_output_533_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193201600))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193332736))))[name = string("layers_6_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_533_cast_fp16 = conv(dilations = dense_output_533_dilations_1, groups = dense_output_533_groups_1, pad = dense_output_533_pad_1, pad_type = dense_output_533_pad_type_1, strides = dense_output_533_strides_1, weight = layers_6_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_533_cast_fp16")];
            string sparse_output_533_pad_type_1 = const()[name = string("sparse_output_533_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_533_strides_1 = const()[name = string("sparse_output_533_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_533_pad_1 = const()[name = string("sparse_output_533_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_533_dilations_1 = const()[name = string("sparse_output_533_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_533_groups_1 = const()[name = string("sparse_output_533_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193336000))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193333312))))[name = string("layers_6_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_533_cast_fp16 = conv(dilations = sparse_output_533_dilations_1, groups = sparse_output_533_groups_1, pad = sparse_output_533_pad_1, pad_type = sparse_output_533_pad_type_1, strides = sparse_output_533_strides_1, weight = layers_6_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_533_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8505_cast_fp16 = add(x = dense_output_533_cast_fp16, y = sparse_output_533_cast_fp16)[name = string("op_8505_cast_fp16")];
            tensor<int32, [4]> var_8506 = const()[name = string("op_8506"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8508 = const()[name = string("op_8508"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8507_cast_fp16 = transpose(perm = var_8506, x = var_8505_cast_fp16)[name = string("transpose_716")];
            tensor<fp16, [1, 51, 128]> v_head_209_cast_fp16 = reshape(shape = var_8508, x = var_8507_cast_fp16)[name = string("v_head_209_cast_fp16")];
            string dense_output_535_pad_type_1 = const()[name = string("dense_output_535_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_535_strides_1 = const()[name = string("dense_output_535_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_535_pad_1 = const()[name = string("dense_output_535_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_535_dilations_1 = const()[name = string("dense_output_535_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_535_groups_1 = const()[name = string("dense_output_535_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193352448))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193483584))))[name = string("layers_6_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_535_cast_fp16 = conv(dilations = dense_output_535_dilations_1, groups = dense_output_535_groups_1, pad = dense_output_535_pad_1, pad_type = dense_output_535_pad_type_1, strides = dense_output_535_strides_1, weight = layers_6_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_535_cast_fp16")];
            string sparse_output_535_pad_type_1 = const()[name = string("sparse_output_535_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_535_strides_1 = const()[name = string("sparse_output_535_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_535_pad_1 = const()[name = string("sparse_output_535_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_535_dilations_1 = const()[name = string("sparse_output_535_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_535_groups_1 = const()[name = string("sparse_output_535_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193486848))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193484160))))[name = string("layers_6_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_535_cast_fp16 = conv(dilations = sparse_output_535_dilations_1, groups = sparse_output_535_groups_1, pad = sparse_output_535_pad_1, pad_type = sparse_output_535_pad_type_1, strides = sparse_output_535_strides_1, weight = layers_6_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_535_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_8524_cast_fp16 = add(x = dense_output_535_cast_fp16, y = sparse_output_535_cast_fp16)[name = string("op_8524_cast_fp16")];
            tensor<int32, [4]> var_8525 = const()[name = string("op_8525"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8527 = const()[name = string("op_8527"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_8526_cast_fp16 = transpose(perm = var_8525, x = var_8524_cast_fp16)[name = string("transpose_715")];
            tensor<fp16, [1, 101, 128]> p_head_209_cast_fp16 = reshape(shape = var_8527, x = var_8526_cast_fp16)[name = string("p_head_209_cast_fp16")];
            tensor<fp16, [128]> var_8529_to_fp16 = const()[name = string("op_8529_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193503296)))];
            tensor<fp16, [1, 51, 128]> var_8530_cast_fp16 = add(x = q_head_105_cast_fp16, y = var_8529_to_fp16)[name = string("op_8530_cast_fp16")];
            tensor<int32, [1]> q_u_105_axes_1 = const()[name = string("q_u_105_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_105_cast_fp16 = expand_dims(axes = q_u_105_axes_1, x = var_8530_cast_fp16)[name = string("q_u_105_cast_fp16")];
            tensor<fp16, [128]> var_8532_to_fp16 = const()[name = string("op_8532_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193503616)))];
            tensor<fp16, [1, 51, 128]> var_8533_cast_fp16 = add(x = q_head_105_cast_fp16, y = var_8532_to_fp16)[name = string("op_8533_cast_fp16")];
            tensor<int32, [1]> q_v_105_axes_1 = const()[name = string("q_v_105_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_105_cast_fp16 = expand_dims(axes = q_v_105_axes_1, x = var_8533_cast_fp16)[name = string("q_v_105_cast_fp16")];
            tensor<int32, [1]> k_head_211_axes_1 = const()[name = string("k_head_211_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_211_cast_fp16 = expand_dims(axes = k_head_211_axes_1, x = k_head_209_cast_fp16)[name = string("k_head_211_cast_fp16")];
            tensor<int32, [1]> v_head_211_axes_1 = const()[name = string("v_head_211_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_211_cast_fp16 = expand_dims(axes = v_head_211_axes_1, x = v_head_209_cast_fp16)[name = string("v_head_211_cast_fp16")];
            tensor<int32, [1]> p_head_211_axes_1 = const()[name = string("p_head_211_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_211_cast_fp16 = expand_dims(axes = p_head_211_axes_1, x = p_head_209_cast_fp16)[name = string("p_head_211_cast_fp16")];
            bool var_8539_transpose_x_3 = const()[name = string("op_8539_transpose_x_3"), val = bool(false)];
            bool var_8539_transpose_y_3 = const()[name = string("op_8539_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_8539_cast_fp16 = matmul(transpose_x = var_8539_transpose_x_3, transpose_y = var_8539_transpose_y_3, x = q_u_105_cast_fp16, y = k_head_211_cast_fp16)[name = string("op_8539_cast_fp16")];
            fp16 var_8540_to_fp16 = const()[name = string("op_8540_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_105_cast_fp16 = mul(x = var_8539_cast_fp16, y = var_8540_to_fp16)[name = string("scores_content_105_cast_fp16")];
            bool x_557_transpose_x_3 = const()[name = string("x_557_transpose_x_3"), val = bool(false)];
            bool x_557_transpose_y_3 = const()[name = string("x_557_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_557_cast_fp16 = matmul(transpose_x = x_557_transpose_x_3, transpose_y = x_557_transpose_y_3, x = q_v_105_cast_fp16, y = p_head_211_cast_fp16)[name = string("x_557_cast_fp16")];
            tensor<int32, [8]> x_559_pad_1 = const()[name = string("x_559_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_559_mode_1 = const()[name = string("x_559_mode_1"), val = string("constant")];
            fp16 const_1657_to_fp16 = const()[name = string("const_1657_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_559_cast_fp16 = pad(constant_val = const_1657_to_fp16, mode = x_559_mode_1, pad = x_559_pad_1, x = x_557_cast_fp16)[name = string("x_559_cast_fp16")];
            tensor<int32, [4]> var_8554 = const()[name = string("op_8554"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_561_cast_fp16 = reshape(shape = var_8554, x = x_559_cast_fp16)[name = string("x_561_cast_fp16")];
            tensor<int32, [4]> var_8558_begin_1 = const()[name = string("op_8558_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8558_end_1 = const()[name = string("op_8558_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_8558_end_mask_1 = const()[name = string("op_8558_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_8558_cast_fp16 = slice_by_index(begin = var_8558_begin_1, end = var_8558_end_1, end_mask = var_8558_end_mask_1, x = x_561_cast_fp16)[name = string("op_8558_cast_fp16")];
            tensor<int32, [4]> var_8560 = const()[name = string("op_8560"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_8561_cast_fp16 = reshape(shape = var_8560, x = var_8558_cast_fp16)[name = string("op_8561_cast_fp16")];
            tensor<int32, [4]> var_8566_begin_1 = const()[name = string("op_8566_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8566_end_1 = const()[name = string("op_8566_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_8566_end_mask_1 = const()[name = string("op_8566_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_8566_cast_fp16 = slice_by_index(begin = var_8566_begin_1, end = var_8566_end_1, end_mask = var_8566_end_mask_1, x = var_8561_cast_fp16)[name = string("op_8566_cast_fp16")];
            fp16 var_8567_to_fp16 = const()[name = string("op_8567_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_105_cast_fp16 = mul(x = var_8566_cast_fp16, y = var_8567_to_fp16)[name = string("scores_pos_105_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_105_cast_fp16 = add(x = scores_content_105_cast_fp16, y = scores_pos_105_cast_fp16)[name = string("logits_105_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_8570_cast_fp16 = softmax(axis = var_7811, x = logits_105_cast_fp16)[name = string("op_8570_cast_fp16")];
            bool var_8572_transpose_x_1 = const()[name = string("op_8572_transpose_x_1"), val = bool(false)];
            bool var_8572_transpose_y_1 = const()[name = string("op_8572_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_8572_cast_fp16 = matmul(transpose_x = var_8572_transpose_x_1, transpose_y = var_8572_transpose_y_1, x = var_8570_cast_fp16, y = v_head_211_cast_fp16)[name = string("op_8572_cast_fp16")];
            tensor<int32, [1]> var_8573_axes_1 = const()[name = string("op_8573_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_8573_cast_fp16 = squeeze(axes = var_8573_axes_1, x = var_8572_cast_fp16)[name = string("op_8573_cast_fp16")];
            string dense_output_537_pad_type_1 = const()[name = string("dense_output_537_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_537_strides_1 = const()[name = string("dense_output_537_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_537_pad_1 = const()[name = string("dense_output_537_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_537_dilations_1 = const()[name = string("dense_output_537_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_537_groups_1 = const()[name = string("dense_output_537_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193503936))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193635072))))[name = string("layers_6_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_537_cast_fp16 = conv(dilations = dense_output_537_dilations_1, groups = dense_output_537_groups_1, pad = dense_output_537_pad_1, pad_type = dense_output_537_pad_type_1, strides = dense_output_537_strides_1, weight = layers_6_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_537_cast_fp16")];
            string sparse_output_537_pad_type_1 = const()[name = string("sparse_output_537_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_537_strides_1 = const()[name = string("sparse_output_537_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_537_pad_1 = const()[name = string("sparse_output_537_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_537_dilations_1 = const()[name = string("sparse_output_537_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_537_groups_1 = const()[name = string("sparse_output_537_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193638336))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193635648))))[name = string("layers_6_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_537_cast_fp16 = conv(dilations = sparse_output_537_dilations_1, groups = sparse_output_537_groups_1, pad = sparse_output_537_pad_1, pad_type = sparse_output_537_pad_type_1, strides = sparse_output_537_strides_1, weight = layers_6_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_537_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8588_cast_fp16 = add(x = dense_output_537_cast_fp16, y = sparse_output_537_cast_fp16)[name = string("op_8588_cast_fp16")];
            tensor<int32, [4]> var_8589 = const()[name = string("op_8589"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8591 = const()[name = string("op_8591"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8590_cast_fp16 = transpose(perm = var_8589, x = var_8588_cast_fp16)[name = string("transpose_714")];
            tensor<fp16, [1, 51, 128]> q_head_107_cast_fp16 = reshape(shape = var_8591, x = var_8590_cast_fp16)[name = string("q_head_107_cast_fp16")];
            string dense_output_539_pad_type_1 = const()[name = string("dense_output_539_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_539_strides_1 = const()[name = string("dense_output_539_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_539_pad_1 = const()[name = string("dense_output_539_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_539_dilations_1 = const()[name = string("dense_output_539_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_539_groups_1 = const()[name = string("dense_output_539_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193654784))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193785920))))[name = string("layers_6_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_539_cast_fp16 = conv(dilations = dense_output_539_dilations_1, groups = dense_output_539_groups_1, pad = dense_output_539_pad_1, pad_type = dense_output_539_pad_type_1, strides = dense_output_539_strides_1, weight = layers_6_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_539_cast_fp16")];
            string sparse_output_539_pad_type_1 = const()[name = string("sparse_output_539_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_539_strides_1 = const()[name = string("sparse_output_539_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_539_pad_1 = const()[name = string("sparse_output_539_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_539_dilations_1 = const()[name = string("sparse_output_539_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_539_groups_1 = const()[name = string("sparse_output_539_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193789184))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193786496))))[name = string("layers_6_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_539_cast_fp16 = conv(dilations = sparse_output_539_dilations_1, groups = sparse_output_539_groups_1, pad = sparse_output_539_pad_1, pad_type = sparse_output_539_pad_type_1, strides = sparse_output_539_strides_1, weight = layers_6_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_539_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8607_cast_fp16 = add(x = dense_output_539_cast_fp16, y = sparse_output_539_cast_fp16)[name = string("op_8607_cast_fp16")];
            tensor<int32, [4]> var_8608 = const()[name = string("op_8608"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8610 = const()[name = string("op_8610"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8609_cast_fp16 = transpose(perm = var_8608, x = var_8607_cast_fp16)[name = string("transpose_713")];
            tensor<fp16, [1, 51, 128]> k_head_213_cast_fp16 = reshape(shape = var_8610, x = var_8609_cast_fp16)[name = string("k_head_213_cast_fp16")];
            string dense_output_541_pad_type_1 = const()[name = string("dense_output_541_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_541_strides_1 = const()[name = string("dense_output_541_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_541_pad_1 = const()[name = string("dense_output_541_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_541_dilations_1 = const()[name = string("dense_output_541_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_541_groups_1 = const()[name = string("dense_output_541_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193805632))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193936768))))[name = string("layers_6_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_541_cast_fp16 = conv(dilations = dense_output_541_dilations_1, groups = dense_output_541_groups_1, pad = dense_output_541_pad_1, pad_type = dense_output_541_pad_type_1, strides = dense_output_541_strides_1, weight = layers_6_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_541_cast_fp16")];
            string sparse_output_541_pad_type_1 = const()[name = string("sparse_output_541_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_541_strides_1 = const()[name = string("sparse_output_541_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_541_pad_1 = const()[name = string("sparse_output_541_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_541_dilations_1 = const()[name = string("sparse_output_541_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_541_groups_1 = const()[name = string("sparse_output_541_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193940032))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193937344))))[name = string("layers_6_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_541_cast_fp16 = conv(dilations = sparse_output_541_dilations_1, groups = sparse_output_541_groups_1, pad = sparse_output_541_pad_1, pad_type = sparse_output_541_pad_type_1, strides = sparse_output_541_strides_1, weight = layers_6_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_541_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8626_cast_fp16 = add(x = dense_output_541_cast_fp16, y = sparse_output_541_cast_fp16)[name = string("op_8626_cast_fp16")];
            tensor<int32, [4]> var_8627 = const()[name = string("op_8627"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8629 = const()[name = string("op_8629"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8628_cast_fp16 = transpose(perm = var_8627, x = var_8626_cast_fp16)[name = string("transpose_712")];
            tensor<fp16, [1, 51, 128]> v_head_213_cast_fp16 = reshape(shape = var_8629, x = var_8628_cast_fp16)[name = string("v_head_213_cast_fp16")];
            string dense_output_543_pad_type_1 = const()[name = string("dense_output_543_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_543_strides_1 = const()[name = string("dense_output_543_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_543_pad_1 = const()[name = string("dense_output_543_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_543_dilations_1 = const()[name = string("dense_output_543_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_543_groups_1 = const()[name = string("dense_output_543_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193956480))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194087616))))[name = string("layers_6_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_543_cast_fp16 = conv(dilations = dense_output_543_dilations_1, groups = dense_output_543_groups_1, pad = dense_output_543_pad_1, pad_type = dense_output_543_pad_type_1, strides = dense_output_543_strides_1, weight = layers_6_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_543_cast_fp16")];
            string sparse_output_543_pad_type_1 = const()[name = string("sparse_output_543_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_543_strides_1 = const()[name = string("sparse_output_543_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_543_pad_1 = const()[name = string("sparse_output_543_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_543_dilations_1 = const()[name = string("sparse_output_543_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_543_groups_1 = const()[name = string("sparse_output_543_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194090880))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194088192))))[name = string("layers_6_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_543_cast_fp16 = conv(dilations = sparse_output_543_dilations_1, groups = sparse_output_543_groups_1, pad = sparse_output_543_pad_1, pad_type = sparse_output_543_pad_type_1, strides = sparse_output_543_strides_1, weight = layers_6_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_543_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_8645_cast_fp16 = add(x = dense_output_543_cast_fp16, y = sparse_output_543_cast_fp16)[name = string("op_8645_cast_fp16")];
            tensor<int32, [4]> var_8646 = const()[name = string("op_8646"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8648 = const()[name = string("op_8648"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_8647_cast_fp16 = transpose(perm = var_8646, x = var_8645_cast_fp16)[name = string("transpose_711")];
            tensor<fp16, [1, 101, 128]> p_head_213_cast_fp16 = reshape(shape = var_8648, x = var_8647_cast_fp16)[name = string("p_head_213_cast_fp16")];
            tensor<fp16, [128]> var_8650_to_fp16 = const()[name = string("op_8650_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194107328)))];
            tensor<fp16, [1, 51, 128]> var_8651_cast_fp16 = add(x = q_head_107_cast_fp16, y = var_8650_to_fp16)[name = string("op_8651_cast_fp16")];
            tensor<int32, [1]> q_u_107_axes_1 = const()[name = string("q_u_107_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_107_cast_fp16 = expand_dims(axes = q_u_107_axes_1, x = var_8651_cast_fp16)[name = string("q_u_107_cast_fp16")];
            tensor<fp16, [128]> var_8653_to_fp16 = const()[name = string("op_8653_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194107648)))];
            tensor<fp16, [1, 51, 128]> var_8654_cast_fp16 = add(x = q_head_107_cast_fp16, y = var_8653_to_fp16)[name = string("op_8654_cast_fp16")];
            tensor<int32, [1]> q_v_107_axes_1 = const()[name = string("q_v_107_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_107_cast_fp16 = expand_dims(axes = q_v_107_axes_1, x = var_8654_cast_fp16)[name = string("q_v_107_cast_fp16")];
            tensor<int32, [1]> k_head_215_axes_1 = const()[name = string("k_head_215_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_215_cast_fp16 = expand_dims(axes = k_head_215_axes_1, x = k_head_213_cast_fp16)[name = string("k_head_215_cast_fp16")];
            tensor<int32, [1]> v_head_215_axes_1 = const()[name = string("v_head_215_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_215_cast_fp16 = expand_dims(axes = v_head_215_axes_1, x = v_head_213_cast_fp16)[name = string("v_head_215_cast_fp16")];
            tensor<int32, [1]> p_head_215_axes_1 = const()[name = string("p_head_215_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_215_cast_fp16 = expand_dims(axes = p_head_215_axes_1, x = p_head_213_cast_fp16)[name = string("p_head_215_cast_fp16")];
            bool var_8660_transpose_x_3 = const()[name = string("op_8660_transpose_x_3"), val = bool(false)];
            bool var_8660_transpose_y_3 = const()[name = string("op_8660_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_8660_cast_fp16 = matmul(transpose_x = var_8660_transpose_x_3, transpose_y = var_8660_transpose_y_3, x = q_u_107_cast_fp16, y = k_head_215_cast_fp16)[name = string("op_8660_cast_fp16")];
            fp16 var_8661_to_fp16 = const()[name = string("op_8661_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_107_cast_fp16 = mul(x = var_8660_cast_fp16, y = var_8661_to_fp16)[name = string("scores_content_107_cast_fp16")];
            bool x_565_transpose_x_3 = const()[name = string("x_565_transpose_x_3"), val = bool(false)];
            bool x_565_transpose_y_3 = const()[name = string("x_565_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_565_cast_fp16 = matmul(transpose_x = x_565_transpose_x_3, transpose_y = x_565_transpose_y_3, x = q_v_107_cast_fp16, y = p_head_215_cast_fp16)[name = string("x_565_cast_fp16")];
            tensor<int32, [8]> x_567_pad_1 = const()[name = string("x_567_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_567_mode_1 = const()[name = string("x_567_mode_1"), val = string("constant")];
            fp16 const_1663_to_fp16 = const()[name = string("const_1663_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_567_cast_fp16 = pad(constant_val = const_1663_to_fp16, mode = x_567_mode_1, pad = x_567_pad_1, x = x_565_cast_fp16)[name = string("x_567_cast_fp16")];
            tensor<int32, [4]> var_8675 = const()[name = string("op_8675"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_569_cast_fp16 = reshape(shape = var_8675, x = x_567_cast_fp16)[name = string("x_569_cast_fp16")];
            tensor<int32, [4]> var_8679_begin_1 = const()[name = string("op_8679_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8679_end_1 = const()[name = string("op_8679_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_8679_end_mask_1 = const()[name = string("op_8679_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_8679_cast_fp16 = slice_by_index(begin = var_8679_begin_1, end = var_8679_end_1, end_mask = var_8679_end_mask_1, x = x_569_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<int32, [4]> var_8681 = const()[name = string("op_8681"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_8682_cast_fp16 = reshape(shape = var_8681, x = var_8679_cast_fp16)[name = string("op_8682_cast_fp16")];
            tensor<int32, [4]> var_8687_begin_1 = const()[name = string("op_8687_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8687_end_1 = const()[name = string("op_8687_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_8687_end_mask_1 = const()[name = string("op_8687_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_8687_cast_fp16 = slice_by_index(begin = var_8687_begin_1, end = var_8687_end_1, end_mask = var_8687_end_mask_1, x = var_8682_cast_fp16)[name = string("op_8687_cast_fp16")];
            fp16 var_8688_to_fp16 = const()[name = string("op_8688_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_107_cast_fp16 = mul(x = var_8687_cast_fp16, y = var_8688_to_fp16)[name = string("scores_pos_107_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_107_cast_fp16 = add(x = scores_content_107_cast_fp16, y = scores_pos_107_cast_fp16)[name = string("logits_107_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_8691_cast_fp16 = softmax(axis = var_7811, x = logits_107_cast_fp16)[name = string("op_8691_cast_fp16")];
            bool var_8693_transpose_x_1 = const()[name = string("op_8693_transpose_x_1"), val = bool(false)];
            bool var_8693_transpose_y_1 = const()[name = string("op_8693_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_8693_cast_fp16 = matmul(transpose_x = var_8693_transpose_x_1, transpose_y = var_8693_transpose_y_1, x = var_8691_cast_fp16, y = v_head_215_cast_fp16)[name = string("op_8693_cast_fp16")];
            tensor<int32, [1]> var_8694_axes_1 = const()[name = string("op_8694_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_8694_cast_fp16 = squeeze(axes = var_8694_axes_1, x = var_8693_cast_fp16)[name = string("op_8694_cast_fp16")];
            string dense_output_545_pad_type_1 = const()[name = string("dense_output_545_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_545_strides_1 = const()[name = string("dense_output_545_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_545_pad_1 = const()[name = string("dense_output_545_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_545_dilations_1 = const()[name = string("dense_output_545_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_545_groups_1 = const()[name = string("dense_output_545_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194107968))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194239104))))[name = string("layers_6_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_545_cast_fp16 = conv(dilations = dense_output_545_dilations_1, groups = dense_output_545_groups_1, pad = dense_output_545_pad_1, pad_type = dense_output_545_pad_type_1, strides = dense_output_545_strides_1, weight = layers_6_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_545_cast_fp16")];
            string sparse_output_545_pad_type_1 = const()[name = string("sparse_output_545_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_545_strides_1 = const()[name = string("sparse_output_545_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_545_pad_1 = const()[name = string("sparse_output_545_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_545_dilations_1 = const()[name = string("sparse_output_545_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_545_groups_1 = const()[name = string("sparse_output_545_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194242368))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194239680))))[name = string("layers_6_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_545_cast_fp16 = conv(dilations = sparse_output_545_dilations_1, groups = sparse_output_545_groups_1, pad = sparse_output_545_pad_1, pad_type = sparse_output_545_pad_type_1, strides = sparse_output_545_strides_1, weight = layers_6_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_545_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8709_cast_fp16 = add(x = dense_output_545_cast_fp16, y = sparse_output_545_cast_fp16)[name = string("op_8709_cast_fp16")];
            tensor<int32, [4]> var_8710 = const()[name = string("op_8710"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8712 = const()[name = string("op_8712"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8711_cast_fp16 = transpose(perm = var_8710, x = var_8709_cast_fp16)[name = string("transpose_710")];
            tensor<fp16, [1, 51, 128]> q_head_109_cast_fp16 = reshape(shape = var_8712, x = var_8711_cast_fp16)[name = string("q_head_109_cast_fp16")];
            string dense_output_547_pad_type_1 = const()[name = string("dense_output_547_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_547_strides_1 = const()[name = string("dense_output_547_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_547_pad_1 = const()[name = string("dense_output_547_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_547_dilations_1 = const()[name = string("dense_output_547_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_547_groups_1 = const()[name = string("dense_output_547_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194258816))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194389952))))[name = string("layers_6_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_547_cast_fp16 = conv(dilations = dense_output_547_dilations_1, groups = dense_output_547_groups_1, pad = dense_output_547_pad_1, pad_type = dense_output_547_pad_type_1, strides = dense_output_547_strides_1, weight = layers_6_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_547_cast_fp16")];
            string sparse_output_547_pad_type_1 = const()[name = string("sparse_output_547_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_547_strides_1 = const()[name = string("sparse_output_547_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_547_pad_1 = const()[name = string("sparse_output_547_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_547_dilations_1 = const()[name = string("sparse_output_547_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_547_groups_1 = const()[name = string("sparse_output_547_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194393216))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194390528))))[name = string("layers_6_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_547_cast_fp16 = conv(dilations = sparse_output_547_dilations_1, groups = sparse_output_547_groups_1, pad = sparse_output_547_pad_1, pad_type = sparse_output_547_pad_type_1, strides = sparse_output_547_strides_1, weight = layers_6_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_547_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8728_cast_fp16 = add(x = dense_output_547_cast_fp16, y = sparse_output_547_cast_fp16)[name = string("op_8728_cast_fp16")];
            tensor<int32, [4]> var_8729 = const()[name = string("op_8729"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8731 = const()[name = string("op_8731"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8730_cast_fp16 = transpose(perm = var_8729, x = var_8728_cast_fp16)[name = string("transpose_709")];
            tensor<fp16, [1, 51, 128]> k_head_217_cast_fp16 = reshape(shape = var_8731, x = var_8730_cast_fp16)[name = string("k_head_217_cast_fp16")];
            string dense_output_549_pad_type_1 = const()[name = string("dense_output_549_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_549_strides_1 = const()[name = string("dense_output_549_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_549_pad_1 = const()[name = string("dense_output_549_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_549_dilations_1 = const()[name = string("dense_output_549_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_549_groups_1 = const()[name = string("dense_output_549_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194409664))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194540800))))[name = string("layers_6_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_549_cast_fp16 = conv(dilations = dense_output_549_dilations_1, groups = dense_output_549_groups_1, pad = dense_output_549_pad_1, pad_type = dense_output_549_pad_type_1, strides = dense_output_549_strides_1, weight = layers_6_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_549_cast_fp16")];
            string sparse_output_549_pad_type_1 = const()[name = string("sparse_output_549_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_549_strides_1 = const()[name = string("sparse_output_549_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_549_pad_1 = const()[name = string("sparse_output_549_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_549_dilations_1 = const()[name = string("sparse_output_549_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_549_groups_1 = const()[name = string("sparse_output_549_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194544064))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194541376))))[name = string("layers_6_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_549_cast_fp16 = conv(dilations = sparse_output_549_dilations_1, groups = sparse_output_549_groups_1, pad = sparse_output_549_pad_1, pad_type = sparse_output_549_pad_type_1, strides = sparse_output_549_strides_1, weight = layers_6_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_549_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8747_cast_fp16 = add(x = dense_output_549_cast_fp16, y = sparse_output_549_cast_fp16)[name = string("op_8747_cast_fp16")];
            tensor<int32, [4]> var_8748 = const()[name = string("op_8748"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8750 = const()[name = string("op_8750"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8749_cast_fp16 = transpose(perm = var_8748, x = var_8747_cast_fp16)[name = string("transpose_708")];
            tensor<fp16, [1, 51, 128]> v_head_217_cast_fp16 = reshape(shape = var_8750, x = var_8749_cast_fp16)[name = string("v_head_217_cast_fp16")];
            string dense_output_551_pad_type_1 = const()[name = string("dense_output_551_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_551_strides_1 = const()[name = string("dense_output_551_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_551_pad_1 = const()[name = string("dense_output_551_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_551_dilations_1 = const()[name = string("dense_output_551_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_551_groups_1 = const()[name = string("dense_output_551_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194560512))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194691648))))[name = string("layers_6_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_551_cast_fp16 = conv(dilations = dense_output_551_dilations_1, groups = dense_output_551_groups_1, pad = dense_output_551_pad_1, pad_type = dense_output_551_pad_type_1, strides = dense_output_551_strides_1, weight = layers_6_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_551_cast_fp16")];
            string sparse_output_551_pad_type_1 = const()[name = string("sparse_output_551_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_551_strides_1 = const()[name = string("sparse_output_551_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_551_pad_1 = const()[name = string("sparse_output_551_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_551_dilations_1 = const()[name = string("sparse_output_551_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_551_groups_1 = const()[name = string("sparse_output_551_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194694912))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194692224))))[name = string("layers_6_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_551_cast_fp16 = conv(dilations = sparse_output_551_dilations_1, groups = sparse_output_551_groups_1, pad = sparse_output_551_pad_1, pad_type = sparse_output_551_pad_type_1, strides = sparse_output_551_strides_1, weight = layers_6_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_551_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_8766_cast_fp16 = add(x = dense_output_551_cast_fp16, y = sparse_output_551_cast_fp16)[name = string("op_8766_cast_fp16")];
            tensor<int32, [4]> var_8767 = const()[name = string("op_8767"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8769 = const()[name = string("op_8769"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_8768_cast_fp16 = transpose(perm = var_8767, x = var_8766_cast_fp16)[name = string("transpose_707")];
            tensor<fp16, [1, 101, 128]> p_head_217_cast_fp16 = reshape(shape = var_8769, x = var_8768_cast_fp16)[name = string("p_head_217_cast_fp16")];
            tensor<fp16, [128]> var_8771_to_fp16 = const()[name = string("op_8771_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194711360)))];
            tensor<fp16, [1, 51, 128]> var_8772_cast_fp16 = add(x = q_head_109_cast_fp16, y = var_8771_to_fp16)[name = string("op_8772_cast_fp16")];
            tensor<int32, [1]> q_u_109_axes_1 = const()[name = string("q_u_109_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_109_cast_fp16 = expand_dims(axes = q_u_109_axes_1, x = var_8772_cast_fp16)[name = string("q_u_109_cast_fp16")];
            tensor<fp16, [128]> var_8774_to_fp16 = const()[name = string("op_8774_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194711680)))];
            tensor<fp16, [1, 51, 128]> var_8775_cast_fp16 = add(x = q_head_109_cast_fp16, y = var_8774_to_fp16)[name = string("op_8775_cast_fp16")];
            tensor<int32, [1]> q_v_109_axes_1 = const()[name = string("q_v_109_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_109_cast_fp16 = expand_dims(axes = q_v_109_axes_1, x = var_8775_cast_fp16)[name = string("q_v_109_cast_fp16")];
            tensor<int32, [1]> k_head_219_axes_1 = const()[name = string("k_head_219_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_219_cast_fp16 = expand_dims(axes = k_head_219_axes_1, x = k_head_217_cast_fp16)[name = string("k_head_219_cast_fp16")];
            tensor<int32, [1]> v_head_219_axes_1 = const()[name = string("v_head_219_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_219_cast_fp16 = expand_dims(axes = v_head_219_axes_1, x = v_head_217_cast_fp16)[name = string("v_head_219_cast_fp16")];
            tensor<int32, [1]> p_head_219_axes_1 = const()[name = string("p_head_219_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_219_cast_fp16 = expand_dims(axes = p_head_219_axes_1, x = p_head_217_cast_fp16)[name = string("p_head_219_cast_fp16")];
            bool var_8781_transpose_x_3 = const()[name = string("op_8781_transpose_x_3"), val = bool(false)];
            bool var_8781_transpose_y_3 = const()[name = string("op_8781_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_8781_cast_fp16 = matmul(transpose_x = var_8781_transpose_x_3, transpose_y = var_8781_transpose_y_3, x = q_u_109_cast_fp16, y = k_head_219_cast_fp16)[name = string("op_8781_cast_fp16")];
            fp16 var_8782_to_fp16 = const()[name = string("op_8782_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_109_cast_fp16 = mul(x = var_8781_cast_fp16, y = var_8782_to_fp16)[name = string("scores_content_109_cast_fp16")];
            bool x_573_transpose_x_3 = const()[name = string("x_573_transpose_x_3"), val = bool(false)];
            bool x_573_transpose_y_3 = const()[name = string("x_573_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_573_cast_fp16 = matmul(transpose_x = x_573_transpose_x_3, transpose_y = x_573_transpose_y_3, x = q_v_109_cast_fp16, y = p_head_219_cast_fp16)[name = string("x_573_cast_fp16")];
            tensor<int32, [8]> x_575_pad_1 = const()[name = string("x_575_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_575_mode_1 = const()[name = string("x_575_mode_1"), val = string("constant")];
            fp16 const_1669_to_fp16 = const()[name = string("const_1669_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_575_cast_fp16 = pad(constant_val = const_1669_to_fp16, mode = x_575_mode_1, pad = x_575_pad_1, x = x_573_cast_fp16)[name = string("x_575_cast_fp16")];
            tensor<int32, [4]> var_8796 = const()[name = string("op_8796"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_577_cast_fp16 = reshape(shape = var_8796, x = x_575_cast_fp16)[name = string("x_577_cast_fp16")];
            tensor<int32, [4]> var_8800_begin_1 = const()[name = string("op_8800_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8800_end_1 = const()[name = string("op_8800_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_8800_end_mask_1 = const()[name = string("op_8800_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_8800_cast_fp16 = slice_by_index(begin = var_8800_begin_1, end = var_8800_end_1, end_mask = var_8800_end_mask_1, x = x_577_cast_fp16)[name = string("op_8800_cast_fp16")];
            tensor<int32, [4]> var_8802 = const()[name = string("op_8802"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_8803_cast_fp16 = reshape(shape = var_8802, x = var_8800_cast_fp16)[name = string("op_8803_cast_fp16")];
            tensor<int32, [4]> var_8808_begin_1 = const()[name = string("op_8808_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8808_end_1 = const()[name = string("op_8808_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_8808_end_mask_1 = const()[name = string("op_8808_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_8808_cast_fp16 = slice_by_index(begin = var_8808_begin_1, end = var_8808_end_1, end_mask = var_8808_end_mask_1, x = var_8803_cast_fp16)[name = string("op_8808_cast_fp16")];
            fp16 var_8809_to_fp16 = const()[name = string("op_8809_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_109_cast_fp16 = mul(x = var_8808_cast_fp16, y = var_8809_to_fp16)[name = string("scores_pos_109_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_109_cast_fp16 = add(x = scores_content_109_cast_fp16, y = scores_pos_109_cast_fp16)[name = string("logits_109_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_8812_cast_fp16 = softmax(axis = var_7811, x = logits_109_cast_fp16)[name = string("op_8812_cast_fp16")];
            bool var_8814_transpose_x_1 = const()[name = string("op_8814_transpose_x_1"), val = bool(false)];
            bool var_8814_transpose_y_1 = const()[name = string("op_8814_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_8814_cast_fp16 = matmul(transpose_x = var_8814_transpose_x_1, transpose_y = var_8814_transpose_y_1, x = var_8812_cast_fp16, y = v_head_219_cast_fp16)[name = string("op_8814_cast_fp16")];
            tensor<int32, [1]> var_8815_axes_1 = const()[name = string("op_8815_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_8815_cast_fp16 = squeeze(axes = var_8815_axes_1, x = var_8814_cast_fp16)[name = string("op_8815_cast_fp16")];
            string dense_output_553_pad_type_1 = const()[name = string("dense_output_553_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_553_strides_1 = const()[name = string("dense_output_553_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_553_pad_1 = const()[name = string("dense_output_553_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_553_dilations_1 = const()[name = string("dense_output_553_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_553_groups_1 = const()[name = string("dense_output_553_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194712000))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194843136))))[name = string("layers_6_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_553_cast_fp16 = conv(dilations = dense_output_553_dilations_1, groups = dense_output_553_groups_1, pad = dense_output_553_pad_1, pad_type = dense_output_553_pad_type_1, strides = dense_output_553_strides_1, weight = layers_6_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_553_cast_fp16")];
            string sparse_output_553_pad_type_1 = const()[name = string("sparse_output_553_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_553_strides_1 = const()[name = string("sparse_output_553_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_553_pad_1 = const()[name = string("sparse_output_553_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_553_dilations_1 = const()[name = string("sparse_output_553_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_553_groups_1 = const()[name = string("sparse_output_553_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194846400))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194843712))))[name = string("layers_6_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_553_cast_fp16 = conv(dilations = sparse_output_553_dilations_1, groups = sparse_output_553_groups_1, pad = sparse_output_553_pad_1, pad_type = sparse_output_553_pad_type_1, strides = sparse_output_553_strides_1, weight = layers_6_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_553_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8830_cast_fp16 = add(x = dense_output_553_cast_fp16, y = sparse_output_553_cast_fp16)[name = string("op_8830_cast_fp16")];
            tensor<int32, [4]> var_8831 = const()[name = string("op_8831"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8833 = const()[name = string("op_8833"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8832_cast_fp16 = transpose(perm = var_8831, x = var_8830_cast_fp16)[name = string("transpose_706")];
            tensor<fp16, [1, 51, 128]> q_head_111_cast_fp16 = reshape(shape = var_8833, x = var_8832_cast_fp16)[name = string("q_head_111_cast_fp16")];
            string dense_output_555_pad_type_1 = const()[name = string("dense_output_555_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_555_strides_1 = const()[name = string("dense_output_555_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_555_pad_1 = const()[name = string("dense_output_555_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_555_dilations_1 = const()[name = string("dense_output_555_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_555_groups_1 = const()[name = string("dense_output_555_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194862848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194993984))))[name = string("layers_6_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_555_cast_fp16 = conv(dilations = dense_output_555_dilations_1, groups = dense_output_555_groups_1, pad = dense_output_555_pad_1, pad_type = dense_output_555_pad_type_1, strides = dense_output_555_strides_1, weight = layers_6_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_555_cast_fp16")];
            string sparse_output_555_pad_type_1 = const()[name = string("sparse_output_555_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_555_strides_1 = const()[name = string("sparse_output_555_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_555_pad_1 = const()[name = string("sparse_output_555_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_555_dilations_1 = const()[name = string("sparse_output_555_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_555_groups_1 = const()[name = string("sparse_output_555_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194997248))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194994560))))[name = string("layers_6_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_555_cast_fp16 = conv(dilations = sparse_output_555_dilations_1, groups = sparse_output_555_groups_1, pad = sparse_output_555_pad_1, pad_type = sparse_output_555_pad_type_1, strides = sparse_output_555_strides_1, weight = layers_6_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_555_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8849_cast_fp16 = add(x = dense_output_555_cast_fp16, y = sparse_output_555_cast_fp16)[name = string("op_8849_cast_fp16")];
            tensor<int32, [4]> var_8850 = const()[name = string("op_8850"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8852 = const()[name = string("op_8852"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8851_cast_fp16 = transpose(perm = var_8850, x = var_8849_cast_fp16)[name = string("transpose_705")];
            tensor<fp16, [1, 51, 128]> k_head_221_cast_fp16 = reshape(shape = var_8852, x = var_8851_cast_fp16)[name = string("k_head_221_cast_fp16")];
            string dense_output_557_pad_type_1 = const()[name = string("dense_output_557_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_557_strides_1 = const()[name = string("dense_output_557_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_557_pad_1 = const()[name = string("dense_output_557_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_557_dilations_1 = const()[name = string("dense_output_557_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_557_groups_1 = const()[name = string("dense_output_557_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195013696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195144832))))[name = string("layers_6_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_557_cast_fp16 = conv(dilations = dense_output_557_dilations_1, groups = dense_output_557_groups_1, pad = dense_output_557_pad_1, pad_type = dense_output_557_pad_type_1, strides = dense_output_557_strides_1, weight = layers_6_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_305_cast_fp16)[name = string("dense_output_557_cast_fp16")];
            string sparse_output_557_pad_type_1 = const()[name = string("sparse_output_557_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_557_strides_1 = const()[name = string("sparse_output_557_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_557_pad_1 = const()[name = string("sparse_output_557_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_557_dilations_1 = const()[name = string("sparse_output_557_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_557_groups_1 = const()[name = string("sparse_output_557_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195148096))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195145408))))[name = string("layers_6_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_557_cast_fp16 = conv(dilations = sparse_output_557_dilations_1, groups = sparse_output_557_groups_1, pad = sparse_output_557_pad_1, pad_type = sparse_output_557_pad_type_1, strides = sparse_output_557_strides_1, weight = layers_6_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_305_cast_fp16)[name = string("sparse_output_557_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_8868_cast_fp16 = add(x = dense_output_557_cast_fp16, y = sparse_output_557_cast_fp16)[name = string("op_8868_cast_fp16")];
            tensor<int32, [4]> var_8869 = const()[name = string("op_8869"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8871 = const()[name = string("op_8871"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_8870_cast_fp16 = transpose(perm = var_8869, x = var_8868_cast_fp16)[name = string("transpose_704")];
            tensor<fp16, [1, 51, 128]> v_head_221_cast_fp16 = reshape(shape = var_8871, x = var_8870_cast_fp16)[name = string("v_head_221_cast_fp16")];
            string dense_output_559_pad_type_1 = const()[name = string("dense_output_559_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_559_strides_1 = const()[name = string("dense_output_559_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_559_pad_1 = const()[name = string("dense_output_559_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_559_dilations_1 = const()[name = string("dense_output_559_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_559_groups_1 = const()[name = string("dense_output_559_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195164544))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195295680))))[name = string("layers_6_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_559_cast_fp16 = conv(dilations = dense_output_559_dilations_1, groups = dense_output_559_groups_1, pad = dense_output_559_pad_1, pad_type = dense_output_559_pad_type_1, strides = dense_output_559_strides_1, weight = layers_6_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_559_cast_fp16")];
            string sparse_output_559_pad_type_1 = const()[name = string("sparse_output_559_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_559_strides_1 = const()[name = string("sparse_output_559_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_559_pad_1 = const()[name = string("sparse_output_559_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_559_dilations_1 = const()[name = string("sparse_output_559_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_559_groups_1 = const()[name = string("sparse_output_559_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_6_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195298944))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195296256))))[name = string("layers_6_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_559_cast_fp16 = conv(dilations = sparse_output_559_dilations_1, groups = sparse_output_559_groups_1, pad = sparse_output_559_pad_1, pad_type = sparse_output_559_pad_type_1, strides = sparse_output_559_strides_1, weight = layers_6_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_559_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_8887_cast_fp16 = add(x = dense_output_559_cast_fp16, y = sparse_output_559_cast_fp16)[name = string("op_8887_cast_fp16")];
            tensor<int32, [4]> var_8888 = const()[name = string("op_8888"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_8890 = const()[name = string("op_8890"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_8889_cast_fp16 = transpose(perm = var_8888, x = var_8887_cast_fp16)[name = string("transpose_703")];
            tensor<fp16, [1, 101, 128]> p_head_221_cast_fp16 = reshape(shape = var_8890, x = var_8889_cast_fp16)[name = string("p_head_221_cast_fp16")];
            tensor<fp16, [128]> var_8892_to_fp16 = const()[name = string("op_8892_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195315392)))];
            tensor<fp16, [1, 51, 128]> var_8893_cast_fp16 = add(x = q_head_111_cast_fp16, y = var_8892_to_fp16)[name = string("op_8893_cast_fp16")];
            tensor<int32, [1]> q_u_111_axes_1 = const()[name = string("q_u_111_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_111_cast_fp16 = expand_dims(axes = q_u_111_axes_1, x = var_8893_cast_fp16)[name = string("q_u_111_cast_fp16")];
            tensor<fp16, [128]> var_8895_to_fp16 = const()[name = string("op_8895_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195315712)))];
            tensor<fp16, [1, 51, 128]> var_8896_cast_fp16 = add(x = q_head_111_cast_fp16, y = var_8895_to_fp16)[name = string("op_8896_cast_fp16")];
            tensor<int32, [1]> q_v_111_axes_1 = const()[name = string("q_v_111_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_111_cast_fp16 = expand_dims(axes = q_v_111_axes_1, x = var_8896_cast_fp16)[name = string("q_v_111_cast_fp16")];
            tensor<int32, [1]> k_head_223_axes_1 = const()[name = string("k_head_223_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_223_cast_fp16 = expand_dims(axes = k_head_223_axes_1, x = k_head_221_cast_fp16)[name = string("k_head_223_cast_fp16")];
            tensor<int32, [1]> v_head_223_axes_1 = const()[name = string("v_head_223_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_223_cast_fp16 = expand_dims(axes = v_head_223_axes_1, x = v_head_221_cast_fp16)[name = string("v_head_223_cast_fp16")];
            tensor<int32, [1]> p_head_223_axes_1 = const()[name = string("p_head_223_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_223_cast_fp16 = expand_dims(axes = p_head_223_axes_1, x = p_head_221_cast_fp16)[name = string("p_head_223_cast_fp16")];
            bool var_8902_transpose_x_3 = const()[name = string("op_8902_transpose_x_3"), val = bool(false)];
            bool var_8902_transpose_y_3 = const()[name = string("op_8902_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_8902_cast_fp16 = matmul(transpose_x = var_8902_transpose_x_3, transpose_y = var_8902_transpose_y_3, x = q_u_111_cast_fp16, y = k_head_223_cast_fp16)[name = string("op_8902_cast_fp16")];
            fp16 var_8903_to_fp16 = const()[name = string("op_8903_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_111_cast_fp16 = mul(x = var_8902_cast_fp16, y = var_8903_to_fp16)[name = string("scores_content_111_cast_fp16")];
            bool x_581_transpose_x_3 = const()[name = string("x_581_transpose_x_3"), val = bool(false)];
            bool x_581_transpose_y_3 = const()[name = string("x_581_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_581_cast_fp16 = matmul(transpose_x = x_581_transpose_x_3, transpose_y = x_581_transpose_y_3, x = q_v_111_cast_fp16, y = p_head_223_cast_fp16)[name = string("x_581_cast_fp16")];
            tensor<int32, [8]> x_583_pad_1 = const()[name = string("x_583_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_583_mode_1 = const()[name = string("x_583_mode_1"), val = string("constant")];
            fp16 const_1675_to_fp16 = const()[name = string("const_1675_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_583_cast_fp16 = pad(constant_val = const_1675_to_fp16, mode = x_583_mode_1, pad = x_583_pad_1, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")];
            tensor<int32, [4]> var_8917 = const()[name = string("op_8917"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_585_cast_fp16 = reshape(shape = var_8917, x = x_583_cast_fp16)[name = string("x_585_cast_fp16")];
            tensor<int32, [4]> var_8921_begin_1 = const()[name = string("op_8921_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_8921_end_1 = const()[name = string("op_8921_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_8921_end_mask_1 = const()[name = string("op_8921_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_8921_cast_fp16 = slice_by_index(begin = var_8921_begin_1, end = var_8921_end_1, end_mask = var_8921_end_mask_1, x = x_585_cast_fp16)[name = string("op_8921_cast_fp16")];
            tensor<int32, [4]> var_8923 = const()[name = string("op_8923"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_8924_cast_fp16 = reshape(shape = var_8923, x = var_8921_cast_fp16)[name = string("op_8924_cast_fp16")];
            tensor<int32, [4]> var_8929_begin_1 = const()[name = string("op_8929_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8929_end_1 = const()[name = string("op_8929_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_8929_end_mask_1 = const()[name = string("op_8929_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_1, end = var_8929_end_1, end_mask = var_8929_end_mask_1, x = var_8924_cast_fp16)[name = string("op_8929_cast_fp16")];
            fp16 var_8930_to_fp16 = const()[name = string("op_8930_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_111_cast_fp16 = mul(x = var_8929_cast_fp16, y = var_8930_to_fp16)[name = string("scores_pos_111_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_111_cast_fp16 = add(x = scores_content_111_cast_fp16, y = scores_pos_111_cast_fp16)[name = string("logits_111_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_8933_cast_fp16 = softmax(axis = var_7811, x = logits_111_cast_fp16)[name = string("op_8933_cast_fp16")];
            bool var_8935_transpose_x_1 = const()[name = string("op_8935_transpose_x_1"), val = bool(false)];
            bool var_8935_transpose_y_1 = const()[name = string("op_8935_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_8935_cast_fp16 = matmul(transpose_x = var_8935_transpose_x_1, transpose_y = var_8935_transpose_y_1, x = var_8933_cast_fp16, y = v_head_223_cast_fp16)[name = string("op_8935_cast_fp16")];
            tensor<int32, [1]> o_head_13_axes_1 = const()[name = string("o_head_13_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_13_cast_fp16 = squeeze(axes = o_head_13_axes_1, x = var_8935_cast_fp16)[name = string("o_head_13_cast_fp16")];
            bool out_13_interleave_1 = const()[name = string("out_13_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_13_cast_fp16 = concat(axis = var_7811, interleave = out_13_interleave_1, values = (var_8089_cast_fp16, var_8210_cast_fp16, var_8331_cast_fp16, var_8452_cast_fp16, var_8573_cast_fp16, var_8694_cast_fp16, var_8815_cast_fp16, o_head_13_cast_fp16))[name = string("out_13_cast_fp16")];
            tensor<int32, [3]> var_8939_perm_1 = const()[name = string("op_8939_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_313_axes_1 = const()[name = string("input_313_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_8939_cast_fp16 = transpose(perm = var_8939_perm_1, x = out_13_cast_fp16)[name = string("transpose_702")];
            tensor<fp16, [1, 1024, 51, 1]> input_313_cast_fp16 = expand_dims(axes = input_313_axes_1, x = var_8939_cast_fp16)[name = string("input_313_cast_fp16")];
            string dense_output_561_pad_type_1 = const()[name = string("dense_output_561_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_561_strides_1 = const()[name = string("dense_output_561_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_561_pad_1 = const()[name = string("dense_output_561_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_561_dilations_1 = const()[name = string("dense_output_561_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_561_groups_1 = const()[name = string("dense_output_561_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195316032))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196364672))))[name = string("layers_6_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_561_cast_fp16 = conv(dilations = dense_output_561_dilations_1, groups = dense_output_561_groups_1, pad = dense_output_561_pad_1, pad_type = dense_output_561_pad_type_1, strides = dense_output_561_strides_1, weight = layers_6_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_313_cast_fp16)[name = string("dense_output_561_cast_fp16")];
            string sparse_output_561_pad_type_1 = const()[name = string("sparse_output_561_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_561_strides_1 = const()[name = string("sparse_output_561_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_561_pad_1 = const()[name = string("sparse_output_561_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_561_dilations_1 = const()[name = string("sparse_output_561_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_561_groups_1 = const()[name = string("sparse_output_561_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196386304))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196365248))))[name = string("layers_6_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_561_cast_fp16 = conv(dilations = sparse_output_561_dilations_1, groups = sparse_output_561_groups_1, pad = sparse_output_561_pad_1, pad_type = sparse_output_561_pad_type_1, strides = sparse_output_561_strides_1, weight = layers_6_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_313_cast_fp16)[name = string("sparse_output_561_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_13_cast_fp16 = add(x = dense_output_561_cast_fp16, y = sparse_output_561_cast_fp16)[name = string("out_conv_13_cast_fp16")];
            tensor<int32, [1]> var_8956_axes_1 = const()[name = string("op_8956_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_8956_cast_fp16 = squeeze(axes = var_8956_axes_1, x = out_conv_13_cast_fp16)[name = string("op_8956_cast_fp16")];
            tensor<int32, [3]> var_8957_perm_1 = const()[name = string("op_8957_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_8957_cast_fp16 = transpose(perm = var_8957_perm_1, x = var_8956_cast_fp16)[name = string("transpose_701")];
            tensor<fp16, [1, 51, 1024]> input_315_cast_fp16 = add(x = input_303_cast_fp16, y = var_8957_cast_fp16)[name = string("input_315_cast_fp16")];
            tensor<int32, [1]> x_589_axes_1 = const()[name = string("x_589_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_6_norm_conv_weight_to_fp16 = const()[name = string("layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196517440)))];
            tensor<fp16, [1024]> layers_6_norm_conv_bias_to_fp16 = const()[name = string("layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196519552)))];
            tensor<fp16, [1, 51, 1024]> x_589_cast_fp16 = layer_norm(axes = x_589_axes_1, beta = layers_6_norm_conv_bias_to_fp16, epsilon = var_7826_to_fp16, gamma = layers_6_norm_conv_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_589_cast_fp16")];
            tensor<int32, [3]> var_8967_perm_1 = const()[name = string("op_8967_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_317_axes_1 = const()[name = string("input_317_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_8967_cast_fp16 = transpose(perm = var_8967_perm_1, x = x_589_cast_fp16)[name = string("transpose_700")];
            tensor<fp16, [1, 1024, 51, 1]> input_317_cast_fp16 = expand_dims(axes = input_317_axes_1, x = var_8967_cast_fp16)[name = string("input_317_cast_fp16")];
            string dense_output_563_pad_type_1 = const()[name = string("dense_output_563_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_563_strides_1 = const()[name = string("dense_output_563_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_563_pad_1 = const()[name = string("dense_output_563_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_563_dilations_1 = const()[name = string("dense_output_563_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_563_groups_1 = const()[name = string("dense_output_563_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_6_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196521664))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198618880))))[name = string("layers_6_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_563_cast_fp16 = conv(dilations = dense_output_563_dilations_1, groups = dense_output_563_groups_1, pad = dense_output_563_pad_1, pad_type = dense_output_563_pad_type_1, strides = dense_output_563_strides_1, weight = layers_6_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_317_cast_fp16)[name = string("dense_output_563_cast_fp16")];
            string sparse_output_563_pad_type_1 = const()[name = string("sparse_output_563_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_563_strides_1 = const()[name = string("sparse_output_563_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_563_pad_1 = const()[name = string("sparse_output_563_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_563_dilations_1 = const()[name = string("sparse_output_563_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_563_groups_1 = const()[name = string("sparse_output_563_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_6_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198661504))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198619456))))[name = string("layers_6_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_563_cast_fp16 = conv(dilations = sparse_output_563_dilations_1, groups = sparse_output_563_groups_1, pad = sparse_output_563_pad_1, pad_type = sparse_output_563_pad_type_1, strides = sparse_output_563_strides_1, weight = layers_6_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_317_cast_fp16)[name = string("sparse_output_563_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_319_cast_fp16 = add(x = dense_output_563_cast_fp16, y = sparse_output_563_cast_fp16)[name = string("input_319_cast_fp16")];
            int32 input_321_split_num_splits_1 = const()[name = string("input_321_split_num_splits_1"), val = int32(2)];
            int32 input_321_split_axis_1 = const()[name = string("input_321_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_321_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_321_split_cast_fp16_1 = split(axis = input_321_split_axis_1, num_splits = input_321_split_num_splits_1, x = input_319_cast_fp16)[name = string("input_321_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_321_split_1_sigmoid_cast_fp16 = sigmoid(x = input_321_split_cast_fp16_1)[name = string("input_321_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_321_cast_fp16 = mul(x = input_321_split_cast_fp16_0, y = input_321_split_1_sigmoid_cast_fp16)[name = string("input_321_cast_fp16")];
            tensor<int32, [8]> input_323_pad_1 = const()[name = string("input_323_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_323_mode_1 = const()[name = string("input_323_mode_1"), val = string("constant")];
            fp16 const_1677_to_fp16 = const()[name = string("const_1677_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_323_cast_fp16 = pad(constant_val = const_1677_to_fp16, mode = input_323_mode_1, pad = input_323_pad_1, x = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
            string dense_output_565_pad_type_1 = const()[name = string("dense_output_565_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_565_strides_1 = const()[name = string("dense_output_565_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_565_pad_1 = const()[name = string("dense_output_565_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_565_dilations_1 = const()[name = string("dense_output_565_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_565_groups_1 = const()[name = string("dense_output_565_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_565_cast_fp16 = conv(dilations = dense_output_565_dilations_1, groups = dense_output_565_groups_1, pad = dense_output_565_pad_1, pad_type = dense_output_565_pad_type_1, strides = dense_output_565_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_323_cast_fp16)[name = string("dense_output_565_cast_fp16")];
            string sparse_output_565_pad_type_1 = const()[name = string("sparse_output_565_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_565_strides_1 = const()[name = string("sparse_output_565_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_565_pad_1 = const()[name = string("sparse_output_565_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_565_dilations_1 = const()[name = string("sparse_output_565_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_565_groups_1 = const()[name = string("sparse_output_565_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_6_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198923712))))[name = string("layers_6_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_565_cast_fp16 = conv(dilations = sparse_output_565_dilations_1, groups = sparse_output_565_groups_1, pad = sparse_output_565_pad_1, pad_type = sparse_output_565_pad_type_1, strides = sparse_output_565_strides_1, weight = layers_6_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_323_cast_fp16)[name = string("sparse_output_565_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_325_cast_fp16 = add(x = dense_output_565_cast_fp16, y = sparse_output_565_cast_fp16)[name = string("input_325_cast_fp16")];
            tensor<fp16, [1024]> layers_6_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_6_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198942208)))];
            tensor<fp16, [1024]> layers_6_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_6_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198944320)))];
            tensor<fp16, [1024]> layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198946432)))];
            tensor<fp16, [1024]> layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198948544)))];
            tensor<fp16, [1, 1024, 51, 1]> input_327_cast_fp16 = batch_norm(beta = layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_7826_to_fp16, gamma = layers_6_conv_batch_norm_weight_to_fp16, mean = layers_6_conv_batch_norm_running_mean_to_fp16, variance = layers_6_conv_batch_norm_running_var_to_fp16, x = input_325_cast_fp16)[name = string("input_327_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_329_cast_fp16 = silu(x = input_327_cast_fp16)[name = string("input_329_cast_fp16")];
            string dense_output_567_pad_type_1 = const()[name = string("dense_output_567_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_567_strides_1 = const()[name = string("dense_output_567_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_567_pad_1 = const()[name = string("dense_output_567_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_567_dilations_1 = const()[name = string("dense_output_567_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_567_groups_1 = const()[name = string("dense_output_567_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198950656))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199999296))))[name = string("layers_6_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_567_cast_fp16 = conv(dilations = dense_output_567_dilations_1, groups = dense_output_567_groups_1, pad = dense_output_567_pad_1, pad_type = dense_output_567_pad_type_1, strides = dense_output_567_strides_1, weight = layers_6_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("dense_output_567_cast_fp16")];
            string sparse_output_567_pad_type_1 = const()[name = string("sparse_output_567_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_567_strides_1 = const()[name = string("sparse_output_567_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_567_pad_1 = const()[name = string("sparse_output_567_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_567_dilations_1 = const()[name = string("sparse_output_567_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_567_groups_1 = const()[name = string("sparse_output_567_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_6_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200020928))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199999872))))[name = string("layers_6_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_567_cast_fp16 = conv(dilations = sparse_output_567_dilations_1, groups = sparse_output_567_groups_1, pad = sparse_output_567_pad_1, pad_type = sparse_output_567_pad_type_1, strides = sparse_output_567_strides_1, weight = layers_6_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_329_cast_fp16)[name = string("sparse_output_567_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_591_cast_fp16 = add(x = dense_output_567_cast_fp16, y = sparse_output_567_cast_fp16)[name = string("x_591_cast_fp16")];
            tensor<int32, [1]> var_9023_axes_1 = const()[name = string("op_9023_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_9023_cast_fp16 = squeeze(axes = var_9023_axes_1, x = x_591_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<int32, [3]> var_9024_perm_1 = const()[name = string("op_9024_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_9024_cast_fp16 = transpose(perm = var_9024_perm_1, x = var_9023_cast_fp16)[name = string("transpose_699")];
            tensor<fp16, [1, 51, 1024]> input_331_cast_fp16 = add(x = input_315_cast_fp16, y = var_9024_cast_fp16)[name = string("input_331_cast_fp16")];
            tensor<int32, [1]> x_593_axes_1 = const()[name = string("x_593_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200152064)))];
            tensor<fp16, [1024]> layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200154176)))];
            tensor<fp16, [1, 51, 1024]> x_593_cast_fp16 = layer_norm(axes = x_593_axes_1, beta = layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_7826_to_fp16, gamma = layers_6_norm_feed_forward2_weight_to_fp16, x = input_331_cast_fp16)[name = string("x_593_cast_fp16")];
            tensor<int32, [4]> var_9034 = const()[name = string("op_9034"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_595_cast_fp16 = reshape(shape = var_9034, x = x_593_cast_fp16)[name = string("x_595_cast_fp16")];
            tensor<int32, [4]> input_333_perm_1 = const()[name = string("input_333_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_569_pad_type_1 = const()[name = string("dense_output_569_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_569_strides_1 = const()[name = string("dense_output_569_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_569_pad_1 = const()[name = string("dense_output_569_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_569_dilations_1 = const()[name = string("dense_output_569_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_569_groups_1 = const()[name = string("dense_output_569_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_6_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200156288))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204350656))))[name = string("layers_6_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_333_cast_fp16 = transpose(perm = input_333_perm_1, x = x_595_cast_fp16)[name = string("transpose_698")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_569_cast_fp16 = conv(dilations = dense_output_569_dilations_1, groups = dense_output_569_groups_1, pad = dense_output_569_pad_1, pad_type = dense_output_569_pad_type_1, strides = dense_output_569_strides_1, weight = layers_6_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_333_cast_fp16)[name = string("dense_output_569_cast_fp16")];
            string sparse_output_569_pad_type_1 = const()[name = string("sparse_output_569_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_569_strides_1 = const()[name = string("sparse_output_569_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_569_pad_1 = const()[name = string("sparse_output_569_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_569_dilations_1 = const()[name = string("sparse_output_569_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_569_groups_1 = const()[name = string("sparse_output_569_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_6_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204435200))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204351232))))[name = string("layers_6_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_569_cast_fp16 = conv(dilations = sparse_output_569_dilations_1, groups = sparse_output_569_groups_1, pad = sparse_output_569_pad_1, pad_type = sparse_output_569_pad_type_1, strides = sparse_output_569_strides_1, weight = layers_6_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_333_cast_fp16)[name = string("sparse_output_569_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_335_cast_fp16 = add(x = dense_output_569_cast_fp16, y = sparse_output_569_cast_fp16)[name = string("input_335_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_337_cast_fp16 = silu(x = input_335_cast_fp16)[name = string("input_337_cast_fp16")];
            string dense_output_571_pad_type_1 = const()[name = string("dense_output_571_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_571_strides_1 = const()[name = string("dense_output_571_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_571_pad_1 = const()[name = string("dense_output_571_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_571_dilations_1 = const()[name = string("dense_output_571_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_571_groups_1 = const()[name = string("dense_output_571_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_6_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204959552))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209153920))))[name = string("layers_6_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_571_cast_fp16 = conv(dilations = dense_output_571_dilations_1, groups = dense_output_571_groups_1, pad = dense_output_571_pad_1, pad_type = dense_output_571_pad_type_1, strides = dense_output_571_strides_1, weight = layers_6_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_337_cast_fp16)[name = string("dense_output_571_cast_fp16")];
            string sparse_output_571_pad_type_1 = const()[name = string("sparse_output_571_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_571_strides_1 = const()[name = string("sparse_output_571_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_571_pad_1 = const()[name = string("sparse_output_571_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_571_dilations_1 = const()[name = string("sparse_output_571_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_571_groups_1 = const()[name = string("sparse_output_571_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_6_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209238464))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209154496))))[name = string("layers_6_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_571_cast_fp16 = conv(dilations = sparse_output_571_dilations_1, groups = sparse_output_571_groups_1, pad = sparse_output_571_pad_1, pad_type = sparse_output_571_pad_type_1, strides = sparse_output_571_strides_1, weight = layers_6_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_337_cast_fp16)[name = string("sparse_output_571_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_597_cast_fp16 = add(x = dense_output_571_cast_fp16, y = sparse_output_571_cast_fp16)[name = string("x_597_cast_fp16")];
            tensor<int32, [4]> x_599_perm_1 = const()[name = string("x_599_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_9069 = const()[name = string("op_9069"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_599_cast_fp16 = transpose(perm = x_599_perm_1, x = x_597_cast_fp16)[name = string("transpose_697")];
            tensor<fp16, [1, 51, 1024]> var_9070_cast_fp16 = reshape(shape = var_9069, x = x_599_cast_fp16)[name = string("op_9070_cast_fp16")];
            fp16 var_9071_to_fp16 = const()[name = string("op_9071_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_9072_cast_fp16 = mul(x = var_9070_cast_fp16, y = var_9071_to_fp16)[name = string("op_9072_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_339_cast_fp16 = add(x = input_331_cast_fp16, y = var_9072_cast_fp16)[name = string("input_339_cast_fp16")];
            tensor<int32, [1]> input_341_axes_1 = const()[name = string("input_341_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_6_norm_out_weight_to_fp16 = const()[name = string("layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209762816)))];
            tensor<fp16, [1024]> layers_6_norm_out_bias_to_fp16 = const()[name = string("layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209764928)))];
            tensor<fp16, [1, 51, 1024]> input_341_cast_fp16 = layer_norm(axes = input_341_axes_1, beta = layers_6_norm_out_bias_to_fp16, epsilon = var_7826_to_fp16, gamma = layers_6_norm_out_weight_to_fp16, x = input_339_cast_fp16)[name = string("input_341_cast_fp16")];
            int32 var_9080 = const()[name = string("op_9080"), val = int32(-1)];
            tensor<int32, [1]> x_601_axes_1 = const()[name = string("x_601_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209767040)))];
            tensor<fp16, [1024]> layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209769152)))];
            fp16 var_9095_to_fp16 = const()[name = string("op_9095_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_601_cast_fp16 = layer_norm(axes = x_601_axes_1, beta = layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_9095_to_fp16, gamma = layers_7_norm_feed_forward1_weight_to_fp16, x = input_341_cast_fp16)[name = string("x_601_cast_fp16")];
            tensor<int32, [4]> var_9114 = const()[name = string("op_9114"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_603_cast_fp16 = reshape(shape = var_9114, x = x_601_cast_fp16)[name = string("x_603_cast_fp16")];
            tensor<int32, [4]> input_343_perm_1 = const()[name = string("input_343_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_573_pad_type_1 = const()[name = string("dense_output_573_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_573_strides_1 = const()[name = string("dense_output_573_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_573_pad_1 = const()[name = string("dense_output_573_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_573_dilations_1 = const()[name = string("dense_output_573_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_573_groups_1 = const()[name = string("dense_output_573_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_7_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209771264))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213965632))))[name = string("layers_7_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_343_cast_fp16 = transpose(perm = input_343_perm_1, x = x_603_cast_fp16)[name = string("transpose_696")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_573_cast_fp16 = conv(dilations = dense_output_573_dilations_1, groups = dense_output_573_groups_1, pad = dense_output_573_pad_1, pad_type = dense_output_573_pad_type_1, strides = dense_output_573_strides_1, weight = layers_7_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("dense_output_573_cast_fp16")];
            string sparse_output_573_pad_type_1 = const()[name = string("sparse_output_573_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_573_strides_1 = const()[name = string("sparse_output_573_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_573_pad_1 = const()[name = string("sparse_output_573_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_573_dilations_1 = const()[name = string("sparse_output_573_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_573_groups_1 = const()[name = string("sparse_output_573_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_7_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214050176))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213966208))))[name = string("layers_7_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_573_cast_fp16 = conv(dilations = sparse_output_573_dilations_1, groups = sparse_output_573_groups_1, pad = sparse_output_573_pad_1, pad_type = sparse_output_573_pad_type_1, strides = sparse_output_573_strides_1, weight = layers_7_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_343_cast_fp16)[name = string("sparse_output_573_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_345_cast_fp16 = add(x = dense_output_573_cast_fp16, y = sparse_output_573_cast_fp16)[name = string("input_345_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_347_cast_fp16 = silu(x = input_345_cast_fp16)[name = string("input_347_cast_fp16")];
            string dense_output_575_pad_type_1 = const()[name = string("dense_output_575_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_575_strides_1 = const()[name = string("dense_output_575_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_575_pad_1 = const()[name = string("dense_output_575_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_575_dilations_1 = const()[name = string("dense_output_575_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_575_groups_1 = const()[name = string("dense_output_575_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_7_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214574528))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218768896))))[name = string("layers_7_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_575_cast_fp16 = conv(dilations = dense_output_575_dilations_1, groups = dense_output_575_groups_1, pad = dense_output_575_pad_1, pad_type = dense_output_575_pad_type_1, strides = dense_output_575_strides_1, weight = layers_7_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_347_cast_fp16)[name = string("dense_output_575_cast_fp16")];
            string sparse_output_575_pad_type_1 = const()[name = string("sparse_output_575_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_575_strides_1 = const()[name = string("sparse_output_575_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_575_pad_1 = const()[name = string("sparse_output_575_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_575_dilations_1 = const()[name = string("sparse_output_575_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_575_groups_1 = const()[name = string("sparse_output_575_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_7_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218853440))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218769472))))[name = string("layers_7_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_575_cast_fp16 = conv(dilations = sparse_output_575_dilations_1, groups = sparse_output_575_groups_1, pad = sparse_output_575_pad_1, pad_type = sparse_output_575_pad_type_1, strides = sparse_output_575_strides_1, weight = layers_7_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_347_cast_fp16)[name = string("sparse_output_575_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_605_cast_fp16 = add(x = dense_output_575_cast_fp16, y = sparse_output_575_cast_fp16)[name = string("x_605_cast_fp16")];
            tensor<int32, [4]> x_607_perm_1 = const()[name = string("x_607_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_9149 = const()[name = string("op_9149"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_607_cast_fp16 = transpose(perm = x_607_perm_1, x = x_605_cast_fp16)[name = string("transpose_695")];
            tensor<fp16, [1, 51, 1024]> var_9150_cast_fp16 = reshape(shape = var_9149, x = x_607_cast_fp16)[name = string("op_9150_cast_fp16")];
            fp16 var_9151_to_fp16 = const()[name = string("op_9151_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_9152_cast_fp16 = mul(x = var_9150_cast_fp16, y = var_9151_to_fp16)[name = string("op_9152_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_349_cast_fp16 = add(x = input_341_cast_fp16, y = var_9152_cast_fp16)[name = string("input_349_cast_fp16")];
            tensor<int32, [1]> q_15_axes_1 = const()[name = string("q_15_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_7_norm_self_att_weight_to_fp16 = const()[name = string("layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219377792)))];
            tensor<fp16, [1024]> layers_7_norm_self_att_bias_to_fp16 = const()[name = string("layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219379904)))];
            tensor<fp16, [1, 51, 1024]> q_15_cast_fp16 = layer_norm(axes = q_15_axes_1, beta = layers_7_norm_self_att_bias_to_fp16, epsilon = var_9095_to_fp16, gamma = layers_7_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("q_15_cast_fp16")];
            tensor<int32, [3]> var_9226 = const()[name = string("op_9226"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_351_axes_1 = const()[name = string("input_351_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_9227_cast_fp16 = transpose(perm = var_9226, x = q_15_cast_fp16)[name = string("transpose_694")];
            tensor<fp16, [1, 1024, 51, 1]> input_351_cast_fp16 = expand_dims(axes = input_351_axes_1, x = var_9227_cast_fp16)[name = string("input_351_cast_fp16")];
            string dense_output_577_pad_type_1 = const()[name = string("dense_output_577_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_577_strides_1 = const()[name = string("dense_output_577_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_577_pad_1 = const()[name = string("dense_output_577_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_577_dilations_1 = const()[name = string("dense_output_577_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_577_groups_1 = const()[name = string("dense_output_577_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219382016))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219513152))))[name = string("layers_7_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_577_cast_fp16 = conv(dilations = dense_output_577_dilations_1, groups = dense_output_577_groups_1, pad = dense_output_577_pad_1, pad_type = dense_output_577_pad_type_1, strides = dense_output_577_strides_1, weight = layers_7_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_577_cast_fp16")];
            string sparse_output_577_pad_type_1 = const()[name = string("sparse_output_577_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_577_strides_1 = const()[name = string("sparse_output_577_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_577_pad_1 = const()[name = string("sparse_output_577_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_577_dilations_1 = const()[name = string("sparse_output_577_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_577_groups_1 = const()[name = string("sparse_output_577_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219516416))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219513728))))[name = string("layers_7_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_577_cast_fp16 = conv(dilations = sparse_output_577_dilations_1, groups = sparse_output_577_groups_1, pad = sparse_output_577_pad_1, pad_type = sparse_output_577_pad_type_1, strides = sparse_output_577_strides_1, weight = layers_7_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_577_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9252_cast_fp16 = add(x = dense_output_577_cast_fp16, y = sparse_output_577_cast_fp16)[name = string("op_9252_cast_fp16")];
            tensor<int32, [4]> var_9253 = const()[name = string("op_9253"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9255 = const()[name = string("op_9255"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9254_cast_fp16 = transpose(perm = var_9253, x = var_9252_cast_fp16)[name = string("transpose_693")];
            tensor<fp16, [1, 51, 128]> q_head_113_cast_fp16 = reshape(shape = var_9255, x = var_9254_cast_fp16)[name = string("q_head_113_cast_fp16")];
            string dense_output_579_pad_type_1 = const()[name = string("dense_output_579_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_579_strides_1 = const()[name = string("dense_output_579_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_579_pad_1 = const()[name = string("dense_output_579_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_579_dilations_1 = const()[name = string("dense_output_579_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_579_groups_1 = const()[name = string("dense_output_579_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219532864))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219664000))))[name = string("layers_7_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_579_cast_fp16 = conv(dilations = dense_output_579_dilations_1, groups = dense_output_579_groups_1, pad = dense_output_579_pad_1, pad_type = dense_output_579_pad_type_1, strides = dense_output_579_strides_1, weight = layers_7_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_579_cast_fp16")];
            string sparse_output_579_pad_type_1 = const()[name = string("sparse_output_579_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_579_strides_1 = const()[name = string("sparse_output_579_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_579_pad_1 = const()[name = string("sparse_output_579_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_579_dilations_1 = const()[name = string("sparse_output_579_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_579_groups_1 = const()[name = string("sparse_output_579_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219667264))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219664576))))[name = string("layers_7_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_579_cast_fp16 = conv(dilations = sparse_output_579_dilations_1, groups = sparse_output_579_groups_1, pad = sparse_output_579_pad_1, pad_type = sparse_output_579_pad_type_1, strides = sparse_output_579_strides_1, weight = layers_7_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_579_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9271_cast_fp16 = add(x = dense_output_579_cast_fp16, y = sparse_output_579_cast_fp16)[name = string("op_9271_cast_fp16")];
            tensor<int32, [4]> var_9272 = const()[name = string("op_9272"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9274 = const()[name = string("op_9274"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9273_cast_fp16 = transpose(perm = var_9272, x = var_9271_cast_fp16)[name = string("transpose_692")];
            tensor<fp16, [1, 51, 128]> k_head_225_cast_fp16 = reshape(shape = var_9274, x = var_9273_cast_fp16)[name = string("k_head_225_cast_fp16")];
            string dense_output_581_pad_type_1 = const()[name = string("dense_output_581_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_581_strides_1 = const()[name = string("dense_output_581_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_581_pad_1 = const()[name = string("dense_output_581_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_581_dilations_1 = const()[name = string("dense_output_581_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_581_groups_1 = const()[name = string("dense_output_581_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219683712))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219814848))))[name = string("layers_7_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_581_cast_fp16 = conv(dilations = dense_output_581_dilations_1, groups = dense_output_581_groups_1, pad = dense_output_581_pad_1, pad_type = dense_output_581_pad_type_1, strides = dense_output_581_strides_1, weight = layers_7_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_581_cast_fp16")];
            string sparse_output_581_pad_type_1 = const()[name = string("sparse_output_581_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_581_strides_1 = const()[name = string("sparse_output_581_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_581_pad_1 = const()[name = string("sparse_output_581_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_581_dilations_1 = const()[name = string("sparse_output_581_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_581_groups_1 = const()[name = string("sparse_output_581_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219818112))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219815424))))[name = string("layers_7_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_581_cast_fp16 = conv(dilations = sparse_output_581_dilations_1, groups = sparse_output_581_groups_1, pad = sparse_output_581_pad_1, pad_type = sparse_output_581_pad_type_1, strides = sparse_output_581_strides_1, weight = layers_7_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_581_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9290_cast_fp16 = add(x = dense_output_581_cast_fp16, y = sparse_output_581_cast_fp16)[name = string("op_9290_cast_fp16")];
            tensor<int32, [4]> var_9291 = const()[name = string("op_9291"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9293 = const()[name = string("op_9293"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9292_cast_fp16 = transpose(perm = var_9291, x = var_9290_cast_fp16)[name = string("transpose_691")];
            tensor<fp16, [1, 51, 128]> v_head_225_cast_fp16 = reshape(shape = var_9293, x = var_9292_cast_fp16)[name = string("v_head_225_cast_fp16")];
            string dense_output_583_pad_type_1 = const()[name = string("dense_output_583_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_583_strides_1 = const()[name = string("dense_output_583_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_583_pad_1 = const()[name = string("dense_output_583_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_583_dilations_1 = const()[name = string("dense_output_583_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_583_groups_1 = const()[name = string("dense_output_583_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219834560))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219965696))))[name = string("layers_7_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_583_cast_fp16 = conv(dilations = dense_output_583_dilations_1, groups = dense_output_583_groups_1, pad = dense_output_583_pad_1, pad_type = dense_output_583_pad_type_1, strides = dense_output_583_strides_1, weight = layers_7_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_583_cast_fp16")];
            string sparse_output_583_pad_type_1 = const()[name = string("sparse_output_583_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_583_strides_1 = const()[name = string("sparse_output_583_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_583_pad_1 = const()[name = string("sparse_output_583_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_583_dilations_1 = const()[name = string("sparse_output_583_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_583_groups_1 = const()[name = string("sparse_output_583_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219968960))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219966272))))[name = string("layers_7_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_583_cast_fp16 = conv(dilations = sparse_output_583_dilations_1, groups = sparse_output_583_groups_1, pad = sparse_output_583_pad_1, pad_type = sparse_output_583_pad_type_1, strides = sparse_output_583_strides_1, weight = layers_7_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_583_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_9309_cast_fp16 = add(x = dense_output_583_cast_fp16, y = sparse_output_583_cast_fp16)[name = string("op_9309_cast_fp16")];
            tensor<int32, [4]> var_9310 = const()[name = string("op_9310"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9312 = const()[name = string("op_9312"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_9311_cast_fp16 = transpose(perm = var_9310, x = var_9309_cast_fp16)[name = string("transpose_690")];
            tensor<fp16, [1, 101, 128]> p_head_225_cast_fp16 = reshape(shape = var_9312, x = var_9311_cast_fp16)[name = string("p_head_225_cast_fp16")];
            tensor<fp16, [128]> var_9314_to_fp16 = const()[name = string("op_9314_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219985408)))];
            tensor<fp16, [1, 51, 128]> var_9315_cast_fp16 = add(x = q_head_113_cast_fp16, y = var_9314_to_fp16)[name = string("op_9315_cast_fp16")];
            tensor<int32, [1]> q_u_113_axes_1 = const()[name = string("q_u_113_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_113_cast_fp16 = expand_dims(axes = q_u_113_axes_1, x = var_9315_cast_fp16)[name = string("q_u_113_cast_fp16")];
            tensor<fp16, [128]> var_9317_to_fp16 = const()[name = string("op_9317_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219985728)))];
            tensor<fp16, [1, 51, 128]> var_9318_cast_fp16 = add(x = q_head_113_cast_fp16, y = var_9317_to_fp16)[name = string("op_9318_cast_fp16")];
            tensor<int32, [1]> q_v_113_axes_1 = const()[name = string("q_v_113_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_113_cast_fp16 = expand_dims(axes = q_v_113_axes_1, x = var_9318_cast_fp16)[name = string("q_v_113_cast_fp16")];
            tensor<int32, [1]> k_head_227_axes_1 = const()[name = string("k_head_227_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_227_cast_fp16 = expand_dims(axes = k_head_227_axes_1, x = k_head_225_cast_fp16)[name = string("k_head_227_cast_fp16")];
            tensor<int32, [1]> v_head_227_axes_1 = const()[name = string("v_head_227_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_227_cast_fp16 = expand_dims(axes = v_head_227_axes_1, x = v_head_225_cast_fp16)[name = string("v_head_227_cast_fp16")];
            tensor<int32, [1]> p_head_227_axes_1 = const()[name = string("p_head_227_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_227_cast_fp16 = expand_dims(axes = p_head_227_axes_1, x = p_head_225_cast_fp16)[name = string("p_head_227_cast_fp16")];
            bool var_9324_transpose_x_3 = const()[name = string("op_9324_transpose_x_3"), val = bool(false)];
            bool var_9324_transpose_y_3 = const()[name = string("op_9324_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_9324_cast_fp16 = matmul(transpose_x = var_9324_transpose_x_3, transpose_y = var_9324_transpose_y_3, x = q_u_113_cast_fp16, y = k_head_227_cast_fp16)[name = string("op_9324_cast_fp16")];
            fp16 var_9325_to_fp16 = const()[name = string("op_9325_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_113_cast_fp16 = mul(x = var_9324_cast_fp16, y = var_9325_to_fp16)[name = string("scores_content_113_cast_fp16")];
            bool x_609_transpose_x_3 = const()[name = string("x_609_transpose_x_3"), val = bool(false)];
            bool x_609_transpose_y_3 = const()[name = string("x_609_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_609_cast_fp16 = matmul(transpose_x = x_609_transpose_x_3, transpose_y = x_609_transpose_y_3, x = q_v_113_cast_fp16, y = p_head_227_cast_fp16)[name = string("x_609_cast_fp16")];
            tensor<int32, [8]> x_611_pad_1 = const()[name = string("x_611_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_611_mode_1 = const()[name = string("x_611_mode_1"), val = string("constant")];
            fp16 const_1687_to_fp16 = const()[name = string("const_1687_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_611_cast_fp16 = pad(constant_val = const_1687_to_fp16, mode = x_611_mode_1, pad = x_611_pad_1, x = x_609_cast_fp16)[name = string("x_611_cast_fp16")];
            tensor<int32, [4]> var_9339 = const()[name = string("op_9339"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_613_cast_fp16 = reshape(shape = var_9339, x = x_611_cast_fp16)[name = string("x_613_cast_fp16")];
            tensor<int32, [4]> var_9343_begin_1 = const()[name = string("op_9343_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9343_end_1 = const()[name = string("op_9343_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_9343_end_mask_1 = const()[name = string("op_9343_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_9343_cast_fp16 = slice_by_index(begin = var_9343_begin_1, end = var_9343_end_1, end_mask = var_9343_end_mask_1, x = x_613_cast_fp16)[name = string("op_9343_cast_fp16")];
            tensor<int32, [4]> var_9345 = const()[name = string("op_9345"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_9346_cast_fp16 = reshape(shape = var_9345, x = var_9343_cast_fp16)[name = string("op_9346_cast_fp16")];
            tensor<int32, [4]> var_9351_begin_1 = const()[name = string("op_9351_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9351_end_1 = const()[name = string("op_9351_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_9351_end_mask_1 = const()[name = string("op_9351_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_9351_cast_fp16 = slice_by_index(begin = var_9351_begin_1, end = var_9351_end_1, end_mask = var_9351_end_mask_1, x = var_9346_cast_fp16)[name = string("op_9351_cast_fp16")];
            fp16 var_9352_to_fp16 = const()[name = string("op_9352_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_113_cast_fp16 = mul(x = var_9351_cast_fp16, y = var_9352_to_fp16)[name = string("scores_pos_113_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_113_cast_fp16 = add(x = scores_content_113_cast_fp16, y = scores_pos_113_cast_fp16)[name = string("logits_113_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_9355_cast_fp16 = softmax(axis = var_9080, x = logits_113_cast_fp16)[name = string("op_9355_cast_fp16")];
            bool var_9357_transpose_x_1 = const()[name = string("op_9357_transpose_x_1"), val = bool(false)];
            bool var_9357_transpose_y_1 = const()[name = string("op_9357_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_9357_cast_fp16 = matmul(transpose_x = var_9357_transpose_x_1, transpose_y = var_9357_transpose_y_1, x = var_9355_cast_fp16, y = v_head_227_cast_fp16)[name = string("op_9357_cast_fp16")];
            tensor<int32, [1]> var_9358_axes_1 = const()[name = string("op_9358_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_9358_cast_fp16 = squeeze(axes = var_9358_axes_1, x = var_9357_cast_fp16)[name = string("op_9358_cast_fp16")];
            string dense_output_585_pad_type_1 = const()[name = string("dense_output_585_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_585_strides_1 = const()[name = string("dense_output_585_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_585_pad_1 = const()[name = string("dense_output_585_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_585_dilations_1 = const()[name = string("dense_output_585_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_585_groups_1 = const()[name = string("dense_output_585_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219986048))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220117184))))[name = string("layers_7_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_585_cast_fp16 = conv(dilations = dense_output_585_dilations_1, groups = dense_output_585_groups_1, pad = dense_output_585_pad_1, pad_type = dense_output_585_pad_type_1, strides = dense_output_585_strides_1, weight = layers_7_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_585_cast_fp16")];
            string sparse_output_585_pad_type_1 = const()[name = string("sparse_output_585_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_585_strides_1 = const()[name = string("sparse_output_585_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_585_pad_1 = const()[name = string("sparse_output_585_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_585_dilations_1 = const()[name = string("sparse_output_585_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_585_groups_1 = const()[name = string("sparse_output_585_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220120448))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220117760))))[name = string("layers_7_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_585_cast_fp16 = conv(dilations = sparse_output_585_dilations_1, groups = sparse_output_585_groups_1, pad = sparse_output_585_pad_1, pad_type = sparse_output_585_pad_type_1, strides = sparse_output_585_strides_1, weight = layers_7_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_585_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9373_cast_fp16 = add(x = dense_output_585_cast_fp16, y = sparse_output_585_cast_fp16)[name = string("op_9373_cast_fp16")];
            tensor<int32, [4]> var_9374 = const()[name = string("op_9374"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9376 = const()[name = string("op_9376"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9375_cast_fp16 = transpose(perm = var_9374, x = var_9373_cast_fp16)[name = string("transpose_689")];
            tensor<fp16, [1, 51, 128]> q_head_115_cast_fp16 = reshape(shape = var_9376, x = var_9375_cast_fp16)[name = string("q_head_115_cast_fp16")];
            string dense_output_587_pad_type_1 = const()[name = string("dense_output_587_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_587_strides_1 = const()[name = string("dense_output_587_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_587_pad_1 = const()[name = string("dense_output_587_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_587_dilations_1 = const()[name = string("dense_output_587_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_587_groups_1 = const()[name = string("dense_output_587_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220136896))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220268032))))[name = string("layers_7_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_587_cast_fp16 = conv(dilations = dense_output_587_dilations_1, groups = dense_output_587_groups_1, pad = dense_output_587_pad_1, pad_type = dense_output_587_pad_type_1, strides = dense_output_587_strides_1, weight = layers_7_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_587_cast_fp16")];
            string sparse_output_587_pad_type_1 = const()[name = string("sparse_output_587_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_587_strides_1 = const()[name = string("sparse_output_587_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_587_pad_1 = const()[name = string("sparse_output_587_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_587_dilations_1 = const()[name = string("sparse_output_587_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_587_groups_1 = const()[name = string("sparse_output_587_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220271296))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220268608))))[name = string("layers_7_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_587_cast_fp16 = conv(dilations = sparse_output_587_dilations_1, groups = sparse_output_587_groups_1, pad = sparse_output_587_pad_1, pad_type = sparse_output_587_pad_type_1, strides = sparse_output_587_strides_1, weight = layers_7_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_587_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9392_cast_fp16 = add(x = dense_output_587_cast_fp16, y = sparse_output_587_cast_fp16)[name = string("op_9392_cast_fp16")];
            tensor<int32, [4]> var_9393 = const()[name = string("op_9393"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9395 = const()[name = string("op_9395"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9394_cast_fp16 = transpose(perm = var_9393, x = var_9392_cast_fp16)[name = string("transpose_688")];
            tensor<fp16, [1, 51, 128]> k_head_229_cast_fp16 = reshape(shape = var_9395, x = var_9394_cast_fp16)[name = string("k_head_229_cast_fp16")];
            string dense_output_589_pad_type_1 = const()[name = string("dense_output_589_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_589_strides_1 = const()[name = string("dense_output_589_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_589_pad_1 = const()[name = string("dense_output_589_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_589_dilations_1 = const()[name = string("dense_output_589_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_589_groups_1 = const()[name = string("dense_output_589_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220287744))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220418880))))[name = string("layers_7_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_589_cast_fp16 = conv(dilations = dense_output_589_dilations_1, groups = dense_output_589_groups_1, pad = dense_output_589_pad_1, pad_type = dense_output_589_pad_type_1, strides = dense_output_589_strides_1, weight = layers_7_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_589_cast_fp16")];
            string sparse_output_589_pad_type_1 = const()[name = string("sparse_output_589_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_589_strides_1 = const()[name = string("sparse_output_589_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_589_pad_1 = const()[name = string("sparse_output_589_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_589_dilations_1 = const()[name = string("sparse_output_589_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_589_groups_1 = const()[name = string("sparse_output_589_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220422144))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220419456))))[name = string("layers_7_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_589_cast_fp16 = conv(dilations = sparse_output_589_dilations_1, groups = sparse_output_589_groups_1, pad = sparse_output_589_pad_1, pad_type = sparse_output_589_pad_type_1, strides = sparse_output_589_strides_1, weight = layers_7_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_589_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9411_cast_fp16 = add(x = dense_output_589_cast_fp16, y = sparse_output_589_cast_fp16)[name = string("op_9411_cast_fp16")];
            tensor<int32, [4]> var_9412 = const()[name = string("op_9412"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9414 = const()[name = string("op_9414"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9413_cast_fp16 = transpose(perm = var_9412, x = var_9411_cast_fp16)[name = string("transpose_687")];
            tensor<fp16, [1, 51, 128]> v_head_229_cast_fp16 = reshape(shape = var_9414, x = var_9413_cast_fp16)[name = string("v_head_229_cast_fp16")];
            string dense_output_591_pad_type_1 = const()[name = string("dense_output_591_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_591_strides_1 = const()[name = string("dense_output_591_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_591_pad_1 = const()[name = string("dense_output_591_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_591_dilations_1 = const()[name = string("dense_output_591_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_591_groups_1 = const()[name = string("dense_output_591_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220438592))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220569728))))[name = string("layers_7_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_591_cast_fp16 = conv(dilations = dense_output_591_dilations_1, groups = dense_output_591_groups_1, pad = dense_output_591_pad_1, pad_type = dense_output_591_pad_type_1, strides = dense_output_591_strides_1, weight = layers_7_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_591_cast_fp16")];
            string sparse_output_591_pad_type_1 = const()[name = string("sparse_output_591_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_591_strides_1 = const()[name = string("sparse_output_591_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_591_pad_1 = const()[name = string("sparse_output_591_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_591_dilations_1 = const()[name = string("sparse_output_591_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_591_groups_1 = const()[name = string("sparse_output_591_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220572992))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220570304))))[name = string("layers_7_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_591_cast_fp16 = conv(dilations = sparse_output_591_dilations_1, groups = sparse_output_591_groups_1, pad = sparse_output_591_pad_1, pad_type = sparse_output_591_pad_type_1, strides = sparse_output_591_strides_1, weight = layers_7_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_591_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_9430_cast_fp16 = add(x = dense_output_591_cast_fp16, y = sparse_output_591_cast_fp16)[name = string("op_9430_cast_fp16")];
            tensor<int32, [4]> var_9431 = const()[name = string("op_9431"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9433 = const()[name = string("op_9433"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_9432_cast_fp16 = transpose(perm = var_9431, x = var_9430_cast_fp16)[name = string("transpose_686")];
            tensor<fp16, [1, 101, 128]> p_head_229_cast_fp16 = reshape(shape = var_9433, x = var_9432_cast_fp16)[name = string("p_head_229_cast_fp16")];
            tensor<fp16, [128]> var_9435_to_fp16 = const()[name = string("op_9435_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220589440)))];
            tensor<fp16, [1, 51, 128]> var_9436_cast_fp16 = add(x = q_head_115_cast_fp16, y = var_9435_to_fp16)[name = string("op_9436_cast_fp16")];
            tensor<int32, [1]> q_u_115_axes_1 = const()[name = string("q_u_115_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_115_cast_fp16 = expand_dims(axes = q_u_115_axes_1, x = var_9436_cast_fp16)[name = string("q_u_115_cast_fp16")];
            tensor<fp16, [128]> var_9438_to_fp16 = const()[name = string("op_9438_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220589760)))];
            tensor<fp16, [1, 51, 128]> var_9439_cast_fp16 = add(x = q_head_115_cast_fp16, y = var_9438_to_fp16)[name = string("op_9439_cast_fp16")];
            tensor<int32, [1]> q_v_115_axes_1 = const()[name = string("q_v_115_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_115_cast_fp16 = expand_dims(axes = q_v_115_axes_1, x = var_9439_cast_fp16)[name = string("q_v_115_cast_fp16")];
            tensor<int32, [1]> k_head_231_axes_1 = const()[name = string("k_head_231_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_231_cast_fp16 = expand_dims(axes = k_head_231_axes_1, x = k_head_229_cast_fp16)[name = string("k_head_231_cast_fp16")];
            tensor<int32, [1]> v_head_231_axes_1 = const()[name = string("v_head_231_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_231_cast_fp16 = expand_dims(axes = v_head_231_axes_1, x = v_head_229_cast_fp16)[name = string("v_head_231_cast_fp16")];
            tensor<int32, [1]> p_head_231_axes_1 = const()[name = string("p_head_231_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_231_cast_fp16 = expand_dims(axes = p_head_231_axes_1, x = p_head_229_cast_fp16)[name = string("p_head_231_cast_fp16")];
            bool var_9445_transpose_x_3 = const()[name = string("op_9445_transpose_x_3"), val = bool(false)];
            bool var_9445_transpose_y_3 = const()[name = string("op_9445_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_9445_cast_fp16 = matmul(transpose_x = var_9445_transpose_x_3, transpose_y = var_9445_transpose_y_3, x = q_u_115_cast_fp16, y = k_head_231_cast_fp16)[name = string("op_9445_cast_fp16")];
            fp16 var_9446_to_fp16 = const()[name = string("op_9446_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_115_cast_fp16 = mul(x = var_9445_cast_fp16, y = var_9446_to_fp16)[name = string("scores_content_115_cast_fp16")];
            bool x_617_transpose_x_3 = const()[name = string("x_617_transpose_x_3"), val = bool(false)];
            bool x_617_transpose_y_3 = const()[name = string("x_617_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_617_cast_fp16 = matmul(transpose_x = x_617_transpose_x_3, transpose_y = x_617_transpose_y_3, x = q_v_115_cast_fp16, y = p_head_231_cast_fp16)[name = string("x_617_cast_fp16")];
            tensor<int32, [8]> x_619_pad_1 = const()[name = string("x_619_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_619_mode_1 = const()[name = string("x_619_mode_1"), val = string("constant")];
            fp16 const_1693_to_fp16 = const()[name = string("const_1693_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_619_cast_fp16 = pad(constant_val = const_1693_to_fp16, mode = x_619_mode_1, pad = x_619_pad_1, x = x_617_cast_fp16)[name = string("x_619_cast_fp16")];
            tensor<int32, [4]> var_9460 = const()[name = string("op_9460"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_621_cast_fp16 = reshape(shape = var_9460, x = x_619_cast_fp16)[name = string("x_621_cast_fp16")];
            tensor<int32, [4]> var_9464_begin_1 = const()[name = string("op_9464_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9464_end_1 = const()[name = string("op_9464_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_9464_end_mask_1 = const()[name = string("op_9464_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_9464_cast_fp16 = slice_by_index(begin = var_9464_begin_1, end = var_9464_end_1, end_mask = var_9464_end_mask_1, x = x_621_cast_fp16)[name = string("op_9464_cast_fp16")];
            tensor<int32, [4]> var_9466 = const()[name = string("op_9466"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_9467_cast_fp16 = reshape(shape = var_9466, x = var_9464_cast_fp16)[name = string("op_9467_cast_fp16")];
            tensor<int32, [4]> var_9472_begin_1 = const()[name = string("op_9472_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9472_end_1 = const()[name = string("op_9472_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_9472_end_mask_1 = const()[name = string("op_9472_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_9472_cast_fp16 = slice_by_index(begin = var_9472_begin_1, end = var_9472_end_1, end_mask = var_9472_end_mask_1, x = var_9467_cast_fp16)[name = string("op_9472_cast_fp16")];
            fp16 var_9473_to_fp16 = const()[name = string("op_9473_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_115_cast_fp16 = mul(x = var_9472_cast_fp16, y = var_9473_to_fp16)[name = string("scores_pos_115_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_115_cast_fp16 = add(x = scores_content_115_cast_fp16, y = scores_pos_115_cast_fp16)[name = string("logits_115_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_9476_cast_fp16 = softmax(axis = var_9080, x = logits_115_cast_fp16)[name = string("op_9476_cast_fp16")];
            bool var_9478_transpose_x_1 = const()[name = string("op_9478_transpose_x_1"), val = bool(false)];
            bool var_9478_transpose_y_1 = const()[name = string("op_9478_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_9478_cast_fp16 = matmul(transpose_x = var_9478_transpose_x_1, transpose_y = var_9478_transpose_y_1, x = var_9476_cast_fp16, y = v_head_231_cast_fp16)[name = string("op_9478_cast_fp16")];
            tensor<int32, [1]> var_9479_axes_1 = const()[name = string("op_9479_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_9479_cast_fp16 = squeeze(axes = var_9479_axes_1, x = var_9478_cast_fp16)[name = string("op_9479_cast_fp16")];
            string dense_output_593_pad_type_1 = const()[name = string("dense_output_593_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_593_strides_1 = const()[name = string("dense_output_593_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_593_pad_1 = const()[name = string("dense_output_593_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_593_dilations_1 = const()[name = string("dense_output_593_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_593_groups_1 = const()[name = string("dense_output_593_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220590080))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220721216))))[name = string("layers_7_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_593_cast_fp16 = conv(dilations = dense_output_593_dilations_1, groups = dense_output_593_groups_1, pad = dense_output_593_pad_1, pad_type = dense_output_593_pad_type_1, strides = dense_output_593_strides_1, weight = layers_7_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_593_cast_fp16")];
            string sparse_output_593_pad_type_1 = const()[name = string("sparse_output_593_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_593_strides_1 = const()[name = string("sparse_output_593_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_593_pad_1 = const()[name = string("sparse_output_593_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_593_dilations_1 = const()[name = string("sparse_output_593_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_593_groups_1 = const()[name = string("sparse_output_593_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220724480))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220721792))))[name = string("layers_7_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_593_cast_fp16 = conv(dilations = sparse_output_593_dilations_1, groups = sparse_output_593_groups_1, pad = sparse_output_593_pad_1, pad_type = sparse_output_593_pad_type_1, strides = sparse_output_593_strides_1, weight = layers_7_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_593_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9494_cast_fp16 = add(x = dense_output_593_cast_fp16, y = sparse_output_593_cast_fp16)[name = string("op_9494_cast_fp16")];
            tensor<int32, [4]> var_9495 = const()[name = string("op_9495"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9497 = const()[name = string("op_9497"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9496_cast_fp16 = transpose(perm = var_9495, x = var_9494_cast_fp16)[name = string("transpose_685")];
            tensor<fp16, [1, 51, 128]> q_head_117_cast_fp16 = reshape(shape = var_9497, x = var_9496_cast_fp16)[name = string("q_head_117_cast_fp16")];
            string dense_output_595_pad_type_1 = const()[name = string("dense_output_595_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_595_strides_1 = const()[name = string("dense_output_595_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_595_pad_1 = const()[name = string("dense_output_595_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_595_dilations_1 = const()[name = string("dense_output_595_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_595_groups_1 = const()[name = string("dense_output_595_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220740928))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220872064))))[name = string("layers_7_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_595_cast_fp16 = conv(dilations = dense_output_595_dilations_1, groups = dense_output_595_groups_1, pad = dense_output_595_pad_1, pad_type = dense_output_595_pad_type_1, strides = dense_output_595_strides_1, weight = layers_7_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_595_cast_fp16")];
            string sparse_output_595_pad_type_1 = const()[name = string("sparse_output_595_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_595_strides_1 = const()[name = string("sparse_output_595_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_595_pad_1 = const()[name = string("sparse_output_595_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_595_dilations_1 = const()[name = string("sparse_output_595_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_595_groups_1 = const()[name = string("sparse_output_595_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220875328))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220872640))))[name = string("layers_7_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_595_cast_fp16 = conv(dilations = sparse_output_595_dilations_1, groups = sparse_output_595_groups_1, pad = sparse_output_595_pad_1, pad_type = sparse_output_595_pad_type_1, strides = sparse_output_595_strides_1, weight = layers_7_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_595_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9513_cast_fp16 = add(x = dense_output_595_cast_fp16, y = sparse_output_595_cast_fp16)[name = string("op_9513_cast_fp16")];
            tensor<int32, [4]> var_9514 = const()[name = string("op_9514"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9516 = const()[name = string("op_9516"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9515_cast_fp16 = transpose(perm = var_9514, x = var_9513_cast_fp16)[name = string("transpose_684")];
            tensor<fp16, [1, 51, 128]> k_head_233_cast_fp16 = reshape(shape = var_9516, x = var_9515_cast_fp16)[name = string("k_head_233_cast_fp16")];
            string dense_output_597_pad_type_1 = const()[name = string("dense_output_597_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_597_strides_1 = const()[name = string("dense_output_597_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_597_pad_1 = const()[name = string("dense_output_597_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_597_dilations_1 = const()[name = string("dense_output_597_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_597_groups_1 = const()[name = string("dense_output_597_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220891776))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221022912))))[name = string("layers_7_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_597_cast_fp16 = conv(dilations = dense_output_597_dilations_1, groups = dense_output_597_groups_1, pad = dense_output_597_pad_1, pad_type = dense_output_597_pad_type_1, strides = dense_output_597_strides_1, weight = layers_7_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_597_cast_fp16")];
            string sparse_output_597_pad_type_1 = const()[name = string("sparse_output_597_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_597_strides_1 = const()[name = string("sparse_output_597_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_597_pad_1 = const()[name = string("sparse_output_597_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_597_dilations_1 = const()[name = string("sparse_output_597_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_597_groups_1 = const()[name = string("sparse_output_597_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221026176))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221023488))))[name = string("layers_7_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_597_cast_fp16 = conv(dilations = sparse_output_597_dilations_1, groups = sparse_output_597_groups_1, pad = sparse_output_597_pad_1, pad_type = sparse_output_597_pad_type_1, strides = sparse_output_597_strides_1, weight = layers_7_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_597_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9532_cast_fp16 = add(x = dense_output_597_cast_fp16, y = sparse_output_597_cast_fp16)[name = string("op_9532_cast_fp16")];
            tensor<int32, [4]> var_9533 = const()[name = string("op_9533"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9535 = const()[name = string("op_9535"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9534_cast_fp16 = transpose(perm = var_9533, x = var_9532_cast_fp16)[name = string("transpose_683")];
            tensor<fp16, [1, 51, 128]> v_head_233_cast_fp16 = reshape(shape = var_9535, x = var_9534_cast_fp16)[name = string("v_head_233_cast_fp16")];
            string dense_output_599_pad_type_1 = const()[name = string("dense_output_599_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_599_strides_1 = const()[name = string("dense_output_599_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_599_pad_1 = const()[name = string("dense_output_599_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_599_dilations_1 = const()[name = string("dense_output_599_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_599_groups_1 = const()[name = string("dense_output_599_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221042624))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221173760))))[name = string("layers_7_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_599_cast_fp16 = conv(dilations = dense_output_599_dilations_1, groups = dense_output_599_groups_1, pad = dense_output_599_pad_1, pad_type = dense_output_599_pad_type_1, strides = dense_output_599_strides_1, weight = layers_7_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_599_cast_fp16")];
            string sparse_output_599_pad_type_1 = const()[name = string("sparse_output_599_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_599_strides_1 = const()[name = string("sparse_output_599_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_599_pad_1 = const()[name = string("sparse_output_599_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_599_dilations_1 = const()[name = string("sparse_output_599_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_599_groups_1 = const()[name = string("sparse_output_599_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221177024))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221174336))))[name = string("layers_7_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_599_cast_fp16 = conv(dilations = sparse_output_599_dilations_1, groups = sparse_output_599_groups_1, pad = sparse_output_599_pad_1, pad_type = sparse_output_599_pad_type_1, strides = sparse_output_599_strides_1, weight = layers_7_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_599_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_9551_cast_fp16 = add(x = dense_output_599_cast_fp16, y = sparse_output_599_cast_fp16)[name = string("op_9551_cast_fp16")];
            tensor<int32, [4]> var_9552 = const()[name = string("op_9552"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9554 = const()[name = string("op_9554"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_9553_cast_fp16 = transpose(perm = var_9552, x = var_9551_cast_fp16)[name = string("transpose_682")];
            tensor<fp16, [1, 101, 128]> p_head_233_cast_fp16 = reshape(shape = var_9554, x = var_9553_cast_fp16)[name = string("p_head_233_cast_fp16")];
            tensor<fp16, [128]> var_9556_to_fp16 = const()[name = string("op_9556_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221193472)))];
            tensor<fp16, [1, 51, 128]> var_9557_cast_fp16 = add(x = q_head_117_cast_fp16, y = var_9556_to_fp16)[name = string("op_9557_cast_fp16")];
            tensor<int32, [1]> q_u_117_axes_1 = const()[name = string("q_u_117_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_117_cast_fp16 = expand_dims(axes = q_u_117_axes_1, x = var_9557_cast_fp16)[name = string("q_u_117_cast_fp16")];
            tensor<fp16, [128]> var_9559_to_fp16 = const()[name = string("op_9559_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221193792)))];
            tensor<fp16, [1, 51, 128]> var_9560_cast_fp16 = add(x = q_head_117_cast_fp16, y = var_9559_to_fp16)[name = string("op_9560_cast_fp16")];
            tensor<int32, [1]> q_v_117_axes_1 = const()[name = string("q_v_117_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_117_cast_fp16 = expand_dims(axes = q_v_117_axes_1, x = var_9560_cast_fp16)[name = string("q_v_117_cast_fp16")];
            tensor<int32, [1]> k_head_235_axes_1 = const()[name = string("k_head_235_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_235_cast_fp16 = expand_dims(axes = k_head_235_axes_1, x = k_head_233_cast_fp16)[name = string("k_head_235_cast_fp16")];
            tensor<int32, [1]> v_head_235_axes_1 = const()[name = string("v_head_235_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_235_cast_fp16 = expand_dims(axes = v_head_235_axes_1, x = v_head_233_cast_fp16)[name = string("v_head_235_cast_fp16")];
            tensor<int32, [1]> p_head_235_axes_1 = const()[name = string("p_head_235_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_235_cast_fp16 = expand_dims(axes = p_head_235_axes_1, x = p_head_233_cast_fp16)[name = string("p_head_235_cast_fp16")];
            bool var_9566_transpose_x_3 = const()[name = string("op_9566_transpose_x_3"), val = bool(false)];
            bool var_9566_transpose_y_3 = const()[name = string("op_9566_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_9566_cast_fp16 = matmul(transpose_x = var_9566_transpose_x_3, transpose_y = var_9566_transpose_y_3, x = q_u_117_cast_fp16, y = k_head_235_cast_fp16)[name = string("op_9566_cast_fp16")];
            fp16 var_9567_to_fp16 = const()[name = string("op_9567_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_117_cast_fp16 = mul(x = var_9566_cast_fp16, y = var_9567_to_fp16)[name = string("scores_content_117_cast_fp16")];
            bool x_625_transpose_x_3 = const()[name = string("x_625_transpose_x_3"), val = bool(false)];
            bool x_625_transpose_y_3 = const()[name = string("x_625_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_625_cast_fp16 = matmul(transpose_x = x_625_transpose_x_3, transpose_y = x_625_transpose_y_3, x = q_v_117_cast_fp16, y = p_head_235_cast_fp16)[name = string("x_625_cast_fp16")];
            tensor<int32, [8]> x_627_pad_1 = const()[name = string("x_627_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_627_mode_1 = const()[name = string("x_627_mode_1"), val = string("constant")];
            fp16 const_1699_to_fp16 = const()[name = string("const_1699_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_627_cast_fp16 = pad(constant_val = const_1699_to_fp16, mode = x_627_mode_1, pad = x_627_pad_1, x = x_625_cast_fp16)[name = string("x_627_cast_fp16")];
            tensor<int32, [4]> var_9581 = const()[name = string("op_9581"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_629_cast_fp16 = reshape(shape = var_9581, x = x_627_cast_fp16)[name = string("x_629_cast_fp16")];
            tensor<int32, [4]> var_9585_begin_1 = const()[name = string("op_9585_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9585_end_1 = const()[name = string("op_9585_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_9585_end_mask_1 = const()[name = string("op_9585_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_9585_cast_fp16 = slice_by_index(begin = var_9585_begin_1, end = var_9585_end_1, end_mask = var_9585_end_mask_1, x = x_629_cast_fp16)[name = string("op_9585_cast_fp16")];
            tensor<int32, [4]> var_9587 = const()[name = string("op_9587"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_9588_cast_fp16 = reshape(shape = var_9587, x = var_9585_cast_fp16)[name = string("op_9588_cast_fp16")];
            tensor<int32, [4]> var_9593_begin_1 = const()[name = string("op_9593_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9593_end_1 = const()[name = string("op_9593_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_9593_end_mask_1 = const()[name = string("op_9593_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_9593_cast_fp16 = slice_by_index(begin = var_9593_begin_1, end = var_9593_end_1, end_mask = var_9593_end_mask_1, x = var_9588_cast_fp16)[name = string("op_9593_cast_fp16")];
            fp16 var_9594_to_fp16 = const()[name = string("op_9594_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_117_cast_fp16 = mul(x = var_9593_cast_fp16, y = var_9594_to_fp16)[name = string("scores_pos_117_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_117_cast_fp16 = add(x = scores_content_117_cast_fp16, y = scores_pos_117_cast_fp16)[name = string("logits_117_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_9597_cast_fp16 = softmax(axis = var_9080, x = logits_117_cast_fp16)[name = string("op_9597_cast_fp16")];
            bool var_9599_transpose_x_1 = const()[name = string("op_9599_transpose_x_1"), val = bool(false)];
            bool var_9599_transpose_y_1 = const()[name = string("op_9599_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_9599_cast_fp16 = matmul(transpose_x = var_9599_transpose_x_1, transpose_y = var_9599_transpose_y_1, x = var_9597_cast_fp16, y = v_head_235_cast_fp16)[name = string("op_9599_cast_fp16")];
            tensor<int32, [1]> var_9600_axes_1 = const()[name = string("op_9600_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_9600_cast_fp16 = squeeze(axes = var_9600_axes_1, x = var_9599_cast_fp16)[name = string("op_9600_cast_fp16")];
            string dense_output_601_pad_type_1 = const()[name = string("dense_output_601_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_601_strides_1 = const()[name = string("dense_output_601_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_601_pad_1 = const()[name = string("dense_output_601_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_601_dilations_1 = const()[name = string("dense_output_601_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_601_groups_1 = const()[name = string("dense_output_601_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221194112))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221325248))))[name = string("layers_7_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_601_cast_fp16 = conv(dilations = dense_output_601_dilations_1, groups = dense_output_601_groups_1, pad = dense_output_601_pad_1, pad_type = dense_output_601_pad_type_1, strides = dense_output_601_strides_1, weight = layers_7_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_601_cast_fp16")];
            string sparse_output_601_pad_type_1 = const()[name = string("sparse_output_601_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_601_strides_1 = const()[name = string("sparse_output_601_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_601_pad_1 = const()[name = string("sparse_output_601_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_601_dilations_1 = const()[name = string("sparse_output_601_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_601_groups_1 = const()[name = string("sparse_output_601_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221328512))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221325824))))[name = string("layers_7_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_601_cast_fp16 = conv(dilations = sparse_output_601_dilations_1, groups = sparse_output_601_groups_1, pad = sparse_output_601_pad_1, pad_type = sparse_output_601_pad_type_1, strides = sparse_output_601_strides_1, weight = layers_7_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_601_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9615_cast_fp16 = add(x = dense_output_601_cast_fp16, y = sparse_output_601_cast_fp16)[name = string("op_9615_cast_fp16")];
            tensor<int32, [4]> var_9616 = const()[name = string("op_9616"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9618 = const()[name = string("op_9618"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9617_cast_fp16 = transpose(perm = var_9616, x = var_9615_cast_fp16)[name = string("transpose_681")];
            tensor<fp16, [1, 51, 128]> q_head_119_cast_fp16 = reshape(shape = var_9618, x = var_9617_cast_fp16)[name = string("q_head_119_cast_fp16")];
            string dense_output_603_pad_type_1 = const()[name = string("dense_output_603_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_603_strides_1 = const()[name = string("dense_output_603_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_603_pad_1 = const()[name = string("dense_output_603_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_603_dilations_1 = const()[name = string("dense_output_603_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_603_groups_1 = const()[name = string("dense_output_603_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221344960))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221476096))))[name = string("layers_7_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_603_cast_fp16 = conv(dilations = dense_output_603_dilations_1, groups = dense_output_603_groups_1, pad = dense_output_603_pad_1, pad_type = dense_output_603_pad_type_1, strides = dense_output_603_strides_1, weight = layers_7_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_603_cast_fp16")];
            string sparse_output_603_pad_type_1 = const()[name = string("sparse_output_603_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_603_strides_1 = const()[name = string("sparse_output_603_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_603_pad_1 = const()[name = string("sparse_output_603_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_603_dilations_1 = const()[name = string("sparse_output_603_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_603_groups_1 = const()[name = string("sparse_output_603_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221479360))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221476672))))[name = string("layers_7_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_603_cast_fp16 = conv(dilations = sparse_output_603_dilations_1, groups = sparse_output_603_groups_1, pad = sparse_output_603_pad_1, pad_type = sparse_output_603_pad_type_1, strides = sparse_output_603_strides_1, weight = layers_7_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_603_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9634_cast_fp16 = add(x = dense_output_603_cast_fp16, y = sparse_output_603_cast_fp16)[name = string("op_9634_cast_fp16")];
            tensor<int32, [4]> var_9635 = const()[name = string("op_9635"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9637 = const()[name = string("op_9637"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9636_cast_fp16 = transpose(perm = var_9635, x = var_9634_cast_fp16)[name = string("transpose_680")];
            tensor<fp16, [1, 51, 128]> k_head_237_cast_fp16 = reshape(shape = var_9637, x = var_9636_cast_fp16)[name = string("k_head_237_cast_fp16")];
            string dense_output_605_pad_type_1 = const()[name = string("dense_output_605_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_605_strides_1 = const()[name = string("dense_output_605_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_605_pad_1 = const()[name = string("dense_output_605_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_605_dilations_1 = const()[name = string("dense_output_605_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_605_groups_1 = const()[name = string("dense_output_605_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221495808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221626944))))[name = string("layers_7_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_605_cast_fp16 = conv(dilations = dense_output_605_dilations_1, groups = dense_output_605_groups_1, pad = dense_output_605_pad_1, pad_type = dense_output_605_pad_type_1, strides = dense_output_605_strides_1, weight = layers_7_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_605_cast_fp16")];
            string sparse_output_605_pad_type_1 = const()[name = string("sparse_output_605_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_605_strides_1 = const()[name = string("sparse_output_605_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_605_pad_1 = const()[name = string("sparse_output_605_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_605_dilations_1 = const()[name = string("sparse_output_605_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_605_groups_1 = const()[name = string("sparse_output_605_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221630208))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221627520))))[name = string("layers_7_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_605_cast_fp16 = conv(dilations = sparse_output_605_dilations_1, groups = sparse_output_605_groups_1, pad = sparse_output_605_pad_1, pad_type = sparse_output_605_pad_type_1, strides = sparse_output_605_strides_1, weight = layers_7_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_605_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9653_cast_fp16 = add(x = dense_output_605_cast_fp16, y = sparse_output_605_cast_fp16)[name = string("op_9653_cast_fp16")];
            tensor<int32, [4]> var_9654 = const()[name = string("op_9654"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9656 = const()[name = string("op_9656"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9655_cast_fp16 = transpose(perm = var_9654, x = var_9653_cast_fp16)[name = string("transpose_679")];
            tensor<fp16, [1, 51, 128]> v_head_237_cast_fp16 = reshape(shape = var_9656, x = var_9655_cast_fp16)[name = string("v_head_237_cast_fp16")];
            string dense_output_607_pad_type_1 = const()[name = string("dense_output_607_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_607_strides_1 = const()[name = string("dense_output_607_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_607_pad_1 = const()[name = string("dense_output_607_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_607_dilations_1 = const()[name = string("dense_output_607_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_607_groups_1 = const()[name = string("dense_output_607_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221646656))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221777792))))[name = string("layers_7_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_607_cast_fp16 = conv(dilations = dense_output_607_dilations_1, groups = dense_output_607_groups_1, pad = dense_output_607_pad_1, pad_type = dense_output_607_pad_type_1, strides = dense_output_607_strides_1, weight = layers_7_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_607_cast_fp16")];
            string sparse_output_607_pad_type_1 = const()[name = string("sparse_output_607_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_607_strides_1 = const()[name = string("sparse_output_607_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_607_pad_1 = const()[name = string("sparse_output_607_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_607_dilations_1 = const()[name = string("sparse_output_607_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_607_groups_1 = const()[name = string("sparse_output_607_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221781056))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221778368))))[name = string("layers_7_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_607_cast_fp16 = conv(dilations = sparse_output_607_dilations_1, groups = sparse_output_607_groups_1, pad = sparse_output_607_pad_1, pad_type = sparse_output_607_pad_type_1, strides = sparse_output_607_strides_1, weight = layers_7_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_607_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_9672_cast_fp16 = add(x = dense_output_607_cast_fp16, y = sparse_output_607_cast_fp16)[name = string("op_9672_cast_fp16")];
            tensor<int32, [4]> var_9673 = const()[name = string("op_9673"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9675 = const()[name = string("op_9675"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_9674_cast_fp16 = transpose(perm = var_9673, x = var_9672_cast_fp16)[name = string("transpose_678")];
            tensor<fp16, [1, 101, 128]> p_head_237_cast_fp16 = reshape(shape = var_9675, x = var_9674_cast_fp16)[name = string("p_head_237_cast_fp16")];
            tensor<fp16, [128]> var_9677_to_fp16 = const()[name = string("op_9677_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221797504)))];
            tensor<fp16, [1, 51, 128]> var_9678_cast_fp16 = add(x = q_head_119_cast_fp16, y = var_9677_to_fp16)[name = string("op_9678_cast_fp16")];
            tensor<int32, [1]> q_u_119_axes_1 = const()[name = string("q_u_119_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_119_cast_fp16 = expand_dims(axes = q_u_119_axes_1, x = var_9678_cast_fp16)[name = string("q_u_119_cast_fp16")];
            tensor<fp16, [128]> var_9680_to_fp16 = const()[name = string("op_9680_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221797824)))];
            tensor<fp16, [1, 51, 128]> var_9681_cast_fp16 = add(x = q_head_119_cast_fp16, y = var_9680_to_fp16)[name = string("op_9681_cast_fp16")];
            tensor<int32, [1]> q_v_119_axes_1 = const()[name = string("q_v_119_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_119_cast_fp16 = expand_dims(axes = q_v_119_axes_1, x = var_9681_cast_fp16)[name = string("q_v_119_cast_fp16")];
            tensor<int32, [1]> k_head_239_axes_1 = const()[name = string("k_head_239_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_239_cast_fp16 = expand_dims(axes = k_head_239_axes_1, x = k_head_237_cast_fp16)[name = string("k_head_239_cast_fp16")];
            tensor<int32, [1]> v_head_239_axes_1 = const()[name = string("v_head_239_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_239_cast_fp16 = expand_dims(axes = v_head_239_axes_1, x = v_head_237_cast_fp16)[name = string("v_head_239_cast_fp16")];
            tensor<int32, [1]> p_head_239_axes_1 = const()[name = string("p_head_239_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_239_cast_fp16 = expand_dims(axes = p_head_239_axes_1, x = p_head_237_cast_fp16)[name = string("p_head_239_cast_fp16")];
            bool var_9687_transpose_x_3 = const()[name = string("op_9687_transpose_x_3"), val = bool(false)];
            bool var_9687_transpose_y_3 = const()[name = string("op_9687_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_9687_cast_fp16 = matmul(transpose_x = var_9687_transpose_x_3, transpose_y = var_9687_transpose_y_3, x = q_u_119_cast_fp16, y = k_head_239_cast_fp16)[name = string("op_9687_cast_fp16")];
            fp16 var_9688_to_fp16 = const()[name = string("op_9688_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_119_cast_fp16 = mul(x = var_9687_cast_fp16, y = var_9688_to_fp16)[name = string("scores_content_119_cast_fp16")];
            bool x_633_transpose_x_3 = const()[name = string("x_633_transpose_x_3"), val = bool(false)];
            bool x_633_transpose_y_3 = const()[name = string("x_633_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_633_cast_fp16 = matmul(transpose_x = x_633_transpose_x_3, transpose_y = x_633_transpose_y_3, x = q_v_119_cast_fp16, y = p_head_239_cast_fp16)[name = string("x_633_cast_fp16")];
            tensor<int32, [8]> x_635_pad_1 = const()[name = string("x_635_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_635_mode_1 = const()[name = string("x_635_mode_1"), val = string("constant")];
            fp16 const_1705_to_fp16 = const()[name = string("const_1705_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_635_cast_fp16 = pad(constant_val = const_1705_to_fp16, mode = x_635_mode_1, pad = x_635_pad_1, x = x_633_cast_fp16)[name = string("x_635_cast_fp16")];
            tensor<int32, [4]> var_9702 = const()[name = string("op_9702"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_637_cast_fp16 = reshape(shape = var_9702, x = x_635_cast_fp16)[name = string("x_637_cast_fp16")];
            tensor<int32, [4]> var_9706_begin_1 = const()[name = string("op_9706_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9706_end_1 = const()[name = string("op_9706_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_9706_end_mask_1 = const()[name = string("op_9706_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_9706_cast_fp16 = slice_by_index(begin = var_9706_begin_1, end = var_9706_end_1, end_mask = var_9706_end_mask_1, x = x_637_cast_fp16)[name = string("op_9706_cast_fp16")];
            tensor<int32, [4]> var_9708 = const()[name = string("op_9708"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_9709_cast_fp16 = reshape(shape = var_9708, x = var_9706_cast_fp16)[name = string("op_9709_cast_fp16")];
            tensor<int32, [4]> var_9714_begin_1 = const()[name = string("op_9714_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9714_end_1 = const()[name = string("op_9714_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_9714_end_mask_1 = const()[name = string("op_9714_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_9714_cast_fp16 = slice_by_index(begin = var_9714_begin_1, end = var_9714_end_1, end_mask = var_9714_end_mask_1, x = var_9709_cast_fp16)[name = string("op_9714_cast_fp16")];
            fp16 var_9715_to_fp16 = const()[name = string("op_9715_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_119_cast_fp16 = mul(x = var_9714_cast_fp16, y = var_9715_to_fp16)[name = string("scores_pos_119_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_119_cast_fp16 = add(x = scores_content_119_cast_fp16, y = scores_pos_119_cast_fp16)[name = string("logits_119_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_9718_cast_fp16 = softmax(axis = var_9080, x = logits_119_cast_fp16)[name = string("op_9718_cast_fp16")];
            bool var_9720_transpose_x_1 = const()[name = string("op_9720_transpose_x_1"), val = bool(false)];
            bool var_9720_transpose_y_1 = const()[name = string("op_9720_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_9720_cast_fp16 = matmul(transpose_x = var_9720_transpose_x_1, transpose_y = var_9720_transpose_y_1, x = var_9718_cast_fp16, y = v_head_239_cast_fp16)[name = string("op_9720_cast_fp16")];
            tensor<int32, [1]> var_9721_axes_1 = const()[name = string("op_9721_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_9721_cast_fp16 = squeeze(axes = var_9721_axes_1, x = var_9720_cast_fp16)[name = string("op_9721_cast_fp16")];
            string dense_output_609_pad_type_1 = const()[name = string("dense_output_609_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_609_strides_1 = const()[name = string("dense_output_609_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_609_pad_1 = const()[name = string("dense_output_609_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_609_dilations_1 = const()[name = string("dense_output_609_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_609_groups_1 = const()[name = string("dense_output_609_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221798144))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221929280))))[name = string("layers_7_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_609_cast_fp16 = conv(dilations = dense_output_609_dilations_1, groups = dense_output_609_groups_1, pad = dense_output_609_pad_1, pad_type = dense_output_609_pad_type_1, strides = dense_output_609_strides_1, weight = layers_7_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_609_cast_fp16")];
            string sparse_output_609_pad_type_1 = const()[name = string("sparse_output_609_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_609_strides_1 = const()[name = string("sparse_output_609_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_609_pad_1 = const()[name = string("sparse_output_609_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_609_dilations_1 = const()[name = string("sparse_output_609_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_609_groups_1 = const()[name = string("sparse_output_609_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221932544))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221929856))))[name = string("layers_7_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_609_cast_fp16 = conv(dilations = sparse_output_609_dilations_1, groups = sparse_output_609_groups_1, pad = sparse_output_609_pad_1, pad_type = sparse_output_609_pad_type_1, strides = sparse_output_609_strides_1, weight = layers_7_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_609_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9736_cast_fp16 = add(x = dense_output_609_cast_fp16, y = sparse_output_609_cast_fp16)[name = string("op_9736_cast_fp16")];
            tensor<int32, [4]> var_9737 = const()[name = string("op_9737"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9739 = const()[name = string("op_9739"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9738_cast_fp16 = transpose(perm = var_9737, x = var_9736_cast_fp16)[name = string("transpose_677")];
            tensor<fp16, [1, 51, 128]> q_head_121_cast_fp16 = reshape(shape = var_9739, x = var_9738_cast_fp16)[name = string("q_head_121_cast_fp16")];
            string dense_output_611_pad_type_1 = const()[name = string("dense_output_611_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_611_strides_1 = const()[name = string("dense_output_611_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_611_pad_1 = const()[name = string("dense_output_611_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_611_dilations_1 = const()[name = string("dense_output_611_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_611_groups_1 = const()[name = string("dense_output_611_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221948992))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222080128))))[name = string("layers_7_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_611_cast_fp16 = conv(dilations = dense_output_611_dilations_1, groups = dense_output_611_groups_1, pad = dense_output_611_pad_1, pad_type = dense_output_611_pad_type_1, strides = dense_output_611_strides_1, weight = layers_7_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_611_cast_fp16")];
            string sparse_output_611_pad_type_1 = const()[name = string("sparse_output_611_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_611_strides_1 = const()[name = string("sparse_output_611_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_611_pad_1 = const()[name = string("sparse_output_611_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_611_dilations_1 = const()[name = string("sparse_output_611_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_611_groups_1 = const()[name = string("sparse_output_611_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222083392))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222080704))))[name = string("layers_7_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_611_cast_fp16 = conv(dilations = sparse_output_611_dilations_1, groups = sparse_output_611_groups_1, pad = sparse_output_611_pad_1, pad_type = sparse_output_611_pad_type_1, strides = sparse_output_611_strides_1, weight = layers_7_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_611_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9755_cast_fp16 = add(x = dense_output_611_cast_fp16, y = sparse_output_611_cast_fp16)[name = string("op_9755_cast_fp16")];
            tensor<int32, [4]> var_9756 = const()[name = string("op_9756"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9758 = const()[name = string("op_9758"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9757_cast_fp16 = transpose(perm = var_9756, x = var_9755_cast_fp16)[name = string("transpose_676")];
            tensor<fp16, [1, 51, 128]> k_head_241_cast_fp16 = reshape(shape = var_9758, x = var_9757_cast_fp16)[name = string("k_head_241_cast_fp16")];
            string dense_output_613_pad_type_1 = const()[name = string("dense_output_613_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_613_strides_1 = const()[name = string("dense_output_613_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_613_pad_1 = const()[name = string("dense_output_613_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_613_dilations_1 = const()[name = string("dense_output_613_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_613_groups_1 = const()[name = string("dense_output_613_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222099840))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222230976))))[name = string("layers_7_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_613_cast_fp16 = conv(dilations = dense_output_613_dilations_1, groups = dense_output_613_groups_1, pad = dense_output_613_pad_1, pad_type = dense_output_613_pad_type_1, strides = dense_output_613_strides_1, weight = layers_7_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_613_cast_fp16")];
            string sparse_output_613_pad_type_1 = const()[name = string("sparse_output_613_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_613_strides_1 = const()[name = string("sparse_output_613_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_613_pad_1 = const()[name = string("sparse_output_613_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_613_dilations_1 = const()[name = string("sparse_output_613_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_613_groups_1 = const()[name = string("sparse_output_613_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222234240))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222231552))))[name = string("layers_7_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_613_cast_fp16 = conv(dilations = sparse_output_613_dilations_1, groups = sparse_output_613_groups_1, pad = sparse_output_613_pad_1, pad_type = sparse_output_613_pad_type_1, strides = sparse_output_613_strides_1, weight = layers_7_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_613_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9774_cast_fp16 = add(x = dense_output_613_cast_fp16, y = sparse_output_613_cast_fp16)[name = string("op_9774_cast_fp16")];
            tensor<int32, [4]> var_9775 = const()[name = string("op_9775"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9777 = const()[name = string("op_9777"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9776_cast_fp16 = transpose(perm = var_9775, x = var_9774_cast_fp16)[name = string("transpose_675")];
            tensor<fp16, [1, 51, 128]> v_head_241_cast_fp16 = reshape(shape = var_9777, x = var_9776_cast_fp16)[name = string("v_head_241_cast_fp16")];
            string dense_output_615_pad_type_1 = const()[name = string("dense_output_615_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_615_strides_1 = const()[name = string("dense_output_615_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_615_pad_1 = const()[name = string("dense_output_615_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_615_dilations_1 = const()[name = string("dense_output_615_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_615_groups_1 = const()[name = string("dense_output_615_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222250688))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222381824))))[name = string("layers_7_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_615_cast_fp16 = conv(dilations = dense_output_615_dilations_1, groups = dense_output_615_groups_1, pad = dense_output_615_pad_1, pad_type = dense_output_615_pad_type_1, strides = dense_output_615_strides_1, weight = layers_7_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_615_cast_fp16")];
            string sparse_output_615_pad_type_1 = const()[name = string("sparse_output_615_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_615_strides_1 = const()[name = string("sparse_output_615_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_615_pad_1 = const()[name = string("sparse_output_615_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_615_dilations_1 = const()[name = string("sparse_output_615_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_615_groups_1 = const()[name = string("sparse_output_615_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222385088))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222382400))))[name = string("layers_7_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_615_cast_fp16 = conv(dilations = sparse_output_615_dilations_1, groups = sparse_output_615_groups_1, pad = sparse_output_615_pad_1, pad_type = sparse_output_615_pad_type_1, strides = sparse_output_615_strides_1, weight = layers_7_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_615_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_9793_cast_fp16 = add(x = dense_output_615_cast_fp16, y = sparse_output_615_cast_fp16)[name = string("op_9793_cast_fp16")];
            tensor<int32, [4]> var_9794 = const()[name = string("op_9794"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9796 = const()[name = string("op_9796"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_9795_cast_fp16 = transpose(perm = var_9794, x = var_9793_cast_fp16)[name = string("transpose_674")];
            tensor<fp16, [1, 101, 128]> p_head_241_cast_fp16 = reshape(shape = var_9796, x = var_9795_cast_fp16)[name = string("p_head_241_cast_fp16")];
            tensor<fp16, [128]> var_9798_to_fp16 = const()[name = string("op_9798_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222401536)))];
            tensor<fp16, [1, 51, 128]> var_9799_cast_fp16 = add(x = q_head_121_cast_fp16, y = var_9798_to_fp16)[name = string("op_9799_cast_fp16")];
            tensor<int32, [1]> q_u_121_axes_1 = const()[name = string("q_u_121_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_121_cast_fp16 = expand_dims(axes = q_u_121_axes_1, x = var_9799_cast_fp16)[name = string("q_u_121_cast_fp16")];
            tensor<fp16, [128]> var_9801_to_fp16 = const()[name = string("op_9801_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222401856)))];
            tensor<fp16, [1, 51, 128]> var_9802_cast_fp16 = add(x = q_head_121_cast_fp16, y = var_9801_to_fp16)[name = string("op_9802_cast_fp16")];
            tensor<int32, [1]> q_v_121_axes_1 = const()[name = string("q_v_121_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_121_cast_fp16 = expand_dims(axes = q_v_121_axes_1, x = var_9802_cast_fp16)[name = string("q_v_121_cast_fp16")];
            tensor<int32, [1]> k_head_243_axes_1 = const()[name = string("k_head_243_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_243_cast_fp16 = expand_dims(axes = k_head_243_axes_1, x = k_head_241_cast_fp16)[name = string("k_head_243_cast_fp16")];
            tensor<int32, [1]> v_head_243_axes_1 = const()[name = string("v_head_243_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_243_cast_fp16 = expand_dims(axes = v_head_243_axes_1, x = v_head_241_cast_fp16)[name = string("v_head_243_cast_fp16")];
            tensor<int32, [1]> p_head_243_axes_1 = const()[name = string("p_head_243_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_243_cast_fp16 = expand_dims(axes = p_head_243_axes_1, x = p_head_241_cast_fp16)[name = string("p_head_243_cast_fp16")];
            bool var_9808_transpose_x_3 = const()[name = string("op_9808_transpose_x_3"), val = bool(false)];
            bool var_9808_transpose_y_3 = const()[name = string("op_9808_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_9808_cast_fp16 = matmul(transpose_x = var_9808_transpose_x_3, transpose_y = var_9808_transpose_y_3, x = q_u_121_cast_fp16, y = k_head_243_cast_fp16)[name = string("op_9808_cast_fp16")];
            fp16 var_9809_to_fp16 = const()[name = string("op_9809_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_121_cast_fp16 = mul(x = var_9808_cast_fp16, y = var_9809_to_fp16)[name = string("scores_content_121_cast_fp16")];
            bool x_641_transpose_x_3 = const()[name = string("x_641_transpose_x_3"), val = bool(false)];
            bool x_641_transpose_y_3 = const()[name = string("x_641_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_641_cast_fp16 = matmul(transpose_x = x_641_transpose_x_3, transpose_y = x_641_transpose_y_3, x = q_v_121_cast_fp16, y = p_head_243_cast_fp16)[name = string("x_641_cast_fp16")];
            tensor<int32, [8]> x_643_pad_1 = const()[name = string("x_643_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_643_mode_1 = const()[name = string("x_643_mode_1"), val = string("constant")];
            fp16 const_1711_to_fp16 = const()[name = string("const_1711_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_643_cast_fp16 = pad(constant_val = const_1711_to_fp16, mode = x_643_mode_1, pad = x_643_pad_1, x = x_641_cast_fp16)[name = string("x_643_cast_fp16")];
            tensor<int32, [4]> var_9823 = const()[name = string("op_9823"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_645_cast_fp16 = reshape(shape = var_9823, x = x_643_cast_fp16)[name = string("x_645_cast_fp16")];
            tensor<int32, [4]> var_9827_begin_1 = const()[name = string("op_9827_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9827_end_1 = const()[name = string("op_9827_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_9827_end_mask_1 = const()[name = string("op_9827_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_9827_cast_fp16 = slice_by_index(begin = var_9827_begin_1, end = var_9827_end_1, end_mask = var_9827_end_mask_1, x = x_645_cast_fp16)[name = string("op_9827_cast_fp16")];
            tensor<int32, [4]> var_9829 = const()[name = string("op_9829"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_9830_cast_fp16 = reshape(shape = var_9829, x = var_9827_cast_fp16)[name = string("op_9830_cast_fp16")];
            tensor<int32, [4]> var_9835_begin_1 = const()[name = string("op_9835_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9835_end_1 = const()[name = string("op_9835_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_9835_end_mask_1 = const()[name = string("op_9835_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_9835_cast_fp16 = slice_by_index(begin = var_9835_begin_1, end = var_9835_end_1, end_mask = var_9835_end_mask_1, x = var_9830_cast_fp16)[name = string("op_9835_cast_fp16")];
            fp16 var_9836_to_fp16 = const()[name = string("op_9836_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_121_cast_fp16 = mul(x = var_9835_cast_fp16, y = var_9836_to_fp16)[name = string("scores_pos_121_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_121_cast_fp16 = add(x = scores_content_121_cast_fp16, y = scores_pos_121_cast_fp16)[name = string("logits_121_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_9839_cast_fp16 = softmax(axis = var_9080, x = logits_121_cast_fp16)[name = string("op_9839_cast_fp16")];
            bool var_9841_transpose_x_1 = const()[name = string("op_9841_transpose_x_1"), val = bool(false)];
            bool var_9841_transpose_y_1 = const()[name = string("op_9841_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_9841_cast_fp16 = matmul(transpose_x = var_9841_transpose_x_1, transpose_y = var_9841_transpose_y_1, x = var_9839_cast_fp16, y = v_head_243_cast_fp16)[name = string("op_9841_cast_fp16")];
            tensor<int32, [1]> var_9842_axes_1 = const()[name = string("op_9842_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_9842_cast_fp16 = squeeze(axes = var_9842_axes_1, x = var_9841_cast_fp16)[name = string("op_9842_cast_fp16")];
            string dense_output_617_pad_type_1 = const()[name = string("dense_output_617_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_617_strides_1 = const()[name = string("dense_output_617_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_617_pad_1 = const()[name = string("dense_output_617_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_617_dilations_1 = const()[name = string("dense_output_617_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_617_groups_1 = const()[name = string("dense_output_617_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222402176))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222533312))))[name = string("layers_7_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_617_cast_fp16 = conv(dilations = dense_output_617_dilations_1, groups = dense_output_617_groups_1, pad = dense_output_617_pad_1, pad_type = dense_output_617_pad_type_1, strides = dense_output_617_strides_1, weight = layers_7_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_617_cast_fp16")];
            string sparse_output_617_pad_type_1 = const()[name = string("sparse_output_617_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_617_strides_1 = const()[name = string("sparse_output_617_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_617_pad_1 = const()[name = string("sparse_output_617_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_617_dilations_1 = const()[name = string("sparse_output_617_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_617_groups_1 = const()[name = string("sparse_output_617_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222536576))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222533888))))[name = string("layers_7_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_617_cast_fp16 = conv(dilations = sparse_output_617_dilations_1, groups = sparse_output_617_groups_1, pad = sparse_output_617_pad_1, pad_type = sparse_output_617_pad_type_1, strides = sparse_output_617_strides_1, weight = layers_7_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_617_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9857_cast_fp16 = add(x = dense_output_617_cast_fp16, y = sparse_output_617_cast_fp16)[name = string("op_9857_cast_fp16")];
            tensor<int32, [4]> var_9858 = const()[name = string("op_9858"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9860 = const()[name = string("op_9860"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9859_cast_fp16 = transpose(perm = var_9858, x = var_9857_cast_fp16)[name = string("transpose_673")];
            tensor<fp16, [1, 51, 128]> q_head_123_cast_fp16 = reshape(shape = var_9860, x = var_9859_cast_fp16)[name = string("q_head_123_cast_fp16")];
            string dense_output_619_pad_type_1 = const()[name = string("dense_output_619_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_619_strides_1 = const()[name = string("dense_output_619_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_619_pad_1 = const()[name = string("dense_output_619_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_619_dilations_1 = const()[name = string("dense_output_619_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_619_groups_1 = const()[name = string("dense_output_619_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222553024))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222684160))))[name = string("layers_7_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_619_cast_fp16 = conv(dilations = dense_output_619_dilations_1, groups = dense_output_619_groups_1, pad = dense_output_619_pad_1, pad_type = dense_output_619_pad_type_1, strides = dense_output_619_strides_1, weight = layers_7_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_619_cast_fp16")];
            string sparse_output_619_pad_type_1 = const()[name = string("sparse_output_619_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_619_strides_1 = const()[name = string("sparse_output_619_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_619_pad_1 = const()[name = string("sparse_output_619_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_619_dilations_1 = const()[name = string("sparse_output_619_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_619_groups_1 = const()[name = string("sparse_output_619_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222687424))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222684736))))[name = string("layers_7_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_619_cast_fp16 = conv(dilations = sparse_output_619_dilations_1, groups = sparse_output_619_groups_1, pad = sparse_output_619_pad_1, pad_type = sparse_output_619_pad_type_1, strides = sparse_output_619_strides_1, weight = layers_7_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_619_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9876_cast_fp16 = add(x = dense_output_619_cast_fp16, y = sparse_output_619_cast_fp16)[name = string("op_9876_cast_fp16")];
            tensor<int32, [4]> var_9877 = const()[name = string("op_9877"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9879 = const()[name = string("op_9879"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9878_cast_fp16 = transpose(perm = var_9877, x = var_9876_cast_fp16)[name = string("transpose_672")];
            tensor<fp16, [1, 51, 128]> k_head_245_cast_fp16 = reshape(shape = var_9879, x = var_9878_cast_fp16)[name = string("k_head_245_cast_fp16")];
            string dense_output_621_pad_type_1 = const()[name = string("dense_output_621_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_621_strides_1 = const()[name = string("dense_output_621_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_621_pad_1 = const()[name = string("dense_output_621_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_621_dilations_1 = const()[name = string("dense_output_621_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_621_groups_1 = const()[name = string("dense_output_621_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222703872))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222835008))))[name = string("layers_7_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_621_cast_fp16 = conv(dilations = dense_output_621_dilations_1, groups = dense_output_621_groups_1, pad = dense_output_621_pad_1, pad_type = dense_output_621_pad_type_1, strides = dense_output_621_strides_1, weight = layers_7_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_621_cast_fp16")];
            string sparse_output_621_pad_type_1 = const()[name = string("sparse_output_621_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_621_strides_1 = const()[name = string("sparse_output_621_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_621_pad_1 = const()[name = string("sparse_output_621_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_621_dilations_1 = const()[name = string("sparse_output_621_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_621_groups_1 = const()[name = string("sparse_output_621_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222838272))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222835584))))[name = string("layers_7_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_621_cast_fp16 = conv(dilations = sparse_output_621_dilations_1, groups = sparse_output_621_groups_1, pad = sparse_output_621_pad_1, pad_type = sparse_output_621_pad_type_1, strides = sparse_output_621_strides_1, weight = layers_7_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_621_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9895_cast_fp16 = add(x = dense_output_621_cast_fp16, y = sparse_output_621_cast_fp16)[name = string("op_9895_cast_fp16")];
            tensor<int32, [4]> var_9896 = const()[name = string("op_9896"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9898 = const()[name = string("op_9898"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9897_cast_fp16 = transpose(perm = var_9896, x = var_9895_cast_fp16)[name = string("transpose_671")];
            tensor<fp16, [1, 51, 128]> v_head_245_cast_fp16 = reshape(shape = var_9898, x = var_9897_cast_fp16)[name = string("v_head_245_cast_fp16")];
            string dense_output_623_pad_type_1 = const()[name = string("dense_output_623_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_623_strides_1 = const()[name = string("dense_output_623_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_623_pad_1 = const()[name = string("dense_output_623_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_623_dilations_1 = const()[name = string("dense_output_623_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_623_groups_1 = const()[name = string("dense_output_623_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222854720))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222985856))))[name = string("layers_7_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_623_cast_fp16 = conv(dilations = dense_output_623_dilations_1, groups = dense_output_623_groups_1, pad = dense_output_623_pad_1, pad_type = dense_output_623_pad_type_1, strides = dense_output_623_strides_1, weight = layers_7_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_623_cast_fp16")];
            string sparse_output_623_pad_type_1 = const()[name = string("sparse_output_623_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_623_strides_1 = const()[name = string("sparse_output_623_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_623_pad_1 = const()[name = string("sparse_output_623_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_623_dilations_1 = const()[name = string("sparse_output_623_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_623_groups_1 = const()[name = string("sparse_output_623_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222989120))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222986432))))[name = string("layers_7_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_623_cast_fp16 = conv(dilations = sparse_output_623_dilations_1, groups = sparse_output_623_groups_1, pad = sparse_output_623_pad_1, pad_type = sparse_output_623_pad_type_1, strides = sparse_output_623_strides_1, weight = layers_7_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_623_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_9914_cast_fp16 = add(x = dense_output_623_cast_fp16, y = sparse_output_623_cast_fp16)[name = string("op_9914_cast_fp16")];
            tensor<int32, [4]> var_9915 = const()[name = string("op_9915"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9917 = const()[name = string("op_9917"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_9916_cast_fp16 = transpose(perm = var_9915, x = var_9914_cast_fp16)[name = string("transpose_670")];
            tensor<fp16, [1, 101, 128]> p_head_245_cast_fp16 = reshape(shape = var_9917, x = var_9916_cast_fp16)[name = string("p_head_245_cast_fp16")];
            tensor<fp16, [128]> var_9919_to_fp16 = const()[name = string("op_9919_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223005568)))];
            tensor<fp16, [1, 51, 128]> var_9920_cast_fp16 = add(x = q_head_123_cast_fp16, y = var_9919_to_fp16)[name = string("op_9920_cast_fp16")];
            tensor<int32, [1]> q_u_123_axes_1 = const()[name = string("q_u_123_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_123_cast_fp16 = expand_dims(axes = q_u_123_axes_1, x = var_9920_cast_fp16)[name = string("q_u_123_cast_fp16")];
            tensor<fp16, [128]> var_9922_to_fp16 = const()[name = string("op_9922_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223005888)))];
            tensor<fp16, [1, 51, 128]> var_9923_cast_fp16 = add(x = q_head_123_cast_fp16, y = var_9922_to_fp16)[name = string("op_9923_cast_fp16")];
            tensor<int32, [1]> q_v_123_axes_1 = const()[name = string("q_v_123_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_123_cast_fp16 = expand_dims(axes = q_v_123_axes_1, x = var_9923_cast_fp16)[name = string("q_v_123_cast_fp16")];
            tensor<int32, [1]> k_head_247_axes_1 = const()[name = string("k_head_247_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_247_cast_fp16 = expand_dims(axes = k_head_247_axes_1, x = k_head_245_cast_fp16)[name = string("k_head_247_cast_fp16")];
            tensor<int32, [1]> v_head_247_axes_1 = const()[name = string("v_head_247_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_247_cast_fp16 = expand_dims(axes = v_head_247_axes_1, x = v_head_245_cast_fp16)[name = string("v_head_247_cast_fp16")];
            tensor<int32, [1]> p_head_247_axes_1 = const()[name = string("p_head_247_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_247_cast_fp16 = expand_dims(axes = p_head_247_axes_1, x = p_head_245_cast_fp16)[name = string("p_head_247_cast_fp16")];
            bool var_9929_transpose_x_3 = const()[name = string("op_9929_transpose_x_3"), val = bool(false)];
            bool var_9929_transpose_y_3 = const()[name = string("op_9929_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_9929_cast_fp16 = matmul(transpose_x = var_9929_transpose_x_3, transpose_y = var_9929_transpose_y_3, x = q_u_123_cast_fp16, y = k_head_247_cast_fp16)[name = string("op_9929_cast_fp16")];
            fp16 var_9930_to_fp16 = const()[name = string("op_9930_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_123_cast_fp16 = mul(x = var_9929_cast_fp16, y = var_9930_to_fp16)[name = string("scores_content_123_cast_fp16")];
            bool x_649_transpose_x_3 = const()[name = string("x_649_transpose_x_3"), val = bool(false)];
            bool x_649_transpose_y_3 = const()[name = string("x_649_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_649_cast_fp16 = matmul(transpose_x = x_649_transpose_x_3, transpose_y = x_649_transpose_y_3, x = q_v_123_cast_fp16, y = p_head_247_cast_fp16)[name = string("x_649_cast_fp16")];
            tensor<int32, [8]> x_651_pad_1 = const()[name = string("x_651_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_651_mode_1 = const()[name = string("x_651_mode_1"), val = string("constant")];
            fp16 const_1717_to_fp16 = const()[name = string("const_1717_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_651_cast_fp16 = pad(constant_val = const_1717_to_fp16, mode = x_651_mode_1, pad = x_651_pad_1, x = x_649_cast_fp16)[name = string("x_651_cast_fp16")];
            tensor<int32, [4]> var_9944 = const()[name = string("op_9944"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_653_cast_fp16 = reshape(shape = var_9944, x = x_651_cast_fp16)[name = string("x_653_cast_fp16")];
            tensor<int32, [4]> var_9948_begin_1 = const()[name = string("op_9948_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_9948_end_1 = const()[name = string("op_9948_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_9948_end_mask_1 = const()[name = string("op_9948_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_9948_cast_fp16 = slice_by_index(begin = var_9948_begin_1, end = var_9948_end_1, end_mask = var_9948_end_mask_1, x = x_653_cast_fp16)[name = string("op_9948_cast_fp16")];
            tensor<int32, [4]> var_9950 = const()[name = string("op_9950"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_9951_cast_fp16 = reshape(shape = var_9950, x = var_9948_cast_fp16)[name = string("op_9951_cast_fp16")];
            tensor<int32, [4]> var_9956_begin_1 = const()[name = string("op_9956_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9956_end_1 = const()[name = string("op_9956_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_9956_end_mask_1 = const()[name = string("op_9956_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_9956_cast_fp16 = slice_by_index(begin = var_9956_begin_1, end = var_9956_end_1, end_mask = var_9956_end_mask_1, x = var_9951_cast_fp16)[name = string("op_9956_cast_fp16")];
            fp16 var_9957_to_fp16 = const()[name = string("op_9957_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_123_cast_fp16 = mul(x = var_9956_cast_fp16, y = var_9957_to_fp16)[name = string("scores_pos_123_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_123_cast_fp16 = add(x = scores_content_123_cast_fp16, y = scores_pos_123_cast_fp16)[name = string("logits_123_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_9960_cast_fp16 = softmax(axis = var_9080, x = logits_123_cast_fp16)[name = string("op_9960_cast_fp16")];
            bool var_9962_transpose_x_1 = const()[name = string("op_9962_transpose_x_1"), val = bool(false)];
            bool var_9962_transpose_y_1 = const()[name = string("op_9962_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_9962_cast_fp16 = matmul(transpose_x = var_9962_transpose_x_1, transpose_y = var_9962_transpose_y_1, x = var_9960_cast_fp16, y = v_head_247_cast_fp16)[name = string("op_9962_cast_fp16")];
            tensor<int32, [1]> var_9963_axes_1 = const()[name = string("op_9963_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_9963_cast_fp16 = squeeze(axes = var_9963_axes_1, x = var_9962_cast_fp16)[name = string("op_9963_cast_fp16")];
            string dense_output_625_pad_type_1 = const()[name = string("dense_output_625_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_625_strides_1 = const()[name = string("dense_output_625_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_625_pad_1 = const()[name = string("dense_output_625_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_625_dilations_1 = const()[name = string("dense_output_625_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_625_groups_1 = const()[name = string("dense_output_625_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223006208))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223137344))))[name = string("layers_7_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_625_cast_fp16 = conv(dilations = dense_output_625_dilations_1, groups = dense_output_625_groups_1, pad = dense_output_625_pad_1, pad_type = dense_output_625_pad_type_1, strides = dense_output_625_strides_1, weight = layers_7_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_625_cast_fp16")];
            string sparse_output_625_pad_type_1 = const()[name = string("sparse_output_625_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_625_strides_1 = const()[name = string("sparse_output_625_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_625_pad_1 = const()[name = string("sparse_output_625_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_625_dilations_1 = const()[name = string("sparse_output_625_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_625_groups_1 = const()[name = string("sparse_output_625_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223140608))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223137920))))[name = string("layers_7_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_625_cast_fp16 = conv(dilations = sparse_output_625_dilations_1, groups = sparse_output_625_groups_1, pad = sparse_output_625_pad_1, pad_type = sparse_output_625_pad_type_1, strides = sparse_output_625_strides_1, weight = layers_7_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_625_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9978_cast_fp16 = add(x = dense_output_625_cast_fp16, y = sparse_output_625_cast_fp16)[name = string("op_9978_cast_fp16")];
            tensor<int32, [4]> var_9979 = const()[name = string("op_9979"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_9981 = const()[name = string("op_9981"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9980_cast_fp16 = transpose(perm = var_9979, x = var_9978_cast_fp16)[name = string("transpose_669")];
            tensor<fp16, [1, 51, 128]> q_head_125_cast_fp16 = reshape(shape = var_9981, x = var_9980_cast_fp16)[name = string("q_head_125_cast_fp16")];
            string dense_output_627_pad_type_1 = const()[name = string("dense_output_627_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_627_strides_1 = const()[name = string("dense_output_627_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_627_pad_1 = const()[name = string("dense_output_627_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_627_dilations_1 = const()[name = string("dense_output_627_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_627_groups_1 = const()[name = string("dense_output_627_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223157056))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223288192))))[name = string("layers_7_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_627_cast_fp16 = conv(dilations = dense_output_627_dilations_1, groups = dense_output_627_groups_1, pad = dense_output_627_pad_1, pad_type = dense_output_627_pad_type_1, strides = dense_output_627_strides_1, weight = layers_7_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_627_cast_fp16")];
            string sparse_output_627_pad_type_1 = const()[name = string("sparse_output_627_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_627_strides_1 = const()[name = string("sparse_output_627_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_627_pad_1 = const()[name = string("sparse_output_627_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_627_dilations_1 = const()[name = string("sparse_output_627_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_627_groups_1 = const()[name = string("sparse_output_627_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223291456))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223288768))))[name = string("layers_7_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_627_cast_fp16 = conv(dilations = sparse_output_627_dilations_1, groups = sparse_output_627_groups_1, pad = sparse_output_627_pad_1, pad_type = sparse_output_627_pad_type_1, strides = sparse_output_627_strides_1, weight = layers_7_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_627_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_9997_cast_fp16 = add(x = dense_output_627_cast_fp16, y = sparse_output_627_cast_fp16)[name = string("op_9997_cast_fp16")];
            tensor<int32, [4]> var_9998 = const()[name = string("op_9998"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10000 = const()[name = string("op_10000"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_9999_cast_fp16 = transpose(perm = var_9998, x = var_9997_cast_fp16)[name = string("transpose_668")];
            tensor<fp16, [1, 51, 128]> k_head_249_cast_fp16 = reshape(shape = var_10000, x = var_9999_cast_fp16)[name = string("k_head_249_cast_fp16")];
            string dense_output_629_pad_type_1 = const()[name = string("dense_output_629_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_629_strides_1 = const()[name = string("dense_output_629_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_629_pad_1 = const()[name = string("dense_output_629_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_629_dilations_1 = const()[name = string("dense_output_629_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_629_groups_1 = const()[name = string("dense_output_629_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223307904))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223439040))))[name = string("layers_7_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_629_cast_fp16 = conv(dilations = dense_output_629_dilations_1, groups = dense_output_629_groups_1, pad = dense_output_629_pad_1, pad_type = dense_output_629_pad_type_1, strides = dense_output_629_strides_1, weight = layers_7_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_629_cast_fp16")];
            string sparse_output_629_pad_type_1 = const()[name = string("sparse_output_629_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_629_strides_1 = const()[name = string("sparse_output_629_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_629_pad_1 = const()[name = string("sparse_output_629_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_629_dilations_1 = const()[name = string("sparse_output_629_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_629_groups_1 = const()[name = string("sparse_output_629_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223442304))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223439616))))[name = string("layers_7_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_629_cast_fp16 = conv(dilations = sparse_output_629_dilations_1, groups = sparse_output_629_groups_1, pad = sparse_output_629_pad_1, pad_type = sparse_output_629_pad_type_1, strides = sparse_output_629_strides_1, weight = layers_7_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_629_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10016_cast_fp16 = add(x = dense_output_629_cast_fp16, y = sparse_output_629_cast_fp16)[name = string("op_10016_cast_fp16")];
            tensor<int32, [4]> var_10017 = const()[name = string("op_10017"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10019 = const()[name = string("op_10019"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10018_cast_fp16 = transpose(perm = var_10017, x = var_10016_cast_fp16)[name = string("transpose_667")];
            tensor<fp16, [1, 51, 128]> v_head_249_cast_fp16 = reshape(shape = var_10019, x = var_10018_cast_fp16)[name = string("v_head_249_cast_fp16")];
            string dense_output_631_pad_type_1 = const()[name = string("dense_output_631_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_631_strides_1 = const()[name = string("dense_output_631_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_631_pad_1 = const()[name = string("dense_output_631_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_631_dilations_1 = const()[name = string("dense_output_631_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_631_groups_1 = const()[name = string("dense_output_631_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223458752))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223589888))))[name = string("layers_7_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_631_cast_fp16 = conv(dilations = dense_output_631_dilations_1, groups = dense_output_631_groups_1, pad = dense_output_631_pad_1, pad_type = dense_output_631_pad_type_1, strides = dense_output_631_strides_1, weight = layers_7_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_631_cast_fp16")];
            string sparse_output_631_pad_type_1 = const()[name = string("sparse_output_631_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_631_strides_1 = const()[name = string("sparse_output_631_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_631_pad_1 = const()[name = string("sparse_output_631_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_631_dilations_1 = const()[name = string("sparse_output_631_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_631_groups_1 = const()[name = string("sparse_output_631_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223593152))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223590464))))[name = string("layers_7_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_631_cast_fp16 = conv(dilations = sparse_output_631_dilations_1, groups = sparse_output_631_groups_1, pad = sparse_output_631_pad_1, pad_type = sparse_output_631_pad_type_1, strides = sparse_output_631_strides_1, weight = layers_7_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_631_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_10035_cast_fp16 = add(x = dense_output_631_cast_fp16, y = sparse_output_631_cast_fp16)[name = string("op_10035_cast_fp16")];
            tensor<int32, [4]> var_10036 = const()[name = string("op_10036"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10038 = const()[name = string("op_10038"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_10037_cast_fp16 = transpose(perm = var_10036, x = var_10035_cast_fp16)[name = string("transpose_666")];
            tensor<fp16, [1, 101, 128]> p_head_249_cast_fp16 = reshape(shape = var_10038, x = var_10037_cast_fp16)[name = string("p_head_249_cast_fp16")];
            tensor<fp16, [128]> var_10040_to_fp16 = const()[name = string("op_10040_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223609600)))];
            tensor<fp16, [1, 51, 128]> var_10041_cast_fp16 = add(x = q_head_125_cast_fp16, y = var_10040_to_fp16)[name = string("op_10041_cast_fp16")];
            tensor<int32, [1]> q_u_125_axes_1 = const()[name = string("q_u_125_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_125_cast_fp16 = expand_dims(axes = q_u_125_axes_1, x = var_10041_cast_fp16)[name = string("q_u_125_cast_fp16")];
            tensor<fp16, [128]> var_10043_to_fp16 = const()[name = string("op_10043_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223609920)))];
            tensor<fp16, [1, 51, 128]> var_10044_cast_fp16 = add(x = q_head_125_cast_fp16, y = var_10043_to_fp16)[name = string("op_10044_cast_fp16")];
            tensor<int32, [1]> q_v_125_axes_1 = const()[name = string("q_v_125_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_125_cast_fp16 = expand_dims(axes = q_v_125_axes_1, x = var_10044_cast_fp16)[name = string("q_v_125_cast_fp16")];
            tensor<int32, [1]> k_head_251_axes_1 = const()[name = string("k_head_251_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_251_cast_fp16 = expand_dims(axes = k_head_251_axes_1, x = k_head_249_cast_fp16)[name = string("k_head_251_cast_fp16")];
            tensor<int32, [1]> v_head_251_axes_1 = const()[name = string("v_head_251_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_251_cast_fp16 = expand_dims(axes = v_head_251_axes_1, x = v_head_249_cast_fp16)[name = string("v_head_251_cast_fp16")];
            tensor<int32, [1]> p_head_251_axes_1 = const()[name = string("p_head_251_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_251_cast_fp16 = expand_dims(axes = p_head_251_axes_1, x = p_head_249_cast_fp16)[name = string("p_head_251_cast_fp16")];
            bool var_10050_transpose_x_3 = const()[name = string("op_10050_transpose_x_3"), val = bool(false)];
            bool var_10050_transpose_y_3 = const()[name = string("op_10050_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_10050_cast_fp16 = matmul(transpose_x = var_10050_transpose_x_3, transpose_y = var_10050_transpose_y_3, x = q_u_125_cast_fp16, y = k_head_251_cast_fp16)[name = string("op_10050_cast_fp16")];
            fp16 var_10051_to_fp16 = const()[name = string("op_10051_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_125_cast_fp16 = mul(x = var_10050_cast_fp16, y = var_10051_to_fp16)[name = string("scores_content_125_cast_fp16")];
            bool x_657_transpose_x_3 = const()[name = string("x_657_transpose_x_3"), val = bool(false)];
            bool x_657_transpose_y_3 = const()[name = string("x_657_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_657_cast_fp16 = matmul(transpose_x = x_657_transpose_x_3, transpose_y = x_657_transpose_y_3, x = q_v_125_cast_fp16, y = p_head_251_cast_fp16)[name = string("x_657_cast_fp16")];
            tensor<int32, [8]> x_659_pad_1 = const()[name = string("x_659_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_659_mode_1 = const()[name = string("x_659_mode_1"), val = string("constant")];
            fp16 const_1723_to_fp16 = const()[name = string("const_1723_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_659_cast_fp16 = pad(constant_val = const_1723_to_fp16, mode = x_659_mode_1, pad = x_659_pad_1, x = x_657_cast_fp16)[name = string("x_659_cast_fp16")];
            tensor<int32, [4]> var_10065 = const()[name = string("op_10065"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_661_cast_fp16 = reshape(shape = var_10065, x = x_659_cast_fp16)[name = string("x_661_cast_fp16")];
            tensor<int32, [4]> var_10069_begin_1 = const()[name = string("op_10069_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10069_end_1 = const()[name = string("op_10069_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_10069_end_mask_1 = const()[name = string("op_10069_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_10069_cast_fp16 = slice_by_index(begin = var_10069_begin_1, end = var_10069_end_1, end_mask = var_10069_end_mask_1, x = x_661_cast_fp16)[name = string("op_10069_cast_fp16")];
            tensor<int32, [4]> var_10071 = const()[name = string("op_10071"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_10072_cast_fp16 = reshape(shape = var_10071, x = var_10069_cast_fp16)[name = string("op_10072_cast_fp16")];
            tensor<int32, [4]> var_10077_begin_1 = const()[name = string("op_10077_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10077_end_1 = const()[name = string("op_10077_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_10077_end_mask_1 = const()[name = string("op_10077_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_10077_cast_fp16 = slice_by_index(begin = var_10077_begin_1, end = var_10077_end_1, end_mask = var_10077_end_mask_1, x = var_10072_cast_fp16)[name = string("op_10077_cast_fp16")];
            fp16 var_10078_to_fp16 = const()[name = string("op_10078_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_125_cast_fp16 = mul(x = var_10077_cast_fp16, y = var_10078_to_fp16)[name = string("scores_pos_125_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_125_cast_fp16 = add(x = scores_content_125_cast_fp16, y = scores_pos_125_cast_fp16)[name = string("logits_125_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_10081_cast_fp16 = softmax(axis = var_9080, x = logits_125_cast_fp16)[name = string("op_10081_cast_fp16")];
            bool var_10083_transpose_x_1 = const()[name = string("op_10083_transpose_x_1"), val = bool(false)];
            bool var_10083_transpose_y_1 = const()[name = string("op_10083_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_10083_cast_fp16 = matmul(transpose_x = var_10083_transpose_x_1, transpose_y = var_10083_transpose_y_1, x = var_10081_cast_fp16, y = v_head_251_cast_fp16)[name = string("op_10083_cast_fp16")];
            tensor<int32, [1]> var_10084_axes_1 = const()[name = string("op_10084_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_10084_cast_fp16 = squeeze(axes = var_10084_axes_1, x = var_10083_cast_fp16)[name = string("op_10084_cast_fp16")];
            string dense_output_633_pad_type_1 = const()[name = string("dense_output_633_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_633_strides_1 = const()[name = string("dense_output_633_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_633_pad_1 = const()[name = string("dense_output_633_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_633_dilations_1 = const()[name = string("dense_output_633_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_633_groups_1 = const()[name = string("dense_output_633_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223610240))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223741376))))[name = string("layers_7_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_633_cast_fp16 = conv(dilations = dense_output_633_dilations_1, groups = dense_output_633_groups_1, pad = dense_output_633_pad_1, pad_type = dense_output_633_pad_type_1, strides = dense_output_633_strides_1, weight = layers_7_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_633_cast_fp16")];
            string sparse_output_633_pad_type_1 = const()[name = string("sparse_output_633_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_633_strides_1 = const()[name = string("sparse_output_633_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_633_pad_1 = const()[name = string("sparse_output_633_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_633_dilations_1 = const()[name = string("sparse_output_633_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_633_groups_1 = const()[name = string("sparse_output_633_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223744640))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223741952))))[name = string("layers_7_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_633_cast_fp16 = conv(dilations = sparse_output_633_dilations_1, groups = sparse_output_633_groups_1, pad = sparse_output_633_pad_1, pad_type = sparse_output_633_pad_type_1, strides = sparse_output_633_strides_1, weight = layers_7_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_633_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10099_cast_fp16 = add(x = dense_output_633_cast_fp16, y = sparse_output_633_cast_fp16)[name = string("op_10099_cast_fp16")];
            tensor<int32, [4]> var_10100 = const()[name = string("op_10100"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10102 = const()[name = string("op_10102"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10101_cast_fp16 = transpose(perm = var_10100, x = var_10099_cast_fp16)[name = string("transpose_665")];
            tensor<fp16, [1, 51, 128]> q_head_127_cast_fp16 = reshape(shape = var_10102, x = var_10101_cast_fp16)[name = string("q_head_127_cast_fp16")];
            string dense_output_635_pad_type_1 = const()[name = string("dense_output_635_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_635_strides_1 = const()[name = string("dense_output_635_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_635_pad_1 = const()[name = string("dense_output_635_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_635_dilations_1 = const()[name = string("dense_output_635_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_635_groups_1 = const()[name = string("dense_output_635_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223761088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223892224))))[name = string("layers_7_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_635_cast_fp16 = conv(dilations = dense_output_635_dilations_1, groups = dense_output_635_groups_1, pad = dense_output_635_pad_1, pad_type = dense_output_635_pad_type_1, strides = dense_output_635_strides_1, weight = layers_7_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_635_cast_fp16")];
            string sparse_output_635_pad_type_1 = const()[name = string("sparse_output_635_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_635_strides_1 = const()[name = string("sparse_output_635_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_635_pad_1 = const()[name = string("sparse_output_635_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_635_dilations_1 = const()[name = string("sparse_output_635_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_635_groups_1 = const()[name = string("sparse_output_635_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223895488))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223892800))))[name = string("layers_7_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_635_cast_fp16 = conv(dilations = sparse_output_635_dilations_1, groups = sparse_output_635_groups_1, pad = sparse_output_635_pad_1, pad_type = sparse_output_635_pad_type_1, strides = sparse_output_635_strides_1, weight = layers_7_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_635_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10118_cast_fp16 = add(x = dense_output_635_cast_fp16, y = sparse_output_635_cast_fp16)[name = string("op_10118_cast_fp16")];
            tensor<int32, [4]> var_10119 = const()[name = string("op_10119"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10121 = const()[name = string("op_10121"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10120_cast_fp16 = transpose(perm = var_10119, x = var_10118_cast_fp16)[name = string("transpose_664")];
            tensor<fp16, [1, 51, 128]> k_head_253_cast_fp16 = reshape(shape = var_10121, x = var_10120_cast_fp16)[name = string("k_head_253_cast_fp16")];
            string dense_output_637_pad_type_1 = const()[name = string("dense_output_637_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_637_strides_1 = const()[name = string("dense_output_637_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_637_pad_1 = const()[name = string("dense_output_637_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_637_dilations_1 = const()[name = string("dense_output_637_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_637_groups_1 = const()[name = string("dense_output_637_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223911936))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224043072))))[name = string("layers_7_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_637_cast_fp16 = conv(dilations = dense_output_637_dilations_1, groups = dense_output_637_groups_1, pad = dense_output_637_pad_1, pad_type = dense_output_637_pad_type_1, strides = dense_output_637_strides_1, weight = layers_7_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("dense_output_637_cast_fp16")];
            string sparse_output_637_pad_type_1 = const()[name = string("sparse_output_637_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_637_strides_1 = const()[name = string("sparse_output_637_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_637_pad_1 = const()[name = string("sparse_output_637_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_637_dilations_1 = const()[name = string("sparse_output_637_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_637_groups_1 = const()[name = string("sparse_output_637_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224046336))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224043648))))[name = string("layers_7_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_637_cast_fp16 = conv(dilations = sparse_output_637_dilations_1, groups = sparse_output_637_groups_1, pad = sparse_output_637_pad_1, pad_type = sparse_output_637_pad_type_1, strides = sparse_output_637_strides_1, weight = layers_7_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_351_cast_fp16)[name = string("sparse_output_637_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10137_cast_fp16 = add(x = dense_output_637_cast_fp16, y = sparse_output_637_cast_fp16)[name = string("op_10137_cast_fp16")];
            tensor<int32, [4]> var_10138 = const()[name = string("op_10138"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10140 = const()[name = string("op_10140"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10139_cast_fp16 = transpose(perm = var_10138, x = var_10137_cast_fp16)[name = string("transpose_663")];
            tensor<fp16, [1, 51, 128]> v_head_253_cast_fp16 = reshape(shape = var_10140, x = var_10139_cast_fp16)[name = string("v_head_253_cast_fp16")];
            string dense_output_639_pad_type_1 = const()[name = string("dense_output_639_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_639_strides_1 = const()[name = string("dense_output_639_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_639_pad_1 = const()[name = string("dense_output_639_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_639_dilations_1 = const()[name = string("dense_output_639_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_639_groups_1 = const()[name = string("dense_output_639_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224062784))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224193920))))[name = string("layers_7_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_639_cast_fp16 = conv(dilations = dense_output_639_dilations_1, groups = dense_output_639_groups_1, pad = dense_output_639_pad_1, pad_type = dense_output_639_pad_type_1, strides = dense_output_639_strides_1, weight = layers_7_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_639_cast_fp16")];
            string sparse_output_639_pad_type_1 = const()[name = string("sparse_output_639_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_639_strides_1 = const()[name = string("sparse_output_639_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_639_pad_1 = const()[name = string("sparse_output_639_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_639_dilations_1 = const()[name = string("sparse_output_639_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_639_groups_1 = const()[name = string("sparse_output_639_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_7_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224197184))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224194496))))[name = string("layers_7_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_639_cast_fp16 = conv(dilations = sparse_output_639_dilations_1, groups = sparse_output_639_groups_1, pad = sparse_output_639_pad_1, pad_type = sparse_output_639_pad_type_1, strides = sparse_output_639_strides_1, weight = layers_7_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_639_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_10156_cast_fp16 = add(x = dense_output_639_cast_fp16, y = sparse_output_639_cast_fp16)[name = string("op_10156_cast_fp16")];
            tensor<int32, [4]> var_10157 = const()[name = string("op_10157"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10159 = const()[name = string("op_10159"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_10158_cast_fp16 = transpose(perm = var_10157, x = var_10156_cast_fp16)[name = string("transpose_662")];
            tensor<fp16, [1, 101, 128]> p_head_253_cast_fp16 = reshape(shape = var_10159, x = var_10158_cast_fp16)[name = string("p_head_253_cast_fp16")];
            tensor<fp16, [128]> var_10161_to_fp16 = const()[name = string("op_10161_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224213632)))];
            tensor<fp16, [1, 51, 128]> var_10162_cast_fp16 = add(x = q_head_127_cast_fp16, y = var_10161_to_fp16)[name = string("op_10162_cast_fp16")];
            tensor<int32, [1]> q_u_127_axes_1 = const()[name = string("q_u_127_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_127_cast_fp16 = expand_dims(axes = q_u_127_axes_1, x = var_10162_cast_fp16)[name = string("q_u_127_cast_fp16")];
            tensor<fp16, [128]> var_10164_to_fp16 = const()[name = string("op_10164_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224213952)))];
            tensor<fp16, [1, 51, 128]> var_10165_cast_fp16 = add(x = q_head_127_cast_fp16, y = var_10164_to_fp16)[name = string("op_10165_cast_fp16")];
            tensor<int32, [1]> q_v_127_axes_1 = const()[name = string("q_v_127_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_127_cast_fp16 = expand_dims(axes = q_v_127_axes_1, x = var_10165_cast_fp16)[name = string("q_v_127_cast_fp16")];
            tensor<int32, [1]> k_head_255_axes_1 = const()[name = string("k_head_255_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_255_cast_fp16 = expand_dims(axes = k_head_255_axes_1, x = k_head_253_cast_fp16)[name = string("k_head_255_cast_fp16")];
            tensor<int32, [1]> v_head_255_axes_1 = const()[name = string("v_head_255_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_255_cast_fp16 = expand_dims(axes = v_head_255_axes_1, x = v_head_253_cast_fp16)[name = string("v_head_255_cast_fp16")];
            tensor<int32, [1]> p_head_255_axes_1 = const()[name = string("p_head_255_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_255_cast_fp16 = expand_dims(axes = p_head_255_axes_1, x = p_head_253_cast_fp16)[name = string("p_head_255_cast_fp16")];
            bool var_10171_transpose_x_3 = const()[name = string("op_10171_transpose_x_3"), val = bool(false)];
            bool var_10171_transpose_y_3 = const()[name = string("op_10171_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_10171_cast_fp16 = matmul(transpose_x = var_10171_transpose_x_3, transpose_y = var_10171_transpose_y_3, x = q_u_127_cast_fp16, y = k_head_255_cast_fp16)[name = string("op_10171_cast_fp16")];
            fp16 var_10172_to_fp16 = const()[name = string("op_10172_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_127_cast_fp16 = mul(x = var_10171_cast_fp16, y = var_10172_to_fp16)[name = string("scores_content_127_cast_fp16")];
            bool x_665_transpose_x_3 = const()[name = string("x_665_transpose_x_3"), val = bool(false)];
            bool x_665_transpose_y_3 = const()[name = string("x_665_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_665_cast_fp16 = matmul(transpose_x = x_665_transpose_x_3, transpose_y = x_665_transpose_y_3, x = q_v_127_cast_fp16, y = p_head_255_cast_fp16)[name = string("x_665_cast_fp16")];
            tensor<int32, [8]> x_667_pad_1 = const()[name = string("x_667_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_667_mode_1 = const()[name = string("x_667_mode_1"), val = string("constant")];
            fp16 const_1729_to_fp16 = const()[name = string("const_1729_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_667_cast_fp16 = pad(constant_val = const_1729_to_fp16, mode = x_667_mode_1, pad = x_667_pad_1, x = x_665_cast_fp16)[name = string("x_667_cast_fp16")];
            tensor<int32, [4]> var_10186 = const()[name = string("op_10186"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_669_cast_fp16 = reshape(shape = var_10186, x = x_667_cast_fp16)[name = string("x_669_cast_fp16")];
            tensor<int32, [4]> var_10190_begin_1 = const()[name = string("op_10190_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10190_end_1 = const()[name = string("op_10190_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_10190_end_mask_1 = const()[name = string("op_10190_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_10190_cast_fp16 = slice_by_index(begin = var_10190_begin_1, end = var_10190_end_1, end_mask = var_10190_end_mask_1, x = x_669_cast_fp16)[name = string("op_10190_cast_fp16")];
            tensor<int32, [4]> var_10192 = const()[name = string("op_10192"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_10193_cast_fp16 = reshape(shape = var_10192, x = var_10190_cast_fp16)[name = string("op_10193_cast_fp16")];
            tensor<int32, [4]> var_10198_begin_1 = const()[name = string("op_10198_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10198_end_1 = const()[name = string("op_10198_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_10198_end_mask_1 = const()[name = string("op_10198_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_10198_cast_fp16 = slice_by_index(begin = var_10198_begin_1, end = var_10198_end_1, end_mask = var_10198_end_mask_1, x = var_10193_cast_fp16)[name = string("op_10198_cast_fp16")];
            fp16 var_10199_to_fp16 = const()[name = string("op_10199_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_127_cast_fp16 = mul(x = var_10198_cast_fp16, y = var_10199_to_fp16)[name = string("scores_pos_127_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_127_cast_fp16 = add(x = scores_content_127_cast_fp16, y = scores_pos_127_cast_fp16)[name = string("logits_127_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_10202_cast_fp16 = softmax(axis = var_9080, x = logits_127_cast_fp16)[name = string("op_10202_cast_fp16")];
            bool var_10204_transpose_x_1 = const()[name = string("op_10204_transpose_x_1"), val = bool(false)];
            bool var_10204_transpose_y_1 = const()[name = string("op_10204_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_10204_cast_fp16 = matmul(transpose_x = var_10204_transpose_x_1, transpose_y = var_10204_transpose_y_1, x = var_10202_cast_fp16, y = v_head_255_cast_fp16)[name = string("op_10204_cast_fp16")];
            tensor<int32, [1]> o_head_15_axes_1 = const()[name = string("o_head_15_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_15_cast_fp16 = squeeze(axes = o_head_15_axes_1, x = var_10204_cast_fp16)[name = string("o_head_15_cast_fp16")];
            bool out_15_interleave_1 = const()[name = string("out_15_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_15_cast_fp16 = concat(axis = var_9080, interleave = out_15_interleave_1, values = (var_9358_cast_fp16, var_9479_cast_fp16, var_9600_cast_fp16, var_9721_cast_fp16, var_9842_cast_fp16, var_9963_cast_fp16, var_10084_cast_fp16, o_head_15_cast_fp16))[name = string("out_15_cast_fp16")];
            tensor<int32, [3]> var_10208_perm_1 = const()[name = string("op_10208_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_359_axes_1 = const()[name = string("input_359_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_10208_cast_fp16 = transpose(perm = var_10208_perm_1, x = out_15_cast_fp16)[name = string("transpose_661")];
            tensor<fp16, [1, 1024, 51, 1]> input_359_cast_fp16 = expand_dims(axes = input_359_axes_1, x = var_10208_cast_fp16)[name = string("input_359_cast_fp16")];
            string dense_output_641_pad_type_1 = const()[name = string("dense_output_641_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_641_strides_1 = const()[name = string("dense_output_641_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_641_pad_1 = const()[name = string("dense_output_641_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_641_dilations_1 = const()[name = string("dense_output_641_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_641_groups_1 = const()[name = string("dense_output_641_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224214272))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225262912))))[name = string("layers_7_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_641_cast_fp16 = conv(dilations = dense_output_641_dilations_1, groups = dense_output_641_groups_1, pad = dense_output_641_pad_1, pad_type = dense_output_641_pad_type_1, strides = dense_output_641_strides_1, weight = layers_7_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_359_cast_fp16)[name = string("dense_output_641_cast_fp16")];
            string sparse_output_641_pad_type_1 = const()[name = string("sparse_output_641_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_641_strides_1 = const()[name = string("sparse_output_641_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_641_pad_1 = const()[name = string("sparse_output_641_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_641_dilations_1 = const()[name = string("sparse_output_641_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_641_groups_1 = const()[name = string("sparse_output_641_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225284544))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225263488))))[name = string("layers_7_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_641_cast_fp16 = conv(dilations = sparse_output_641_dilations_1, groups = sparse_output_641_groups_1, pad = sparse_output_641_pad_1, pad_type = sparse_output_641_pad_type_1, strides = sparse_output_641_strides_1, weight = layers_7_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_359_cast_fp16)[name = string("sparse_output_641_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_15_cast_fp16 = add(x = dense_output_641_cast_fp16, y = sparse_output_641_cast_fp16)[name = string("out_conv_15_cast_fp16")];
            tensor<int32, [1]> var_10225_axes_1 = const()[name = string("op_10225_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_10225_cast_fp16 = squeeze(axes = var_10225_axes_1, x = out_conv_15_cast_fp16)[name = string("op_10225_cast_fp16")];
            tensor<int32, [3]> var_10226_perm_1 = const()[name = string("op_10226_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_10226_cast_fp16 = transpose(perm = var_10226_perm_1, x = var_10225_cast_fp16)[name = string("transpose_660")];
            tensor<fp16, [1, 51, 1024]> input_361_cast_fp16 = add(x = input_349_cast_fp16, y = var_10226_cast_fp16)[name = string("input_361_cast_fp16")];
            tensor<int32, [1]> x_673_axes_1 = const()[name = string("x_673_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_7_norm_conv_weight_to_fp16 = const()[name = string("layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225415680)))];
            tensor<fp16, [1024]> layers_7_norm_conv_bias_to_fp16 = const()[name = string("layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225417792)))];
            tensor<fp16, [1, 51, 1024]> x_673_cast_fp16 = layer_norm(axes = x_673_axes_1, beta = layers_7_norm_conv_bias_to_fp16, epsilon = var_9095_to_fp16, gamma = layers_7_norm_conv_weight_to_fp16, x = input_361_cast_fp16)[name = string("x_673_cast_fp16")];
            tensor<int32, [3]> var_10236_perm_1 = const()[name = string("op_10236_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_363_axes_1 = const()[name = string("input_363_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_10236_cast_fp16 = transpose(perm = var_10236_perm_1, x = x_673_cast_fp16)[name = string("transpose_659")];
            tensor<fp16, [1, 1024, 51, 1]> input_363_cast_fp16 = expand_dims(axes = input_363_axes_1, x = var_10236_cast_fp16)[name = string("input_363_cast_fp16")];
            string dense_output_643_pad_type_1 = const()[name = string("dense_output_643_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_643_strides_1 = const()[name = string("dense_output_643_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_643_pad_1 = const()[name = string("dense_output_643_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_643_dilations_1 = const()[name = string("dense_output_643_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_643_groups_1 = const()[name = string("dense_output_643_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_7_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225419904))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227517120))))[name = string("layers_7_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_643_cast_fp16 = conv(dilations = dense_output_643_dilations_1, groups = dense_output_643_groups_1, pad = dense_output_643_pad_1, pad_type = dense_output_643_pad_type_1, strides = dense_output_643_strides_1, weight = layers_7_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_363_cast_fp16)[name = string("dense_output_643_cast_fp16")];
            string sparse_output_643_pad_type_1 = const()[name = string("sparse_output_643_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_643_strides_1 = const()[name = string("sparse_output_643_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_643_pad_1 = const()[name = string("sparse_output_643_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_643_dilations_1 = const()[name = string("sparse_output_643_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_643_groups_1 = const()[name = string("sparse_output_643_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_7_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227559744))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227517696))))[name = string("layers_7_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_643_cast_fp16 = conv(dilations = sparse_output_643_dilations_1, groups = sparse_output_643_groups_1, pad = sparse_output_643_pad_1, pad_type = sparse_output_643_pad_type_1, strides = sparse_output_643_strides_1, weight = layers_7_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_363_cast_fp16)[name = string("sparse_output_643_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_365_cast_fp16 = add(x = dense_output_643_cast_fp16, y = sparse_output_643_cast_fp16)[name = string("input_365_cast_fp16")];
            int32 input_367_split_num_splits_1 = const()[name = string("input_367_split_num_splits_1"), val = int32(2)];
            int32 input_367_split_axis_1 = const()[name = string("input_367_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_367_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_367_split_cast_fp16_1 = split(axis = input_367_split_axis_1, num_splits = input_367_split_num_splits_1, x = input_365_cast_fp16)[name = string("input_367_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_367_split_1_sigmoid_cast_fp16 = sigmoid(x = input_367_split_cast_fp16_1)[name = string("input_367_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_367_cast_fp16 = mul(x = input_367_split_cast_fp16_0, y = input_367_split_1_sigmoid_cast_fp16)[name = string("input_367_cast_fp16")];
            tensor<int32, [8]> input_369_pad_1 = const()[name = string("input_369_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_369_mode_1 = const()[name = string("input_369_mode_1"), val = string("constant")];
            fp16 const_1731_to_fp16 = const()[name = string("const_1731_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_369_cast_fp16 = pad(constant_val = const_1731_to_fp16, mode = input_369_mode_1, pad = input_369_pad_1, x = input_367_cast_fp16)[name = string("input_369_cast_fp16")];
            string dense_output_645_pad_type_1 = const()[name = string("dense_output_645_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_645_strides_1 = const()[name = string("dense_output_645_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_645_pad_1 = const()[name = string("dense_output_645_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_645_dilations_1 = const()[name = string("dense_output_645_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_645_groups_1 = const()[name = string("dense_output_645_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_645_cast_fp16 = conv(dilations = dense_output_645_dilations_1, groups = dense_output_645_groups_1, pad = dense_output_645_pad_1, pad_type = dense_output_645_pad_type_1, strides = dense_output_645_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_369_cast_fp16)[name = string("dense_output_645_cast_fp16")];
            string sparse_output_645_pad_type_1 = const()[name = string("sparse_output_645_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_645_strides_1 = const()[name = string("sparse_output_645_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_645_pad_1 = const()[name = string("sparse_output_645_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_645_dilations_1 = const()[name = string("sparse_output_645_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_645_groups_1 = const()[name = string("sparse_output_645_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_7_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227840448))), nonzero_data = tensor<fp16, [9215]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227821952))))[name = string("layers_7_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_645_cast_fp16 = conv(dilations = sparse_output_645_dilations_1, groups = sparse_output_645_groups_1, pad = sparse_output_645_pad_1, pad_type = sparse_output_645_pad_type_1, strides = sparse_output_645_strides_1, weight = layers_7_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_369_cast_fp16)[name = string("sparse_output_645_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_371_cast_fp16 = add(x = dense_output_645_cast_fp16, y = sparse_output_645_cast_fp16)[name = string("input_371_cast_fp16")];
            tensor<fp16, [1024]> layers_7_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_7_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229020160)))];
            tensor<fp16, [1024]> layers_7_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_7_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229022272)))];
            tensor<fp16, [1024]> layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229024384)))];
            tensor<fp16, [1024]> layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229026496)))];
            tensor<fp16, [1, 1024, 51, 1]> input_373_cast_fp16 = batch_norm(beta = layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_9095_to_fp16, gamma = layers_7_conv_batch_norm_weight_to_fp16, mean = layers_7_conv_batch_norm_running_mean_to_fp16, variance = layers_7_conv_batch_norm_running_var_to_fp16, x = input_371_cast_fp16)[name = string("input_373_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_375_cast_fp16 = silu(x = input_373_cast_fp16)[name = string("input_375_cast_fp16")];
            string dense_output_647_pad_type_1 = const()[name = string("dense_output_647_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_647_strides_1 = const()[name = string("dense_output_647_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_647_pad_1 = const()[name = string("dense_output_647_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_647_dilations_1 = const()[name = string("dense_output_647_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_647_groups_1 = const()[name = string("dense_output_647_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(229028608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230077248))))[name = string("layers_7_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_647_cast_fp16 = conv(dilations = dense_output_647_dilations_1, groups = dense_output_647_groups_1, pad = dense_output_647_pad_1, pad_type = dense_output_647_pad_type_1, strides = dense_output_647_strides_1, weight = layers_7_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_375_cast_fp16)[name = string("dense_output_647_cast_fp16")];
            string sparse_output_647_pad_type_1 = const()[name = string("sparse_output_647_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_647_strides_1 = const()[name = string("sparse_output_647_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_647_pad_1 = const()[name = string("sparse_output_647_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_647_dilations_1 = const()[name = string("sparse_output_647_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_647_groups_1 = const()[name = string("sparse_output_647_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_7_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230098880))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230077824))))[name = string("layers_7_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_647_cast_fp16 = conv(dilations = sparse_output_647_dilations_1, groups = sparse_output_647_groups_1, pad = sparse_output_647_pad_1, pad_type = sparse_output_647_pad_type_1, strides = sparse_output_647_strides_1, weight = layers_7_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_375_cast_fp16)[name = string("sparse_output_647_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_675_cast_fp16 = add(x = dense_output_647_cast_fp16, y = sparse_output_647_cast_fp16)[name = string("x_675_cast_fp16")];
            tensor<int32, [1]> var_10292_axes_1 = const()[name = string("op_10292_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_10292_cast_fp16 = squeeze(axes = var_10292_axes_1, x = x_675_cast_fp16)[name = string("op_10292_cast_fp16")];
            tensor<int32, [3]> var_10293_perm_1 = const()[name = string("op_10293_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_10293_cast_fp16 = transpose(perm = var_10293_perm_1, x = var_10292_cast_fp16)[name = string("transpose_658")];
            tensor<fp16, [1, 51, 1024]> input_377_cast_fp16 = add(x = input_361_cast_fp16, y = var_10293_cast_fp16)[name = string("input_377_cast_fp16")];
            tensor<int32, [1]> x_677_axes_1 = const()[name = string("x_677_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230230016)))];
            tensor<fp16, [1024]> layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230232128)))];
            tensor<fp16, [1, 51, 1024]> x_677_cast_fp16 = layer_norm(axes = x_677_axes_1, beta = layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_9095_to_fp16, gamma = layers_7_norm_feed_forward2_weight_to_fp16, x = input_377_cast_fp16)[name = string("x_677_cast_fp16")];
            tensor<int32, [4]> var_10303 = const()[name = string("op_10303"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_679_cast_fp16 = reshape(shape = var_10303, x = x_677_cast_fp16)[name = string("x_679_cast_fp16")];
            tensor<int32, [4]> input_379_perm_1 = const()[name = string("input_379_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_649_pad_type_1 = const()[name = string("dense_output_649_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_649_strides_1 = const()[name = string("dense_output_649_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_649_pad_1 = const()[name = string("dense_output_649_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_649_dilations_1 = const()[name = string("dense_output_649_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_649_groups_1 = const()[name = string("dense_output_649_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_7_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230234240))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234428608))))[name = string("layers_7_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_379_cast_fp16 = transpose(perm = input_379_perm_1, x = x_679_cast_fp16)[name = string("transpose_657")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_649_cast_fp16 = conv(dilations = dense_output_649_dilations_1, groups = dense_output_649_groups_1, pad = dense_output_649_pad_1, pad_type = dense_output_649_pad_type_1, strides = dense_output_649_strides_1, weight = layers_7_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_379_cast_fp16)[name = string("dense_output_649_cast_fp16")];
            string sparse_output_649_pad_type_1 = const()[name = string("sparse_output_649_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_649_strides_1 = const()[name = string("sparse_output_649_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_649_pad_1 = const()[name = string("sparse_output_649_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_649_dilations_1 = const()[name = string("sparse_output_649_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_649_groups_1 = const()[name = string("sparse_output_649_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_7_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234513152))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234429184))))[name = string("layers_7_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_649_cast_fp16 = conv(dilations = sparse_output_649_dilations_1, groups = sparse_output_649_groups_1, pad = sparse_output_649_pad_1, pad_type = sparse_output_649_pad_type_1, strides = sparse_output_649_strides_1, weight = layers_7_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_379_cast_fp16)[name = string("sparse_output_649_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_381_cast_fp16 = add(x = dense_output_649_cast_fp16, y = sparse_output_649_cast_fp16)[name = string("input_381_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_383_cast_fp16 = silu(x = input_381_cast_fp16)[name = string("input_383_cast_fp16")];
            string dense_output_651_pad_type_1 = const()[name = string("dense_output_651_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_651_strides_1 = const()[name = string("dense_output_651_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_651_pad_1 = const()[name = string("dense_output_651_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_651_dilations_1 = const()[name = string("dense_output_651_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_651_groups_1 = const()[name = string("dense_output_651_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_7_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235037504))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239231872))))[name = string("layers_7_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_651_cast_fp16 = conv(dilations = dense_output_651_dilations_1, groups = dense_output_651_groups_1, pad = dense_output_651_pad_1, pad_type = dense_output_651_pad_type_1, strides = dense_output_651_strides_1, weight = layers_7_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_383_cast_fp16)[name = string("dense_output_651_cast_fp16")];
            string sparse_output_651_pad_type_1 = const()[name = string("sparse_output_651_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_651_strides_1 = const()[name = string("sparse_output_651_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_651_pad_1 = const()[name = string("sparse_output_651_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_651_dilations_1 = const()[name = string("sparse_output_651_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_651_groups_1 = const()[name = string("sparse_output_651_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_7_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239316416))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239232448))))[name = string("layers_7_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_651_cast_fp16 = conv(dilations = sparse_output_651_dilations_1, groups = sparse_output_651_groups_1, pad = sparse_output_651_pad_1, pad_type = sparse_output_651_pad_type_1, strides = sparse_output_651_strides_1, weight = layers_7_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_383_cast_fp16)[name = string("sparse_output_651_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_681_cast_fp16 = add(x = dense_output_651_cast_fp16, y = sparse_output_651_cast_fp16)[name = string("x_681_cast_fp16")];
            tensor<int32, [4]> x_683_perm_1 = const()[name = string("x_683_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_10338 = const()[name = string("op_10338"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_683_cast_fp16 = transpose(perm = x_683_perm_1, x = x_681_cast_fp16)[name = string("transpose_656")];
            tensor<fp16, [1, 51, 1024]> var_10339_cast_fp16 = reshape(shape = var_10338, x = x_683_cast_fp16)[name = string("op_10339_cast_fp16")];
            fp16 var_10340_to_fp16 = const()[name = string("op_10340_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_10341_cast_fp16 = mul(x = var_10339_cast_fp16, y = var_10340_to_fp16)[name = string("op_10341_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_385_cast_fp16 = add(x = input_377_cast_fp16, y = var_10341_cast_fp16)[name = string("input_385_cast_fp16")];
            tensor<int32, [1]> input_387_axes_1 = const()[name = string("input_387_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_7_norm_out_weight_to_fp16 = const()[name = string("layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239840768)))];
            tensor<fp16, [1024]> layers_7_norm_out_bias_to_fp16 = const()[name = string("layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239842880)))];
            tensor<fp16, [1, 51, 1024]> input_387_cast_fp16 = layer_norm(axes = input_387_axes_1, beta = layers_7_norm_out_bias_to_fp16, epsilon = var_9095_to_fp16, gamma = layers_7_norm_out_weight_to_fp16, x = input_385_cast_fp16)[name = string("input_387_cast_fp16")];
            int32 var_10349 = const()[name = string("op_10349"), val = int32(-1)];
            tensor<int32, [1]> x_685_axes_1 = const()[name = string("x_685_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239844992)))];
            tensor<fp16, [1024]> layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239847104)))];
            fp16 var_10364_to_fp16 = const()[name = string("op_10364_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_685_cast_fp16 = layer_norm(axes = x_685_axes_1, beta = layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_10364_to_fp16, gamma = layers_8_norm_feed_forward1_weight_to_fp16, x = input_387_cast_fp16)[name = string("x_685_cast_fp16")];
            tensor<int32, [4]> var_10383 = const()[name = string("op_10383"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_687_cast_fp16 = reshape(shape = var_10383, x = x_685_cast_fp16)[name = string("x_687_cast_fp16")];
            tensor<int32, [4]> input_389_perm_1 = const()[name = string("input_389_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_653_pad_type_1 = const()[name = string("dense_output_653_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_653_strides_1 = const()[name = string("dense_output_653_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_653_pad_1 = const()[name = string("dense_output_653_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_653_dilations_1 = const()[name = string("dense_output_653_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_653_groups_1 = const()[name = string("dense_output_653_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_8_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239849216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244043584))))[name = string("layers_8_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_389_cast_fp16 = transpose(perm = input_389_perm_1, x = x_687_cast_fp16)[name = string("transpose_655")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_653_cast_fp16 = conv(dilations = dense_output_653_dilations_1, groups = dense_output_653_groups_1, pad = dense_output_653_pad_1, pad_type = dense_output_653_pad_type_1, strides = dense_output_653_strides_1, weight = layers_8_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_389_cast_fp16)[name = string("dense_output_653_cast_fp16")];
            string sparse_output_653_pad_type_1 = const()[name = string("sparse_output_653_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_653_strides_1 = const()[name = string("sparse_output_653_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_653_pad_1 = const()[name = string("sparse_output_653_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_653_dilations_1 = const()[name = string("sparse_output_653_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_653_groups_1 = const()[name = string("sparse_output_653_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_8_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244128128))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244044160))))[name = string("layers_8_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_653_cast_fp16 = conv(dilations = sparse_output_653_dilations_1, groups = sparse_output_653_groups_1, pad = sparse_output_653_pad_1, pad_type = sparse_output_653_pad_type_1, strides = sparse_output_653_strides_1, weight = layers_8_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_389_cast_fp16)[name = string("sparse_output_653_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_391_cast_fp16 = add(x = dense_output_653_cast_fp16, y = sparse_output_653_cast_fp16)[name = string("input_391_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_393_cast_fp16 = silu(x = input_391_cast_fp16)[name = string("input_393_cast_fp16")];
            string dense_output_655_pad_type_1 = const()[name = string("dense_output_655_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_655_strides_1 = const()[name = string("dense_output_655_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_655_pad_1 = const()[name = string("dense_output_655_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_655_dilations_1 = const()[name = string("dense_output_655_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_655_groups_1 = const()[name = string("dense_output_655_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_8_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244652480))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248846848))))[name = string("layers_8_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_655_cast_fp16 = conv(dilations = dense_output_655_dilations_1, groups = dense_output_655_groups_1, pad = dense_output_655_pad_1, pad_type = dense_output_655_pad_type_1, strides = dense_output_655_strides_1, weight = layers_8_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_393_cast_fp16)[name = string("dense_output_655_cast_fp16")];
            string sparse_output_655_pad_type_1 = const()[name = string("sparse_output_655_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_655_strides_1 = const()[name = string("sparse_output_655_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_655_pad_1 = const()[name = string("sparse_output_655_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_655_dilations_1 = const()[name = string("sparse_output_655_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_655_groups_1 = const()[name = string("sparse_output_655_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_8_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248931392))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248847424))))[name = string("layers_8_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_655_cast_fp16 = conv(dilations = sparse_output_655_dilations_1, groups = sparse_output_655_groups_1, pad = sparse_output_655_pad_1, pad_type = sparse_output_655_pad_type_1, strides = sparse_output_655_strides_1, weight = layers_8_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_393_cast_fp16)[name = string("sparse_output_655_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_689_cast_fp16 = add(x = dense_output_655_cast_fp16, y = sparse_output_655_cast_fp16)[name = string("x_689_cast_fp16")];
            tensor<int32, [4]> x_691_perm_1 = const()[name = string("x_691_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_10418 = const()[name = string("op_10418"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_691_cast_fp16 = transpose(perm = x_691_perm_1, x = x_689_cast_fp16)[name = string("transpose_654")];
            tensor<fp16, [1, 51, 1024]> var_10419_cast_fp16 = reshape(shape = var_10418, x = x_691_cast_fp16)[name = string("op_10419_cast_fp16")];
            fp16 var_10420_to_fp16 = const()[name = string("op_10420_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_10421_cast_fp16 = mul(x = var_10419_cast_fp16, y = var_10420_to_fp16)[name = string("op_10421_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_395_cast_fp16 = add(x = input_387_cast_fp16, y = var_10421_cast_fp16)[name = string("input_395_cast_fp16")];
            tensor<int32, [1]> q_17_axes_1 = const()[name = string("q_17_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_8_norm_self_att_weight_to_fp16 = const()[name = string("layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249455744)))];
            tensor<fp16, [1024]> layers_8_norm_self_att_bias_to_fp16 = const()[name = string("layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249457856)))];
            tensor<fp16, [1, 51, 1024]> q_17_cast_fp16 = layer_norm(axes = q_17_axes_1, beta = layers_8_norm_self_att_bias_to_fp16, epsilon = var_10364_to_fp16, gamma = layers_8_norm_self_att_weight_to_fp16, x = input_395_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<int32, [3]> var_10495 = const()[name = string("op_10495"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_397_axes_1 = const()[name = string("input_397_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_10496_cast_fp16 = transpose(perm = var_10495, x = q_17_cast_fp16)[name = string("transpose_653")];
            tensor<fp16, [1, 1024, 51, 1]> input_397_cast_fp16 = expand_dims(axes = input_397_axes_1, x = var_10496_cast_fp16)[name = string("input_397_cast_fp16")];
            string dense_output_657_pad_type_1 = const()[name = string("dense_output_657_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_657_strides_1 = const()[name = string("dense_output_657_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_657_pad_1 = const()[name = string("dense_output_657_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_657_dilations_1 = const()[name = string("dense_output_657_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_657_groups_1 = const()[name = string("dense_output_657_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249459968))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249591104))))[name = string("layers_8_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_657_cast_fp16 = conv(dilations = dense_output_657_dilations_1, groups = dense_output_657_groups_1, pad = dense_output_657_pad_1, pad_type = dense_output_657_pad_type_1, strides = dense_output_657_strides_1, weight = layers_8_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_657_cast_fp16")];
            string sparse_output_657_pad_type_1 = const()[name = string("sparse_output_657_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_657_strides_1 = const()[name = string("sparse_output_657_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_657_pad_1 = const()[name = string("sparse_output_657_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_657_dilations_1 = const()[name = string("sparse_output_657_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_657_groups_1 = const()[name = string("sparse_output_657_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249594368))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249591680))))[name = string("layers_8_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_657_cast_fp16 = conv(dilations = sparse_output_657_dilations_1, groups = sparse_output_657_groups_1, pad = sparse_output_657_pad_1, pad_type = sparse_output_657_pad_type_1, strides = sparse_output_657_strides_1, weight = layers_8_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_657_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10521_cast_fp16 = add(x = dense_output_657_cast_fp16, y = sparse_output_657_cast_fp16)[name = string("op_10521_cast_fp16")];
            tensor<int32, [4]> var_10522 = const()[name = string("op_10522"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10524 = const()[name = string("op_10524"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10523_cast_fp16 = transpose(perm = var_10522, x = var_10521_cast_fp16)[name = string("transpose_652")];
            tensor<fp16, [1, 51, 128]> q_head_129_cast_fp16 = reshape(shape = var_10524, x = var_10523_cast_fp16)[name = string("q_head_129_cast_fp16")];
            string dense_output_659_pad_type_1 = const()[name = string("dense_output_659_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_659_strides_1 = const()[name = string("dense_output_659_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_659_pad_1 = const()[name = string("dense_output_659_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_659_dilations_1 = const()[name = string("dense_output_659_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_659_groups_1 = const()[name = string("dense_output_659_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249610816))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249741952))))[name = string("layers_8_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_659_cast_fp16 = conv(dilations = dense_output_659_dilations_1, groups = dense_output_659_groups_1, pad = dense_output_659_pad_1, pad_type = dense_output_659_pad_type_1, strides = dense_output_659_strides_1, weight = layers_8_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_659_cast_fp16")];
            string sparse_output_659_pad_type_1 = const()[name = string("sparse_output_659_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_659_strides_1 = const()[name = string("sparse_output_659_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_659_pad_1 = const()[name = string("sparse_output_659_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_659_dilations_1 = const()[name = string("sparse_output_659_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_659_groups_1 = const()[name = string("sparse_output_659_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249745216))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249742528))))[name = string("layers_8_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_659_cast_fp16 = conv(dilations = sparse_output_659_dilations_1, groups = sparse_output_659_groups_1, pad = sparse_output_659_pad_1, pad_type = sparse_output_659_pad_type_1, strides = sparse_output_659_strides_1, weight = layers_8_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_659_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10540_cast_fp16 = add(x = dense_output_659_cast_fp16, y = sparse_output_659_cast_fp16)[name = string("op_10540_cast_fp16")];
            tensor<int32, [4]> var_10541 = const()[name = string("op_10541"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10543 = const()[name = string("op_10543"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10542_cast_fp16 = transpose(perm = var_10541, x = var_10540_cast_fp16)[name = string("transpose_651")];
            tensor<fp16, [1, 51, 128]> k_head_257_cast_fp16 = reshape(shape = var_10543, x = var_10542_cast_fp16)[name = string("k_head_257_cast_fp16")];
            string dense_output_661_pad_type_1 = const()[name = string("dense_output_661_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_661_strides_1 = const()[name = string("dense_output_661_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_661_pad_1 = const()[name = string("dense_output_661_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_661_dilations_1 = const()[name = string("dense_output_661_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_661_groups_1 = const()[name = string("dense_output_661_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249761664))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249892800))))[name = string("layers_8_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_661_cast_fp16 = conv(dilations = dense_output_661_dilations_1, groups = dense_output_661_groups_1, pad = dense_output_661_pad_1, pad_type = dense_output_661_pad_type_1, strides = dense_output_661_strides_1, weight = layers_8_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_661_cast_fp16")];
            string sparse_output_661_pad_type_1 = const()[name = string("sparse_output_661_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_661_strides_1 = const()[name = string("sparse_output_661_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_661_pad_1 = const()[name = string("sparse_output_661_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_661_dilations_1 = const()[name = string("sparse_output_661_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_661_groups_1 = const()[name = string("sparse_output_661_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249896064))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249893376))))[name = string("layers_8_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_661_cast_fp16 = conv(dilations = sparse_output_661_dilations_1, groups = sparse_output_661_groups_1, pad = sparse_output_661_pad_1, pad_type = sparse_output_661_pad_type_1, strides = sparse_output_661_strides_1, weight = layers_8_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_661_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10559_cast_fp16 = add(x = dense_output_661_cast_fp16, y = sparse_output_661_cast_fp16)[name = string("op_10559_cast_fp16")];
            tensor<int32, [4]> var_10560 = const()[name = string("op_10560"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10562 = const()[name = string("op_10562"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10561_cast_fp16 = transpose(perm = var_10560, x = var_10559_cast_fp16)[name = string("transpose_650")];
            tensor<fp16, [1, 51, 128]> v_head_257_cast_fp16 = reshape(shape = var_10562, x = var_10561_cast_fp16)[name = string("v_head_257_cast_fp16")];
            string dense_output_663_pad_type_1 = const()[name = string("dense_output_663_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_663_strides_1 = const()[name = string("dense_output_663_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_663_pad_1 = const()[name = string("dense_output_663_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_663_dilations_1 = const()[name = string("dense_output_663_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_663_groups_1 = const()[name = string("dense_output_663_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249912512))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250043648))))[name = string("layers_8_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_663_cast_fp16 = conv(dilations = dense_output_663_dilations_1, groups = dense_output_663_groups_1, pad = dense_output_663_pad_1, pad_type = dense_output_663_pad_type_1, strides = dense_output_663_strides_1, weight = layers_8_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_663_cast_fp16")];
            string sparse_output_663_pad_type_1 = const()[name = string("sparse_output_663_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_663_strides_1 = const()[name = string("sparse_output_663_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_663_pad_1 = const()[name = string("sparse_output_663_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_663_dilations_1 = const()[name = string("sparse_output_663_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_663_groups_1 = const()[name = string("sparse_output_663_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250046912))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250044224))))[name = string("layers_8_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_663_cast_fp16 = conv(dilations = sparse_output_663_dilations_1, groups = sparse_output_663_groups_1, pad = sparse_output_663_pad_1, pad_type = sparse_output_663_pad_type_1, strides = sparse_output_663_strides_1, weight = layers_8_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_663_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_10578_cast_fp16 = add(x = dense_output_663_cast_fp16, y = sparse_output_663_cast_fp16)[name = string("op_10578_cast_fp16")];
            tensor<int32, [4]> var_10579 = const()[name = string("op_10579"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10581 = const()[name = string("op_10581"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_10580_cast_fp16 = transpose(perm = var_10579, x = var_10578_cast_fp16)[name = string("transpose_649")];
            tensor<fp16, [1, 101, 128]> p_head_257_cast_fp16 = reshape(shape = var_10581, x = var_10580_cast_fp16)[name = string("p_head_257_cast_fp16")];
            tensor<fp16, [128]> var_10583_to_fp16 = const()[name = string("op_10583_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250063360)))];
            tensor<fp16, [1, 51, 128]> var_10584_cast_fp16 = add(x = q_head_129_cast_fp16, y = var_10583_to_fp16)[name = string("op_10584_cast_fp16")];
            tensor<int32, [1]> q_u_129_axes_1 = const()[name = string("q_u_129_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_129_cast_fp16 = expand_dims(axes = q_u_129_axes_1, x = var_10584_cast_fp16)[name = string("q_u_129_cast_fp16")];
            tensor<fp16, [128]> var_10586_to_fp16 = const()[name = string("op_10586_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250063680)))];
            tensor<fp16, [1, 51, 128]> var_10587_cast_fp16 = add(x = q_head_129_cast_fp16, y = var_10586_to_fp16)[name = string("op_10587_cast_fp16")];
            tensor<int32, [1]> q_v_129_axes_1 = const()[name = string("q_v_129_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_129_cast_fp16 = expand_dims(axes = q_v_129_axes_1, x = var_10587_cast_fp16)[name = string("q_v_129_cast_fp16")];
            tensor<int32, [1]> k_head_259_axes_1 = const()[name = string("k_head_259_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_259_cast_fp16 = expand_dims(axes = k_head_259_axes_1, x = k_head_257_cast_fp16)[name = string("k_head_259_cast_fp16")];
            tensor<int32, [1]> v_head_259_axes_1 = const()[name = string("v_head_259_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_259_cast_fp16 = expand_dims(axes = v_head_259_axes_1, x = v_head_257_cast_fp16)[name = string("v_head_259_cast_fp16")];
            tensor<int32, [1]> p_head_259_axes_1 = const()[name = string("p_head_259_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_259_cast_fp16 = expand_dims(axes = p_head_259_axes_1, x = p_head_257_cast_fp16)[name = string("p_head_259_cast_fp16")];
            bool var_10593_transpose_x_3 = const()[name = string("op_10593_transpose_x_3"), val = bool(false)];
            bool var_10593_transpose_y_3 = const()[name = string("op_10593_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_10593_cast_fp16 = matmul(transpose_x = var_10593_transpose_x_3, transpose_y = var_10593_transpose_y_3, x = q_u_129_cast_fp16, y = k_head_259_cast_fp16)[name = string("op_10593_cast_fp16")];
            fp16 var_10594_to_fp16 = const()[name = string("op_10594_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_129_cast_fp16 = mul(x = var_10593_cast_fp16, y = var_10594_to_fp16)[name = string("scores_content_129_cast_fp16")];
            bool x_693_transpose_x_3 = const()[name = string("x_693_transpose_x_3"), val = bool(false)];
            bool x_693_transpose_y_3 = const()[name = string("x_693_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_693_cast_fp16 = matmul(transpose_x = x_693_transpose_x_3, transpose_y = x_693_transpose_y_3, x = q_v_129_cast_fp16, y = p_head_259_cast_fp16)[name = string("x_693_cast_fp16")];
            tensor<int32, [8]> x_695_pad_1 = const()[name = string("x_695_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_695_mode_1 = const()[name = string("x_695_mode_1"), val = string("constant")];
            fp16 const_1741_to_fp16 = const()[name = string("const_1741_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_695_cast_fp16 = pad(constant_val = const_1741_to_fp16, mode = x_695_mode_1, pad = x_695_pad_1, x = x_693_cast_fp16)[name = string("x_695_cast_fp16")];
            tensor<int32, [4]> var_10608 = const()[name = string("op_10608"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_697_cast_fp16 = reshape(shape = var_10608, x = x_695_cast_fp16)[name = string("x_697_cast_fp16")];
            tensor<int32, [4]> var_10612_begin_1 = const()[name = string("op_10612_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10612_end_1 = const()[name = string("op_10612_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_10612_end_mask_1 = const()[name = string("op_10612_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_10612_cast_fp16 = slice_by_index(begin = var_10612_begin_1, end = var_10612_end_1, end_mask = var_10612_end_mask_1, x = x_697_cast_fp16)[name = string("op_10612_cast_fp16")];
            tensor<int32, [4]> var_10614 = const()[name = string("op_10614"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_10615_cast_fp16 = reshape(shape = var_10614, x = var_10612_cast_fp16)[name = string("op_10615_cast_fp16")];
            tensor<int32, [4]> var_10620_begin_1 = const()[name = string("op_10620_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10620_end_1 = const()[name = string("op_10620_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_10620_end_mask_1 = const()[name = string("op_10620_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_10620_cast_fp16 = slice_by_index(begin = var_10620_begin_1, end = var_10620_end_1, end_mask = var_10620_end_mask_1, x = var_10615_cast_fp16)[name = string("op_10620_cast_fp16")];
            fp16 var_10621_to_fp16 = const()[name = string("op_10621_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_129_cast_fp16 = mul(x = var_10620_cast_fp16, y = var_10621_to_fp16)[name = string("scores_pos_129_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_129_cast_fp16 = add(x = scores_content_129_cast_fp16, y = scores_pos_129_cast_fp16)[name = string("logits_129_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_10624_cast_fp16 = softmax(axis = var_10349, x = logits_129_cast_fp16)[name = string("op_10624_cast_fp16")];
            bool var_10626_transpose_x_1 = const()[name = string("op_10626_transpose_x_1"), val = bool(false)];
            bool var_10626_transpose_y_1 = const()[name = string("op_10626_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_10626_cast_fp16 = matmul(transpose_x = var_10626_transpose_x_1, transpose_y = var_10626_transpose_y_1, x = var_10624_cast_fp16, y = v_head_259_cast_fp16)[name = string("op_10626_cast_fp16")];
            tensor<int32, [1]> var_10627_axes_1 = const()[name = string("op_10627_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_10627_cast_fp16 = squeeze(axes = var_10627_axes_1, x = var_10626_cast_fp16)[name = string("op_10627_cast_fp16")];
            string dense_output_665_pad_type_1 = const()[name = string("dense_output_665_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_665_strides_1 = const()[name = string("dense_output_665_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_665_pad_1 = const()[name = string("dense_output_665_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_665_dilations_1 = const()[name = string("dense_output_665_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_665_groups_1 = const()[name = string("dense_output_665_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250064000))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250195136))))[name = string("layers_8_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_665_cast_fp16 = conv(dilations = dense_output_665_dilations_1, groups = dense_output_665_groups_1, pad = dense_output_665_pad_1, pad_type = dense_output_665_pad_type_1, strides = dense_output_665_strides_1, weight = layers_8_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_665_cast_fp16")];
            string sparse_output_665_pad_type_1 = const()[name = string("sparse_output_665_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_665_strides_1 = const()[name = string("sparse_output_665_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_665_pad_1 = const()[name = string("sparse_output_665_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_665_dilations_1 = const()[name = string("sparse_output_665_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_665_groups_1 = const()[name = string("sparse_output_665_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250198400))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250195712))))[name = string("layers_8_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_665_cast_fp16 = conv(dilations = sparse_output_665_dilations_1, groups = sparse_output_665_groups_1, pad = sparse_output_665_pad_1, pad_type = sparse_output_665_pad_type_1, strides = sparse_output_665_strides_1, weight = layers_8_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_665_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10642_cast_fp16 = add(x = dense_output_665_cast_fp16, y = sparse_output_665_cast_fp16)[name = string("op_10642_cast_fp16")];
            tensor<int32, [4]> var_10643 = const()[name = string("op_10643"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10645 = const()[name = string("op_10645"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10644_cast_fp16 = transpose(perm = var_10643, x = var_10642_cast_fp16)[name = string("transpose_648")];
            tensor<fp16, [1, 51, 128]> q_head_131_cast_fp16 = reshape(shape = var_10645, x = var_10644_cast_fp16)[name = string("q_head_131_cast_fp16")];
            string dense_output_667_pad_type_1 = const()[name = string("dense_output_667_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_667_strides_1 = const()[name = string("dense_output_667_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_667_pad_1 = const()[name = string("dense_output_667_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_667_dilations_1 = const()[name = string("dense_output_667_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_667_groups_1 = const()[name = string("dense_output_667_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250214848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250345984))))[name = string("layers_8_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_667_cast_fp16 = conv(dilations = dense_output_667_dilations_1, groups = dense_output_667_groups_1, pad = dense_output_667_pad_1, pad_type = dense_output_667_pad_type_1, strides = dense_output_667_strides_1, weight = layers_8_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_667_cast_fp16")];
            string sparse_output_667_pad_type_1 = const()[name = string("sparse_output_667_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_667_strides_1 = const()[name = string("sparse_output_667_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_667_pad_1 = const()[name = string("sparse_output_667_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_667_dilations_1 = const()[name = string("sparse_output_667_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_667_groups_1 = const()[name = string("sparse_output_667_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250349248))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250346560))))[name = string("layers_8_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_667_cast_fp16 = conv(dilations = sparse_output_667_dilations_1, groups = sparse_output_667_groups_1, pad = sparse_output_667_pad_1, pad_type = sparse_output_667_pad_type_1, strides = sparse_output_667_strides_1, weight = layers_8_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_667_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10661_cast_fp16 = add(x = dense_output_667_cast_fp16, y = sparse_output_667_cast_fp16)[name = string("op_10661_cast_fp16")];
            tensor<int32, [4]> var_10662 = const()[name = string("op_10662"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10664 = const()[name = string("op_10664"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10663_cast_fp16 = transpose(perm = var_10662, x = var_10661_cast_fp16)[name = string("transpose_647")];
            tensor<fp16, [1, 51, 128]> k_head_261_cast_fp16 = reshape(shape = var_10664, x = var_10663_cast_fp16)[name = string("k_head_261_cast_fp16")];
            string dense_output_669_pad_type_1 = const()[name = string("dense_output_669_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_669_strides_1 = const()[name = string("dense_output_669_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_669_pad_1 = const()[name = string("dense_output_669_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_669_dilations_1 = const()[name = string("dense_output_669_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_669_groups_1 = const()[name = string("dense_output_669_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250365696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250496832))))[name = string("layers_8_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_669_cast_fp16 = conv(dilations = dense_output_669_dilations_1, groups = dense_output_669_groups_1, pad = dense_output_669_pad_1, pad_type = dense_output_669_pad_type_1, strides = dense_output_669_strides_1, weight = layers_8_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_669_cast_fp16")];
            string sparse_output_669_pad_type_1 = const()[name = string("sparse_output_669_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_669_strides_1 = const()[name = string("sparse_output_669_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_669_pad_1 = const()[name = string("sparse_output_669_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_669_dilations_1 = const()[name = string("sparse_output_669_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_669_groups_1 = const()[name = string("sparse_output_669_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250500096))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250497408))))[name = string("layers_8_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_669_cast_fp16 = conv(dilations = sparse_output_669_dilations_1, groups = sparse_output_669_groups_1, pad = sparse_output_669_pad_1, pad_type = sparse_output_669_pad_type_1, strides = sparse_output_669_strides_1, weight = layers_8_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_669_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10680_cast_fp16 = add(x = dense_output_669_cast_fp16, y = sparse_output_669_cast_fp16)[name = string("op_10680_cast_fp16")];
            tensor<int32, [4]> var_10681 = const()[name = string("op_10681"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10683 = const()[name = string("op_10683"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10682_cast_fp16 = transpose(perm = var_10681, x = var_10680_cast_fp16)[name = string("transpose_646")];
            tensor<fp16, [1, 51, 128]> v_head_261_cast_fp16 = reshape(shape = var_10683, x = var_10682_cast_fp16)[name = string("v_head_261_cast_fp16")];
            string dense_output_671_pad_type_1 = const()[name = string("dense_output_671_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_671_strides_1 = const()[name = string("dense_output_671_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_671_pad_1 = const()[name = string("dense_output_671_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_671_dilations_1 = const()[name = string("dense_output_671_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_671_groups_1 = const()[name = string("dense_output_671_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250516544))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250647680))))[name = string("layers_8_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_671_cast_fp16 = conv(dilations = dense_output_671_dilations_1, groups = dense_output_671_groups_1, pad = dense_output_671_pad_1, pad_type = dense_output_671_pad_type_1, strides = dense_output_671_strides_1, weight = layers_8_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_671_cast_fp16")];
            string sparse_output_671_pad_type_1 = const()[name = string("sparse_output_671_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_671_strides_1 = const()[name = string("sparse_output_671_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_671_pad_1 = const()[name = string("sparse_output_671_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_671_dilations_1 = const()[name = string("sparse_output_671_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_671_groups_1 = const()[name = string("sparse_output_671_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250650944))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250648256))))[name = string("layers_8_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_671_cast_fp16 = conv(dilations = sparse_output_671_dilations_1, groups = sparse_output_671_groups_1, pad = sparse_output_671_pad_1, pad_type = sparse_output_671_pad_type_1, strides = sparse_output_671_strides_1, weight = layers_8_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_671_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_10699_cast_fp16 = add(x = dense_output_671_cast_fp16, y = sparse_output_671_cast_fp16)[name = string("op_10699_cast_fp16")];
            tensor<int32, [4]> var_10700 = const()[name = string("op_10700"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10702 = const()[name = string("op_10702"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_10701_cast_fp16 = transpose(perm = var_10700, x = var_10699_cast_fp16)[name = string("transpose_645")];
            tensor<fp16, [1, 101, 128]> p_head_261_cast_fp16 = reshape(shape = var_10702, x = var_10701_cast_fp16)[name = string("p_head_261_cast_fp16")];
            tensor<fp16, [128]> var_10704_to_fp16 = const()[name = string("op_10704_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250667392)))];
            tensor<fp16, [1, 51, 128]> var_10705_cast_fp16 = add(x = q_head_131_cast_fp16, y = var_10704_to_fp16)[name = string("op_10705_cast_fp16")];
            tensor<int32, [1]> q_u_131_axes_1 = const()[name = string("q_u_131_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_131_cast_fp16 = expand_dims(axes = q_u_131_axes_1, x = var_10705_cast_fp16)[name = string("q_u_131_cast_fp16")];
            tensor<fp16, [128]> var_10707_to_fp16 = const()[name = string("op_10707_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250667712)))];
            tensor<fp16, [1, 51, 128]> var_10708_cast_fp16 = add(x = q_head_131_cast_fp16, y = var_10707_to_fp16)[name = string("op_10708_cast_fp16")];
            tensor<int32, [1]> q_v_131_axes_1 = const()[name = string("q_v_131_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_131_cast_fp16 = expand_dims(axes = q_v_131_axes_1, x = var_10708_cast_fp16)[name = string("q_v_131_cast_fp16")];
            tensor<int32, [1]> k_head_263_axes_1 = const()[name = string("k_head_263_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_263_cast_fp16 = expand_dims(axes = k_head_263_axes_1, x = k_head_261_cast_fp16)[name = string("k_head_263_cast_fp16")];
            tensor<int32, [1]> v_head_263_axes_1 = const()[name = string("v_head_263_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_263_cast_fp16 = expand_dims(axes = v_head_263_axes_1, x = v_head_261_cast_fp16)[name = string("v_head_263_cast_fp16")];
            tensor<int32, [1]> p_head_263_axes_1 = const()[name = string("p_head_263_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_263_cast_fp16 = expand_dims(axes = p_head_263_axes_1, x = p_head_261_cast_fp16)[name = string("p_head_263_cast_fp16")];
            bool var_10714_transpose_x_3 = const()[name = string("op_10714_transpose_x_3"), val = bool(false)];
            bool var_10714_transpose_y_3 = const()[name = string("op_10714_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_10714_cast_fp16 = matmul(transpose_x = var_10714_transpose_x_3, transpose_y = var_10714_transpose_y_3, x = q_u_131_cast_fp16, y = k_head_263_cast_fp16)[name = string("op_10714_cast_fp16")];
            fp16 var_10715_to_fp16 = const()[name = string("op_10715_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_131_cast_fp16 = mul(x = var_10714_cast_fp16, y = var_10715_to_fp16)[name = string("scores_content_131_cast_fp16")];
            bool x_701_transpose_x_3 = const()[name = string("x_701_transpose_x_3"), val = bool(false)];
            bool x_701_transpose_y_3 = const()[name = string("x_701_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_701_cast_fp16 = matmul(transpose_x = x_701_transpose_x_3, transpose_y = x_701_transpose_y_3, x = q_v_131_cast_fp16, y = p_head_263_cast_fp16)[name = string("x_701_cast_fp16")];
            tensor<int32, [8]> x_703_pad_1 = const()[name = string("x_703_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_703_mode_1 = const()[name = string("x_703_mode_1"), val = string("constant")];
            fp16 const_1747_to_fp16 = const()[name = string("const_1747_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_703_cast_fp16 = pad(constant_val = const_1747_to_fp16, mode = x_703_mode_1, pad = x_703_pad_1, x = x_701_cast_fp16)[name = string("x_703_cast_fp16")];
            tensor<int32, [4]> var_10729 = const()[name = string("op_10729"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_705_cast_fp16 = reshape(shape = var_10729, x = x_703_cast_fp16)[name = string("x_705_cast_fp16")];
            tensor<int32, [4]> var_10733_begin_1 = const()[name = string("op_10733_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10733_end_1 = const()[name = string("op_10733_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_10733_end_mask_1 = const()[name = string("op_10733_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_10733_cast_fp16 = slice_by_index(begin = var_10733_begin_1, end = var_10733_end_1, end_mask = var_10733_end_mask_1, x = x_705_cast_fp16)[name = string("op_10733_cast_fp16")];
            tensor<int32, [4]> var_10735 = const()[name = string("op_10735"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_10736_cast_fp16 = reshape(shape = var_10735, x = var_10733_cast_fp16)[name = string("op_10736_cast_fp16")];
            tensor<int32, [4]> var_10741_begin_1 = const()[name = string("op_10741_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10741_end_1 = const()[name = string("op_10741_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_10741_end_mask_1 = const()[name = string("op_10741_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_10741_cast_fp16 = slice_by_index(begin = var_10741_begin_1, end = var_10741_end_1, end_mask = var_10741_end_mask_1, x = var_10736_cast_fp16)[name = string("op_10741_cast_fp16")];
            fp16 var_10742_to_fp16 = const()[name = string("op_10742_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_131_cast_fp16 = mul(x = var_10741_cast_fp16, y = var_10742_to_fp16)[name = string("scores_pos_131_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_131_cast_fp16 = add(x = scores_content_131_cast_fp16, y = scores_pos_131_cast_fp16)[name = string("logits_131_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_10745_cast_fp16 = softmax(axis = var_10349, x = logits_131_cast_fp16)[name = string("op_10745_cast_fp16")];
            bool var_10747_transpose_x_1 = const()[name = string("op_10747_transpose_x_1"), val = bool(false)];
            bool var_10747_transpose_y_1 = const()[name = string("op_10747_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_10747_cast_fp16 = matmul(transpose_x = var_10747_transpose_x_1, transpose_y = var_10747_transpose_y_1, x = var_10745_cast_fp16, y = v_head_263_cast_fp16)[name = string("op_10747_cast_fp16")];
            tensor<int32, [1]> var_10748_axes_1 = const()[name = string("op_10748_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_10748_cast_fp16 = squeeze(axes = var_10748_axes_1, x = var_10747_cast_fp16)[name = string("op_10748_cast_fp16")];
            string dense_output_673_pad_type_1 = const()[name = string("dense_output_673_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_673_strides_1 = const()[name = string("dense_output_673_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_673_pad_1 = const()[name = string("dense_output_673_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_673_dilations_1 = const()[name = string("dense_output_673_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_673_groups_1 = const()[name = string("dense_output_673_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250668032))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250799168))))[name = string("layers_8_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_673_cast_fp16 = conv(dilations = dense_output_673_dilations_1, groups = dense_output_673_groups_1, pad = dense_output_673_pad_1, pad_type = dense_output_673_pad_type_1, strides = dense_output_673_strides_1, weight = layers_8_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_673_cast_fp16")];
            string sparse_output_673_pad_type_1 = const()[name = string("sparse_output_673_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_673_strides_1 = const()[name = string("sparse_output_673_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_673_pad_1 = const()[name = string("sparse_output_673_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_673_dilations_1 = const()[name = string("sparse_output_673_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_673_groups_1 = const()[name = string("sparse_output_673_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250802432))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250799744))))[name = string("layers_8_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_673_cast_fp16 = conv(dilations = sparse_output_673_dilations_1, groups = sparse_output_673_groups_1, pad = sparse_output_673_pad_1, pad_type = sparse_output_673_pad_type_1, strides = sparse_output_673_strides_1, weight = layers_8_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_673_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10763_cast_fp16 = add(x = dense_output_673_cast_fp16, y = sparse_output_673_cast_fp16)[name = string("op_10763_cast_fp16")];
            tensor<int32, [4]> var_10764 = const()[name = string("op_10764"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10766 = const()[name = string("op_10766"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10765_cast_fp16 = transpose(perm = var_10764, x = var_10763_cast_fp16)[name = string("transpose_644")];
            tensor<fp16, [1, 51, 128]> q_head_133_cast_fp16 = reshape(shape = var_10766, x = var_10765_cast_fp16)[name = string("q_head_133_cast_fp16")];
            string dense_output_675_pad_type_1 = const()[name = string("dense_output_675_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_675_strides_1 = const()[name = string("dense_output_675_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_675_pad_1 = const()[name = string("dense_output_675_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_675_dilations_1 = const()[name = string("dense_output_675_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_675_groups_1 = const()[name = string("dense_output_675_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250818880))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250950016))))[name = string("layers_8_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_675_cast_fp16 = conv(dilations = dense_output_675_dilations_1, groups = dense_output_675_groups_1, pad = dense_output_675_pad_1, pad_type = dense_output_675_pad_type_1, strides = dense_output_675_strides_1, weight = layers_8_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_675_cast_fp16")];
            string sparse_output_675_pad_type_1 = const()[name = string("sparse_output_675_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_675_strides_1 = const()[name = string("sparse_output_675_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_675_pad_1 = const()[name = string("sparse_output_675_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_675_dilations_1 = const()[name = string("sparse_output_675_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_675_groups_1 = const()[name = string("sparse_output_675_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250953280))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250950592))))[name = string("layers_8_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_675_cast_fp16 = conv(dilations = sparse_output_675_dilations_1, groups = sparse_output_675_groups_1, pad = sparse_output_675_pad_1, pad_type = sparse_output_675_pad_type_1, strides = sparse_output_675_strides_1, weight = layers_8_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_675_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10782_cast_fp16 = add(x = dense_output_675_cast_fp16, y = sparse_output_675_cast_fp16)[name = string("op_10782_cast_fp16")];
            tensor<int32, [4]> var_10783 = const()[name = string("op_10783"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10785 = const()[name = string("op_10785"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10784_cast_fp16 = transpose(perm = var_10783, x = var_10782_cast_fp16)[name = string("transpose_643")];
            tensor<fp16, [1, 51, 128]> k_head_265_cast_fp16 = reshape(shape = var_10785, x = var_10784_cast_fp16)[name = string("k_head_265_cast_fp16")];
            string dense_output_677_pad_type_1 = const()[name = string("dense_output_677_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_677_strides_1 = const()[name = string("dense_output_677_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_677_pad_1 = const()[name = string("dense_output_677_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_677_dilations_1 = const()[name = string("dense_output_677_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_677_groups_1 = const()[name = string("dense_output_677_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250969728))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251100864))))[name = string("layers_8_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_677_cast_fp16 = conv(dilations = dense_output_677_dilations_1, groups = dense_output_677_groups_1, pad = dense_output_677_pad_1, pad_type = dense_output_677_pad_type_1, strides = dense_output_677_strides_1, weight = layers_8_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_677_cast_fp16")];
            string sparse_output_677_pad_type_1 = const()[name = string("sparse_output_677_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_677_strides_1 = const()[name = string("sparse_output_677_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_677_pad_1 = const()[name = string("sparse_output_677_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_677_dilations_1 = const()[name = string("sparse_output_677_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_677_groups_1 = const()[name = string("sparse_output_677_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251104128))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251101440))))[name = string("layers_8_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_677_cast_fp16 = conv(dilations = sparse_output_677_dilations_1, groups = sparse_output_677_groups_1, pad = sparse_output_677_pad_1, pad_type = sparse_output_677_pad_type_1, strides = sparse_output_677_strides_1, weight = layers_8_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_677_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10801_cast_fp16 = add(x = dense_output_677_cast_fp16, y = sparse_output_677_cast_fp16)[name = string("op_10801_cast_fp16")];
            tensor<int32, [4]> var_10802 = const()[name = string("op_10802"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10804 = const()[name = string("op_10804"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10803_cast_fp16 = transpose(perm = var_10802, x = var_10801_cast_fp16)[name = string("transpose_642")];
            tensor<fp16, [1, 51, 128]> v_head_265_cast_fp16 = reshape(shape = var_10804, x = var_10803_cast_fp16)[name = string("v_head_265_cast_fp16")];
            string dense_output_679_pad_type_1 = const()[name = string("dense_output_679_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_679_strides_1 = const()[name = string("dense_output_679_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_679_pad_1 = const()[name = string("dense_output_679_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_679_dilations_1 = const()[name = string("dense_output_679_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_679_groups_1 = const()[name = string("dense_output_679_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251120576))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251251712))))[name = string("layers_8_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_679_cast_fp16 = conv(dilations = dense_output_679_dilations_1, groups = dense_output_679_groups_1, pad = dense_output_679_pad_1, pad_type = dense_output_679_pad_type_1, strides = dense_output_679_strides_1, weight = layers_8_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_679_cast_fp16")];
            string sparse_output_679_pad_type_1 = const()[name = string("sparse_output_679_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_679_strides_1 = const()[name = string("sparse_output_679_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_679_pad_1 = const()[name = string("sparse_output_679_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_679_dilations_1 = const()[name = string("sparse_output_679_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_679_groups_1 = const()[name = string("sparse_output_679_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251254976))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251252288))))[name = string("layers_8_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_679_cast_fp16 = conv(dilations = sparse_output_679_dilations_1, groups = sparse_output_679_groups_1, pad = sparse_output_679_pad_1, pad_type = sparse_output_679_pad_type_1, strides = sparse_output_679_strides_1, weight = layers_8_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_679_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_10820_cast_fp16 = add(x = dense_output_679_cast_fp16, y = sparse_output_679_cast_fp16)[name = string("op_10820_cast_fp16")];
            tensor<int32, [4]> var_10821 = const()[name = string("op_10821"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10823 = const()[name = string("op_10823"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_10822_cast_fp16 = transpose(perm = var_10821, x = var_10820_cast_fp16)[name = string("transpose_641")];
            tensor<fp16, [1, 101, 128]> p_head_265_cast_fp16 = reshape(shape = var_10823, x = var_10822_cast_fp16)[name = string("p_head_265_cast_fp16")];
            tensor<fp16, [128]> var_10825_to_fp16 = const()[name = string("op_10825_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251271424)))];
            tensor<fp16, [1, 51, 128]> var_10826_cast_fp16 = add(x = q_head_133_cast_fp16, y = var_10825_to_fp16)[name = string("op_10826_cast_fp16")];
            tensor<int32, [1]> q_u_133_axes_1 = const()[name = string("q_u_133_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_133_cast_fp16 = expand_dims(axes = q_u_133_axes_1, x = var_10826_cast_fp16)[name = string("q_u_133_cast_fp16")];
            tensor<fp16, [128]> var_10828_to_fp16 = const()[name = string("op_10828_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251271744)))];
            tensor<fp16, [1, 51, 128]> var_10829_cast_fp16 = add(x = q_head_133_cast_fp16, y = var_10828_to_fp16)[name = string("op_10829_cast_fp16")];
            tensor<int32, [1]> q_v_133_axes_1 = const()[name = string("q_v_133_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_133_cast_fp16 = expand_dims(axes = q_v_133_axes_1, x = var_10829_cast_fp16)[name = string("q_v_133_cast_fp16")];
            tensor<int32, [1]> k_head_267_axes_1 = const()[name = string("k_head_267_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_267_cast_fp16 = expand_dims(axes = k_head_267_axes_1, x = k_head_265_cast_fp16)[name = string("k_head_267_cast_fp16")];
            tensor<int32, [1]> v_head_267_axes_1 = const()[name = string("v_head_267_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_267_cast_fp16 = expand_dims(axes = v_head_267_axes_1, x = v_head_265_cast_fp16)[name = string("v_head_267_cast_fp16")];
            tensor<int32, [1]> p_head_267_axes_1 = const()[name = string("p_head_267_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_267_cast_fp16 = expand_dims(axes = p_head_267_axes_1, x = p_head_265_cast_fp16)[name = string("p_head_267_cast_fp16")];
            bool var_10835_transpose_x_3 = const()[name = string("op_10835_transpose_x_3"), val = bool(false)];
            bool var_10835_transpose_y_3 = const()[name = string("op_10835_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_10835_cast_fp16 = matmul(transpose_x = var_10835_transpose_x_3, transpose_y = var_10835_transpose_y_3, x = q_u_133_cast_fp16, y = k_head_267_cast_fp16)[name = string("op_10835_cast_fp16")];
            fp16 var_10836_to_fp16 = const()[name = string("op_10836_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_133_cast_fp16 = mul(x = var_10835_cast_fp16, y = var_10836_to_fp16)[name = string("scores_content_133_cast_fp16")];
            bool x_709_transpose_x_3 = const()[name = string("x_709_transpose_x_3"), val = bool(false)];
            bool x_709_transpose_y_3 = const()[name = string("x_709_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_709_cast_fp16 = matmul(transpose_x = x_709_transpose_x_3, transpose_y = x_709_transpose_y_3, x = q_v_133_cast_fp16, y = p_head_267_cast_fp16)[name = string("x_709_cast_fp16")];
            tensor<int32, [8]> x_711_pad_1 = const()[name = string("x_711_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_711_mode_1 = const()[name = string("x_711_mode_1"), val = string("constant")];
            fp16 const_1753_to_fp16 = const()[name = string("const_1753_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_711_cast_fp16 = pad(constant_val = const_1753_to_fp16, mode = x_711_mode_1, pad = x_711_pad_1, x = x_709_cast_fp16)[name = string("x_711_cast_fp16")];
            tensor<int32, [4]> var_10850 = const()[name = string("op_10850"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_713_cast_fp16 = reshape(shape = var_10850, x = x_711_cast_fp16)[name = string("x_713_cast_fp16")];
            tensor<int32, [4]> var_10854_begin_1 = const()[name = string("op_10854_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10854_end_1 = const()[name = string("op_10854_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_10854_end_mask_1 = const()[name = string("op_10854_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_10854_cast_fp16 = slice_by_index(begin = var_10854_begin_1, end = var_10854_end_1, end_mask = var_10854_end_mask_1, x = x_713_cast_fp16)[name = string("op_10854_cast_fp16")];
            tensor<int32, [4]> var_10856 = const()[name = string("op_10856"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_10857_cast_fp16 = reshape(shape = var_10856, x = var_10854_cast_fp16)[name = string("op_10857_cast_fp16")];
            tensor<int32, [4]> var_10862_begin_1 = const()[name = string("op_10862_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10862_end_1 = const()[name = string("op_10862_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_10862_end_mask_1 = const()[name = string("op_10862_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_10862_cast_fp16 = slice_by_index(begin = var_10862_begin_1, end = var_10862_end_1, end_mask = var_10862_end_mask_1, x = var_10857_cast_fp16)[name = string("op_10862_cast_fp16")];
            fp16 var_10863_to_fp16 = const()[name = string("op_10863_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_133_cast_fp16 = mul(x = var_10862_cast_fp16, y = var_10863_to_fp16)[name = string("scores_pos_133_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_133_cast_fp16 = add(x = scores_content_133_cast_fp16, y = scores_pos_133_cast_fp16)[name = string("logits_133_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_10866_cast_fp16 = softmax(axis = var_10349, x = logits_133_cast_fp16)[name = string("op_10866_cast_fp16")];
            bool var_10868_transpose_x_1 = const()[name = string("op_10868_transpose_x_1"), val = bool(false)];
            bool var_10868_transpose_y_1 = const()[name = string("op_10868_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_10868_cast_fp16 = matmul(transpose_x = var_10868_transpose_x_1, transpose_y = var_10868_transpose_y_1, x = var_10866_cast_fp16, y = v_head_267_cast_fp16)[name = string("op_10868_cast_fp16")];
            tensor<int32, [1]> var_10869_axes_1 = const()[name = string("op_10869_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_10869_cast_fp16 = squeeze(axes = var_10869_axes_1, x = var_10868_cast_fp16)[name = string("op_10869_cast_fp16")];
            string dense_output_681_pad_type_1 = const()[name = string("dense_output_681_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_681_strides_1 = const()[name = string("dense_output_681_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_681_pad_1 = const()[name = string("dense_output_681_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_681_dilations_1 = const()[name = string("dense_output_681_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_681_groups_1 = const()[name = string("dense_output_681_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251272064))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251403200))))[name = string("layers_8_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_681_cast_fp16 = conv(dilations = dense_output_681_dilations_1, groups = dense_output_681_groups_1, pad = dense_output_681_pad_1, pad_type = dense_output_681_pad_type_1, strides = dense_output_681_strides_1, weight = layers_8_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_681_cast_fp16")];
            string sparse_output_681_pad_type_1 = const()[name = string("sparse_output_681_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_681_strides_1 = const()[name = string("sparse_output_681_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_681_pad_1 = const()[name = string("sparse_output_681_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_681_dilations_1 = const()[name = string("sparse_output_681_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_681_groups_1 = const()[name = string("sparse_output_681_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251406464))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251403776))))[name = string("layers_8_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_681_cast_fp16 = conv(dilations = sparse_output_681_dilations_1, groups = sparse_output_681_groups_1, pad = sparse_output_681_pad_1, pad_type = sparse_output_681_pad_type_1, strides = sparse_output_681_strides_1, weight = layers_8_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_681_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10884_cast_fp16 = add(x = dense_output_681_cast_fp16, y = sparse_output_681_cast_fp16)[name = string("op_10884_cast_fp16")];
            tensor<int32, [4]> var_10885 = const()[name = string("op_10885"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10887 = const()[name = string("op_10887"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10886_cast_fp16 = transpose(perm = var_10885, x = var_10884_cast_fp16)[name = string("transpose_640")];
            tensor<fp16, [1, 51, 128]> q_head_135_cast_fp16 = reshape(shape = var_10887, x = var_10886_cast_fp16)[name = string("q_head_135_cast_fp16")];
            string dense_output_683_pad_type_1 = const()[name = string("dense_output_683_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_683_strides_1 = const()[name = string("dense_output_683_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_683_pad_1 = const()[name = string("dense_output_683_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_683_dilations_1 = const()[name = string("dense_output_683_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_683_groups_1 = const()[name = string("dense_output_683_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251422912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251554048))))[name = string("layers_8_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_683_cast_fp16 = conv(dilations = dense_output_683_dilations_1, groups = dense_output_683_groups_1, pad = dense_output_683_pad_1, pad_type = dense_output_683_pad_type_1, strides = dense_output_683_strides_1, weight = layers_8_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_683_cast_fp16")];
            string sparse_output_683_pad_type_1 = const()[name = string("sparse_output_683_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_683_strides_1 = const()[name = string("sparse_output_683_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_683_pad_1 = const()[name = string("sparse_output_683_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_683_dilations_1 = const()[name = string("sparse_output_683_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_683_groups_1 = const()[name = string("sparse_output_683_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251557312))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251554624))))[name = string("layers_8_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_683_cast_fp16 = conv(dilations = sparse_output_683_dilations_1, groups = sparse_output_683_groups_1, pad = sparse_output_683_pad_1, pad_type = sparse_output_683_pad_type_1, strides = sparse_output_683_strides_1, weight = layers_8_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_683_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10903_cast_fp16 = add(x = dense_output_683_cast_fp16, y = sparse_output_683_cast_fp16)[name = string("op_10903_cast_fp16")];
            tensor<int32, [4]> var_10904 = const()[name = string("op_10904"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10906 = const()[name = string("op_10906"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10905_cast_fp16 = transpose(perm = var_10904, x = var_10903_cast_fp16)[name = string("transpose_639")];
            tensor<fp16, [1, 51, 128]> k_head_269_cast_fp16 = reshape(shape = var_10906, x = var_10905_cast_fp16)[name = string("k_head_269_cast_fp16")];
            string dense_output_685_pad_type_1 = const()[name = string("dense_output_685_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_685_strides_1 = const()[name = string("dense_output_685_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_685_pad_1 = const()[name = string("dense_output_685_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_685_dilations_1 = const()[name = string("dense_output_685_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_685_groups_1 = const()[name = string("dense_output_685_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251573760))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251704896))))[name = string("layers_8_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_685_cast_fp16 = conv(dilations = dense_output_685_dilations_1, groups = dense_output_685_groups_1, pad = dense_output_685_pad_1, pad_type = dense_output_685_pad_type_1, strides = dense_output_685_strides_1, weight = layers_8_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_685_cast_fp16")];
            string sparse_output_685_pad_type_1 = const()[name = string("sparse_output_685_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_685_strides_1 = const()[name = string("sparse_output_685_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_685_pad_1 = const()[name = string("sparse_output_685_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_685_dilations_1 = const()[name = string("sparse_output_685_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_685_groups_1 = const()[name = string("sparse_output_685_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251708160))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251705472))))[name = string("layers_8_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_685_cast_fp16 = conv(dilations = sparse_output_685_dilations_1, groups = sparse_output_685_groups_1, pad = sparse_output_685_pad_1, pad_type = sparse_output_685_pad_type_1, strides = sparse_output_685_strides_1, weight = layers_8_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_685_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_10922_cast_fp16 = add(x = dense_output_685_cast_fp16, y = sparse_output_685_cast_fp16)[name = string("op_10922_cast_fp16")];
            tensor<int32, [4]> var_10923 = const()[name = string("op_10923"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10925 = const()[name = string("op_10925"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_10924_cast_fp16 = transpose(perm = var_10923, x = var_10922_cast_fp16)[name = string("transpose_638")];
            tensor<fp16, [1, 51, 128]> v_head_269_cast_fp16 = reshape(shape = var_10925, x = var_10924_cast_fp16)[name = string("v_head_269_cast_fp16")];
            string dense_output_687_pad_type_1 = const()[name = string("dense_output_687_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_687_strides_1 = const()[name = string("dense_output_687_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_687_pad_1 = const()[name = string("dense_output_687_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_687_dilations_1 = const()[name = string("dense_output_687_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_687_groups_1 = const()[name = string("dense_output_687_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251724608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251855744))))[name = string("layers_8_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_687_cast_fp16 = conv(dilations = dense_output_687_dilations_1, groups = dense_output_687_groups_1, pad = dense_output_687_pad_1, pad_type = dense_output_687_pad_type_1, strides = dense_output_687_strides_1, weight = layers_8_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_687_cast_fp16")];
            string sparse_output_687_pad_type_1 = const()[name = string("sparse_output_687_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_687_strides_1 = const()[name = string("sparse_output_687_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_687_pad_1 = const()[name = string("sparse_output_687_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_687_dilations_1 = const()[name = string("sparse_output_687_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_687_groups_1 = const()[name = string("sparse_output_687_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251859008))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251856320))))[name = string("layers_8_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_687_cast_fp16 = conv(dilations = sparse_output_687_dilations_1, groups = sparse_output_687_groups_1, pad = sparse_output_687_pad_1, pad_type = sparse_output_687_pad_type_1, strides = sparse_output_687_strides_1, weight = layers_8_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_687_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_10941_cast_fp16 = add(x = dense_output_687_cast_fp16, y = sparse_output_687_cast_fp16)[name = string("op_10941_cast_fp16")];
            tensor<int32, [4]> var_10942 = const()[name = string("op_10942"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_10944 = const()[name = string("op_10944"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_10943_cast_fp16 = transpose(perm = var_10942, x = var_10941_cast_fp16)[name = string("transpose_637")];
            tensor<fp16, [1, 101, 128]> p_head_269_cast_fp16 = reshape(shape = var_10944, x = var_10943_cast_fp16)[name = string("p_head_269_cast_fp16")];
            tensor<fp16, [128]> var_10946_to_fp16 = const()[name = string("op_10946_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251875456)))];
            tensor<fp16, [1, 51, 128]> var_10947_cast_fp16 = add(x = q_head_135_cast_fp16, y = var_10946_to_fp16)[name = string("op_10947_cast_fp16")];
            tensor<int32, [1]> q_u_135_axes_1 = const()[name = string("q_u_135_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_135_cast_fp16 = expand_dims(axes = q_u_135_axes_1, x = var_10947_cast_fp16)[name = string("q_u_135_cast_fp16")];
            tensor<fp16, [128]> var_10949_to_fp16 = const()[name = string("op_10949_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251875776)))];
            tensor<fp16, [1, 51, 128]> var_10950_cast_fp16 = add(x = q_head_135_cast_fp16, y = var_10949_to_fp16)[name = string("op_10950_cast_fp16")];
            tensor<int32, [1]> q_v_135_axes_1 = const()[name = string("q_v_135_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_135_cast_fp16 = expand_dims(axes = q_v_135_axes_1, x = var_10950_cast_fp16)[name = string("q_v_135_cast_fp16")];
            tensor<int32, [1]> k_head_271_axes_1 = const()[name = string("k_head_271_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_271_cast_fp16 = expand_dims(axes = k_head_271_axes_1, x = k_head_269_cast_fp16)[name = string("k_head_271_cast_fp16")];
            tensor<int32, [1]> v_head_271_axes_1 = const()[name = string("v_head_271_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_271_cast_fp16 = expand_dims(axes = v_head_271_axes_1, x = v_head_269_cast_fp16)[name = string("v_head_271_cast_fp16")];
            tensor<int32, [1]> p_head_271_axes_1 = const()[name = string("p_head_271_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_271_cast_fp16 = expand_dims(axes = p_head_271_axes_1, x = p_head_269_cast_fp16)[name = string("p_head_271_cast_fp16")];
            bool var_10956_transpose_x_3 = const()[name = string("op_10956_transpose_x_3"), val = bool(false)];
            bool var_10956_transpose_y_3 = const()[name = string("op_10956_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_10956_cast_fp16 = matmul(transpose_x = var_10956_transpose_x_3, transpose_y = var_10956_transpose_y_3, x = q_u_135_cast_fp16, y = k_head_271_cast_fp16)[name = string("op_10956_cast_fp16")];
            fp16 var_10957_to_fp16 = const()[name = string("op_10957_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_135_cast_fp16 = mul(x = var_10956_cast_fp16, y = var_10957_to_fp16)[name = string("scores_content_135_cast_fp16")];
            bool x_717_transpose_x_3 = const()[name = string("x_717_transpose_x_3"), val = bool(false)];
            bool x_717_transpose_y_3 = const()[name = string("x_717_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_717_cast_fp16 = matmul(transpose_x = x_717_transpose_x_3, transpose_y = x_717_transpose_y_3, x = q_v_135_cast_fp16, y = p_head_271_cast_fp16)[name = string("x_717_cast_fp16")];
            tensor<int32, [8]> x_719_pad_1 = const()[name = string("x_719_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_719_mode_1 = const()[name = string("x_719_mode_1"), val = string("constant")];
            fp16 const_1759_to_fp16 = const()[name = string("const_1759_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_719_cast_fp16 = pad(constant_val = const_1759_to_fp16, mode = x_719_mode_1, pad = x_719_pad_1, x = x_717_cast_fp16)[name = string("x_719_cast_fp16")];
            tensor<int32, [4]> var_10971 = const()[name = string("op_10971"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_721_cast_fp16 = reshape(shape = var_10971, x = x_719_cast_fp16)[name = string("x_721_cast_fp16")];
            tensor<int32, [4]> var_10975_begin_1 = const()[name = string("op_10975_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_10975_end_1 = const()[name = string("op_10975_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_10975_end_mask_1 = const()[name = string("op_10975_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_10975_cast_fp16 = slice_by_index(begin = var_10975_begin_1, end = var_10975_end_1, end_mask = var_10975_end_mask_1, x = x_721_cast_fp16)[name = string("op_10975_cast_fp16")];
            tensor<int32, [4]> var_10977 = const()[name = string("op_10977"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_10978_cast_fp16 = reshape(shape = var_10977, x = var_10975_cast_fp16)[name = string("op_10978_cast_fp16")];
            tensor<int32, [4]> var_10983_begin_1 = const()[name = string("op_10983_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10983_end_1 = const()[name = string("op_10983_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_10983_end_mask_1 = const()[name = string("op_10983_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_10983_cast_fp16 = slice_by_index(begin = var_10983_begin_1, end = var_10983_end_1, end_mask = var_10983_end_mask_1, x = var_10978_cast_fp16)[name = string("op_10983_cast_fp16")];
            fp16 var_10984_to_fp16 = const()[name = string("op_10984_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_135_cast_fp16 = mul(x = var_10983_cast_fp16, y = var_10984_to_fp16)[name = string("scores_pos_135_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_135_cast_fp16 = add(x = scores_content_135_cast_fp16, y = scores_pos_135_cast_fp16)[name = string("logits_135_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_10987_cast_fp16 = softmax(axis = var_10349, x = logits_135_cast_fp16)[name = string("op_10987_cast_fp16")];
            bool var_10989_transpose_x_1 = const()[name = string("op_10989_transpose_x_1"), val = bool(false)];
            bool var_10989_transpose_y_1 = const()[name = string("op_10989_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_10989_cast_fp16 = matmul(transpose_x = var_10989_transpose_x_1, transpose_y = var_10989_transpose_y_1, x = var_10987_cast_fp16, y = v_head_271_cast_fp16)[name = string("op_10989_cast_fp16")];
            tensor<int32, [1]> var_10990_axes_1 = const()[name = string("op_10990_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_10990_cast_fp16 = squeeze(axes = var_10990_axes_1, x = var_10989_cast_fp16)[name = string("op_10990_cast_fp16")];
            string dense_output_689_pad_type_1 = const()[name = string("dense_output_689_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_689_strides_1 = const()[name = string("dense_output_689_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_689_pad_1 = const()[name = string("dense_output_689_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_689_dilations_1 = const()[name = string("dense_output_689_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_689_groups_1 = const()[name = string("dense_output_689_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251876096))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252007232))))[name = string("layers_8_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_689_cast_fp16 = conv(dilations = dense_output_689_dilations_1, groups = dense_output_689_groups_1, pad = dense_output_689_pad_1, pad_type = dense_output_689_pad_type_1, strides = dense_output_689_strides_1, weight = layers_8_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_689_cast_fp16")];
            string sparse_output_689_pad_type_1 = const()[name = string("sparse_output_689_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_689_strides_1 = const()[name = string("sparse_output_689_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_689_pad_1 = const()[name = string("sparse_output_689_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_689_dilations_1 = const()[name = string("sparse_output_689_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_689_groups_1 = const()[name = string("sparse_output_689_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252010496))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252007808))))[name = string("layers_8_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_689_cast_fp16 = conv(dilations = sparse_output_689_dilations_1, groups = sparse_output_689_groups_1, pad = sparse_output_689_pad_1, pad_type = sparse_output_689_pad_type_1, strides = sparse_output_689_strides_1, weight = layers_8_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_689_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11005_cast_fp16 = add(x = dense_output_689_cast_fp16, y = sparse_output_689_cast_fp16)[name = string("op_11005_cast_fp16")];
            tensor<int32, [4]> var_11006 = const()[name = string("op_11006"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11008 = const()[name = string("op_11008"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11007_cast_fp16 = transpose(perm = var_11006, x = var_11005_cast_fp16)[name = string("transpose_636")];
            tensor<fp16, [1, 51, 128]> q_head_137_cast_fp16 = reshape(shape = var_11008, x = var_11007_cast_fp16)[name = string("q_head_137_cast_fp16")];
            string dense_output_691_pad_type_1 = const()[name = string("dense_output_691_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_691_strides_1 = const()[name = string("dense_output_691_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_691_pad_1 = const()[name = string("dense_output_691_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_691_dilations_1 = const()[name = string("dense_output_691_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_691_groups_1 = const()[name = string("dense_output_691_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252026944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252158080))))[name = string("layers_8_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_691_cast_fp16 = conv(dilations = dense_output_691_dilations_1, groups = dense_output_691_groups_1, pad = dense_output_691_pad_1, pad_type = dense_output_691_pad_type_1, strides = dense_output_691_strides_1, weight = layers_8_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_691_cast_fp16")];
            string sparse_output_691_pad_type_1 = const()[name = string("sparse_output_691_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_691_strides_1 = const()[name = string("sparse_output_691_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_691_pad_1 = const()[name = string("sparse_output_691_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_691_dilations_1 = const()[name = string("sparse_output_691_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_691_groups_1 = const()[name = string("sparse_output_691_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252161344))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252158656))))[name = string("layers_8_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_691_cast_fp16 = conv(dilations = sparse_output_691_dilations_1, groups = sparse_output_691_groups_1, pad = sparse_output_691_pad_1, pad_type = sparse_output_691_pad_type_1, strides = sparse_output_691_strides_1, weight = layers_8_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_691_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11024_cast_fp16 = add(x = dense_output_691_cast_fp16, y = sparse_output_691_cast_fp16)[name = string("op_11024_cast_fp16")];
            tensor<int32, [4]> var_11025 = const()[name = string("op_11025"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11027 = const()[name = string("op_11027"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11026_cast_fp16 = transpose(perm = var_11025, x = var_11024_cast_fp16)[name = string("transpose_635")];
            tensor<fp16, [1, 51, 128]> k_head_273_cast_fp16 = reshape(shape = var_11027, x = var_11026_cast_fp16)[name = string("k_head_273_cast_fp16")];
            string dense_output_693_pad_type_1 = const()[name = string("dense_output_693_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_693_strides_1 = const()[name = string("dense_output_693_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_693_pad_1 = const()[name = string("dense_output_693_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_693_dilations_1 = const()[name = string("dense_output_693_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_693_groups_1 = const()[name = string("dense_output_693_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252177792))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252308928))))[name = string("layers_8_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_693_cast_fp16 = conv(dilations = dense_output_693_dilations_1, groups = dense_output_693_groups_1, pad = dense_output_693_pad_1, pad_type = dense_output_693_pad_type_1, strides = dense_output_693_strides_1, weight = layers_8_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_693_cast_fp16")];
            string sparse_output_693_pad_type_1 = const()[name = string("sparse_output_693_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_693_strides_1 = const()[name = string("sparse_output_693_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_693_pad_1 = const()[name = string("sparse_output_693_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_693_dilations_1 = const()[name = string("sparse_output_693_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_693_groups_1 = const()[name = string("sparse_output_693_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252312192))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252309504))))[name = string("layers_8_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_693_cast_fp16 = conv(dilations = sparse_output_693_dilations_1, groups = sparse_output_693_groups_1, pad = sparse_output_693_pad_1, pad_type = sparse_output_693_pad_type_1, strides = sparse_output_693_strides_1, weight = layers_8_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_693_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11043_cast_fp16 = add(x = dense_output_693_cast_fp16, y = sparse_output_693_cast_fp16)[name = string("op_11043_cast_fp16")];
            tensor<int32, [4]> var_11044 = const()[name = string("op_11044"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11046 = const()[name = string("op_11046"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11045_cast_fp16 = transpose(perm = var_11044, x = var_11043_cast_fp16)[name = string("transpose_634")];
            tensor<fp16, [1, 51, 128]> v_head_273_cast_fp16 = reshape(shape = var_11046, x = var_11045_cast_fp16)[name = string("v_head_273_cast_fp16")];
            string dense_output_695_pad_type_1 = const()[name = string("dense_output_695_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_695_strides_1 = const()[name = string("dense_output_695_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_695_pad_1 = const()[name = string("dense_output_695_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_695_dilations_1 = const()[name = string("dense_output_695_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_695_groups_1 = const()[name = string("dense_output_695_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252328640))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252459776))))[name = string("layers_8_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_695_cast_fp16 = conv(dilations = dense_output_695_dilations_1, groups = dense_output_695_groups_1, pad = dense_output_695_pad_1, pad_type = dense_output_695_pad_type_1, strides = dense_output_695_strides_1, weight = layers_8_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_695_cast_fp16")];
            string sparse_output_695_pad_type_1 = const()[name = string("sparse_output_695_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_695_strides_1 = const()[name = string("sparse_output_695_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_695_pad_1 = const()[name = string("sparse_output_695_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_695_dilations_1 = const()[name = string("sparse_output_695_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_695_groups_1 = const()[name = string("sparse_output_695_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252463040))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252460352))))[name = string("layers_8_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_695_cast_fp16 = conv(dilations = sparse_output_695_dilations_1, groups = sparse_output_695_groups_1, pad = sparse_output_695_pad_1, pad_type = sparse_output_695_pad_type_1, strides = sparse_output_695_strides_1, weight = layers_8_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_695_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_11062_cast_fp16 = add(x = dense_output_695_cast_fp16, y = sparse_output_695_cast_fp16)[name = string("op_11062_cast_fp16")];
            tensor<int32, [4]> var_11063 = const()[name = string("op_11063"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11065 = const()[name = string("op_11065"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_11064_cast_fp16 = transpose(perm = var_11063, x = var_11062_cast_fp16)[name = string("transpose_633")];
            tensor<fp16, [1, 101, 128]> p_head_273_cast_fp16 = reshape(shape = var_11065, x = var_11064_cast_fp16)[name = string("p_head_273_cast_fp16")];
            tensor<fp16, [128]> var_11067_to_fp16 = const()[name = string("op_11067_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252479488)))];
            tensor<fp16, [1, 51, 128]> var_11068_cast_fp16 = add(x = q_head_137_cast_fp16, y = var_11067_to_fp16)[name = string("op_11068_cast_fp16")];
            tensor<int32, [1]> q_u_137_axes_1 = const()[name = string("q_u_137_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_137_cast_fp16 = expand_dims(axes = q_u_137_axes_1, x = var_11068_cast_fp16)[name = string("q_u_137_cast_fp16")];
            tensor<fp16, [128]> var_11070_to_fp16 = const()[name = string("op_11070_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252479808)))];
            tensor<fp16, [1, 51, 128]> var_11071_cast_fp16 = add(x = q_head_137_cast_fp16, y = var_11070_to_fp16)[name = string("op_11071_cast_fp16")];
            tensor<int32, [1]> q_v_137_axes_1 = const()[name = string("q_v_137_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_137_cast_fp16 = expand_dims(axes = q_v_137_axes_1, x = var_11071_cast_fp16)[name = string("q_v_137_cast_fp16")];
            tensor<int32, [1]> k_head_275_axes_1 = const()[name = string("k_head_275_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_275_cast_fp16 = expand_dims(axes = k_head_275_axes_1, x = k_head_273_cast_fp16)[name = string("k_head_275_cast_fp16")];
            tensor<int32, [1]> v_head_275_axes_1 = const()[name = string("v_head_275_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_275_cast_fp16 = expand_dims(axes = v_head_275_axes_1, x = v_head_273_cast_fp16)[name = string("v_head_275_cast_fp16")];
            tensor<int32, [1]> p_head_275_axes_1 = const()[name = string("p_head_275_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_275_cast_fp16 = expand_dims(axes = p_head_275_axes_1, x = p_head_273_cast_fp16)[name = string("p_head_275_cast_fp16")];
            bool var_11077_transpose_x_3 = const()[name = string("op_11077_transpose_x_3"), val = bool(false)];
            bool var_11077_transpose_y_3 = const()[name = string("op_11077_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_11077_cast_fp16 = matmul(transpose_x = var_11077_transpose_x_3, transpose_y = var_11077_transpose_y_3, x = q_u_137_cast_fp16, y = k_head_275_cast_fp16)[name = string("op_11077_cast_fp16")];
            fp16 var_11078_to_fp16 = const()[name = string("op_11078_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_137_cast_fp16 = mul(x = var_11077_cast_fp16, y = var_11078_to_fp16)[name = string("scores_content_137_cast_fp16")];
            bool x_725_transpose_x_3 = const()[name = string("x_725_transpose_x_3"), val = bool(false)];
            bool x_725_transpose_y_3 = const()[name = string("x_725_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_725_cast_fp16 = matmul(transpose_x = x_725_transpose_x_3, transpose_y = x_725_transpose_y_3, x = q_v_137_cast_fp16, y = p_head_275_cast_fp16)[name = string("x_725_cast_fp16")];
            tensor<int32, [8]> x_727_pad_1 = const()[name = string("x_727_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_727_mode_1 = const()[name = string("x_727_mode_1"), val = string("constant")];
            fp16 const_1765_to_fp16 = const()[name = string("const_1765_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_727_cast_fp16 = pad(constant_val = const_1765_to_fp16, mode = x_727_mode_1, pad = x_727_pad_1, x = x_725_cast_fp16)[name = string("x_727_cast_fp16")];
            tensor<int32, [4]> var_11092 = const()[name = string("op_11092"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_729_cast_fp16 = reshape(shape = var_11092, x = x_727_cast_fp16)[name = string("x_729_cast_fp16")];
            tensor<int32, [4]> var_11096_begin_1 = const()[name = string("op_11096_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11096_end_1 = const()[name = string("op_11096_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_11096_end_mask_1 = const()[name = string("op_11096_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_11096_cast_fp16 = slice_by_index(begin = var_11096_begin_1, end = var_11096_end_1, end_mask = var_11096_end_mask_1, x = x_729_cast_fp16)[name = string("op_11096_cast_fp16")];
            tensor<int32, [4]> var_11098 = const()[name = string("op_11098"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_11099_cast_fp16 = reshape(shape = var_11098, x = var_11096_cast_fp16)[name = string("op_11099_cast_fp16")];
            tensor<int32, [4]> var_11104_begin_1 = const()[name = string("op_11104_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11104_end_1 = const()[name = string("op_11104_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_11104_end_mask_1 = const()[name = string("op_11104_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_11104_cast_fp16 = slice_by_index(begin = var_11104_begin_1, end = var_11104_end_1, end_mask = var_11104_end_mask_1, x = var_11099_cast_fp16)[name = string("op_11104_cast_fp16")];
            fp16 var_11105_to_fp16 = const()[name = string("op_11105_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_137_cast_fp16 = mul(x = var_11104_cast_fp16, y = var_11105_to_fp16)[name = string("scores_pos_137_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_137_cast_fp16 = add(x = scores_content_137_cast_fp16, y = scores_pos_137_cast_fp16)[name = string("logits_137_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_11108_cast_fp16 = softmax(axis = var_10349, x = logits_137_cast_fp16)[name = string("op_11108_cast_fp16")];
            bool var_11110_transpose_x_1 = const()[name = string("op_11110_transpose_x_1"), val = bool(false)];
            bool var_11110_transpose_y_1 = const()[name = string("op_11110_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_11110_cast_fp16 = matmul(transpose_x = var_11110_transpose_x_1, transpose_y = var_11110_transpose_y_1, x = var_11108_cast_fp16, y = v_head_275_cast_fp16)[name = string("op_11110_cast_fp16")];
            tensor<int32, [1]> var_11111_axes_1 = const()[name = string("op_11111_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_11111_cast_fp16 = squeeze(axes = var_11111_axes_1, x = var_11110_cast_fp16)[name = string("op_11111_cast_fp16")];
            string dense_output_697_pad_type_1 = const()[name = string("dense_output_697_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_697_strides_1 = const()[name = string("dense_output_697_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_697_pad_1 = const()[name = string("dense_output_697_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_697_dilations_1 = const()[name = string("dense_output_697_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_697_groups_1 = const()[name = string("dense_output_697_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252480128))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252611264))))[name = string("layers_8_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_697_cast_fp16 = conv(dilations = dense_output_697_dilations_1, groups = dense_output_697_groups_1, pad = dense_output_697_pad_1, pad_type = dense_output_697_pad_type_1, strides = dense_output_697_strides_1, weight = layers_8_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_697_cast_fp16")];
            string sparse_output_697_pad_type_1 = const()[name = string("sparse_output_697_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_697_strides_1 = const()[name = string("sparse_output_697_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_697_pad_1 = const()[name = string("sparse_output_697_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_697_dilations_1 = const()[name = string("sparse_output_697_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_697_groups_1 = const()[name = string("sparse_output_697_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252614528))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252611840))))[name = string("layers_8_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_697_cast_fp16 = conv(dilations = sparse_output_697_dilations_1, groups = sparse_output_697_groups_1, pad = sparse_output_697_pad_1, pad_type = sparse_output_697_pad_type_1, strides = sparse_output_697_strides_1, weight = layers_8_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_697_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11126_cast_fp16 = add(x = dense_output_697_cast_fp16, y = sparse_output_697_cast_fp16)[name = string("op_11126_cast_fp16")];
            tensor<int32, [4]> var_11127 = const()[name = string("op_11127"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11129 = const()[name = string("op_11129"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11128_cast_fp16 = transpose(perm = var_11127, x = var_11126_cast_fp16)[name = string("transpose_632")];
            tensor<fp16, [1, 51, 128]> q_head_139_cast_fp16 = reshape(shape = var_11129, x = var_11128_cast_fp16)[name = string("q_head_139_cast_fp16")];
            string dense_output_699_pad_type_1 = const()[name = string("dense_output_699_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_699_strides_1 = const()[name = string("dense_output_699_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_699_pad_1 = const()[name = string("dense_output_699_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_699_dilations_1 = const()[name = string("dense_output_699_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_699_groups_1 = const()[name = string("dense_output_699_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252630976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252762112))))[name = string("layers_8_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_699_cast_fp16 = conv(dilations = dense_output_699_dilations_1, groups = dense_output_699_groups_1, pad = dense_output_699_pad_1, pad_type = dense_output_699_pad_type_1, strides = dense_output_699_strides_1, weight = layers_8_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_699_cast_fp16")];
            string sparse_output_699_pad_type_1 = const()[name = string("sparse_output_699_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_699_strides_1 = const()[name = string("sparse_output_699_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_699_pad_1 = const()[name = string("sparse_output_699_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_699_dilations_1 = const()[name = string("sparse_output_699_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_699_groups_1 = const()[name = string("sparse_output_699_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252765376))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252762688))))[name = string("layers_8_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_699_cast_fp16 = conv(dilations = sparse_output_699_dilations_1, groups = sparse_output_699_groups_1, pad = sparse_output_699_pad_1, pad_type = sparse_output_699_pad_type_1, strides = sparse_output_699_strides_1, weight = layers_8_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_699_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11145_cast_fp16 = add(x = dense_output_699_cast_fp16, y = sparse_output_699_cast_fp16)[name = string("op_11145_cast_fp16")];
            tensor<int32, [4]> var_11146 = const()[name = string("op_11146"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11148 = const()[name = string("op_11148"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11147_cast_fp16 = transpose(perm = var_11146, x = var_11145_cast_fp16)[name = string("transpose_631")];
            tensor<fp16, [1, 51, 128]> k_head_277_cast_fp16 = reshape(shape = var_11148, x = var_11147_cast_fp16)[name = string("k_head_277_cast_fp16")];
            string dense_output_701_pad_type_1 = const()[name = string("dense_output_701_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_701_strides_1 = const()[name = string("dense_output_701_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_701_pad_1 = const()[name = string("dense_output_701_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_701_dilations_1 = const()[name = string("dense_output_701_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_701_groups_1 = const()[name = string("dense_output_701_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252781824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252912960))))[name = string("layers_8_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_701_cast_fp16 = conv(dilations = dense_output_701_dilations_1, groups = dense_output_701_groups_1, pad = dense_output_701_pad_1, pad_type = dense_output_701_pad_type_1, strides = dense_output_701_strides_1, weight = layers_8_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_701_cast_fp16")];
            string sparse_output_701_pad_type_1 = const()[name = string("sparse_output_701_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_701_strides_1 = const()[name = string("sparse_output_701_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_701_pad_1 = const()[name = string("sparse_output_701_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_701_dilations_1 = const()[name = string("sparse_output_701_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_701_groups_1 = const()[name = string("sparse_output_701_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252916224))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252913536))))[name = string("layers_8_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_701_cast_fp16 = conv(dilations = sparse_output_701_dilations_1, groups = sparse_output_701_groups_1, pad = sparse_output_701_pad_1, pad_type = sparse_output_701_pad_type_1, strides = sparse_output_701_strides_1, weight = layers_8_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_701_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11164_cast_fp16 = add(x = dense_output_701_cast_fp16, y = sparse_output_701_cast_fp16)[name = string("op_11164_cast_fp16")];
            tensor<int32, [4]> var_11165 = const()[name = string("op_11165"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11167 = const()[name = string("op_11167"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11166_cast_fp16 = transpose(perm = var_11165, x = var_11164_cast_fp16)[name = string("transpose_630")];
            tensor<fp16, [1, 51, 128]> v_head_277_cast_fp16 = reshape(shape = var_11167, x = var_11166_cast_fp16)[name = string("v_head_277_cast_fp16")];
            string dense_output_703_pad_type_1 = const()[name = string("dense_output_703_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_703_strides_1 = const()[name = string("dense_output_703_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_703_pad_1 = const()[name = string("dense_output_703_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_703_dilations_1 = const()[name = string("dense_output_703_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_703_groups_1 = const()[name = string("dense_output_703_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252932672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253063808))))[name = string("layers_8_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_703_cast_fp16 = conv(dilations = dense_output_703_dilations_1, groups = dense_output_703_groups_1, pad = dense_output_703_pad_1, pad_type = dense_output_703_pad_type_1, strides = dense_output_703_strides_1, weight = layers_8_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_703_cast_fp16")];
            string sparse_output_703_pad_type_1 = const()[name = string("sparse_output_703_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_703_strides_1 = const()[name = string("sparse_output_703_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_703_pad_1 = const()[name = string("sparse_output_703_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_703_dilations_1 = const()[name = string("sparse_output_703_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_703_groups_1 = const()[name = string("sparse_output_703_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253067072))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253064384))))[name = string("layers_8_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_703_cast_fp16 = conv(dilations = sparse_output_703_dilations_1, groups = sparse_output_703_groups_1, pad = sparse_output_703_pad_1, pad_type = sparse_output_703_pad_type_1, strides = sparse_output_703_strides_1, weight = layers_8_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_703_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_11183_cast_fp16 = add(x = dense_output_703_cast_fp16, y = sparse_output_703_cast_fp16)[name = string("op_11183_cast_fp16")];
            tensor<int32, [4]> var_11184 = const()[name = string("op_11184"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11186 = const()[name = string("op_11186"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_11185_cast_fp16 = transpose(perm = var_11184, x = var_11183_cast_fp16)[name = string("transpose_629")];
            tensor<fp16, [1, 101, 128]> p_head_277_cast_fp16 = reshape(shape = var_11186, x = var_11185_cast_fp16)[name = string("p_head_277_cast_fp16")];
            tensor<fp16, [128]> var_11188_to_fp16 = const()[name = string("op_11188_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253083520)))];
            tensor<fp16, [1, 51, 128]> var_11189_cast_fp16 = add(x = q_head_139_cast_fp16, y = var_11188_to_fp16)[name = string("op_11189_cast_fp16")];
            tensor<int32, [1]> q_u_139_axes_1 = const()[name = string("q_u_139_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_139_cast_fp16 = expand_dims(axes = q_u_139_axes_1, x = var_11189_cast_fp16)[name = string("q_u_139_cast_fp16")];
            tensor<fp16, [128]> var_11191_to_fp16 = const()[name = string("op_11191_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253083840)))];
            tensor<fp16, [1, 51, 128]> var_11192_cast_fp16 = add(x = q_head_139_cast_fp16, y = var_11191_to_fp16)[name = string("op_11192_cast_fp16")];
            tensor<int32, [1]> q_v_139_axes_1 = const()[name = string("q_v_139_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_139_cast_fp16 = expand_dims(axes = q_v_139_axes_1, x = var_11192_cast_fp16)[name = string("q_v_139_cast_fp16")];
            tensor<int32, [1]> k_head_279_axes_1 = const()[name = string("k_head_279_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_279_cast_fp16 = expand_dims(axes = k_head_279_axes_1, x = k_head_277_cast_fp16)[name = string("k_head_279_cast_fp16")];
            tensor<int32, [1]> v_head_279_axes_1 = const()[name = string("v_head_279_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_279_cast_fp16 = expand_dims(axes = v_head_279_axes_1, x = v_head_277_cast_fp16)[name = string("v_head_279_cast_fp16")];
            tensor<int32, [1]> p_head_279_axes_1 = const()[name = string("p_head_279_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_279_cast_fp16 = expand_dims(axes = p_head_279_axes_1, x = p_head_277_cast_fp16)[name = string("p_head_279_cast_fp16")];
            bool var_11198_transpose_x_3 = const()[name = string("op_11198_transpose_x_3"), val = bool(false)];
            bool var_11198_transpose_y_3 = const()[name = string("op_11198_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_11198_cast_fp16 = matmul(transpose_x = var_11198_transpose_x_3, transpose_y = var_11198_transpose_y_3, x = q_u_139_cast_fp16, y = k_head_279_cast_fp16)[name = string("op_11198_cast_fp16")];
            fp16 var_11199_to_fp16 = const()[name = string("op_11199_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_139_cast_fp16 = mul(x = var_11198_cast_fp16, y = var_11199_to_fp16)[name = string("scores_content_139_cast_fp16")];
            bool x_733_transpose_x_3 = const()[name = string("x_733_transpose_x_3"), val = bool(false)];
            bool x_733_transpose_y_3 = const()[name = string("x_733_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_733_cast_fp16 = matmul(transpose_x = x_733_transpose_x_3, transpose_y = x_733_transpose_y_3, x = q_v_139_cast_fp16, y = p_head_279_cast_fp16)[name = string("x_733_cast_fp16")];
            tensor<int32, [8]> x_735_pad_1 = const()[name = string("x_735_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_735_mode_1 = const()[name = string("x_735_mode_1"), val = string("constant")];
            fp16 const_1771_to_fp16 = const()[name = string("const_1771_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_735_cast_fp16 = pad(constant_val = const_1771_to_fp16, mode = x_735_mode_1, pad = x_735_pad_1, x = x_733_cast_fp16)[name = string("x_735_cast_fp16")];
            tensor<int32, [4]> var_11213 = const()[name = string("op_11213"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_737_cast_fp16 = reshape(shape = var_11213, x = x_735_cast_fp16)[name = string("x_737_cast_fp16")];
            tensor<int32, [4]> var_11217_begin_1 = const()[name = string("op_11217_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11217_end_1 = const()[name = string("op_11217_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_11217_end_mask_1 = const()[name = string("op_11217_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_11217_cast_fp16 = slice_by_index(begin = var_11217_begin_1, end = var_11217_end_1, end_mask = var_11217_end_mask_1, x = x_737_cast_fp16)[name = string("op_11217_cast_fp16")];
            tensor<int32, [4]> var_11219 = const()[name = string("op_11219"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_11220_cast_fp16 = reshape(shape = var_11219, x = var_11217_cast_fp16)[name = string("op_11220_cast_fp16")];
            tensor<int32, [4]> var_11225_begin_1 = const()[name = string("op_11225_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11225_end_1 = const()[name = string("op_11225_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_11225_end_mask_1 = const()[name = string("op_11225_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_11225_cast_fp16 = slice_by_index(begin = var_11225_begin_1, end = var_11225_end_1, end_mask = var_11225_end_mask_1, x = var_11220_cast_fp16)[name = string("op_11225_cast_fp16")];
            fp16 var_11226_to_fp16 = const()[name = string("op_11226_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_139_cast_fp16 = mul(x = var_11225_cast_fp16, y = var_11226_to_fp16)[name = string("scores_pos_139_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_139_cast_fp16 = add(x = scores_content_139_cast_fp16, y = scores_pos_139_cast_fp16)[name = string("logits_139_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_11229_cast_fp16 = softmax(axis = var_10349, x = logits_139_cast_fp16)[name = string("op_11229_cast_fp16")];
            bool var_11231_transpose_x_1 = const()[name = string("op_11231_transpose_x_1"), val = bool(false)];
            bool var_11231_transpose_y_1 = const()[name = string("op_11231_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_11231_cast_fp16 = matmul(transpose_x = var_11231_transpose_x_1, transpose_y = var_11231_transpose_y_1, x = var_11229_cast_fp16, y = v_head_279_cast_fp16)[name = string("op_11231_cast_fp16")];
            tensor<int32, [1]> var_11232_axes_1 = const()[name = string("op_11232_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_11232_cast_fp16 = squeeze(axes = var_11232_axes_1, x = var_11231_cast_fp16)[name = string("op_11232_cast_fp16")];
            string dense_output_705_pad_type_1 = const()[name = string("dense_output_705_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_705_strides_1 = const()[name = string("dense_output_705_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_705_pad_1 = const()[name = string("dense_output_705_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_705_dilations_1 = const()[name = string("dense_output_705_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_705_groups_1 = const()[name = string("dense_output_705_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253084160))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253215296))))[name = string("layers_8_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_705_cast_fp16 = conv(dilations = dense_output_705_dilations_1, groups = dense_output_705_groups_1, pad = dense_output_705_pad_1, pad_type = dense_output_705_pad_type_1, strides = dense_output_705_strides_1, weight = layers_8_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_705_cast_fp16")];
            string sparse_output_705_pad_type_1 = const()[name = string("sparse_output_705_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_705_strides_1 = const()[name = string("sparse_output_705_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_705_pad_1 = const()[name = string("sparse_output_705_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_705_dilations_1 = const()[name = string("sparse_output_705_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_705_groups_1 = const()[name = string("sparse_output_705_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253218560))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253215872))))[name = string("layers_8_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_705_cast_fp16 = conv(dilations = sparse_output_705_dilations_1, groups = sparse_output_705_groups_1, pad = sparse_output_705_pad_1, pad_type = sparse_output_705_pad_type_1, strides = sparse_output_705_strides_1, weight = layers_8_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_705_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11247_cast_fp16 = add(x = dense_output_705_cast_fp16, y = sparse_output_705_cast_fp16)[name = string("op_11247_cast_fp16")];
            tensor<int32, [4]> var_11248 = const()[name = string("op_11248"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11250 = const()[name = string("op_11250"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11249_cast_fp16 = transpose(perm = var_11248, x = var_11247_cast_fp16)[name = string("transpose_628")];
            tensor<fp16, [1, 51, 128]> q_head_141_cast_fp16 = reshape(shape = var_11250, x = var_11249_cast_fp16)[name = string("q_head_141_cast_fp16")];
            string dense_output_707_pad_type_1 = const()[name = string("dense_output_707_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_707_strides_1 = const()[name = string("dense_output_707_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_707_pad_1 = const()[name = string("dense_output_707_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_707_dilations_1 = const()[name = string("dense_output_707_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_707_groups_1 = const()[name = string("dense_output_707_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253235008))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253366144))))[name = string("layers_8_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_707_cast_fp16 = conv(dilations = dense_output_707_dilations_1, groups = dense_output_707_groups_1, pad = dense_output_707_pad_1, pad_type = dense_output_707_pad_type_1, strides = dense_output_707_strides_1, weight = layers_8_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_707_cast_fp16")];
            string sparse_output_707_pad_type_1 = const()[name = string("sparse_output_707_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_707_strides_1 = const()[name = string("sparse_output_707_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_707_pad_1 = const()[name = string("sparse_output_707_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_707_dilations_1 = const()[name = string("sparse_output_707_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_707_groups_1 = const()[name = string("sparse_output_707_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253369408))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253366720))))[name = string("layers_8_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_707_cast_fp16 = conv(dilations = sparse_output_707_dilations_1, groups = sparse_output_707_groups_1, pad = sparse_output_707_pad_1, pad_type = sparse_output_707_pad_type_1, strides = sparse_output_707_strides_1, weight = layers_8_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_707_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11266_cast_fp16 = add(x = dense_output_707_cast_fp16, y = sparse_output_707_cast_fp16)[name = string("op_11266_cast_fp16")];
            tensor<int32, [4]> var_11267 = const()[name = string("op_11267"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11269 = const()[name = string("op_11269"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11268_cast_fp16 = transpose(perm = var_11267, x = var_11266_cast_fp16)[name = string("transpose_627")];
            tensor<fp16, [1, 51, 128]> k_head_281_cast_fp16 = reshape(shape = var_11269, x = var_11268_cast_fp16)[name = string("k_head_281_cast_fp16")];
            string dense_output_709_pad_type_1 = const()[name = string("dense_output_709_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_709_strides_1 = const()[name = string("dense_output_709_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_709_pad_1 = const()[name = string("dense_output_709_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_709_dilations_1 = const()[name = string("dense_output_709_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_709_groups_1 = const()[name = string("dense_output_709_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253385856))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253516992))))[name = string("layers_8_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_709_cast_fp16 = conv(dilations = dense_output_709_dilations_1, groups = dense_output_709_groups_1, pad = dense_output_709_pad_1, pad_type = dense_output_709_pad_type_1, strides = dense_output_709_strides_1, weight = layers_8_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_709_cast_fp16")];
            string sparse_output_709_pad_type_1 = const()[name = string("sparse_output_709_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_709_strides_1 = const()[name = string("sparse_output_709_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_709_pad_1 = const()[name = string("sparse_output_709_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_709_dilations_1 = const()[name = string("sparse_output_709_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_709_groups_1 = const()[name = string("sparse_output_709_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253520256))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253517568))))[name = string("layers_8_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_709_cast_fp16 = conv(dilations = sparse_output_709_dilations_1, groups = sparse_output_709_groups_1, pad = sparse_output_709_pad_1, pad_type = sparse_output_709_pad_type_1, strides = sparse_output_709_strides_1, weight = layers_8_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_709_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11285_cast_fp16 = add(x = dense_output_709_cast_fp16, y = sparse_output_709_cast_fp16)[name = string("op_11285_cast_fp16")];
            tensor<int32, [4]> var_11286 = const()[name = string("op_11286"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11288 = const()[name = string("op_11288"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11287_cast_fp16 = transpose(perm = var_11286, x = var_11285_cast_fp16)[name = string("transpose_626")];
            tensor<fp16, [1, 51, 128]> v_head_281_cast_fp16 = reshape(shape = var_11288, x = var_11287_cast_fp16)[name = string("v_head_281_cast_fp16")];
            string dense_output_711_pad_type_1 = const()[name = string("dense_output_711_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_711_strides_1 = const()[name = string("dense_output_711_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_711_pad_1 = const()[name = string("dense_output_711_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_711_dilations_1 = const()[name = string("dense_output_711_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_711_groups_1 = const()[name = string("dense_output_711_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253536704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253667840))))[name = string("layers_8_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_711_cast_fp16 = conv(dilations = dense_output_711_dilations_1, groups = dense_output_711_groups_1, pad = dense_output_711_pad_1, pad_type = dense_output_711_pad_type_1, strides = dense_output_711_strides_1, weight = layers_8_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_711_cast_fp16")];
            string sparse_output_711_pad_type_1 = const()[name = string("sparse_output_711_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_711_strides_1 = const()[name = string("sparse_output_711_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_711_pad_1 = const()[name = string("sparse_output_711_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_711_dilations_1 = const()[name = string("sparse_output_711_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_711_groups_1 = const()[name = string("sparse_output_711_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253671104))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253668416))))[name = string("layers_8_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_711_cast_fp16 = conv(dilations = sparse_output_711_dilations_1, groups = sparse_output_711_groups_1, pad = sparse_output_711_pad_1, pad_type = sparse_output_711_pad_type_1, strides = sparse_output_711_strides_1, weight = layers_8_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_711_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_11304_cast_fp16 = add(x = dense_output_711_cast_fp16, y = sparse_output_711_cast_fp16)[name = string("op_11304_cast_fp16")];
            tensor<int32, [4]> var_11305 = const()[name = string("op_11305"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11307 = const()[name = string("op_11307"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_11306_cast_fp16 = transpose(perm = var_11305, x = var_11304_cast_fp16)[name = string("transpose_625")];
            tensor<fp16, [1, 101, 128]> p_head_281_cast_fp16 = reshape(shape = var_11307, x = var_11306_cast_fp16)[name = string("p_head_281_cast_fp16")];
            tensor<fp16, [128]> var_11309_to_fp16 = const()[name = string("op_11309_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253687552)))];
            tensor<fp16, [1, 51, 128]> var_11310_cast_fp16 = add(x = q_head_141_cast_fp16, y = var_11309_to_fp16)[name = string("op_11310_cast_fp16")];
            tensor<int32, [1]> q_u_141_axes_1 = const()[name = string("q_u_141_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_141_cast_fp16 = expand_dims(axes = q_u_141_axes_1, x = var_11310_cast_fp16)[name = string("q_u_141_cast_fp16")];
            tensor<fp16, [128]> var_11312_to_fp16 = const()[name = string("op_11312_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253687872)))];
            tensor<fp16, [1, 51, 128]> var_11313_cast_fp16 = add(x = q_head_141_cast_fp16, y = var_11312_to_fp16)[name = string("op_11313_cast_fp16")];
            tensor<int32, [1]> q_v_141_axes_1 = const()[name = string("q_v_141_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_141_cast_fp16 = expand_dims(axes = q_v_141_axes_1, x = var_11313_cast_fp16)[name = string("q_v_141_cast_fp16")];
            tensor<int32, [1]> k_head_283_axes_1 = const()[name = string("k_head_283_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_283_cast_fp16 = expand_dims(axes = k_head_283_axes_1, x = k_head_281_cast_fp16)[name = string("k_head_283_cast_fp16")];
            tensor<int32, [1]> v_head_283_axes_1 = const()[name = string("v_head_283_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_283_cast_fp16 = expand_dims(axes = v_head_283_axes_1, x = v_head_281_cast_fp16)[name = string("v_head_283_cast_fp16")];
            tensor<int32, [1]> p_head_283_axes_1 = const()[name = string("p_head_283_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_283_cast_fp16 = expand_dims(axes = p_head_283_axes_1, x = p_head_281_cast_fp16)[name = string("p_head_283_cast_fp16")];
            bool var_11319_transpose_x_3 = const()[name = string("op_11319_transpose_x_3"), val = bool(false)];
            bool var_11319_transpose_y_3 = const()[name = string("op_11319_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_11319_cast_fp16 = matmul(transpose_x = var_11319_transpose_x_3, transpose_y = var_11319_transpose_y_3, x = q_u_141_cast_fp16, y = k_head_283_cast_fp16)[name = string("op_11319_cast_fp16")];
            fp16 var_11320_to_fp16 = const()[name = string("op_11320_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_141_cast_fp16 = mul(x = var_11319_cast_fp16, y = var_11320_to_fp16)[name = string("scores_content_141_cast_fp16")];
            bool x_741_transpose_x_3 = const()[name = string("x_741_transpose_x_3"), val = bool(false)];
            bool x_741_transpose_y_3 = const()[name = string("x_741_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_741_cast_fp16 = matmul(transpose_x = x_741_transpose_x_3, transpose_y = x_741_transpose_y_3, x = q_v_141_cast_fp16, y = p_head_283_cast_fp16)[name = string("x_741_cast_fp16")];
            tensor<int32, [8]> x_743_pad_1 = const()[name = string("x_743_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_743_mode_1 = const()[name = string("x_743_mode_1"), val = string("constant")];
            fp16 const_1777_to_fp16 = const()[name = string("const_1777_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_743_cast_fp16 = pad(constant_val = const_1777_to_fp16, mode = x_743_mode_1, pad = x_743_pad_1, x = x_741_cast_fp16)[name = string("x_743_cast_fp16")];
            tensor<int32, [4]> var_11334 = const()[name = string("op_11334"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_745_cast_fp16 = reshape(shape = var_11334, x = x_743_cast_fp16)[name = string("x_745_cast_fp16")];
            tensor<int32, [4]> var_11338_begin_1 = const()[name = string("op_11338_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11338_end_1 = const()[name = string("op_11338_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_11338_end_mask_1 = const()[name = string("op_11338_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_11338_cast_fp16 = slice_by_index(begin = var_11338_begin_1, end = var_11338_end_1, end_mask = var_11338_end_mask_1, x = x_745_cast_fp16)[name = string("op_11338_cast_fp16")];
            tensor<int32, [4]> var_11340 = const()[name = string("op_11340"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_11341_cast_fp16 = reshape(shape = var_11340, x = var_11338_cast_fp16)[name = string("op_11341_cast_fp16")];
            tensor<int32, [4]> var_11346_begin_1 = const()[name = string("op_11346_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11346_end_1 = const()[name = string("op_11346_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_11346_end_mask_1 = const()[name = string("op_11346_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_11346_cast_fp16 = slice_by_index(begin = var_11346_begin_1, end = var_11346_end_1, end_mask = var_11346_end_mask_1, x = var_11341_cast_fp16)[name = string("op_11346_cast_fp16")];
            fp16 var_11347_to_fp16 = const()[name = string("op_11347_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_141_cast_fp16 = mul(x = var_11346_cast_fp16, y = var_11347_to_fp16)[name = string("scores_pos_141_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_141_cast_fp16 = add(x = scores_content_141_cast_fp16, y = scores_pos_141_cast_fp16)[name = string("logits_141_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_11350_cast_fp16 = softmax(axis = var_10349, x = logits_141_cast_fp16)[name = string("op_11350_cast_fp16")];
            bool var_11352_transpose_x_1 = const()[name = string("op_11352_transpose_x_1"), val = bool(false)];
            bool var_11352_transpose_y_1 = const()[name = string("op_11352_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_11352_cast_fp16 = matmul(transpose_x = var_11352_transpose_x_1, transpose_y = var_11352_transpose_y_1, x = var_11350_cast_fp16, y = v_head_283_cast_fp16)[name = string("op_11352_cast_fp16")];
            tensor<int32, [1]> var_11353_axes_1 = const()[name = string("op_11353_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_11353_cast_fp16 = squeeze(axes = var_11353_axes_1, x = var_11352_cast_fp16)[name = string("op_11353_cast_fp16")];
            string dense_output_713_pad_type_1 = const()[name = string("dense_output_713_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_713_strides_1 = const()[name = string("dense_output_713_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_713_pad_1 = const()[name = string("dense_output_713_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_713_dilations_1 = const()[name = string("dense_output_713_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_713_groups_1 = const()[name = string("dense_output_713_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253688192))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253819328))))[name = string("layers_8_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_713_cast_fp16 = conv(dilations = dense_output_713_dilations_1, groups = dense_output_713_groups_1, pad = dense_output_713_pad_1, pad_type = dense_output_713_pad_type_1, strides = dense_output_713_strides_1, weight = layers_8_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_713_cast_fp16")];
            string sparse_output_713_pad_type_1 = const()[name = string("sparse_output_713_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_713_strides_1 = const()[name = string("sparse_output_713_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_713_pad_1 = const()[name = string("sparse_output_713_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_713_dilations_1 = const()[name = string("sparse_output_713_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_713_groups_1 = const()[name = string("sparse_output_713_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253822592))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253819904))))[name = string("layers_8_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_713_cast_fp16 = conv(dilations = sparse_output_713_dilations_1, groups = sparse_output_713_groups_1, pad = sparse_output_713_pad_1, pad_type = sparse_output_713_pad_type_1, strides = sparse_output_713_strides_1, weight = layers_8_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_713_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11368_cast_fp16 = add(x = dense_output_713_cast_fp16, y = sparse_output_713_cast_fp16)[name = string("op_11368_cast_fp16")];
            tensor<int32, [4]> var_11369 = const()[name = string("op_11369"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11371 = const()[name = string("op_11371"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11370_cast_fp16 = transpose(perm = var_11369, x = var_11368_cast_fp16)[name = string("transpose_624")];
            tensor<fp16, [1, 51, 128]> q_head_143_cast_fp16 = reshape(shape = var_11371, x = var_11370_cast_fp16)[name = string("q_head_143_cast_fp16")];
            string dense_output_715_pad_type_1 = const()[name = string("dense_output_715_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_715_strides_1 = const()[name = string("dense_output_715_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_715_pad_1 = const()[name = string("dense_output_715_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_715_dilations_1 = const()[name = string("dense_output_715_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_715_groups_1 = const()[name = string("dense_output_715_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253839040))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253970176))))[name = string("layers_8_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_715_cast_fp16 = conv(dilations = dense_output_715_dilations_1, groups = dense_output_715_groups_1, pad = dense_output_715_pad_1, pad_type = dense_output_715_pad_type_1, strides = dense_output_715_strides_1, weight = layers_8_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_715_cast_fp16")];
            string sparse_output_715_pad_type_1 = const()[name = string("sparse_output_715_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_715_strides_1 = const()[name = string("sparse_output_715_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_715_pad_1 = const()[name = string("sparse_output_715_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_715_dilations_1 = const()[name = string("sparse_output_715_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_715_groups_1 = const()[name = string("sparse_output_715_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253973440))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253970752))))[name = string("layers_8_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_715_cast_fp16 = conv(dilations = sparse_output_715_dilations_1, groups = sparse_output_715_groups_1, pad = sparse_output_715_pad_1, pad_type = sparse_output_715_pad_type_1, strides = sparse_output_715_strides_1, weight = layers_8_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_715_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11387_cast_fp16 = add(x = dense_output_715_cast_fp16, y = sparse_output_715_cast_fp16)[name = string("op_11387_cast_fp16")];
            tensor<int32, [4]> var_11388 = const()[name = string("op_11388"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11390 = const()[name = string("op_11390"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11389_cast_fp16 = transpose(perm = var_11388, x = var_11387_cast_fp16)[name = string("transpose_623")];
            tensor<fp16, [1, 51, 128]> k_head_285_cast_fp16 = reshape(shape = var_11390, x = var_11389_cast_fp16)[name = string("k_head_285_cast_fp16")];
            string dense_output_717_pad_type_1 = const()[name = string("dense_output_717_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_717_strides_1 = const()[name = string("dense_output_717_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_717_pad_1 = const()[name = string("dense_output_717_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_717_dilations_1 = const()[name = string("dense_output_717_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_717_groups_1 = const()[name = string("dense_output_717_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253989888))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254121024))))[name = string("layers_8_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_717_cast_fp16 = conv(dilations = dense_output_717_dilations_1, groups = dense_output_717_groups_1, pad = dense_output_717_pad_1, pad_type = dense_output_717_pad_type_1, strides = dense_output_717_strides_1, weight = layers_8_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = string("dense_output_717_cast_fp16")];
            string sparse_output_717_pad_type_1 = const()[name = string("sparse_output_717_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_717_strides_1 = const()[name = string("sparse_output_717_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_717_pad_1 = const()[name = string("sparse_output_717_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_717_dilations_1 = const()[name = string("sparse_output_717_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_717_groups_1 = const()[name = string("sparse_output_717_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254124288))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254121600))))[name = string("layers_8_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_717_cast_fp16 = conv(dilations = sparse_output_717_dilations_1, groups = sparse_output_717_groups_1, pad = sparse_output_717_pad_1, pad_type = sparse_output_717_pad_type_1, strides = sparse_output_717_strides_1, weight = layers_8_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_397_cast_fp16)[name = string("sparse_output_717_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11406_cast_fp16 = add(x = dense_output_717_cast_fp16, y = sparse_output_717_cast_fp16)[name = string("op_11406_cast_fp16")];
            tensor<int32, [4]> var_11407 = const()[name = string("op_11407"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11409 = const()[name = string("op_11409"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11408_cast_fp16 = transpose(perm = var_11407, x = var_11406_cast_fp16)[name = string("transpose_622")];
            tensor<fp16, [1, 51, 128]> v_head_285_cast_fp16 = reshape(shape = var_11409, x = var_11408_cast_fp16)[name = string("v_head_285_cast_fp16")];
            string dense_output_719_pad_type_1 = const()[name = string("dense_output_719_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_719_strides_1 = const()[name = string("dense_output_719_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_719_pad_1 = const()[name = string("dense_output_719_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_719_dilations_1 = const()[name = string("dense_output_719_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_719_groups_1 = const()[name = string("dense_output_719_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254140736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254271872))))[name = string("layers_8_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_719_cast_fp16 = conv(dilations = dense_output_719_dilations_1, groups = dense_output_719_groups_1, pad = dense_output_719_pad_1, pad_type = dense_output_719_pad_type_1, strides = dense_output_719_strides_1, weight = layers_8_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_719_cast_fp16")];
            string sparse_output_719_pad_type_1 = const()[name = string("sparse_output_719_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_719_strides_1 = const()[name = string("sparse_output_719_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_719_pad_1 = const()[name = string("sparse_output_719_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_719_dilations_1 = const()[name = string("sparse_output_719_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_719_groups_1 = const()[name = string("sparse_output_719_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_8_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254275136))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254272448))))[name = string("layers_8_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_719_cast_fp16 = conv(dilations = sparse_output_719_dilations_1, groups = sparse_output_719_groups_1, pad = sparse_output_719_pad_1, pad_type = sparse_output_719_pad_type_1, strides = sparse_output_719_strides_1, weight = layers_8_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_719_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_11425_cast_fp16 = add(x = dense_output_719_cast_fp16, y = sparse_output_719_cast_fp16)[name = string("op_11425_cast_fp16")];
            tensor<int32, [4]> var_11426 = const()[name = string("op_11426"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11428 = const()[name = string("op_11428"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_11427_cast_fp16 = transpose(perm = var_11426, x = var_11425_cast_fp16)[name = string("transpose_621")];
            tensor<fp16, [1, 101, 128]> p_head_285_cast_fp16 = reshape(shape = var_11428, x = var_11427_cast_fp16)[name = string("p_head_285_cast_fp16")];
            tensor<fp16, [128]> var_11430_to_fp16 = const()[name = string("op_11430_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254291584)))];
            tensor<fp16, [1, 51, 128]> var_11431_cast_fp16 = add(x = q_head_143_cast_fp16, y = var_11430_to_fp16)[name = string("op_11431_cast_fp16")];
            tensor<int32, [1]> q_u_143_axes_1 = const()[name = string("q_u_143_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_143_cast_fp16 = expand_dims(axes = q_u_143_axes_1, x = var_11431_cast_fp16)[name = string("q_u_143_cast_fp16")];
            tensor<fp16, [128]> var_11433_to_fp16 = const()[name = string("op_11433_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254291904)))];
            tensor<fp16, [1, 51, 128]> var_11434_cast_fp16 = add(x = q_head_143_cast_fp16, y = var_11433_to_fp16)[name = string("op_11434_cast_fp16")];
            tensor<int32, [1]> q_v_143_axes_1 = const()[name = string("q_v_143_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_143_cast_fp16 = expand_dims(axes = q_v_143_axes_1, x = var_11434_cast_fp16)[name = string("q_v_143_cast_fp16")];
            tensor<int32, [1]> k_head_287_axes_1 = const()[name = string("k_head_287_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_287_cast_fp16 = expand_dims(axes = k_head_287_axes_1, x = k_head_285_cast_fp16)[name = string("k_head_287_cast_fp16")];
            tensor<int32, [1]> v_head_287_axes_1 = const()[name = string("v_head_287_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_287_cast_fp16 = expand_dims(axes = v_head_287_axes_1, x = v_head_285_cast_fp16)[name = string("v_head_287_cast_fp16")];
            tensor<int32, [1]> p_head_287_axes_1 = const()[name = string("p_head_287_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_287_cast_fp16 = expand_dims(axes = p_head_287_axes_1, x = p_head_285_cast_fp16)[name = string("p_head_287_cast_fp16")];
            bool var_11440_transpose_x_3 = const()[name = string("op_11440_transpose_x_3"), val = bool(false)];
            bool var_11440_transpose_y_3 = const()[name = string("op_11440_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_11440_cast_fp16 = matmul(transpose_x = var_11440_transpose_x_3, transpose_y = var_11440_transpose_y_3, x = q_u_143_cast_fp16, y = k_head_287_cast_fp16)[name = string("op_11440_cast_fp16")];
            fp16 var_11441_to_fp16 = const()[name = string("op_11441_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_143_cast_fp16 = mul(x = var_11440_cast_fp16, y = var_11441_to_fp16)[name = string("scores_content_143_cast_fp16")];
            bool x_749_transpose_x_3 = const()[name = string("x_749_transpose_x_3"), val = bool(false)];
            bool x_749_transpose_y_3 = const()[name = string("x_749_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_749_cast_fp16 = matmul(transpose_x = x_749_transpose_x_3, transpose_y = x_749_transpose_y_3, x = q_v_143_cast_fp16, y = p_head_287_cast_fp16)[name = string("x_749_cast_fp16")];
            tensor<int32, [8]> x_751_pad_1 = const()[name = string("x_751_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_751_mode_1 = const()[name = string("x_751_mode_1"), val = string("constant")];
            fp16 const_1783_to_fp16 = const()[name = string("const_1783_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_751_cast_fp16 = pad(constant_val = const_1783_to_fp16, mode = x_751_mode_1, pad = x_751_pad_1, x = x_749_cast_fp16)[name = string("x_751_cast_fp16")];
            tensor<int32, [4]> var_11455 = const()[name = string("op_11455"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_753_cast_fp16 = reshape(shape = var_11455, x = x_751_cast_fp16)[name = string("x_753_cast_fp16")];
            tensor<int32, [4]> var_11459_begin_1 = const()[name = string("op_11459_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11459_end_1 = const()[name = string("op_11459_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_11459_end_mask_1 = const()[name = string("op_11459_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_11459_cast_fp16 = slice_by_index(begin = var_11459_begin_1, end = var_11459_end_1, end_mask = var_11459_end_mask_1, x = x_753_cast_fp16)[name = string("op_11459_cast_fp16")];
            tensor<int32, [4]> var_11461 = const()[name = string("op_11461"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_11462_cast_fp16 = reshape(shape = var_11461, x = var_11459_cast_fp16)[name = string("op_11462_cast_fp16")];
            tensor<int32, [4]> var_11467_begin_1 = const()[name = string("op_11467_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11467_end_1 = const()[name = string("op_11467_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_11467_end_mask_1 = const()[name = string("op_11467_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_11467_cast_fp16 = slice_by_index(begin = var_11467_begin_1, end = var_11467_end_1, end_mask = var_11467_end_mask_1, x = var_11462_cast_fp16)[name = string("op_11467_cast_fp16")];
            fp16 var_11468_to_fp16 = const()[name = string("op_11468_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_143_cast_fp16 = mul(x = var_11467_cast_fp16, y = var_11468_to_fp16)[name = string("scores_pos_143_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_143_cast_fp16 = add(x = scores_content_143_cast_fp16, y = scores_pos_143_cast_fp16)[name = string("logits_143_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_11471_cast_fp16 = softmax(axis = var_10349, x = logits_143_cast_fp16)[name = string("op_11471_cast_fp16")];
            bool var_11473_transpose_x_1 = const()[name = string("op_11473_transpose_x_1"), val = bool(false)];
            bool var_11473_transpose_y_1 = const()[name = string("op_11473_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_11473_cast_fp16 = matmul(transpose_x = var_11473_transpose_x_1, transpose_y = var_11473_transpose_y_1, x = var_11471_cast_fp16, y = v_head_287_cast_fp16)[name = string("op_11473_cast_fp16")];
            tensor<int32, [1]> o_head_17_axes_1 = const()[name = string("o_head_17_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_17_cast_fp16 = squeeze(axes = o_head_17_axes_1, x = var_11473_cast_fp16)[name = string("o_head_17_cast_fp16")];
            bool out_17_interleave_1 = const()[name = string("out_17_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_17_cast_fp16 = concat(axis = var_10349, interleave = out_17_interleave_1, values = (var_10627_cast_fp16, var_10748_cast_fp16, var_10869_cast_fp16, var_10990_cast_fp16, var_11111_cast_fp16, var_11232_cast_fp16, var_11353_cast_fp16, o_head_17_cast_fp16))[name = string("out_17_cast_fp16")];
            tensor<int32, [3]> var_11477_perm_1 = const()[name = string("op_11477_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_405_axes_1 = const()[name = string("input_405_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_11477_cast_fp16 = transpose(perm = var_11477_perm_1, x = out_17_cast_fp16)[name = string("transpose_620")];
            tensor<fp16, [1, 1024, 51, 1]> input_405_cast_fp16 = expand_dims(axes = input_405_axes_1, x = var_11477_cast_fp16)[name = string("input_405_cast_fp16")];
            string dense_output_721_pad_type_1 = const()[name = string("dense_output_721_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_721_strides_1 = const()[name = string("dense_output_721_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_721_pad_1 = const()[name = string("dense_output_721_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_721_dilations_1 = const()[name = string("dense_output_721_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_721_groups_1 = const()[name = string("dense_output_721_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254292224))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255340864))))[name = string("layers_8_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_721_cast_fp16 = conv(dilations = dense_output_721_dilations_1, groups = dense_output_721_groups_1, pad = dense_output_721_pad_1, pad_type = dense_output_721_pad_type_1, strides = dense_output_721_strides_1, weight = layers_8_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_405_cast_fp16)[name = string("dense_output_721_cast_fp16")];
            string sparse_output_721_pad_type_1 = const()[name = string("sparse_output_721_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_721_strides_1 = const()[name = string("sparse_output_721_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_721_pad_1 = const()[name = string("sparse_output_721_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_721_dilations_1 = const()[name = string("sparse_output_721_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_721_groups_1 = const()[name = string("sparse_output_721_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255362496))), nonzero_data = tensor<fp16, [10487]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255341440))))[name = string("layers_8_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_721_cast_fp16 = conv(dilations = sparse_output_721_dilations_1, groups = sparse_output_721_groups_1, pad = sparse_output_721_pad_1, pad_type = sparse_output_721_pad_type_1, strides = sparse_output_721_strides_1, weight = layers_8_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_405_cast_fp16)[name = string("sparse_output_721_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_17_cast_fp16 = add(x = dense_output_721_cast_fp16, y = sparse_output_721_cast_fp16)[name = string("out_conv_17_cast_fp16")];
            tensor<int32, [1]> var_11494_axes_1 = const()[name = string("op_11494_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_11494_cast_fp16 = squeeze(axes = var_11494_axes_1, x = out_conv_17_cast_fp16)[name = string("op_11494_cast_fp16")];
            tensor<int32, [3]> var_11495_perm_1 = const()[name = string("op_11495_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_11495_cast_fp16 = transpose(perm = var_11495_perm_1, x = var_11494_cast_fp16)[name = string("transpose_619")];
            tensor<fp16, [1, 51, 1024]> input_407_cast_fp16 = add(x = input_395_cast_fp16, y = var_11495_cast_fp16)[name = string("input_407_cast_fp16")];
            tensor<int32, [1]> x_757_axes_1 = const()[name = string("x_757_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_8_norm_conv_weight_to_fp16 = const()[name = string("layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255493632)))];
            tensor<fp16, [1024]> layers_8_norm_conv_bias_to_fp16 = const()[name = string("layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255495744)))];
            tensor<fp16, [1, 51, 1024]> x_757_cast_fp16 = layer_norm(axes = x_757_axes_1, beta = layers_8_norm_conv_bias_to_fp16, epsilon = var_10364_to_fp16, gamma = layers_8_norm_conv_weight_to_fp16, x = input_407_cast_fp16)[name = string("x_757_cast_fp16")];
            tensor<int32, [3]> var_11505_perm_1 = const()[name = string("op_11505_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_409_axes_1 = const()[name = string("input_409_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_11505_cast_fp16 = transpose(perm = var_11505_perm_1, x = x_757_cast_fp16)[name = string("transpose_618")];
            tensor<fp16, [1, 1024, 51, 1]> input_409_cast_fp16 = expand_dims(axes = input_409_axes_1, x = var_11505_cast_fp16)[name = string("input_409_cast_fp16")];
            string dense_output_723_pad_type_1 = const()[name = string("dense_output_723_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_723_strides_1 = const()[name = string("dense_output_723_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_723_pad_1 = const()[name = string("dense_output_723_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_723_dilations_1 = const()[name = string("dense_output_723_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_723_groups_1 = const()[name = string("dense_output_723_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_8_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255497856))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257595072))))[name = string("layers_8_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_723_cast_fp16 = conv(dilations = dense_output_723_dilations_1, groups = dense_output_723_groups_1, pad = dense_output_723_pad_1, pad_type = dense_output_723_pad_type_1, strides = dense_output_723_strides_1, weight = layers_8_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_409_cast_fp16)[name = string("dense_output_723_cast_fp16")];
            string sparse_output_723_pad_type_1 = const()[name = string("sparse_output_723_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_723_strides_1 = const()[name = string("sparse_output_723_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_723_pad_1 = const()[name = string("sparse_output_723_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_723_dilations_1 = const()[name = string("sparse_output_723_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_723_groups_1 = const()[name = string("sparse_output_723_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_8_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257637696))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257595648))))[name = string("layers_8_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_723_cast_fp16 = conv(dilations = sparse_output_723_dilations_1, groups = sparse_output_723_groups_1, pad = sparse_output_723_pad_1, pad_type = sparse_output_723_pad_type_1, strides = sparse_output_723_strides_1, weight = layers_8_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_409_cast_fp16)[name = string("sparse_output_723_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_411_cast_fp16 = add(x = dense_output_723_cast_fp16, y = sparse_output_723_cast_fp16)[name = string("input_411_cast_fp16")];
            int32 input_413_split_num_splits_1 = const()[name = string("input_413_split_num_splits_1"), val = int32(2)];
            int32 input_413_split_axis_1 = const()[name = string("input_413_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_413_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_413_split_cast_fp16_1 = split(axis = input_413_split_axis_1, num_splits = input_413_split_num_splits_1, x = input_411_cast_fp16)[name = string("input_413_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_413_split_1_sigmoid_cast_fp16 = sigmoid(x = input_413_split_cast_fp16_1)[name = string("input_413_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_413_cast_fp16 = mul(x = input_413_split_cast_fp16_0, y = input_413_split_1_sigmoid_cast_fp16)[name = string("input_413_cast_fp16")];
            tensor<int32, [8]> input_415_pad_1 = const()[name = string("input_415_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_415_mode_1 = const()[name = string("input_415_mode_1"), val = string("constant")];
            fp16 const_1785_to_fp16 = const()[name = string("const_1785_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_415_cast_fp16 = pad(constant_val = const_1785_to_fp16, mode = input_415_mode_1, pad = input_415_pad_1, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")];
            string dense_output_725_pad_type_1 = const()[name = string("dense_output_725_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_725_strides_1 = const()[name = string("dense_output_725_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_725_pad_1 = const()[name = string("dense_output_725_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_725_dilations_1 = const()[name = string("dense_output_725_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_725_groups_1 = const()[name = string("dense_output_725_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_725_cast_fp16 = conv(dilations = dense_output_725_dilations_1, groups = dense_output_725_groups_1, pad = dense_output_725_pad_1, pad_type = dense_output_725_pad_type_1, strides = dense_output_725_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_415_cast_fp16)[name = string("dense_output_725_cast_fp16")];
            string sparse_output_725_pad_type_1 = const()[name = string("sparse_output_725_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_725_strides_1 = const()[name = string("sparse_output_725_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_725_pad_1 = const()[name = string("sparse_output_725_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_725_dilations_1 = const()[name = string("sparse_output_725_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_725_groups_1 = const()[name = string("sparse_output_725_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_8_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257899904))))[name = string("layers_8_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_725_cast_fp16 = conv(dilations = sparse_output_725_dilations_1, groups = sparse_output_725_groups_1, pad = sparse_output_725_pad_1, pad_type = sparse_output_725_pad_type_1, strides = sparse_output_725_strides_1, weight = layers_8_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_415_cast_fp16)[name = string("sparse_output_725_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_417_cast_fp16 = add(x = dense_output_725_cast_fp16, y = sparse_output_725_cast_fp16)[name = string("input_417_cast_fp16")];
            tensor<fp16, [1024]> layers_8_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_8_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257918400)))];
            tensor<fp16, [1024]> layers_8_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_8_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257920512)))];
            tensor<fp16, [1024]> layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257922624)))];
            tensor<fp16, [1024]> layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257924736)))];
            tensor<fp16, [1, 1024, 51, 1]> input_419_cast_fp16 = batch_norm(beta = layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_10364_to_fp16, gamma = layers_8_conv_batch_norm_weight_to_fp16, mean = layers_8_conv_batch_norm_running_mean_to_fp16, variance = layers_8_conv_batch_norm_running_var_to_fp16, x = input_417_cast_fp16)[name = string("input_419_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_421_cast_fp16 = silu(x = input_419_cast_fp16)[name = string("input_421_cast_fp16")];
            string dense_output_727_pad_type_1 = const()[name = string("dense_output_727_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_727_strides_1 = const()[name = string("dense_output_727_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_727_pad_1 = const()[name = string("dense_output_727_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_727_dilations_1 = const()[name = string("dense_output_727_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_727_groups_1 = const()[name = string("dense_output_727_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257926848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258975488))))[name = string("layers_8_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_727_cast_fp16 = conv(dilations = dense_output_727_dilations_1, groups = dense_output_727_groups_1, pad = dense_output_727_pad_1, pad_type = dense_output_727_pad_type_1, strides = dense_output_727_strides_1, weight = layers_8_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_421_cast_fp16)[name = string("dense_output_727_cast_fp16")];
            string sparse_output_727_pad_type_1 = const()[name = string("sparse_output_727_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_727_strides_1 = const()[name = string("sparse_output_727_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_727_pad_1 = const()[name = string("sparse_output_727_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_727_dilations_1 = const()[name = string("sparse_output_727_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_727_groups_1 = const()[name = string("sparse_output_727_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_8_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258997120))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258976064))))[name = string("layers_8_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_727_cast_fp16 = conv(dilations = sparse_output_727_dilations_1, groups = sparse_output_727_groups_1, pad = sparse_output_727_pad_1, pad_type = sparse_output_727_pad_type_1, strides = sparse_output_727_strides_1, weight = layers_8_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_421_cast_fp16)[name = string("sparse_output_727_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_759_cast_fp16 = add(x = dense_output_727_cast_fp16, y = sparse_output_727_cast_fp16)[name = string("x_759_cast_fp16")];
            tensor<int32, [1]> var_11561_axes_1 = const()[name = string("op_11561_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_11561_cast_fp16 = squeeze(axes = var_11561_axes_1, x = x_759_cast_fp16)[name = string("op_11561_cast_fp16")];
            tensor<int32, [3]> var_11562_perm_1 = const()[name = string("op_11562_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_11562_cast_fp16 = transpose(perm = var_11562_perm_1, x = var_11561_cast_fp16)[name = string("transpose_617")];
            tensor<fp16, [1, 51, 1024]> input_423_cast_fp16 = add(x = input_407_cast_fp16, y = var_11562_cast_fp16)[name = string("input_423_cast_fp16")];
            tensor<int32, [1]> x_761_axes_1 = const()[name = string("x_761_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259128256)))];
            tensor<fp16, [1024]> layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259130368)))];
            tensor<fp16, [1, 51, 1024]> x_761_cast_fp16 = layer_norm(axes = x_761_axes_1, beta = layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_10364_to_fp16, gamma = layers_8_norm_feed_forward2_weight_to_fp16, x = input_423_cast_fp16)[name = string("x_761_cast_fp16")];
            tensor<int32, [4]> var_11572 = const()[name = string("op_11572"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_763_cast_fp16 = reshape(shape = var_11572, x = x_761_cast_fp16)[name = string("x_763_cast_fp16")];
            tensor<int32, [4]> input_425_perm_1 = const()[name = string("input_425_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_729_pad_type_1 = const()[name = string("dense_output_729_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_729_strides_1 = const()[name = string("dense_output_729_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_729_pad_1 = const()[name = string("dense_output_729_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_729_dilations_1 = const()[name = string("dense_output_729_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_729_groups_1 = const()[name = string("dense_output_729_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_8_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259132480))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263326848))))[name = string("layers_8_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_425_cast_fp16 = transpose(perm = input_425_perm_1, x = x_763_cast_fp16)[name = string("transpose_616")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_729_cast_fp16 = conv(dilations = dense_output_729_dilations_1, groups = dense_output_729_groups_1, pad = dense_output_729_pad_1, pad_type = dense_output_729_pad_type_1, strides = dense_output_729_strides_1, weight = layers_8_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_425_cast_fp16)[name = string("dense_output_729_cast_fp16")];
            string sparse_output_729_pad_type_1 = const()[name = string("sparse_output_729_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_729_strides_1 = const()[name = string("sparse_output_729_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_729_pad_1 = const()[name = string("sparse_output_729_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_729_dilations_1 = const()[name = string("sparse_output_729_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_729_groups_1 = const()[name = string("sparse_output_729_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_8_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263411392))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263327424))))[name = string("layers_8_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_729_cast_fp16 = conv(dilations = sparse_output_729_dilations_1, groups = sparse_output_729_groups_1, pad = sparse_output_729_pad_1, pad_type = sparse_output_729_pad_type_1, strides = sparse_output_729_strides_1, weight = layers_8_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_425_cast_fp16)[name = string("sparse_output_729_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_427_cast_fp16 = add(x = dense_output_729_cast_fp16, y = sparse_output_729_cast_fp16)[name = string("input_427_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_429_cast_fp16 = silu(x = input_427_cast_fp16)[name = string("input_429_cast_fp16")];
            string dense_output_731_pad_type_1 = const()[name = string("dense_output_731_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_731_strides_1 = const()[name = string("dense_output_731_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_731_pad_1 = const()[name = string("dense_output_731_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_731_dilations_1 = const()[name = string("dense_output_731_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_731_groups_1 = const()[name = string("dense_output_731_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_8_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263935744))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268130112))))[name = string("layers_8_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_731_cast_fp16 = conv(dilations = dense_output_731_dilations_1, groups = dense_output_731_groups_1, pad = dense_output_731_pad_1, pad_type = dense_output_731_pad_type_1, strides = dense_output_731_strides_1, weight = layers_8_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("dense_output_731_cast_fp16")];
            string sparse_output_731_pad_type_1 = const()[name = string("sparse_output_731_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_731_strides_1 = const()[name = string("sparse_output_731_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_731_pad_1 = const()[name = string("sparse_output_731_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_731_dilations_1 = const()[name = string("sparse_output_731_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_731_groups_1 = const()[name = string("sparse_output_731_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_8_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268214656))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268130688))))[name = string("layers_8_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_731_cast_fp16 = conv(dilations = sparse_output_731_dilations_1, groups = sparse_output_731_groups_1, pad = sparse_output_731_pad_1, pad_type = sparse_output_731_pad_type_1, strides = sparse_output_731_strides_1, weight = layers_8_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_429_cast_fp16)[name = string("sparse_output_731_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_765_cast_fp16 = add(x = dense_output_731_cast_fp16, y = sparse_output_731_cast_fp16)[name = string("x_765_cast_fp16")];
            tensor<int32, [4]> x_767_perm_1 = const()[name = string("x_767_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_11607 = const()[name = string("op_11607"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_767_cast_fp16 = transpose(perm = x_767_perm_1, x = x_765_cast_fp16)[name = string("transpose_615")];
            tensor<fp16, [1, 51, 1024]> var_11608_cast_fp16 = reshape(shape = var_11607, x = x_767_cast_fp16)[name = string("op_11608_cast_fp16")];
            fp16 var_11609_to_fp16 = const()[name = string("op_11609_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_11610_cast_fp16 = mul(x = var_11608_cast_fp16, y = var_11609_to_fp16)[name = string("op_11610_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_431_cast_fp16 = add(x = input_423_cast_fp16, y = var_11610_cast_fp16)[name = string("input_431_cast_fp16")];
            tensor<int32, [1]> input_433_axes_1 = const()[name = string("input_433_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_8_norm_out_weight_to_fp16 = const()[name = string("layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268739008)))];
            tensor<fp16, [1024]> layers_8_norm_out_bias_to_fp16 = const()[name = string("layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268741120)))];
            tensor<fp16, [1, 51, 1024]> input_433_cast_fp16 = layer_norm(axes = input_433_axes_1, beta = layers_8_norm_out_bias_to_fp16, epsilon = var_10364_to_fp16, gamma = layers_8_norm_out_weight_to_fp16, x = input_431_cast_fp16)[name = string("input_433_cast_fp16")];
            int32 var_11618 = const()[name = string("op_11618"), val = int32(-1)];
            tensor<int32, [1]> x_769_axes_1 = const()[name = string("x_769_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268743232)))];
            tensor<fp16, [1024]> layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268745344)))];
            fp16 var_11633_to_fp16 = const()[name = string("op_11633_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_769_cast_fp16 = layer_norm(axes = x_769_axes_1, beta = layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_11633_to_fp16, gamma = layers_9_norm_feed_forward1_weight_to_fp16, x = input_433_cast_fp16)[name = string("x_769_cast_fp16")];
            tensor<int32, [4]> var_11652 = const()[name = string("op_11652"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_771_cast_fp16 = reshape(shape = var_11652, x = x_769_cast_fp16)[name = string("x_771_cast_fp16")];
            tensor<int32, [4]> input_435_perm_1 = const()[name = string("input_435_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_733_pad_type_1 = const()[name = string("dense_output_733_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_733_strides_1 = const()[name = string("dense_output_733_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_733_pad_1 = const()[name = string("dense_output_733_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_733_dilations_1 = const()[name = string("dense_output_733_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_733_groups_1 = const()[name = string("dense_output_733_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_9_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268747456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272941824))))[name = string("layers_9_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_435_cast_fp16 = transpose(perm = input_435_perm_1, x = x_771_cast_fp16)[name = string("transpose_614")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_733_cast_fp16 = conv(dilations = dense_output_733_dilations_1, groups = dense_output_733_groups_1, pad = dense_output_733_pad_1, pad_type = dense_output_733_pad_type_1, strides = dense_output_733_strides_1, weight = layers_9_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_435_cast_fp16)[name = string("dense_output_733_cast_fp16")];
            string sparse_output_733_pad_type_1 = const()[name = string("sparse_output_733_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_733_strides_1 = const()[name = string("sparse_output_733_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_733_pad_1 = const()[name = string("sparse_output_733_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_733_dilations_1 = const()[name = string("sparse_output_733_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_733_groups_1 = const()[name = string("sparse_output_733_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_9_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273026368))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272942400))))[name = string("layers_9_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_733_cast_fp16 = conv(dilations = sparse_output_733_dilations_1, groups = sparse_output_733_groups_1, pad = sparse_output_733_pad_1, pad_type = sparse_output_733_pad_type_1, strides = sparse_output_733_strides_1, weight = layers_9_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_435_cast_fp16)[name = string("sparse_output_733_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_437_cast_fp16 = add(x = dense_output_733_cast_fp16, y = sparse_output_733_cast_fp16)[name = string("input_437_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_439_cast_fp16 = silu(x = input_437_cast_fp16)[name = string("input_439_cast_fp16")];
            string dense_output_735_pad_type_1 = const()[name = string("dense_output_735_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_735_strides_1 = const()[name = string("dense_output_735_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_735_pad_1 = const()[name = string("dense_output_735_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_735_dilations_1 = const()[name = string("dense_output_735_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_735_groups_1 = const()[name = string("dense_output_735_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_9_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273550720))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277745088))))[name = string("layers_9_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_735_cast_fp16 = conv(dilations = dense_output_735_dilations_1, groups = dense_output_735_groups_1, pad = dense_output_735_pad_1, pad_type = dense_output_735_pad_type_1, strides = dense_output_735_strides_1, weight = layers_9_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_439_cast_fp16)[name = string("dense_output_735_cast_fp16")];
            string sparse_output_735_pad_type_1 = const()[name = string("sparse_output_735_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_735_strides_1 = const()[name = string("sparse_output_735_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_735_pad_1 = const()[name = string("sparse_output_735_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_735_dilations_1 = const()[name = string("sparse_output_735_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_735_groups_1 = const()[name = string("sparse_output_735_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_9_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277829632))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277745664))))[name = string("layers_9_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_735_cast_fp16 = conv(dilations = sparse_output_735_dilations_1, groups = sparse_output_735_groups_1, pad = sparse_output_735_pad_1, pad_type = sparse_output_735_pad_type_1, strides = sparse_output_735_strides_1, weight = layers_9_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_439_cast_fp16)[name = string("sparse_output_735_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_773_cast_fp16 = add(x = dense_output_735_cast_fp16, y = sparse_output_735_cast_fp16)[name = string("x_773_cast_fp16")];
            tensor<int32, [4]> x_775_perm_1 = const()[name = string("x_775_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_11687 = const()[name = string("op_11687"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_775_cast_fp16 = transpose(perm = x_775_perm_1, x = x_773_cast_fp16)[name = string("transpose_613")];
            tensor<fp16, [1, 51, 1024]> var_11688_cast_fp16 = reshape(shape = var_11687, x = x_775_cast_fp16)[name = string("op_11688_cast_fp16")];
            fp16 var_11689_to_fp16 = const()[name = string("op_11689_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_11690_cast_fp16 = mul(x = var_11688_cast_fp16, y = var_11689_to_fp16)[name = string("op_11690_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_441_cast_fp16 = add(x = input_433_cast_fp16, y = var_11690_cast_fp16)[name = string("input_441_cast_fp16")];
            tensor<int32, [1]> q_19_axes_1 = const()[name = string("q_19_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_9_norm_self_att_weight_to_fp16 = const()[name = string("layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278353984)))];
            tensor<fp16, [1024]> layers_9_norm_self_att_bias_to_fp16 = const()[name = string("layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278356096)))];
            tensor<fp16, [1, 51, 1024]> q_19_cast_fp16 = layer_norm(axes = q_19_axes_1, beta = layers_9_norm_self_att_bias_to_fp16, epsilon = var_11633_to_fp16, gamma = layers_9_norm_self_att_weight_to_fp16, x = input_441_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<int32, [3]> var_11764 = const()[name = string("op_11764"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_443_axes_1 = const()[name = string("input_443_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_11765_cast_fp16 = transpose(perm = var_11764, x = q_19_cast_fp16)[name = string("transpose_612")];
            tensor<fp16, [1, 1024, 51, 1]> input_443_cast_fp16 = expand_dims(axes = input_443_axes_1, x = var_11765_cast_fp16)[name = string("input_443_cast_fp16")];
            string dense_output_737_pad_type_1 = const()[name = string("dense_output_737_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_737_strides_1 = const()[name = string("dense_output_737_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_737_pad_1 = const()[name = string("dense_output_737_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_737_dilations_1 = const()[name = string("dense_output_737_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_737_groups_1 = const()[name = string("dense_output_737_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278358208))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278489344))))[name = string("layers_9_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_737_cast_fp16 = conv(dilations = dense_output_737_dilations_1, groups = dense_output_737_groups_1, pad = dense_output_737_pad_1, pad_type = dense_output_737_pad_type_1, strides = dense_output_737_strides_1, weight = layers_9_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_737_cast_fp16")];
            string sparse_output_737_pad_type_1 = const()[name = string("sparse_output_737_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_737_strides_1 = const()[name = string("sparse_output_737_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_737_pad_1 = const()[name = string("sparse_output_737_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_737_dilations_1 = const()[name = string("sparse_output_737_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_737_groups_1 = const()[name = string("sparse_output_737_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278492608))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278489920))))[name = string("layers_9_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_737_cast_fp16 = conv(dilations = sparse_output_737_dilations_1, groups = sparse_output_737_groups_1, pad = sparse_output_737_pad_1, pad_type = sparse_output_737_pad_type_1, strides = sparse_output_737_strides_1, weight = layers_9_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_737_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11790_cast_fp16 = add(x = dense_output_737_cast_fp16, y = sparse_output_737_cast_fp16)[name = string("op_11790_cast_fp16")];
            tensor<int32, [4]> var_11791 = const()[name = string("op_11791"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11793 = const()[name = string("op_11793"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11792_cast_fp16 = transpose(perm = var_11791, x = var_11790_cast_fp16)[name = string("transpose_611")];
            tensor<fp16, [1, 51, 128]> q_head_145_cast_fp16 = reshape(shape = var_11793, x = var_11792_cast_fp16)[name = string("q_head_145_cast_fp16")];
            string dense_output_739_pad_type_1 = const()[name = string("dense_output_739_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_739_strides_1 = const()[name = string("dense_output_739_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_739_pad_1 = const()[name = string("dense_output_739_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_739_dilations_1 = const()[name = string("dense_output_739_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_739_groups_1 = const()[name = string("dense_output_739_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278509056))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278640192))))[name = string("layers_9_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_739_cast_fp16 = conv(dilations = dense_output_739_dilations_1, groups = dense_output_739_groups_1, pad = dense_output_739_pad_1, pad_type = dense_output_739_pad_type_1, strides = dense_output_739_strides_1, weight = layers_9_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_739_cast_fp16")];
            string sparse_output_739_pad_type_1 = const()[name = string("sparse_output_739_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_739_strides_1 = const()[name = string("sparse_output_739_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_739_pad_1 = const()[name = string("sparse_output_739_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_739_dilations_1 = const()[name = string("sparse_output_739_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_739_groups_1 = const()[name = string("sparse_output_739_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278643456))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278640768))))[name = string("layers_9_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_739_cast_fp16 = conv(dilations = sparse_output_739_dilations_1, groups = sparse_output_739_groups_1, pad = sparse_output_739_pad_1, pad_type = sparse_output_739_pad_type_1, strides = sparse_output_739_strides_1, weight = layers_9_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_739_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11809_cast_fp16 = add(x = dense_output_739_cast_fp16, y = sparse_output_739_cast_fp16)[name = string("op_11809_cast_fp16")];
            tensor<int32, [4]> var_11810 = const()[name = string("op_11810"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11812 = const()[name = string("op_11812"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11811_cast_fp16 = transpose(perm = var_11810, x = var_11809_cast_fp16)[name = string("transpose_610")];
            tensor<fp16, [1, 51, 128]> k_head_289_cast_fp16 = reshape(shape = var_11812, x = var_11811_cast_fp16)[name = string("k_head_289_cast_fp16")];
            string dense_output_741_pad_type_1 = const()[name = string("dense_output_741_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_741_strides_1 = const()[name = string("dense_output_741_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_741_pad_1 = const()[name = string("dense_output_741_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_741_dilations_1 = const()[name = string("dense_output_741_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_741_groups_1 = const()[name = string("dense_output_741_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278659904))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278791040))))[name = string("layers_9_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_741_cast_fp16 = conv(dilations = dense_output_741_dilations_1, groups = dense_output_741_groups_1, pad = dense_output_741_pad_1, pad_type = dense_output_741_pad_type_1, strides = dense_output_741_strides_1, weight = layers_9_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_741_cast_fp16")];
            string sparse_output_741_pad_type_1 = const()[name = string("sparse_output_741_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_741_strides_1 = const()[name = string("sparse_output_741_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_741_pad_1 = const()[name = string("sparse_output_741_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_741_dilations_1 = const()[name = string("sparse_output_741_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_741_groups_1 = const()[name = string("sparse_output_741_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278794304))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278791616))))[name = string("layers_9_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_741_cast_fp16 = conv(dilations = sparse_output_741_dilations_1, groups = sparse_output_741_groups_1, pad = sparse_output_741_pad_1, pad_type = sparse_output_741_pad_type_1, strides = sparse_output_741_strides_1, weight = layers_9_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_741_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11828_cast_fp16 = add(x = dense_output_741_cast_fp16, y = sparse_output_741_cast_fp16)[name = string("op_11828_cast_fp16")];
            tensor<int32, [4]> var_11829 = const()[name = string("op_11829"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11831 = const()[name = string("op_11831"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11830_cast_fp16 = transpose(perm = var_11829, x = var_11828_cast_fp16)[name = string("transpose_609")];
            tensor<fp16, [1, 51, 128]> v_head_289_cast_fp16 = reshape(shape = var_11831, x = var_11830_cast_fp16)[name = string("v_head_289_cast_fp16")];
            string dense_output_743_pad_type_1 = const()[name = string("dense_output_743_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_743_strides_1 = const()[name = string("dense_output_743_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_743_pad_1 = const()[name = string("dense_output_743_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_743_dilations_1 = const()[name = string("dense_output_743_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_743_groups_1 = const()[name = string("dense_output_743_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278810752))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278941888))))[name = string("layers_9_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_743_cast_fp16 = conv(dilations = dense_output_743_dilations_1, groups = dense_output_743_groups_1, pad = dense_output_743_pad_1, pad_type = dense_output_743_pad_type_1, strides = dense_output_743_strides_1, weight = layers_9_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_743_cast_fp16")];
            string sparse_output_743_pad_type_1 = const()[name = string("sparse_output_743_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_743_strides_1 = const()[name = string("sparse_output_743_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_743_pad_1 = const()[name = string("sparse_output_743_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_743_dilations_1 = const()[name = string("sparse_output_743_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_743_groups_1 = const()[name = string("sparse_output_743_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278945152))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278942464))))[name = string("layers_9_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_743_cast_fp16 = conv(dilations = sparse_output_743_dilations_1, groups = sparse_output_743_groups_1, pad = sparse_output_743_pad_1, pad_type = sparse_output_743_pad_type_1, strides = sparse_output_743_strides_1, weight = layers_9_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_743_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_11847_cast_fp16 = add(x = dense_output_743_cast_fp16, y = sparse_output_743_cast_fp16)[name = string("op_11847_cast_fp16")];
            tensor<int32, [4]> var_11848 = const()[name = string("op_11848"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11850 = const()[name = string("op_11850"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_11849_cast_fp16 = transpose(perm = var_11848, x = var_11847_cast_fp16)[name = string("transpose_608")];
            tensor<fp16, [1, 101, 128]> p_head_289_cast_fp16 = reshape(shape = var_11850, x = var_11849_cast_fp16)[name = string("p_head_289_cast_fp16")];
            tensor<fp16, [128]> var_11852_to_fp16 = const()[name = string("op_11852_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278961600)))];
            tensor<fp16, [1, 51, 128]> var_11853_cast_fp16 = add(x = q_head_145_cast_fp16, y = var_11852_to_fp16)[name = string("op_11853_cast_fp16")];
            tensor<int32, [1]> q_u_145_axes_1 = const()[name = string("q_u_145_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_145_cast_fp16 = expand_dims(axes = q_u_145_axes_1, x = var_11853_cast_fp16)[name = string("q_u_145_cast_fp16")];
            tensor<fp16, [128]> var_11855_to_fp16 = const()[name = string("op_11855_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278961920)))];
            tensor<fp16, [1, 51, 128]> var_11856_cast_fp16 = add(x = q_head_145_cast_fp16, y = var_11855_to_fp16)[name = string("op_11856_cast_fp16")];
            tensor<int32, [1]> q_v_145_axes_1 = const()[name = string("q_v_145_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_145_cast_fp16 = expand_dims(axes = q_v_145_axes_1, x = var_11856_cast_fp16)[name = string("q_v_145_cast_fp16")];
            tensor<int32, [1]> k_head_291_axes_1 = const()[name = string("k_head_291_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_291_cast_fp16 = expand_dims(axes = k_head_291_axes_1, x = k_head_289_cast_fp16)[name = string("k_head_291_cast_fp16")];
            tensor<int32, [1]> v_head_291_axes_1 = const()[name = string("v_head_291_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_291_cast_fp16 = expand_dims(axes = v_head_291_axes_1, x = v_head_289_cast_fp16)[name = string("v_head_291_cast_fp16")];
            tensor<int32, [1]> p_head_291_axes_1 = const()[name = string("p_head_291_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_291_cast_fp16 = expand_dims(axes = p_head_291_axes_1, x = p_head_289_cast_fp16)[name = string("p_head_291_cast_fp16")];
            bool var_11862_transpose_x_3 = const()[name = string("op_11862_transpose_x_3"), val = bool(false)];
            bool var_11862_transpose_y_3 = const()[name = string("op_11862_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_11862_cast_fp16 = matmul(transpose_x = var_11862_transpose_x_3, transpose_y = var_11862_transpose_y_3, x = q_u_145_cast_fp16, y = k_head_291_cast_fp16)[name = string("op_11862_cast_fp16")];
            fp16 var_11863_to_fp16 = const()[name = string("op_11863_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_145_cast_fp16 = mul(x = var_11862_cast_fp16, y = var_11863_to_fp16)[name = string("scores_content_145_cast_fp16")];
            bool x_777_transpose_x_3 = const()[name = string("x_777_transpose_x_3"), val = bool(false)];
            bool x_777_transpose_y_3 = const()[name = string("x_777_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_777_cast_fp16 = matmul(transpose_x = x_777_transpose_x_3, transpose_y = x_777_transpose_y_3, x = q_v_145_cast_fp16, y = p_head_291_cast_fp16)[name = string("x_777_cast_fp16")];
            tensor<int32, [8]> x_779_pad_1 = const()[name = string("x_779_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_779_mode_1 = const()[name = string("x_779_mode_1"), val = string("constant")];
            fp16 const_1795_to_fp16 = const()[name = string("const_1795_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_779_cast_fp16 = pad(constant_val = const_1795_to_fp16, mode = x_779_mode_1, pad = x_779_pad_1, x = x_777_cast_fp16)[name = string("x_779_cast_fp16")];
            tensor<int32, [4]> var_11877 = const()[name = string("op_11877"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_781_cast_fp16 = reshape(shape = var_11877, x = x_779_cast_fp16)[name = string("x_781_cast_fp16")];
            tensor<int32, [4]> var_11881_begin_1 = const()[name = string("op_11881_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_11881_end_1 = const()[name = string("op_11881_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_11881_end_mask_1 = const()[name = string("op_11881_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_11881_cast_fp16 = slice_by_index(begin = var_11881_begin_1, end = var_11881_end_1, end_mask = var_11881_end_mask_1, x = x_781_cast_fp16)[name = string("op_11881_cast_fp16")];
            tensor<int32, [4]> var_11883 = const()[name = string("op_11883"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_11884_cast_fp16 = reshape(shape = var_11883, x = var_11881_cast_fp16)[name = string("op_11884_cast_fp16")];
            tensor<int32, [4]> var_11889_begin_1 = const()[name = string("op_11889_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_11889_end_1 = const()[name = string("op_11889_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_11889_end_mask_1 = const()[name = string("op_11889_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_11889_cast_fp16 = slice_by_index(begin = var_11889_begin_1, end = var_11889_end_1, end_mask = var_11889_end_mask_1, x = var_11884_cast_fp16)[name = string("op_11889_cast_fp16")];
            fp16 var_11890_to_fp16 = const()[name = string("op_11890_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_145_cast_fp16 = mul(x = var_11889_cast_fp16, y = var_11890_to_fp16)[name = string("scores_pos_145_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_145_cast_fp16 = add(x = scores_content_145_cast_fp16, y = scores_pos_145_cast_fp16)[name = string("logits_145_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_11893_cast_fp16 = softmax(axis = var_11618, x = logits_145_cast_fp16)[name = string("op_11893_cast_fp16")];
            bool var_11895_transpose_x_1 = const()[name = string("op_11895_transpose_x_1"), val = bool(false)];
            bool var_11895_transpose_y_1 = const()[name = string("op_11895_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_11895_cast_fp16 = matmul(transpose_x = var_11895_transpose_x_1, transpose_y = var_11895_transpose_y_1, x = var_11893_cast_fp16, y = v_head_291_cast_fp16)[name = string("op_11895_cast_fp16")];
            tensor<int32, [1]> var_11896_axes_1 = const()[name = string("op_11896_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_11896_cast_fp16 = squeeze(axes = var_11896_axes_1, x = var_11895_cast_fp16)[name = string("op_11896_cast_fp16")];
            string dense_output_745_pad_type_1 = const()[name = string("dense_output_745_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_745_strides_1 = const()[name = string("dense_output_745_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_745_pad_1 = const()[name = string("dense_output_745_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_745_dilations_1 = const()[name = string("dense_output_745_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_745_groups_1 = const()[name = string("dense_output_745_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278962240))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279093376))))[name = string("layers_9_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_745_cast_fp16 = conv(dilations = dense_output_745_dilations_1, groups = dense_output_745_groups_1, pad = dense_output_745_pad_1, pad_type = dense_output_745_pad_type_1, strides = dense_output_745_strides_1, weight = layers_9_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_745_cast_fp16")];
            string sparse_output_745_pad_type_1 = const()[name = string("sparse_output_745_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_745_strides_1 = const()[name = string("sparse_output_745_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_745_pad_1 = const()[name = string("sparse_output_745_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_745_dilations_1 = const()[name = string("sparse_output_745_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_745_groups_1 = const()[name = string("sparse_output_745_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279096640))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279093952))))[name = string("layers_9_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_745_cast_fp16 = conv(dilations = sparse_output_745_dilations_1, groups = sparse_output_745_groups_1, pad = sparse_output_745_pad_1, pad_type = sparse_output_745_pad_type_1, strides = sparse_output_745_strides_1, weight = layers_9_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_745_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11911_cast_fp16 = add(x = dense_output_745_cast_fp16, y = sparse_output_745_cast_fp16)[name = string("op_11911_cast_fp16")];
            tensor<int32, [4]> var_11912 = const()[name = string("op_11912"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11914 = const()[name = string("op_11914"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11913_cast_fp16 = transpose(perm = var_11912, x = var_11911_cast_fp16)[name = string("transpose_607")];
            tensor<fp16, [1, 51, 128]> q_head_147_cast_fp16 = reshape(shape = var_11914, x = var_11913_cast_fp16)[name = string("q_head_147_cast_fp16")];
            string dense_output_747_pad_type_1 = const()[name = string("dense_output_747_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_747_strides_1 = const()[name = string("dense_output_747_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_747_pad_1 = const()[name = string("dense_output_747_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_747_dilations_1 = const()[name = string("dense_output_747_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_747_groups_1 = const()[name = string("dense_output_747_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279113088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279244224))))[name = string("layers_9_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_747_cast_fp16 = conv(dilations = dense_output_747_dilations_1, groups = dense_output_747_groups_1, pad = dense_output_747_pad_1, pad_type = dense_output_747_pad_type_1, strides = dense_output_747_strides_1, weight = layers_9_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_747_cast_fp16")];
            string sparse_output_747_pad_type_1 = const()[name = string("sparse_output_747_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_747_strides_1 = const()[name = string("sparse_output_747_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_747_pad_1 = const()[name = string("sparse_output_747_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_747_dilations_1 = const()[name = string("sparse_output_747_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_747_groups_1 = const()[name = string("sparse_output_747_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279247488))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279244800))))[name = string("layers_9_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_747_cast_fp16 = conv(dilations = sparse_output_747_dilations_1, groups = sparse_output_747_groups_1, pad = sparse_output_747_pad_1, pad_type = sparse_output_747_pad_type_1, strides = sparse_output_747_strides_1, weight = layers_9_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_747_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11930_cast_fp16 = add(x = dense_output_747_cast_fp16, y = sparse_output_747_cast_fp16)[name = string("op_11930_cast_fp16")];
            tensor<int32, [4]> var_11931 = const()[name = string("op_11931"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11933 = const()[name = string("op_11933"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11932_cast_fp16 = transpose(perm = var_11931, x = var_11930_cast_fp16)[name = string("transpose_606")];
            tensor<fp16, [1, 51, 128]> k_head_293_cast_fp16 = reshape(shape = var_11933, x = var_11932_cast_fp16)[name = string("k_head_293_cast_fp16")];
            string dense_output_749_pad_type_1 = const()[name = string("dense_output_749_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_749_strides_1 = const()[name = string("dense_output_749_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_749_pad_1 = const()[name = string("dense_output_749_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_749_dilations_1 = const()[name = string("dense_output_749_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_749_groups_1 = const()[name = string("dense_output_749_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279263936))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279395072))))[name = string("layers_9_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_749_cast_fp16 = conv(dilations = dense_output_749_dilations_1, groups = dense_output_749_groups_1, pad = dense_output_749_pad_1, pad_type = dense_output_749_pad_type_1, strides = dense_output_749_strides_1, weight = layers_9_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_749_cast_fp16")];
            string sparse_output_749_pad_type_1 = const()[name = string("sparse_output_749_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_749_strides_1 = const()[name = string("sparse_output_749_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_749_pad_1 = const()[name = string("sparse_output_749_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_749_dilations_1 = const()[name = string("sparse_output_749_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_749_groups_1 = const()[name = string("sparse_output_749_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279398336))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279395648))))[name = string("layers_9_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_749_cast_fp16 = conv(dilations = sparse_output_749_dilations_1, groups = sparse_output_749_groups_1, pad = sparse_output_749_pad_1, pad_type = sparse_output_749_pad_type_1, strides = sparse_output_749_strides_1, weight = layers_9_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_749_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_11949_cast_fp16 = add(x = dense_output_749_cast_fp16, y = sparse_output_749_cast_fp16)[name = string("op_11949_cast_fp16")];
            tensor<int32, [4]> var_11950 = const()[name = string("op_11950"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11952 = const()[name = string("op_11952"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_11951_cast_fp16 = transpose(perm = var_11950, x = var_11949_cast_fp16)[name = string("transpose_605")];
            tensor<fp16, [1, 51, 128]> v_head_293_cast_fp16 = reshape(shape = var_11952, x = var_11951_cast_fp16)[name = string("v_head_293_cast_fp16")];
            string dense_output_751_pad_type_1 = const()[name = string("dense_output_751_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_751_strides_1 = const()[name = string("dense_output_751_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_751_pad_1 = const()[name = string("dense_output_751_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_751_dilations_1 = const()[name = string("dense_output_751_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_751_groups_1 = const()[name = string("dense_output_751_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279414784))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279545920))))[name = string("layers_9_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_751_cast_fp16 = conv(dilations = dense_output_751_dilations_1, groups = dense_output_751_groups_1, pad = dense_output_751_pad_1, pad_type = dense_output_751_pad_type_1, strides = dense_output_751_strides_1, weight = layers_9_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_751_cast_fp16")];
            string sparse_output_751_pad_type_1 = const()[name = string("sparse_output_751_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_751_strides_1 = const()[name = string("sparse_output_751_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_751_pad_1 = const()[name = string("sparse_output_751_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_751_dilations_1 = const()[name = string("sparse_output_751_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_751_groups_1 = const()[name = string("sparse_output_751_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279549184))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279546496))))[name = string("layers_9_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_751_cast_fp16 = conv(dilations = sparse_output_751_dilations_1, groups = sparse_output_751_groups_1, pad = sparse_output_751_pad_1, pad_type = sparse_output_751_pad_type_1, strides = sparse_output_751_strides_1, weight = layers_9_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_751_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_11968_cast_fp16 = add(x = dense_output_751_cast_fp16, y = sparse_output_751_cast_fp16)[name = string("op_11968_cast_fp16")];
            tensor<int32, [4]> var_11969 = const()[name = string("op_11969"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_11971 = const()[name = string("op_11971"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_11970_cast_fp16 = transpose(perm = var_11969, x = var_11968_cast_fp16)[name = string("transpose_604")];
            tensor<fp16, [1, 101, 128]> p_head_293_cast_fp16 = reshape(shape = var_11971, x = var_11970_cast_fp16)[name = string("p_head_293_cast_fp16")];
            tensor<fp16, [128]> var_11973_to_fp16 = const()[name = string("op_11973_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279565632)))];
            tensor<fp16, [1, 51, 128]> var_11974_cast_fp16 = add(x = q_head_147_cast_fp16, y = var_11973_to_fp16)[name = string("op_11974_cast_fp16")];
            tensor<int32, [1]> q_u_147_axes_1 = const()[name = string("q_u_147_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_147_cast_fp16 = expand_dims(axes = q_u_147_axes_1, x = var_11974_cast_fp16)[name = string("q_u_147_cast_fp16")];
            tensor<fp16, [128]> var_11976_to_fp16 = const()[name = string("op_11976_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279565952)))];
            tensor<fp16, [1, 51, 128]> var_11977_cast_fp16 = add(x = q_head_147_cast_fp16, y = var_11976_to_fp16)[name = string("op_11977_cast_fp16")];
            tensor<int32, [1]> q_v_147_axes_1 = const()[name = string("q_v_147_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_147_cast_fp16 = expand_dims(axes = q_v_147_axes_1, x = var_11977_cast_fp16)[name = string("q_v_147_cast_fp16")];
            tensor<int32, [1]> k_head_295_axes_1 = const()[name = string("k_head_295_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_295_cast_fp16 = expand_dims(axes = k_head_295_axes_1, x = k_head_293_cast_fp16)[name = string("k_head_295_cast_fp16")];
            tensor<int32, [1]> v_head_295_axes_1 = const()[name = string("v_head_295_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_295_cast_fp16 = expand_dims(axes = v_head_295_axes_1, x = v_head_293_cast_fp16)[name = string("v_head_295_cast_fp16")];
            tensor<int32, [1]> p_head_295_axes_1 = const()[name = string("p_head_295_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_295_cast_fp16 = expand_dims(axes = p_head_295_axes_1, x = p_head_293_cast_fp16)[name = string("p_head_295_cast_fp16")];
            bool var_11983_transpose_x_3 = const()[name = string("op_11983_transpose_x_3"), val = bool(false)];
            bool var_11983_transpose_y_3 = const()[name = string("op_11983_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_11983_cast_fp16 = matmul(transpose_x = var_11983_transpose_x_3, transpose_y = var_11983_transpose_y_3, x = q_u_147_cast_fp16, y = k_head_295_cast_fp16)[name = string("op_11983_cast_fp16")];
            fp16 var_11984_to_fp16 = const()[name = string("op_11984_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_147_cast_fp16 = mul(x = var_11983_cast_fp16, y = var_11984_to_fp16)[name = string("scores_content_147_cast_fp16")];
            bool x_785_transpose_x_3 = const()[name = string("x_785_transpose_x_3"), val = bool(false)];
            bool x_785_transpose_y_3 = const()[name = string("x_785_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_785_cast_fp16 = matmul(transpose_x = x_785_transpose_x_3, transpose_y = x_785_transpose_y_3, x = q_v_147_cast_fp16, y = p_head_295_cast_fp16)[name = string("x_785_cast_fp16")];
            tensor<int32, [8]> x_787_pad_1 = const()[name = string("x_787_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_787_mode_1 = const()[name = string("x_787_mode_1"), val = string("constant")];
            fp16 const_1801_to_fp16 = const()[name = string("const_1801_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_787_cast_fp16 = pad(constant_val = const_1801_to_fp16, mode = x_787_mode_1, pad = x_787_pad_1, x = x_785_cast_fp16)[name = string("x_787_cast_fp16")];
            tensor<int32, [4]> var_11998 = const()[name = string("op_11998"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_789_cast_fp16 = reshape(shape = var_11998, x = x_787_cast_fp16)[name = string("x_789_cast_fp16")];
            tensor<int32, [4]> var_12002_begin_1 = const()[name = string("op_12002_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12002_end_1 = const()[name = string("op_12002_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_12002_end_mask_1 = const()[name = string("op_12002_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_12002_cast_fp16 = slice_by_index(begin = var_12002_begin_1, end = var_12002_end_1, end_mask = var_12002_end_mask_1, x = x_789_cast_fp16)[name = string("op_12002_cast_fp16")];
            tensor<int32, [4]> var_12004 = const()[name = string("op_12004"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_12005_cast_fp16 = reshape(shape = var_12004, x = var_12002_cast_fp16)[name = string("op_12005_cast_fp16")];
            tensor<int32, [4]> var_12010_begin_1 = const()[name = string("op_12010_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12010_end_1 = const()[name = string("op_12010_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_12010_end_mask_1 = const()[name = string("op_12010_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_12010_cast_fp16 = slice_by_index(begin = var_12010_begin_1, end = var_12010_end_1, end_mask = var_12010_end_mask_1, x = var_12005_cast_fp16)[name = string("op_12010_cast_fp16")];
            fp16 var_12011_to_fp16 = const()[name = string("op_12011_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_147_cast_fp16 = mul(x = var_12010_cast_fp16, y = var_12011_to_fp16)[name = string("scores_pos_147_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_147_cast_fp16 = add(x = scores_content_147_cast_fp16, y = scores_pos_147_cast_fp16)[name = string("logits_147_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_12014_cast_fp16 = softmax(axis = var_11618, x = logits_147_cast_fp16)[name = string("op_12014_cast_fp16")];
            bool var_12016_transpose_x_1 = const()[name = string("op_12016_transpose_x_1"), val = bool(false)];
            bool var_12016_transpose_y_1 = const()[name = string("op_12016_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_12016_cast_fp16 = matmul(transpose_x = var_12016_transpose_x_1, transpose_y = var_12016_transpose_y_1, x = var_12014_cast_fp16, y = v_head_295_cast_fp16)[name = string("op_12016_cast_fp16")];
            tensor<int32, [1]> var_12017_axes_1 = const()[name = string("op_12017_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_12017_cast_fp16 = squeeze(axes = var_12017_axes_1, x = var_12016_cast_fp16)[name = string("op_12017_cast_fp16")];
            string dense_output_753_pad_type_1 = const()[name = string("dense_output_753_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_753_strides_1 = const()[name = string("dense_output_753_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_753_pad_1 = const()[name = string("dense_output_753_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_753_dilations_1 = const()[name = string("dense_output_753_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_753_groups_1 = const()[name = string("dense_output_753_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279566272))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279697408))))[name = string("layers_9_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_753_cast_fp16 = conv(dilations = dense_output_753_dilations_1, groups = dense_output_753_groups_1, pad = dense_output_753_pad_1, pad_type = dense_output_753_pad_type_1, strides = dense_output_753_strides_1, weight = layers_9_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_753_cast_fp16")];
            string sparse_output_753_pad_type_1 = const()[name = string("sparse_output_753_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_753_strides_1 = const()[name = string("sparse_output_753_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_753_pad_1 = const()[name = string("sparse_output_753_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_753_dilations_1 = const()[name = string("sparse_output_753_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_753_groups_1 = const()[name = string("sparse_output_753_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279700672))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279697984))))[name = string("layers_9_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_753_cast_fp16 = conv(dilations = sparse_output_753_dilations_1, groups = sparse_output_753_groups_1, pad = sparse_output_753_pad_1, pad_type = sparse_output_753_pad_type_1, strides = sparse_output_753_strides_1, weight = layers_9_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_753_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12032_cast_fp16 = add(x = dense_output_753_cast_fp16, y = sparse_output_753_cast_fp16)[name = string("op_12032_cast_fp16")];
            tensor<int32, [4]> var_12033 = const()[name = string("op_12033"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12035 = const()[name = string("op_12035"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12034_cast_fp16 = transpose(perm = var_12033, x = var_12032_cast_fp16)[name = string("transpose_603")];
            tensor<fp16, [1, 51, 128]> q_head_149_cast_fp16 = reshape(shape = var_12035, x = var_12034_cast_fp16)[name = string("q_head_149_cast_fp16")];
            string dense_output_755_pad_type_1 = const()[name = string("dense_output_755_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_755_strides_1 = const()[name = string("dense_output_755_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_755_pad_1 = const()[name = string("dense_output_755_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_755_dilations_1 = const()[name = string("dense_output_755_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_755_groups_1 = const()[name = string("dense_output_755_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279717120))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279848256))))[name = string("layers_9_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_755_cast_fp16 = conv(dilations = dense_output_755_dilations_1, groups = dense_output_755_groups_1, pad = dense_output_755_pad_1, pad_type = dense_output_755_pad_type_1, strides = dense_output_755_strides_1, weight = layers_9_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_755_cast_fp16")];
            string sparse_output_755_pad_type_1 = const()[name = string("sparse_output_755_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_755_strides_1 = const()[name = string("sparse_output_755_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_755_pad_1 = const()[name = string("sparse_output_755_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_755_dilations_1 = const()[name = string("sparse_output_755_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_755_groups_1 = const()[name = string("sparse_output_755_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279851520))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279848832))))[name = string("layers_9_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_755_cast_fp16 = conv(dilations = sparse_output_755_dilations_1, groups = sparse_output_755_groups_1, pad = sparse_output_755_pad_1, pad_type = sparse_output_755_pad_type_1, strides = sparse_output_755_strides_1, weight = layers_9_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_755_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12051_cast_fp16 = add(x = dense_output_755_cast_fp16, y = sparse_output_755_cast_fp16)[name = string("op_12051_cast_fp16")];
            tensor<int32, [4]> var_12052 = const()[name = string("op_12052"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12054 = const()[name = string("op_12054"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12053_cast_fp16 = transpose(perm = var_12052, x = var_12051_cast_fp16)[name = string("transpose_602")];
            tensor<fp16, [1, 51, 128]> k_head_297_cast_fp16 = reshape(shape = var_12054, x = var_12053_cast_fp16)[name = string("k_head_297_cast_fp16")];
            string dense_output_757_pad_type_1 = const()[name = string("dense_output_757_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_757_strides_1 = const()[name = string("dense_output_757_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_757_pad_1 = const()[name = string("dense_output_757_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_757_dilations_1 = const()[name = string("dense_output_757_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_757_groups_1 = const()[name = string("dense_output_757_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279867968))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279999104))))[name = string("layers_9_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_757_cast_fp16 = conv(dilations = dense_output_757_dilations_1, groups = dense_output_757_groups_1, pad = dense_output_757_pad_1, pad_type = dense_output_757_pad_type_1, strides = dense_output_757_strides_1, weight = layers_9_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_757_cast_fp16")];
            string sparse_output_757_pad_type_1 = const()[name = string("sparse_output_757_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_757_strides_1 = const()[name = string("sparse_output_757_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_757_pad_1 = const()[name = string("sparse_output_757_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_757_dilations_1 = const()[name = string("sparse_output_757_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_757_groups_1 = const()[name = string("sparse_output_757_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280002368))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279999680))))[name = string("layers_9_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_757_cast_fp16 = conv(dilations = sparse_output_757_dilations_1, groups = sparse_output_757_groups_1, pad = sparse_output_757_pad_1, pad_type = sparse_output_757_pad_type_1, strides = sparse_output_757_strides_1, weight = layers_9_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_757_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12070_cast_fp16 = add(x = dense_output_757_cast_fp16, y = sparse_output_757_cast_fp16)[name = string("op_12070_cast_fp16")];
            tensor<int32, [4]> var_12071 = const()[name = string("op_12071"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12073 = const()[name = string("op_12073"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12072_cast_fp16 = transpose(perm = var_12071, x = var_12070_cast_fp16)[name = string("transpose_601")];
            tensor<fp16, [1, 51, 128]> v_head_297_cast_fp16 = reshape(shape = var_12073, x = var_12072_cast_fp16)[name = string("v_head_297_cast_fp16")];
            string dense_output_759_pad_type_1 = const()[name = string("dense_output_759_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_759_strides_1 = const()[name = string("dense_output_759_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_759_pad_1 = const()[name = string("dense_output_759_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_759_dilations_1 = const()[name = string("dense_output_759_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_759_groups_1 = const()[name = string("dense_output_759_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280018816))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280149952))))[name = string("layers_9_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_759_cast_fp16 = conv(dilations = dense_output_759_dilations_1, groups = dense_output_759_groups_1, pad = dense_output_759_pad_1, pad_type = dense_output_759_pad_type_1, strides = dense_output_759_strides_1, weight = layers_9_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_759_cast_fp16")];
            string sparse_output_759_pad_type_1 = const()[name = string("sparse_output_759_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_759_strides_1 = const()[name = string("sparse_output_759_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_759_pad_1 = const()[name = string("sparse_output_759_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_759_dilations_1 = const()[name = string("sparse_output_759_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_759_groups_1 = const()[name = string("sparse_output_759_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280153216))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280150528))))[name = string("layers_9_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_759_cast_fp16 = conv(dilations = sparse_output_759_dilations_1, groups = sparse_output_759_groups_1, pad = sparse_output_759_pad_1, pad_type = sparse_output_759_pad_type_1, strides = sparse_output_759_strides_1, weight = layers_9_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_759_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_12089_cast_fp16 = add(x = dense_output_759_cast_fp16, y = sparse_output_759_cast_fp16)[name = string("op_12089_cast_fp16")];
            tensor<int32, [4]> var_12090 = const()[name = string("op_12090"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12092 = const()[name = string("op_12092"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_12091_cast_fp16 = transpose(perm = var_12090, x = var_12089_cast_fp16)[name = string("transpose_600")];
            tensor<fp16, [1, 101, 128]> p_head_297_cast_fp16 = reshape(shape = var_12092, x = var_12091_cast_fp16)[name = string("p_head_297_cast_fp16")];
            tensor<fp16, [128]> var_12094_to_fp16 = const()[name = string("op_12094_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280169664)))];
            tensor<fp16, [1, 51, 128]> var_12095_cast_fp16 = add(x = q_head_149_cast_fp16, y = var_12094_to_fp16)[name = string("op_12095_cast_fp16")];
            tensor<int32, [1]> q_u_149_axes_1 = const()[name = string("q_u_149_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_149_cast_fp16 = expand_dims(axes = q_u_149_axes_1, x = var_12095_cast_fp16)[name = string("q_u_149_cast_fp16")];
            tensor<fp16, [128]> var_12097_to_fp16 = const()[name = string("op_12097_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280169984)))];
            tensor<fp16, [1, 51, 128]> var_12098_cast_fp16 = add(x = q_head_149_cast_fp16, y = var_12097_to_fp16)[name = string("op_12098_cast_fp16")];
            tensor<int32, [1]> q_v_149_axes_1 = const()[name = string("q_v_149_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_149_cast_fp16 = expand_dims(axes = q_v_149_axes_1, x = var_12098_cast_fp16)[name = string("q_v_149_cast_fp16")];
            tensor<int32, [1]> k_head_299_axes_1 = const()[name = string("k_head_299_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_299_cast_fp16 = expand_dims(axes = k_head_299_axes_1, x = k_head_297_cast_fp16)[name = string("k_head_299_cast_fp16")];
            tensor<int32, [1]> v_head_299_axes_1 = const()[name = string("v_head_299_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_299_cast_fp16 = expand_dims(axes = v_head_299_axes_1, x = v_head_297_cast_fp16)[name = string("v_head_299_cast_fp16")];
            tensor<int32, [1]> p_head_299_axes_1 = const()[name = string("p_head_299_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_299_cast_fp16 = expand_dims(axes = p_head_299_axes_1, x = p_head_297_cast_fp16)[name = string("p_head_299_cast_fp16")];
            bool var_12104_transpose_x_3 = const()[name = string("op_12104_transpose_x_3"), val = bool(false)];
            bool var_12104_transpose_y_3 = const()[name = string("op_12104_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_12104_cast_fp16 = matmul(transpose_x = var_12104_transpose_x_3, transpose_y = var_12104_transpose_y_3, x = q_u_149_cast_fp16, y = k_head_299_cast_fp16)[name = string("op_12104_cast_fp16")];
            fp16 var_12105_to_fp16 = const()[name = string("op_12105_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_149_cast_fp16 = mul(x = var_12104_cast_fp16, y = var_12105_to_fp16)[name = string("scores_content_149_cast_fp16")];
            bool x_793_transpose_x_3 = const()[name = string("x_793_transpose_x_3"), val = bool(false)];
            bool x_793_transpose_y_3 = const()[name = string("x_793_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_793_cast_fp16 = matmul(transpose_x = x_793_transpose_x_3, transpose_y = x_793_transpose_y_3, x = q_v_149_cast_fp16, y = p_head_299_cast_fp16)[name = string("x_793_cast_fp16")];
            tensor<int32, [8]> x_795_pad_1 = const()[name = string("x_795_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_795_mode_1 = const()[name = string("x_795_mode_1"), val = string("constant")];
            fp16 const_1807_to_fp16 = const()[name = string("const_1807_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_795_cast_fp16 = pad(constant_val = const_1807_to_fp16, mode = x_795_mode_1, pad = x_795_pad_1, x = x_793_cast_fp16)[name = string("x_795_cast_fp16")];
            tensor<int32, [4]> var_12119 = const()[name = string("op_12119"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_797_cast_fp16 = reshape(shape = var_12119, x = x_795_cast_fp16)[name = string("x_797_cast_fp16")];
            tensor<int32, [4]> var_12123_begin_1 = const()[name = string("op_12123_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12123_end_1 = const()[name = string("op_12123_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_12123_end_mask_1 = const()[name = string("op_12123_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_12123_cast_fp16 = slice_by_index(begin = var_12123_begin_1, end = var_12123_end_1, end_mask = var_12123_end_mask_1, x = x_797_cast_fp16)[name = string("op_12123_cast_fp16")];
            tensor<int32, [4]> var_12125 = const()[name = string("op_12125"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_12126_cast_fp16 = reshape(shape = var_12125, x = var_12123_cast_fp16)[name = string("op_12126_cast_fp16")];
            tensor<int32, [4]> var_12131_begin_1 = const()[name = string("op_12131_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12131_end_1 = const()[name = string("op_12131_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_12131_end_mask_1 = const()[name = string("op_12131_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_12131_cast_fp16 = slice_by_index(begin = var_12131_begin_1, end = var_12131_end_1, end_mask = var_12131_end_mask_1, x = var_12126_cast_fp16)[name = string("op_12131_cast_fp16")];
            fp16 var_12132_to_fp16 = const()[name = string("op_12132_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_149_cast_fp16 = mul(x = var_12131_cast_fp16, y = var_12132_to_fp16)[name = string("scores_pos_149_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_149_cast_fp16 = add(x = scores_content_149_cast_fp16, y = scores_pos_149_cast_fp16)[name = string("logits_149_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_12135_cast_fp16 = softmax(axis = var_11618, x = logits_149_cast_fp16)[name = string("op_12135_cast_fp16")];
            bool var_12137_transpose_x_1 = const()[name = string("op_12137_transpose_x_1"), val = bool(false)];
            bool var_12137_transpose_y_1 = const()[name = string("op_12137_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_12137_cast_fp16 = matmul(transpose_x = var_12137_transpose_x_1, transpose_y = var_12137_transpose_y_1, x = var_12135_cast_fp16, y = v_head_299_cast_fp16)[name = string("op_12137_cast_fp16")];
            tensor<int32, [1]> var_12138_axes_1 = const()[name = string("op_12138_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_12138_cast_fp16 = squeeze(axes = var_12138_axes_1, x = var_12137_cast_fp16)[name = string("op_12138_cast_fp16")];
            string dense_output_761_pad_type_1 = const()[name = string("dense_output_761_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_761_strides_1 = const()[name = string("dense_output_761_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_761_pad_1 = const()[name = string("dense_output_761_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_761_dilations_1 = const()[name = string("dense_output_761_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_761_groups_1 = const()[name = string("dense_output_761_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280170304))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280301440))))[name = string("layers_9_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_761_cast_fp16 = conv(dilations = dense_output_761_dilations_1, groups = dense_output_761_groups_1, pad = dense_output_761_pad_1, pad_type = dense_output_761_pad_type_1, strides = dense_output_761_strides_1, weight = layers_9_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_761_cast_fp16")];
            string sparse_output_761_pad_type_1 = const()[name = string("sparse_output_761_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_761_strides_1 = const()[name = string("sparse_output_761_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_761_pad_1 = const()[name = string("sparse_output_761_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_761_dilations_1 = const()[name = string("sparse_output_761_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_761_groups_1 = const()[name = string("sparse_output_761_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280304704))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280302016))))[name = string("layers_9_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_761_cast_fp16 = conv(dilations = sparse_output_761_dilations_1, groups = sparse_output_761_groups_1, pad = sparse_output_761_pad_1, pad_type = sparse_output_761_pad_type_1, strides = sparse_output_761_strides_1, weight = layers_9_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_761_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12153_cast_fp16 = add(x = dense_output_761_cast_fp16, y = sparse_output_761_cast_fp16)[name = string("op_12153_cast_fp16")];
            tensor<int32, [4]> var_12154 = const()[name = string("op_12154"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12156 = const()[name = string("op_12156"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12155_cast_fp16 = transpose(perm = var_12154, x = var_12153_cast_fp16)[name = string("transpose_599")];
            tensor<fp16, [1, 51, 128]> q_head_151_cast_fp16 = reshape(shape = var_12156, x = var_12155_cast_fp16)[name = string("q_head_151_cast_fp16")];
            string dense_output_763_pad_type_1 = const()[name = string("dense_output_763_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_763_strides_1 = const()[name = string("dense_output_763_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_763_pad_1 = const()[name = string("dense_output_763_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_763_dilations_1 = const()[name = string("dense_output_763_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_763_groups_1 = const()[name = string("dense_output_763_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280321152))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280452288))))[name = string("layers_9_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_763_cast_fp16 = conv(dilations = dense_output_763_dilations_1, groups = dense_output_763_groups_1, pad = dense_output_763_pad_1, pad_type = dense_output_763_pad_type_1, strides = dense_output_763_strides_1, weight = layers_9_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_763_cast_fp16")];
            string sparse_output_763_pad_type_1 = const()[name = string("sparse_output_763_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_763_strides_1 = const()[name = string("sparse_output_763_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_763_pad_1 = const()[name = string("sparse_output_763_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_763_dilations_1 = const()[name = string("sparse_output_763_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_763_groups_1 = const()[name = string("sparse_output_763_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280455552))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280452864))))[name = string("layers_9_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_763_cast_fp16 = conv(dilations = sparse_output_763_dilations_1, groups = sparse_output_763_groups_1, pad = sparse_output_763_pad_1, pad_type = sparse_output_763_pad_type_1, strides = sparse_output_763_strides_1, weight = layers_9_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_763_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12172_cast_fp16 = add(x = dense_output_763_cast_fp16, y = sparse_output_763_cast_fp16)[name = string("op_12172_cast_fp16")];
            tensor<int32, [4]> var_12173 = const()[name = string("op_12173"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12175 = const()[name = string("op_12175"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12174_cast_fp16 = transpose(perm = var_12173, x = var_12172_cast_fp16)[name = string("transpose_598")];
            tensor<fp16, [1, 51, 128]> k_head_301_cast_fp16 = reshape(shape = var_12175, x = var_12174_cast_fp16)[name = string("k_head_301_cast_fp16")];
            string dense_output_765_pad_type_1 = const()[name = string("dense_output_765_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_765_strides_1 = const()[name = string("dense_output_765_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_765_pad_1 = const()[name = string("dense_output_765_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_765_dilations_1 = const()[name = string("dense_output_765_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_765_groups_1 = const()[name = string("dense_output_765_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280472000))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280603136))))[name = string("layers_9_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_765_cast_fp16 = conv(dilations = dense_output_765_dilations_1, groups = dense_output_765_groups_1, pad = dense_output_765_pad_1, pad_type = dense_output_765_pad_type_1, strides = dense_output_765_strides_1, weight = layers_9_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_765_cast_fp16")];
            string sparse_output_765_pad_type_1 = const()[name = string("sparse_output_765_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_765_strides_1 = const()[name = string("sparse_output_765_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_765_pad_1 = const()[name = string("sparse_output_765_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_765_dilations_1 = const()[name = string("sparse_output_765_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_765_groups_1 = const()[name = string("sparse_output_765_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280606400))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280603712))))[name = string("layers_9_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_765_cast_fp16 = conv(dilations = sparse_output_765_dilations_1, groups = sparse_output_765_groups_1, pad = sparse_output_765_pad_1, pad_type = sparse_output_765_pad_type_1, strides = sparse_output_765_strides_1, weight = layers_9_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_765_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12191_cast_fp16 = add(x = dense_output_765_cast_fp16, y = sparse_output_765_cast_fp16)[name = string("op_12191_cast_fp16")];
            tensor<int32, [4]> var_12192 = const()[name = string("op_12192"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12194 = const()[name = string("op_12194"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12193_cast_fp16 = transpose(perm = var_12192, x = var_12191_cast_fp16)[name = string("transpose_597")];
            tensor<fp16, [1, 51, 128]> v_head_301_cast_fp16 = reshape(shape = var_12194, x = var_12193_cast_fp16)[name = string("v_head_301_cast_fp16")];
            string dense_output_767_pad_type_1 = const()[name = string("dense_output_767_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_767_strides_1 = const()[name = string("dense_output_767_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_767_pad_1 = const()[name = string("dense_output_767_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_767_dilations_1 = const()[name = string("dense_output_767_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_767_groups_1 = const()[name = string("dense_output_767_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280622848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280753984))))[name = string("layers_9_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_767_cast_fp16 = conv(dilations = dense_output_767_dilations_1, groups = dense_output_767_groups_1, pad = dense_output_767_pad_1, pad_type = dense_output_767_pad_type_1, strides = dense_output_767_strides_1, weight = layers_9_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_767_cast_fp16")];
            string sparse_output_767_pad_type_1 = const()[name = string("sparse_output_767_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_767_strides_1 = const()[name = string("sparse_output_767_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_767_pad_1 = const()[name = string("sparse_output_767_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_767_dilations_1 = const()[name = string("sparse_output_767_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_767_groups_1 = const()[name = string("sparse_output_767_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280757248))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280754560))))[name = string("layers_9_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_767_cast_fp16 = conv(dilations = sparse_output_767_dilations_1, groups = sparse_output_767_groups_1, pad = sparse_output_767_pad_1, pad_type = sparse_output_767_pad_type_1, strides = sparse_output_767_strides_1, weight = layers_9_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_767_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_12210_cast_fp16 = add(x = dense_output_767_cast_fp16, y = sparse_output_767_cast_fp16)[name = string("op_12210_cast_fp16")];
            tensor<int32, [4]> var_12211 = const()[name = string("op_12211"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12213 = const()[name = string("op_12213"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_12212_cast_fp16 = transpose(perm = var_12211, x = var_12210_cast_fp16)[name = string("transpose_596")];
            tensor<fp16, [1, 101, 128]> p_head_301_cast_fp16 = reshape(shape = var_12213, x = var_12212_cast_fp16)[name = string("p_head_301_cast_fp16")];
            tensor<fp16, [128]> var_12215_to_fp16 = const()[name = string("op_12215_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280773696)))];
            tensor<fp16, [1, 51, 128]> var_12216_cast_fp16 = add(x = q_head_151_cast_fp16, y = var_12215_to_fp16)[name = string("op_12216_cast_fp16")];
            tensor<int32, [1]> q_u_151_axes_1 = const()[name = string("q_u_151_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_151_cast_fp16 = expand_dims(axes = q_u_151_axes_1, x = var_12216_cast_fp16)[name = string("q_u_151_cast_fp16")];
            tensor<fp16, [128]> var_12218_to_fp16 = const()[name = string("op_12218_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280774016)))];
            tensor<fp16, [1, 51, 128]> var_12219_cast_fp16 = add(x = q_head_151_cast_fp16, y = var_12218_to_fp16)[name = string("op_12219_cast_fp16")];
            tensor<int32, [1]> q_v_151_axes_1 = const()[name = string("q_v_151_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_151_cast_fp16 = expand_dims(axes = q_v_151_axes_1, x = var_12219_cast_fp16)[name = string("q_v_151_cast_fp16")];
            tensor<int32, [1]> k_head_303_axes_1 = const()[name = string("k_head_303_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_303_cast_fp16 = expand_dims(axes = k_head_303_axes_1, x = k_head_301_cast_fp16)[name = string("k_head_303_cast_fp16")];
            tensor<int32, [1]> v_head_303_axes_1 = const()[name = string("v_head_303_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_303_cast_fp16 = expand_dims(axes = v_head_303_axes_1, x = v_head_301_cast_fp16)[name = string("v_head_303_cast_fp16")];
            tensor<int32, [1]> p_head_303_axes_1 = const()[name = string("p_head_303_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_303_cast_fp16 = expand_dims(axes = p_head_303_axes_1, x = p_head_301_cast_fp16)[name = string("p_head_303_cast_fp16")];
            bool var_12225_transpose_x_3 = const()[name = string("op_12225_transpose_x_3"), val = bool(false)];
            bool var_12225_transpose_y_3 = const()[name = string("op_12225_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_12225_cast_fp16 = matmul(transpose_x = var_12225_transpose_x_3, transpose_y = var_12225_transpose_y_3, x = q_u_151_cast_fp16, y = k_head_303_cast_fp16)[name = string("op_12225_cast_fp16")];
            fp16 var_12226_to_fp16 = const()[name = string("op_12226_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_151_cast_fp16 = mul(x = var_12225_cast_fp16, y = var_12226_to_fp16)[name = string("scores_content_151_cast_fp16")];
            bool x_801_transpose_x_3 = const()[name = string("x_801_transpose_x_3"), val = bool(false)];
            bool x_801_transpose_y_3 = const()[name = string("x_801_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_801_cast_fp16 = matmul(transpose_x = x_801_transpose_x_3, transpose_y = x_801_transpose_y_3, x = q_v_151_cast_fp16, y = p_head_303_cast_fp16)[name = string("x_801_cast_fp16")];
            tensor<int32, [8]> x_803_pad_1 = const()[name = string("x_803_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_803_mode_1 = const()[name = string("x_803_mode_1"), val = string("constant")];
            fp16 const_1813_to_fp16 = const()[name = string("const_1813_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_803_cast_fp16 = pad(constant_val = const_1813_to_fp16, mode = x_803_mode_1, pad = x_803_pad_1, x = x_801_cast_fp16)[name = string("x_803_cast_fp16")];
            tensor<int32, [4]> var_12240 = const()[name = string("op_12240"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_805_cast_fp16 = reshape(shape = var_12240, x = x_803_cast_fp16)[name = string("x_805_cast_fp16")];
            tensor<int32, [4]> var_12244_begin_1 = const()[name = string("op_12244_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12244_end_1 = const()[name = string("op_12244_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_12244_end_mask_1 = const()[name = string("op_12244_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_12244_cast_fp16 = slice_by_index(begin = var_12244_begin_1, end = var_12244_end_1, end_mask = var_12244_end_mask_1, x = x_805_cast_fp16)[name = string("op_12244_cast_fp16")];
            tensor<int32, [4]> var_12246 = const()[name = string("op_12246"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_12247_cast_fp16 = reshape(shape = var_12246, x = var_12244_cast_fp16)[name = string("op_12247_cast_fp16")];
            tensor<int32, [4]> var_12252_begin_1 = const()[name = string("op_12252_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12252_end_1 = const()[name = string("op_12252_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_12252_end_mask_1 = const()[name = string("op_12252_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_12252_cast_fp16 = slice_by_index(begin = var_12252_begin_1, end = var_12252_end_1, end_mask = var_12252_end_mask_1, x = var_12247_cast_fp16)[name = string("op_12252_cast_fp16")];
            fp16 var_12253_to_fp16 = const()[name = string("op_12253_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_151_cast_fp16 = mul(x = var_12252_cast_fp16, y = var_12253_to_fp16)[name = string("scores_pos_151_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_151_cast_fp16 = add(x = scores_content_151_cast_fp16, y = scores_pos_151_cast_fp16)[name = string("logits_151_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_12256_cast_fp16 = softmax(axis = var_11618, x = logits_151_cast_fp16)[name = string("op_12256_cast_fp16")];
            bool var_12258_transpose_x_1 = const()[name = string("op_12258_transpose_x_1"), val = bool(false)];
            bool var_12258_transpose_y_1 = const()[name = string("op_12258_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_12258_cast_fp16 = matmul(transpose_x = var_12258_transpose_x_1, transpose_y = var_12258_transpose_y_1, x = var_12256_cast_fp16, y = v_head_303_cast_fp16)[name = string("op_12258_cast_fp16")];
            tensor<int32, [1]> var_12259_axes_1 = const()[name = string("op_12259_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_12259_cast_fp16 = squeeze(axes = var_12259_axes_1, x = var_12258_cast_fp16)[name = string("op_12259_cast_fp16")];
            string dense_output_769_pad_type_1 = const()[name = string("dense_output_769_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_769_strides_1 = const()[name = string("dense_output_769_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_769_pad_1 = const()[name = string("dense_output_769_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_769_dilations_1 = const()[name = string("dense_output_769_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_769_groups_1 = const()[name = string("dense_output_769_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280774336))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280905472))))[name = string("layers_9_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_769_cast_fp16 = conv(dilations = dense_output_769_dilations_1, groups = dense_output_769_groups_1, pad = dense_output_769_pad_1, pad_type = dense_output_769_pad_type_1, strides = dense_output_769_strides_1, weight = layers_9_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_769_cast_fp16")];
            string sparse_output_769_pad_type_1 = const()[name = string("sparse_output_769_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_769_strides_1 = const()[name = string("sparse_output_769_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_769_pad_1 = const()[name = string("sparse_output_769_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_769_dilations_1 = const()[name = string("sparse_output_769_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_769_groups_1 = const()[name = string("sparse_output_769_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280908736))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280906048))))[name = string("layers_9_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_769_cast_fp16 = conv(dilations = sparse_output_769_dilations_1, groups = sparse_output_769_groups_1, pad = sparse_output_769_pad_1, pad_type = sparse_output_769_pad_type_1, strides = sparse_output_769_strides_1, weight = layers_9_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_769_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12274_cast_fp16 = add(x = dense_output_769_cast_fp16, y = sparse_output_769_cast_fp16)[name = string("op_12274_cast_fp16")];
            tensor<int32, [4]> var_12275 = const()[name = string("op_12275"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12277 = const()[name = string("op_12277"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12276_cast_fp16 = transpose(perm = var_12275, x = var_12274_cast_fp16)[name = string("transpose_595")];
            tensor<fp16, [1, 51, 128]> q_head_153_cast_fp16 = reshape(shape = var_12277, x = var_12276_cast_fp16)[name = string("q_head_153_cast_fp16")];
            string dense_output_771_pad_type_1 = const()[name = string("dense_output_771_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_771_strides_1 = const()[name = string("dense_output_771_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_771_pad_1 = const()[name = string("dense_output_771_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_771_dilations_1 = const()[name = string("dense_output_771_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_771_groups_1 = const()[name = string("dense_output_771_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280925184))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281056320))))[name = string("layers_9_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_771_cast_fp16 = conv(dilations = dense_output_771_dilations_1, groups = dense_output_771_groups_1, pad = dense_output_771_pad_1, pad_type = dense_output_771_pad_type_1, strides = dense_output_771_strides_1, weight = layers_9_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_771_cast_fp16")];
            string sparse_output_771_pad_type_1 = const()[name = string("sparse_output_771_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_771_strides_1 = const()[name = string("sparse_output_771_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_771_pad_1 = const()[name = string("sparse_output_771_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_771_dilations_1 = const()[name = string("sparse_output_771_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_771_groups_1 = const()[name = string("sparse_output_771_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281059584))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281056896))))[name = string("layers_9_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_771_cast_fp16 = conv(dilations = sparse_output_771_dilations_1, groups = sparse_output_771_groups_1, pad = sparse_output_771_pad_1, pad_type = sparse_output_771_pad_type_1, strides = sparse_output_771_strides_1, weight = layers_9_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_771_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12293_cast_fp16 = add(x = dense_output_771_cast_fp16, y = sparse_output_771_cast_fp16)[name = string("op_12293_cast_fp16")];
            tensor<int32, [4]> var_12294 = const()[name = string("op_12294"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12296 = const()[name = string("op_12296"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12295_cast_fp16 = transpose(perm = var_12294, x = var_12293_cast_fp16)[name = string("transpose_594")];
            tensor<fp16, [1, 51, 128]> k_head_305_cast_fp16 = reshape(shape = var_12296, x = var_12295_cast_fp16)[name = string("k_head_305_cast_fp16")];
            string dense_output_773_pad_type_1 = const()[name = string("dense_output_773_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_773_strides_1 = const()[name = string("dense_output_773_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_773_pad_1 = const()[name = string("dense_output_773_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_773_dilations_1 = const()[name = string("dense_output_773_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_773_groups_1 = const()[name = string("dense_output_773_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281076032))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281207168))))[name = string("layers_9_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_773_cast_fp16 = conv(dilations = dense_output_773_dilations_1, groups = dense_output_773_groups_1, pad = dense_output_773_pad_1, pad_type = dense_output_773_pad_type_1, strides = dense_output_773_strides_1, weight = layers_9_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_773_cast_fp16")];
            string sparse_output_773_pad_type_1 = const()[name = string("sparse_output_773_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_773_strides_1 = const()[name = string("sparse_output_773_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_773_pad_1 = const()[name = string("sparse_output_773_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_773_dilations_1 = const()[name = string("sparse_output_773_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_773_groups_1 = const()[name = string("sparse_output_773_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281210432))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281207744))))[name = string("layers_9_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_773_cast_fp16 = conv(dilations = sparse_output_773_dilations_1, groups = sparse_output_773_groups_1, pad = sparse_output_773_pad_1, pad_type = sparse_output_773_pad_type_1, strides = sparse_output_773_strides_1, weight = layers_9_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_773_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12312_cast_fp16 = add(x = dense_output_773_cast_fp16, y = sparse_output_773_cast_fp16)[name = string("op_12312_cast_fp16")];
            tensor<int32, [4]> var_12313 = const()[name = string("op_12313"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12315 = const()[name = string("op_12315"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12314_cast_fp16 = transpose(perm = var_12313, x = var_12312_cast_fp16)[name = string("transpose_593")];
            tensor<fp16, [1, 51, 128]> v_head_305_cast_fp16 = reshape(shape = var_12315, x = var_12314_cast_fp16)[name = string("v_head_305_cast_fp16")];
            string dense_output_775_pad_type_1 = const()[name = string("dense_output_775_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_775_strides_1 = const()[name = string("dense_output_775_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_775_pad_1 = const()[name = string("dense_output_775_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_775_dilations_1 = const()[name = string("dense_output_775_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_775_groups_1 = const()[name = string("dense_output_775_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281226880))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281358016))))[name = string("layers_9_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_775_cast_fp16 = conv(dilations = dense_output_775_dilations_1, groups = dense_output_775_groups_1, pad = dense_output_775_pad_1, pad_type = dense_output_775_pad_type_1, strides = dense_output_775_strides_1, weight = layers_9_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_775_cast_fp16")];
            string sparse_output_775_pad_type_1 = const()[name = string("sparse_output_775_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_775_strides_1 = const()[name = string("sparse_output_775_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_775_pad_1 = const()[name = string("sparse_output_775_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_775_dilations_1 = const()[name = string("sparse_output_775_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_775_groups_1 = const()[name = string("sparse_output_775_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281361280))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281358592))))[name = string("layers_9_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_775_cast_fp16 = conv(dilations = sparse_output_775_dilations_1, groups = sparse_output_775_groups_1, pad = sparse_output_775_pad_1, pad_type = sparse_output_775_pad_type_1, strides = sparse_output_775_strides_1, weight = layers_9_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_775_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_12331_cast_fp16 = add(x = dense_output_775_cast_fp16, y = sparse_output_775_cast_fp16)[name = string("op_12331_cast_fp16")];
            tensor<int32, [4]> var_12332 = const()[name = string("op_12332"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12334 = const()[name = string("op_12334"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_12333_cast_fp16 = transpose(perm = var_12332, x = var_12331_cast_fp16)[name = string("transpose_592")];
            tensor<fp16, [1, 101, 128]> p_head_305_cast_fp16 = reshape(shape = var_12334, x = var_12333_cast_fp16)[name = string("p_head_305_cast_fp16")];
            tensor<fp16, [128]> var_12336_to_fp16 = const()[name = string("op_12336_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281377728)))];
            tensor<fp16, [1, 51, 128]> var_12337_cast_fp16 = add(x = q_head_153_cast_fp16, y = var_12336_to_fp16)[name = string("op_12337_cast_fp16")];
            tensor<int32, [1]> q_u_153_axes_1 = const()[name = string("q_u_153_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_153_cast_fp16 = expand_dims(axes = q_u_153_axes_1, x = var_12337_cast_fp16)[name = string("q_u_153_cast_fp16")];
            tensor<fp16, [128]> var_12339_to_fp16 = const()[name = string("op_12339_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281378048)))];
            tensor<fp16, [1, 51, 128]> var_12340_cast_fp16 = add(x = q_head_153_cast_fp16, y = var_12339_to_fp16)[name = string("op_12340_cast_fp16")];
            tensor<int32, [1]> q_v_153_axes_1 = const()[name = string("q_v_153_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_153_cast_fp16 = expand_dims(axes = q_v_153_axes_1, x = var_12340_cast_fp16)[name = string("q_v_153_cast_fp16")];
            tensor<int32, [1]> k_head_307_axes_1 = const()[name = string("k_head_307_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_307_cast_fp16 = expand_dims(axes = k_head_307_axes_1, x = k_head_305_cast_fp16)[name = string("k_head_307_cast_fp16")];
            tensor<int32, [1]> v_head_307_axes_1 = const()[name = string("v_head_307_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_307_cast_fp16 = expand_dims(axes = v_head_307_axes_1, x = v_head_305_cast_fp16)[name = string("v_head_307_cast_fp16")];
            tensor<int32, [1]> p_head_307_axes_1 = const()[name = string("p_head_307_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_307_cast_fp16 = expand_dims(axes = p_head_307_axes_1, x = p_head_305_cast_fp16)[name = string("p_head_307_cast_fp16")];
            bool var_12346_transpose_x_3 = const()[name = string("op_12346_transpose_x_3"), val = bool(false)];
            bool var_12346_transpose_y_3 = const()[name = string("op_12346_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_12346_cast_fp16 = matmul(transpose_x = var_12346_transpose_x_3, transpose_y = var_12346_transpose_y_3, x = q_u_153_cast_fp16, y = k_head_307_cast_fp16)[name = string("op_12346_cast_fp16")];
            fp16 var_12347_to_fp16 = const()[name = string("op_12347_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_153_cast_fp16 = mul(x = var_12346_cast_fp16, y = var_12347_to_fp16)[name = string("scores_content_153_cast_fp16")];
            bool x_809_transpose_x_3 = const()[name = string("x_809_transpose_x_3"), val = bool(false)];
            bool x_809_transpose_y_3 = const()[name = string("x_809_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_809_cast_fp16 = matmul(transpose_x = x_809_transpose_x_3, transpose_y = x_809_transpose_y_3, x = q_v_153_cast_fp16, y = p_head_307_cast_fp16)[name = string("x_809_cast_fp16")];
            tensor<int32, [8]> x_811_pad_1 = const()[name = string("x_811_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_811_mode_1 = const()[name = string("x_811_mode_1"), val = string("constant")];
            fp16 const_1819_to_fp16 = const()[name = string("const_1819_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_811_cast_fp16 = pad(constant_val = const_1819_to_fp16, mode = x_811_mode_1, pad = x_811_pad_1, x = x_809_cast_fp16)[name = string("x_811_cast_fp16")];
            tensor<int32, [4]> var_12361 = const()[name = string("op_12361"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_813_cast_fp16 = reshape(shape = var_12361, x = x_811_cast_fp16)[name = string("x_813_cast_fp16")];
            tensor<int32, [4]> var_12365_begin_1 = const()[name = string("op_12365_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12365_end_1 = const()[name = string("op_12365_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_12365_end_mask_1 = const()[name = string("op_12365_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_12365_cast_fp16 = slice_by_index(begin = var_12365_begin_1, end = var_12365_end_1, end_mask = var_12365_end_mask_1, x = x_813_cast_fp16)[name = string("op_12365_cast_fp16")];
            tensor<int32, [4]> var_12367 = const()[name = string("op_12367"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_12368_cast_fp16 = reshape(shape = var_12367, x = var_12365_cast_fp16)[name = string("op_12368_cast_fp16")];
            tensor<int32, [4]> var_12373_begin_1 = const()[name = string("op_12373_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12373_end_1 = const()[name = string("op_12373_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_12373_end_mask_1 = const()[name = string("op_12373_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_12373_cast_fp16 = slice_by_index(begin = var_12373_begin_1, end = var_12373_end_1, end_mask = var_12373_end_mask_1, x = var_12368_cast_fp16)[name = string("op_12373_cast_fp16")];
            fp16 var_12374_to_fp16 = const()[name = string("op_12374_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_153_cast_fp16 = mul(x = var_12373_cast_fp16, y = var_12374_to_fp16)[name = string("scores_pos_153_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_153_cast_fp16 = add(x = scores_content_153_cast_fp16, y = scores_pos_153_cast_fp16)[name = string("logits_153_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_12377_cast_fp16 = softmax(axis = var_11618, x = logits_153_cast_fp16)[name = string("op_12377_cast_fp16")];
            bool var_12379_transpose_x_1 = const()[name = string("op_12379_transpose_x_1"), val = bool(false)];
            bool var_12379_transpose_y_1 = const()[name = string("op_12379_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_12379_cast_fp16 = matmul(transpose_x = var_12379_transpose_x_1, transpose_y = var_12379_transpose_y_1, x = var_12377_cast_fp16, y = v_head_307_cast_fp16)[name = string("op_12379_cast_fp16")];
            tensor<int32, [1]> var_12380_axes_1 = const()[name = string("op_12380_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_12380_cast_fp16 = squeeze(axes = var_12380_axes_1, x = var_12379_cast_fp16)[name = string("op_12380_cast_fp16")];
            string dense_output_777_pad_type_1 = const()[name = string("dense_output_777_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_777_strides_1 = const()[name = string("dense_output_777_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_777_pad_1 = const()[name = string("dense_output_777_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_777_dilations_1 = const()[name = string("dense_output_777_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_777_groups_1 = const()[name = string("dense_output_777_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281378368))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281509504))))[name = string("layers_9_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_777_cast_fp16 = conv(dilations = dense_output_777_dilations_1, groups = dense_output_777_groups_1, pad = dense_output_777_pad_1, pad_type = dense_output_777_pad_type_1, strides = dense_output_777_strides_1, weight = layers_9_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_777_cast_fp16")];
            string sparse_output_777_pad_type_1 = const()[name = string("sparse_output_777_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_777_strides_1 = const()[name = string("sparse_output_777_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_777_pad_1 = const()[name = string("sparse_output_777_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_777_dilations_1 = const()[name = string("sparse_output_777_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_777_groups_1 = const()[name = string("sparse_output_777_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281512768))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281510080))))[name = string("layers_9_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_777_cast_fp16 = conv(dilations = sparse_output_777_dilations_1, groups = sparse_output_777_groups_1, pad = sparse_output_777_pad_1, pad_type = sparse_output_777_pad_type_1, strides = sparse_output_777_strides_1, weight = layers_9_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_777_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12395_cast_fp16 = add(x = dense_output_777_cast_fp16, y = sparse_output_777_cast_fp16)[name = string("op_12395_cast_fp16")];
            tensor<int32, [4]> var_12396 = const()[name = string("op_12396"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12398 = const()[name = string("op_12398"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12397_cast_fp16 = transpose(perm = var_12396, x = var_12395_cast_fp16)[name = string("transpose_591")];
            tensor<fp16, [1, 51, 128]> q_head_155_cast_fp16 = reshape(shape = var_12398, x = var_12397_cast_fp16)[name = string("q_head_155_cast_fp16")];
            string dense_output_779_pad_type_1 = const()[name = string("dense_output_779_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_779_strides_1 = const()[name = string("dense_output_779_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_779_pad_1 = const()[name = string("dense_output_779_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_779_dilations_1 = const()[name = string("dense_output_779_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_779_groups_1 = const()[name = string("dense_output_779_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281529216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281660352))))[name = string("layers_9_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_779_cast_fp16 = conv(dilations = dense_output_779_dilations_1, groups = dense_output_779_groups_1, pad = dense_output_779_pad_1, pad_type = dense_output_779_pad_type_1, strides = dense_output_779_strides_1, weight = layers_9_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_779_cast_fp16")];
            string sparse_output_779_pad_type_1 = const()[name = string("sparse_output_779_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_779_strides_1 = const()[name = string("sparse_output_779_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_779_pad_1 = const()[name = string("sparse_output_779_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_779_dilations_1 = const()[name = string("sparse_output_779_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_779_groups_1 = const()[name = string("sparse_output_779_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281663616))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281660928))))[name = string("layers_9_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_779_cast_fp16 = conv(dilations = sparse_output_779_dilations_1, groups = sparse_output_779_groups_1, pad = sparse_output_779_pad_1, pad_type = sparse_output_779_pad_type_1, strides = sparse_output_779_strides_1, weight = layers_9_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_779_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12414_cast_fp16 = add(x = dense_output_779_cast_fp16, y = sparse_output_779_cast_fp16)[name = string("op_12414_cast_fp16")];
            tensor<int32, [4]> var_12415 = const()[name = string("op_12415"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12417 = const()[name = string("op_12417"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12416_cast_fp16 = transpose(perm = var_12415, x = var_12414_cast_fp16)[name = string("transpose_590")];
            tensor<fp16, [1, 51, 128]> k_head_309_cast_fp16 = reshape(shape = var_12417, x = var_12416_cast_fp16)[name = string("k_head_309_cast_fp16")];
            string dense_output_781_pad_type_1 = const()[name = string("dense_output_781_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_781_strides_1 = const()[name = string("dense_output_781_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_781_pad_1 = const()[name = string("dense_output_781_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_781_dilations_1 = const()[name = string("dense_output_781_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_781_groups_1 = const()[name = string("dense_output_781_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281680064))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281811200))))[name = string("layers_9_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_781_cast_fp16 = conv(dilations = dense_output_781_dilations_1, groups = dense_output_781_groups_1, pad = dense_output_781_pad_1, pad_type = dense_output_781_pad_type_1, strides = dense_output_781_strides_1, weight = layers_9_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_781_cast_fp16")];
            string sparse_output_781_pad_type_1 = const()[name = string("sparse_output_781_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_781_strides_1 = const()[name = string("sparse_output_781_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_781_pad_1 = const()[name = string("sparse_output_781_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_781_dilations_1 = const()[name = string("sparse_output_781_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_781_groups_1 = const()[name = string("sparse_output_781_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281814464))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281811776))))[name = string("layers_9_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_781_cast_fp16 = conv(dilations = sparse_output_781_dilations_1, groups = sparse_output_781_groups_1, pad = sparse_output_781_pad_1, pad_type = sparse_output_781_pad_type_1, strides = sparse_output_781_strides_1, weight = layers_9_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_781_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12433_cast_fp16 = add(x = dense_output_781_cast_fp16, y = sparse_output_781_cast_fp16)[name = string("op_12433_cast_fp16")];
            tensor<int32, [4]> var_12434 = const()[name = string("op_12434"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12436 = const()[name = string("op_12436"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12435_cast_fp16 = transpose(perm = var_12434, x = var_12433_cast_fp16)[name = string("transpose_589")];
            tensor<fp16, [1, 51, 128]> v_head_309_cast_fp16 = reshape(shape = var_12436, x = var_12435_cast_fp16)[name = string("v_head_309_cast_fp16")];
            string dense_output_783_pad_type_1 = const()[name = string("dense_output_783_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_783_strides_1 = const()[name = string("dense_output_783_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_783_pad_1 = const()[name = string("dense_output_783_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_783_dilations_1 = const()[name = string("dense_output_783_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_783_groups_1 = const()[name = string("dense_output_783_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281830912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281962048))))[name = string("layers_9_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_783_cast_fp16 = conv(dilations = dense_output_783_dilations_1, groups = dense_output_783_groups_1, pad = dense_output_783_pad_1, pad_type = dense_output_783_pad_type_1, strides = dense_output_783_strides_1, weight = layers_9_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_783_cast_fp16")];
            string sparse_output_783_pad_type_1 = const()[name = string("sparse_output_783_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_783_strides_1 = const()[name = string("sparse_output_783_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_783_pad_1 = const()[name = string("sparse_output_783_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_783_dilations_1 = const()[name = string("sparse_output_783_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_783_groups_1 = const()[name = string("sparse_output_783_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281965312))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281962624))))[name = string("layers_9_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_783_cast_fp16 = conv(dilations = sparse_output_783_dilations_1, groups = sparse_output_783_groups_1, pad = sparse_output_783_pad_1, pad_type = sparse_output_783_pad_type_1, strides = sparse_output_783_strides_1, weight = layers_9_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_783_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_12452_cast_fp16 = add(x = dense_output_783_cast_fp16, y = sparse_output_783_cast_fp16)[name = string("op_12452_cast_fp16")];
            tensor<int32, [4]> var_12453 = const()[name = string("op_12453"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12455 = const()[name = string("op_12455"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_12454_cast_fp16 = transpose(perm = var_12453, x = var_12452_cast_fp16)[name = string("transpose_588")];
            tensor<fp16, [1, 101, 128]> p_head_309_cast_fp16 = reshape(shape = var_12455, x = var_12454_cast_fp16)[name = string("p_head_309_cast_fp16")];
            tensor<fp16, [128]> var_12457_to_fp16 = const()[name = string("op_12457_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281981760)))];
            tensor<fp16, [1, 51, 128]> var_12458_cast_fp16 = add(x = q_head_155_cast_fp16, y = var_12457_to_fp16)[name = string("op_12458_cast_fp16")];
            tensor<int32, [1]> q_u_155_axes_1 = const()[name = string("q_u_155_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_155_cast_fp16 = expand_dims(axes = q_u_155_axes_1, x = var_12458_cast_fp16)[name = string("q_u_155_cast_fp16")];
            tensor<fp16, [128]> var_12460_to_fp16 = const()[name = string("op_12460_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281982080)))];
            tensor<fp16, [1, 51, 128]> var_12461_cast_fp16 = add(x = q_head_155_cast_fp16, y = var_12460_to_fp16)[name = string("op_12461_cast_fp16")];
            tensor<int32, [1]> q_v_155_axes_1 = const()[name = string("q_v_155_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_155_cast_fp16 = expand_dims(axes = q_v_155_axes_1, x = var_12461_cast_fp16)[name = string("q_v_155_cast_fp16")];
            tensor<int32, [1]> k_head_311_axes_1 = const()[name = string("k_head_311_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_311_cast_fp16 = expand_dims(axes = k_head_311_axes_1, x = k_head_309_cast_fp16)[name = string("k_head_311_cast_fp16")];
            tensor<int32, [1]> v_head_311_axes_1 = const()[name = string("v_head_311_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_311_cast_fp16 = expand_dims(axes = v_head_311_axes_1, x = v_head_309_cast_fp16)[name = string("v_head_311_cast_fp16")];
            tensor<int32, [1]> p_head_311_axes_1 = const()[name = string("p_head_311_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_311_cast_fp16 = expand_dims(axes = p_head_311_axes_1, x = p_head_309_cast_fp16)[name = string("p_head_311_cast_fp16")];
            bool var_12467_transpose_x_3 = const()[name = string("op_12467_transpose_x_3"), val = bool(false)];
            bool var_12467_transpose_y_3 = const()[name = string("op_12467_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_12467_cast_fp16 = matmul(transpose_x = var_12467_transpose_x_3, transpose_y = var_12467_transpose_y_3, x = q_u_155_cast_fp16, y = k_head_311_cast_fp16)[name = string("op_12467_cast_fp16")];
            fp16 var_12468_to_fp16 = const()[name = string("op_12468_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_155_cast_fp16 = mul(x = var_12467_cast_fp16, y = var_12468_to_fp16)[name = string("scores_content_155_cast_fp16")];
            bool x_817_transpose_x_3 = const()[name = string("x_817_transpose_x_3"), val = bool(false)];
            bool x_817_transpose_y_3 = const()[name = string("x_817_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_817_cast_fp16 = matmul(transpose_x = x_817_transpose_x_3, transpose_y = x_817_transpose_y_3, x = q_v_155_cast_fp16, y = p_head_311_cast_fp16)[name = string("x_817_cast_fp16")];
            tensor<int32, [8]> x_819_pad_1 = const()[name = string("x_819_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_819_mode_1 = const()[name = string("x_819_mode_1"), val = string("constant")];
            fp16 const_1825_to_fp16 = const()[name = string("const_1825_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_819_cast_fp16 = pad(constant_val = const_1825_to_fp16, mode = x_819_mode_1, pad = x_819_pad_1, x = x_817_cast_fp16)[name = string("x_819_cast_fp16")];
            tensor<int32, [4]> var_12482 = const()[name = string("op_12482"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_821_cast_fp16 = reshape(shape = var_12482, x = x_819_cast_fp16)[name = string("x_821_cast_fp16")];
            tensor<int32, [4]> var_12486_begin_1 = const()[name = string("op_12486_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12486_end_1 = const()[name = string("op_12486_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_12486_end_mask_1 = const()[name = string("op_12486_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_12486_cast_fp16 = slice_by_index(begin = var_12486_begin_1, end = var_12486_end_1, end_mask = var_12486_end_mask_1, x = x_821_cast_fp16)[name = string("op_12486_cast_fp16")];
            tensor<int32, [4]> var_12488 = const()[name = string("op_12488"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_12489_cast_fp16 = reshape(shape = var_12488, x = var_12486_cast_fp16)[name = string("op_12489_cast_fp16")];
            tensor<int32, [4]> var_12494_begin_1 = const()[name = string("op_12494_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12494_end_1 = const()[name = string("op_12494_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_12494_end_mask_1 = const()[name = string("op_12494_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_12494_cast_fp16 = slice_by_index(begin = var_12494_begin_1, end = var_12494_end_1, end_mask = var_12494_end_mask_1, x = var_12489_cast_fp16)[name = string("op_12494_cast_fp16")];
            fp16 var_12495_to_fp16 = const()[name = string("op_12495_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_155_cast_fp16 = mul(x = var_12494_cast_fp16, y = var_12495_to_fp16)[name = string("scores_pos_155_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_155_cast_fp16 = add(x = scores_content_155_cast_fp16, y = scores_pos_155_cast_fp16)[name = string("logits_155_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_12498_cast_fp16 = softmax(axis = var_11618, x = logits_155_cast_fp16)[name = string("op_12498_cast_fp16")];
            bool var_12500_transpose_x_1 = const()[name = string("op_12500_transpose_x_1"), val = bool(false)];
            bool var_12500_transpose_y_1 = const()[name = string("op_12500_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_12500_cast_fp16 = matmul(transpose_x = var_12500_transpose_x_1, transpose_y = var_12500_transpose_y_1, x = var_12498_cast_fp16, y = v_head_311_cast_fp16)[name = string("op_12500_cast_fp16")];
            tensor<int32, [1]> var_12501_axes_1 = const()[name = string("op_12501_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_12501_cast_fp16 = squeeze(axes = var_12501_axes_1, x = var_12500_cast_fp16)[name = string("op_12501_cast_fp16")];
            string dense_output_785_pad_type_1 = const()[name = string("dense_output_785_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_785_strides_1 = const()[name = string("dense_output_785_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_785_pad_1 = const()[name = string("dense_output_785_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_785_dilations_1 = const()[name = string("dense_output_785_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_785_groups_1 = const()[name = string("dense_output_785_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281982400))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282113536))))[name = string("layers_9_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_785_cast_fp16 = conv(dilations = dense_output_785_dilations_1, groups = dense_output_785_groups_1, pad = dense_output_785_pad_1, pad_type = dense_output_785_pad_type_1, strides = dense_output_785_strides_1, weight = layers_9_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_785_cast_fp16")];
            string sparse_output_785_pad_type_1 = const()[name = string("sparse_output_785_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_785_strides_1 = const()[name = string("sparse_output_785_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_785_pad_1 = const()[name = string("sparse_output_785_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_785_dilations_1 = const()[name = string("sparse_output_785_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_785_groups_1 = const()[name = string("sparse_output_785_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282116800))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282114112))))[name = string("layers_9_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_785_cast_fp16 = conv(dilations = sparse_output_785_dilations_1, groups = sparse_output_785_groups_1, pad = sparse_output_785_pad_1, pad_type = sparse_output_785_pad_type_1, strides = sparse_output_785_strides_1, weight = layers_9_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_785_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12516_cast_fp16 = add(x = dense_output_785_cast_fp16, y = sparse_output_785_cast_fp16)[name = string("op_12516_cast_fp16")];
            tensor<int32, [4]> var_12517 = const()[name = string("op_12517"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12519 = const()[name = string("op_12519"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12518_cast_fp16 = transpose(perm = var_12517, x = var_12516_cast_fp16)[name = string("transpose_587")];
            tensor<fp16, [1, 51, 128]> q_head_157_cast_fp16 = reshape(shape = var_12519, x = var_12518_cast_fp16)[name = string("q_head_157_cast_fp16")];
            string dense_output_787_pad_type_1 = const()[name = string("dense_output_787_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_787_strides_1 = const()[name = string("dense_output_787_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_787_pad_1 = const()[name = string("dense_output_787_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_787_dilations_1 = const()[name = string("dense_output_787_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_787_groups_1 = const()[name = string("dense_output_787_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282133248))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282264384))))[name = string("layers_9_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_787_cast_fp16 = conv(dilations = dense_output_787_dilations_1, groups = dense_output_787_groups_1, pad = dense_output_787_pad_1, pad_type = dense_output_787_pad_type_1, strides = dense_output_787_strides_1, weight = layers_9_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_787_cast_fp16")];
            string sparse_output_787_pad_type_1 = const()[name = string("sparse_output_787_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_787_strides_1 = const()[name = string("sparse_output_787_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_787_pad_1 = const()[name = string("sparse_output_787_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_787_dilations_1 = const()[name = string("sparse_output_787_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_787_groups_1 = const()[name = string("sparse_output_787_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282267648))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282264960))))[name = string("layers_9_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_787_cast_fp16 = conv(dilations = sparse_output_787_dilations_1, groups = sparse_output_787_groups_1, pad = sparse_output_787_pad_1, pad_type = sparse_output_787_pad_type_1, strides = sparse_output_787_strides_1, weight = layers_9_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_787_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12535_cast_fp16 = add(x = dense_output_787_cast_fp16, y = sparse_output_787_cast_fp16)[name = string("op_12535_cast_fp16")];
            tensor<int32, [4]> var_12536 = const()[name = string("op_12536"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12538 = const()[name = string("op_12538"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12537_cast_fp16 = transpose(perm = var_12536, x = var_12535_cast_fp16)[name = string("transpose_586")];
            tensor<fp16, [1, 51, 128]> k_head_313_cast_fp16 = reshape(shape = var_12538, x = var_12537_cast_fp16)[name = string("k_head_313_cast_fp16")];
            string dense_output_789_pad_type_1 = const()[name = string("dense_output_789_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_789_strides_1 = const()[name = string("dense_output_789_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_789_pad_1 = const()[name = string("dense_output_789_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_789_dilations_1 = const()[name = string("dense_output_789_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_789_groups_1 = const()[name = string("dense_output_789_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282284096))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282415232))))[name = string("layers_9_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_789_cast_fp16 = conv(dilations = dense_output_789_dilations_1, groups = dense_output_789_groups_1, pad = dense_output_789_pad_1, pad_type = dense_output_789_pad_type_1, strides = dense_output_789_strides_1, weight = layers_9_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_789_cast_fp16")];
            string sparse_output_789_pad_type_1 = const()[name = string("sparse_output_789_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_789_strides_1 = const()[name = string("sparse_output_789_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_789_pad_1 = const()[name = string("sparse_output_789_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_789_dilations_1 = const()[name = string("sparse_output_789_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_789_groups_1 = const()[name = string("sparse_output_789_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282418496))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282415808))))[name = string("layers_9_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_789_cast_fp16 = conv(dilations = sparse_output_789_dilations_1, groups = sparse_output_789_groups_1, pad = sparse_output_789_pad_1, pad_type = sparse_output_789_pad_type_1, strides = sparse_output_789_strides_1, weight = layers_9_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_789_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12554_cast_fp16 = add(x = dense_output_789_cast_fp16, y = sparse_output_789_cast_fp16)[name = string("op_12554_cast_fp16")];
            tensor<int32, [4]> var_12555 = const()[name = string("op_12555"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12557 = const()[name = string("op_12557"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12556_cast_fp16 = transpose(perm = var_12555, x = var_12554_cast_fp16)[name = string("transpose_585")];
            tensor<fp16, [1, 51, 128]> v_head_313_cast_fp16 = reshape(shape = var_12557, x = var_12556_cast_fp16)[name = string("v_head_313_cast_fp16")];
            string dense_output_791_pad_type_1 = const()[name = string("dense_output_791_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_791_strides_1 = const()[name = string("dense_output_791_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_791_pad_1 = const()[name = string("dense_output_791_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_791_dilations_1 = const()[name = string("dense_output_791_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_791_groups_1 = const()[name = string("dense_output_791_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282434944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282566080))))[name = string("layers_9_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_791_cast_fp16 = conv(dilations = dense_output_791_dilations_1, groups = dense_output_791_groups_1, pad = dense_output_791_pad_1, pad_type = dense_output_791_pad_type_1, strides = dense_output_791_strides_1, weight = layers_9_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_791_cast_fp16")];
            string sparse_output_791_pad_type_1 = const()[name = string("sparse_output_791_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_791_strides_1 = const()[name = string("sparse_output_791_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_791_pad_1 = const()[name = string("sparse_output_791_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_791_dilations_1 = const()[name = string("sparse_output_791_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_791_groups_1 = const()[name = string("sparse_output_791_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282569344))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282566656))))[name = string("layers_9_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_791_cast_fp16 = conv(dilations = sparse_output_791_dilations_1, groups = sparse_output_791_groups_1, pad = sparse_output_791_pad_1, pad_type = sparse_output_791_pad_type_1, strides = sparse_output_791_strides_1, weight = layers_9_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_791_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_12573_cast_fp16 = add(x = dense_output_791_cast_fp16, y = sparse_output_791_cast_fp16)[name = string("op_12573_cast_fp16")];
            tensor<int32, [4]> var_12574 = const()[name = string("op_12574"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12576 = const()[name = string("op_12576"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_12575_cast_fp16 = transpose(perm = var_12574, x = var_12573_cast_fp16)[name = string("transpose_584")];
            tensor<fp16, [1, 101, 128]> p_head_313_cast_fp16 = reshape(shape = var_12576, x = var_12575_cast_fp16)[name = string("p_head_313_cast_fp16")];
            tensor<fp16, [128]> var_12578_to_fp16 = const()[name = string("op_12578_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282585792)))];
            tensor<fp16, [1, 51, 128]> var_12579_cast_fp16 = add(x = q_head_157_cast_fp16, y = var_12578_to_fp16)[name = string("op_12579_cast_fp16")];
            tensor<int32, [1]> q_u_157_axes_1 = const()[name = string("q_u_157_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_157_cast_fp16 = expand_dims(axes = q_u_157_axes_1, x = var_12579_cast_fp16)[name = string("q_u_157_cast_fp16")];
            tensor<fp16, [128]> var_12581_to_fp16 = const()[name = string("op_12581_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282586112)))];
            tensor<fp16, [1, 51, 128]> var_12582_cast_fp16 = add(x = q_head_157_cast_fp16, y = var_12581_to_fp16)[name = string("op_12582_cast_fp16")];
            tensor<int32, [1]> q_v_157_axes_1 = const()[name = string("q_v_157_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_157_cast_fp16 = expand_dims(axes = q_v_157_axes_1, x = var_12582_cast_fp16)[name = string("q_v_157_cast_fp16")];
            tensor<int32, [1]> k_head_315_axes_1 = const()[name = string("k_head_315_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_315_cast_fp16 = expand_dims(axes = k_head_315_axes_1, x = k_head_313_cast_fp16)[name = string("k_head_315_cast_fp16")];
            tensor<int32, [1]> v_head_315_axes_1 = const()[name = string("v_head_315_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_315_cast_fp16 = expand_dims(axes = v_head_315_axes_1, x = v_head_313_cast_fp16)[name = string("v_head_315_cast_fp16")];
            tensor<int32, [1]> p_head_315_axes_1 = const()[name = string("p_head_315_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_315_cast_fp16 = expand_dims(axes = p_head_315_axes_1, x = p_head_313_cast_fp16)[name = string("p_head_315_cast_fp16")];
            bool var_12588_transpose_x_3 = const()[name = string("op_12588_transpose_x_3"), val = bool(false)];
            bool var_12588_transpose_y_3 = const()[name = string("op_12588_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_12588_cast_fp16 = matmul(transpose_x = var_12588_transpose_x_3, transpose_y = var_12588_transpose_y_3, x = q_u_157_cast_fp16, y = k_head_315_cast_fp16)[name = string("op_12588_cast_fp16")];
            fp16 var_12589_to_fp16 = const()[name = string("op_12589_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_157_cast_fp16 = mul(x = var_12588_cast_fp16, y = var_12589_to_fp16)[name = string("scores_content_157_cast_fp16")];
            bool x_825_transpose_x_3 = const()[name = string("x_825_transpose_x_3"), val = bool(false)];
            bool x_825_transpose_y_3 = const()[name = string("x_825_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_825_cast_fp16 = matmul(transpose_x = x_825_transpose_x_3, transpose_y = x_825_transpose_y_3, x = q_v_157_cast_fp16, y = p_head_315_cast_fp16)[name = string("x_825_cast_fp16")];
            tensor<int32, [8]> x_827_pad_1 = const()[name = string("x_827_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_827_mode_1 = const()[name = string("x_827_mode_1"), val = string("constant")];
            fp16 const_1831_to_fp16 = const()[name = string("const_1831_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_827_cast_fp16 = pad(constant_val = const_1831_to_fp16, mode = x_827_mode_1, pad = x_827_pad_1, x = x_825_cast_fp16)[name = string("x_827_cast_fp16")];
            tensor<int32, [4]> var_12603 = const()[name = string("op_12603"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_829_cast_fp16 = reshape(shape = var_12603, x = x_827_cast_fp16)[name = string("x_829_cast_fp16")];
            tensor<int32, [4]> var_12607_begin_1 = const()[name = string("op_12607_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12607_end_1 = const()[name = string("op_12607_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_12607_end_mask_1 = const()[name = string("op_12607_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_12607_cast_fp16 = slice_by_index(begin = var_12607_begin_1, end = var_12607_end_1, end_mask = var_12607_end_mask_1, x = x_829_cast_fp16)[name = string("op_12607_cast_fp16")];
            tensor<int32, [4]> var_12609 = const()[name = string("op_12609"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_12610_cast_fp16 = reshape(shape = var_12609, x = var_12607_cast_fp16)[name = string("op_12610_cast_fp16")];
            tensor<int32, [4]> var_12615_begin_1 = const()[name = string("op_12615_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12615_end_1 = const()[name = string("op_12615_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_12615_end_mask_1 = const()[name = string("op_12615_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_12615_cast_fp16 = slice_by_index(begin = var_12615_begin_1, end = var_12615_end_1, end_mask = var_12615_end_mask_1, x = var_12610_cast_fp16)[name = string("op_12615_cast_fp16")];
            fp16 var_12616_to_fp16 = const()[name = string("op_12616_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_157_cast_fp16 = mul(x = var_12615_cast_fp16, y = var_12616_to_fp16)[name = string("scores_pos_157_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_157_cast_fp16 = add(x = scores_content_157_cast_fp16, y = scores_pos_157_cast_fp16)[name = string("logits_157_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_12619_cast_fp16 = softmax(axis = var_11618, x = logits_157_cast_fp16)[name = string("op_12619_cast_fp16")];
            bool var_12621_transpose_x_1 = const()[name = string("op_12621_transpose_x_1"), val = bool(false)];
            bool var_12621_transpose_y_1 = const()[name = string("op_12621_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_12621_cast_fp16 = matmul(transpose_x = var_12621_transpose_x_1, transpose_y = var_12621_transpose_y_1, x = var_12619_cast_fp16, y = v_head_315_cast_fp16)[name = string("op_12621_cast_fp16")];
            tensor<int32, [1]> var_12622_axes_1 = const()[name = string("op_12622_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_12622_cast_fp16 = squeeze(axes = var_12622_axes_1, x = var_12621_cast_fp16)[name = string("op_12622_cast_fp16")];
            string dense_output_793_pad_type_1 = const()[name = string("dense_output_793_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_793_strides_1 = const()[name = string("dense_output_793_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_793_pad_1 = const()[name = string("dense_output_793_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_793_dilations_1 = const()[name = string("dense_output_793_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_793_groups_1 = const()[name = string("dense_output_793_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282586432))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282717568))))[name = string("layers_9_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_793_cast_fp16 = conv(dilations = dense_output_793_dilations_1, groups = dense_output_793_groups_1, pad = dense_output_793_pad_1, pad_type = dense_output_793_pad_type_1, strides = dense_output_793_strides_1, weight = layers_9_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_793_cast_fp16")];
            string sparse_output_793_pad_type_1 = const()[name = string("sparse_output_793_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_793_strides_1 = const()[name = string("sparse_output_793_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_793_pad_1 = const()[name = string("sparse_output_793_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_793_dilations_1 = const()[name = string("sparse_output_793_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_793_groups_1 = const()[name = string("sparse_output_793_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282720832))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282718144))))[name = string("layers_9_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_793_cast_fp16 = conv(dilations = sparse_output_793_dilations_1, groups = sparse_output_793_groups_1, pad = sparse_output_793_pad_1, pad_type = sparse_output_793_pad_type_1, strides = sparse_output_793_strides_1, weight = layers_9_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_793_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12637_cast_fp16 = add(x = dense_output_793_cast_fp16, y = sparse_output_793_cast_fp16)[name = string("op_12637_cast_fp16")];
            tensor<int32, [4]> var_12638 = const()[name = string("op_12638"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12640 = const()[name = string("op_12640"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12639_cast_fp16 = transpose(perm = var_12638, x = var_12637_cast_fp16)[name = string("transpose_583")];
            tensor<fp16, [1, 51, 128]> q_head_159_cast_fp16 = reshape(shape = var_12640, x = var_12639_cast_fp16)[name = string("q_head_159_cast_fp16")];
            string dense_output_795_pad_type_1 = const()[name = string("dense_output_795_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_795_strides_1 = const()[name = string("dense_output_795_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_795_pad_1 = const()[name = string("dense_output_795_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_795_dilations_1 = const()[name = string("dense_output_795_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_795_groups_1 = const()[name = string("dense_output_795_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282737280))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282868416))))[name = string("layers_9_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_795_cast_fp16 = conv(dilations = dense_output_795_dilations_1, groups = dense_output_795_groups_1, pad = dense_output_795_pad_1, pad_type = dense_output_795_pad_type_1, strides = dense_output_795_strides_1, weight = layers_9_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_795_cast_fp16")];
            string sparse_output_795_pad_type_1 = const()[name = string("sparse_output_795_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_795_strides_1 = const()[name = string("sparse_output_795_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_795_pad_1 = const()[name = string("sparse_output_795_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_795_dilations_1 = const()[name = string("sparse_output_795_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_795_groups_1 = const()[name = string("sparse_output_795_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282871680))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282868992))))[name = string("layers_9_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_795_cast_fp16 = conv(dilations = sparse_output_795_dilations_1, groups = sparse_output_795_groups_1, pad = sparse_output_795_pad_1, pad_type = sparse_output_795_pad_type_1, strides = sparse_output_795_strides_1, weight = layers_9_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_795_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12656_cast_fp16 = add(x = dense_output_795_cast_fp16, y = sparse_output_795_cast_fp16)[name = string("op_12656_cast_fp16")];
            tensor<int32, [4]> var_12657 = const()[name = string("op_12657"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12659 = const()[name = string("op_12659"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12658_cast_fp16 = transpose(perm = var_12657, x = var_12656_cast_fp16)[name = string("transpose_582")];
            tensor<fp16, [1, 51, 128]> k_head_317_cast_fp16 = reshape(shape = var_12659, x = var_12658_cast_fp16)[name = string("k_head_317_cast_fp16")];
            string dense_output_797_pad_type_1 = const()[name = string("dense_output_797_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_797_strides_1 = const()[name = string("dense_output_797_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_797_pad_1 = const()[name = string("dense_output_797_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_797_dilations_1 = const()[name = string("dense_output_797_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_797_groups_1 = const()[name = string("dense_output_797_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282888128))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283019264))))[name = string("layers_9_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_797_cast_fp16 = conv(dilations = dense_output_797_dilations_1, groups = dense_output_797_groups_1, pad = dense_output_797_pad_1, pad_type = dense_output_797_pad_type_1, strides = dense_output_797_strides_1, weight = layers_9_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("dense_output_797_cast_fp16")];
            string sparse_output_797_pad_type_1 = const()[name = string("sparse_output_797_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_797_strides_1 = const()[name = string("sparse_output_797_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_797_pad_1 = const()[name = string("sparse_output_797_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_797_dilations_1 = const()[name = string("sparse_output_797_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_797_groups_1 = const()[name = string("sparse_output_797_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283022528))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283019840))))[name = string("layers_9_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_797_cast_fp16 = conv(dilations = sparse_output_797_dilations_1, groups = sparse_output_797_groups_1, pad = sparse_output_797_pad_1, pad_type = sparse_output_797_pad_type_1, strides = sparse_output_797_strides_1, weight = layers_9_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_443_cast_fp16)[name = string("sparse_output_797_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_12675_cast_fp16 = add(x = dense_output_797_cast_fp16, y = sparse_output_797_cast_fp16)[name = string("op_12675_cast_fp16")];
            tensor<int32, [4]> var_12676 = const()[name = string("op_12676"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12678 = const()[name = string("op_12678"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_12677_cast_fp16 = transpose(perm = var_12676, x = var_12675_cast_fp16)[name = string("transpose_581")];
            tensor<fp16, [1, 51, 128]> v_head_317_cast_fp16 = reshape(shape = var_12678, x = var_12677_cast_fp16)[name = string("v_head_317_cast_fp16")];
            string dense_output_799_pad_type_1 = const()[name = string("dense_output_799_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_799_strides_1 = const()[name = string("dense_output_799_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_799_pad_1 = const()[name = string("dense_output_799_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_799_dilations_1 = const()[name = string("dense_output_799_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_799_groups_1 = const()[name = string("dense_output_799_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283038976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283170112))))[name = string("layers_9_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_799_cast_fp16 = conv(dilations = dense_output_799_dilations_1, groups = dense_output_799_groups_1, pad = dense_output_799_pad_1, pad_type = dense_output_799_pad_type_1, strides = dense_output_799_strides_1, weight = layers_9_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_799_cast_fp16")];
            string sparse_output_799_pad_type_1 = const()[name = string("sparse_output_799_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_799_strides_1 = const()[name = string("sparse_output_799_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_799_pad_1 = const()[name = string("sparse_output_799_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_799_dilations_1 = const()[name = string("sparse_output_799_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_799_groups_1 = const()[name = string("sparse_output_799_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_9_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283173376))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283170688))))[name = string("layers_9_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_799_cast_fp16 = conv(dilations = sparse_output_799_dilations_1, groups = sparse_output_799_groups_1, pad = sparse_output_799_pad_1, pad_type = sparse_output_799_pad_type_1, strides = sparse_output_799_strides_1, weight = layers_9_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_799_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_12694_cast_fp16 = add(x = dense_output_799_cast_fp16, y = sparse_output_799_cast_fp16)[name = string("op_12694_cast_fp16")];
            tensor<int32, [4]> var_12695 = const()[name = string("op_12695"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_12697 = const()[name = string("op_12697"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_12696_cast_fp16 = transpose(perm = var_12695, x = var_12694_cast_fp16)[name = string("transpose_580")];
            tensor<fp16, [1, 101, 128]> p_head_317_cast_fp16 = reshape(shape = var_12697, x = var_12696_cast_fp16)[name = string("p_head_317_cast_fp16")];
            tensor<fp16, [128]> var_12699_to_fp16 = const()[name = string("op_12699_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283189824)))];
            tensor<fp16, [1, 51, 128]> var_12700_cast_fp16 = add(x = q_head_159_cast_fp16, y = var_12699_to_fp16)[name = string("op_12700_cast_fp16")];
            tensor<int32, [1]> q_u_159_axes_1 = const()[name = string("q_u_159_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_159_cast_fp16 = expand_dims(axes = q_u_159_axes_1, x = var_12700_cast_fp16)[name = string("q_u_159_cast_fp16")];
            tensor<fp16, [128]> var_12702_to_fp16 = const()[name = string("op_12702_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283190144)))];
            tensor<fp16, [1, 51, 128]> var_12703_cast_fp16 = add(x = q_head_159_cast_fp16, y = var_12702_to_fp16)[name = string("op_12703_cast_fp16")];
            tensor<int32, [1]> q_v_159_axes_1 = const()[name = string("q_v_159_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_159_cast_fp16 = expand_dims(axes = q_v_159_axes_1, x = var_12703_cast_fp16)[name = string("q_v_159_cast_fp16")];
            tensor<int32, [1]> k_head_319_axes_1 = const()[name = string("k_head_319_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_319_cast_fp16 = expand_dims(axes = k_head_319_axes_1, x = k_head_317_cast_fp16)[name = string("k_head_319_cast_fp16")];
            tensor<int32, [1]> v_head_319_axes_1 = const()[name = string("v_head_319_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_319_cast_fp16 = expand_dims(axes = v_head_319_axes_1, x = v_head_317_cast_fp16)[name = string("v_head_319_cast_fp16")];
            tensor<int32, [1]> p_head_319_axes_1 = const()[name = string("p_head_319_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_319_cast_fp16 = expand_dims(axes = p_head_319_axes_1, x = p_head_317_cast_fp16)[name = string("p_head_319_cast_fp16")];
            bool var_12709_transpose_x_3 = const()[name = string("op_12709_transpose_x_3"), val = bool(false)];
            bool var_12709_transpose_y_3 = const()[name = string("op_12709_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_12709_cast_fp16 = matmul(transpose_x = var_12709_transpose_x_3, transpose_y = var_12709_transpose_y_3, x = q_u_159_cast_fp16, y = k_head_319_cast_fp16)[name = string("op_12709_cast_fp16")];
            fp16 var_12710_to_fp16 = const()[name = string("op_12710_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_159_cast_fp16 = mul(x = var_12709_cast_fp16, y = var_12710_to_fp16)[name = string("scores_content_159_cast_fp16")];
            bool x_833_transpose_x_3 = const()[name = string("x_833_transpose_x_3"), val = bool(false)];
            bool x_833_transpose_y_3 = const()[name = string("x_833_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_833_cast_fp16 = matmul(transpose_x = x_833_transpose_x_3, transpose_y = x_833_transpose_y_3, x = q_v_159_cast_fp16, y = p_head_319_cast_fp16)[name = string("x_833_cast_fp16")];
            tensor<int32, [8]> x_835_pad_1 = const()[name = string("x_835_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_835_mode_1 = const()[name = string("x_835_mode_1"), val = string("constant")];
            fp16 const_1837_to_fp16 = const()[name = string("const_1837_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_835_cast_fp16 = pad(constant_val = const_1837_to_fp16, mode = x_835_mode_1, pad = x_835_pad_1, x = x_833_cast_fp16)[name = string("x_835_cast_fp16")];
            tensor<int32, [4]> var_12724 = const()[name = string("op_12724"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_837_cast_fp16 = reshape(shape = var_12724, x = x_835_cast_fp16)[name = string("x_837_cast_fp16")];
            tensor<int32, [4]> var_12728_begin_1 = const()[name = string("op_12728_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_12728_end_1 = const()[name = string("op_12728_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_12728_end_mask_1 = const()[name = string("op_12728_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_12728_cast_fp16 = slice_by_index(begin = var_12728_begin_1, end = var_12728_end_1, end_mask = var_12728_end_mask_1, x = x_837_cast_fp16)[name = string("op_12728_cast_fp16")];
            tensor<int32, [4]> var_12730 = const()[name = string("op_12730"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_12731_cast_fp16 = reshape(shape = var_12730, x = var_12728_cast_fp16)[name = string("op_12731_cast_fp16")];
            tensor<int32, [4]> var_12736_begin_1 = const()[name = string("op_12736_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_12736_end_1 = const()[name = string("op_12736_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_12736_end_mask_1 = const()[name = string("op_12736_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_12736_cast_fp16 = slice_by_index(begin = var_12736_begin_1, end = var_12736_end_1, end_mask = var_12736_end_mask_1, x = var_12731_cast_fp16)[name = string("op_12736_cast_fp16")];
            fp16 var_12737_to_fp16 = const()[name = string("op_12737_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_159_cast_fp16 = mul(x = var_12736_cast_fp16, y = var_12737_to_fp16)[name = string("scores_pos_159_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_159_cast_fp16 = add(x = scores_content_159_cast_fp16, y = scores_pos_159_cast_fp16)[name = string("logits_159_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_12740_cast_fp16 = softmax(axis = var_11618, x = logits_159_cast_fp16)[name = string("op_12740_cast_fp16")];
            bool var_12742_transpose_x_1 = const()[name = string("op_12742_transpose_x_1"), val = bool(false)];
            bool var_12742_transpose_y_1 = const()[name = string("op_12742_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_12742_cast_fp16 = matmul(transpose_x = var_12742_transpose_x_1, transpose_y = var_12742_transpose_y_1, x = var_12740_cast_fp16, y = v_head_319_cast_fp16)[name = string("op_12742_cast_fp16")];
            tensor<int32, [1]> o_head_19_axes_1 = const()[name = string("o_head_19_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_19_cast_fp16 = squeeze(axes = o_head_19_axes_1, x = var_12742_cast_fp16)[name = string("o_head_19_cast_fp16")];
            bool out_19_interleave_1 = const()[name = string("out_19_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_19_cast_fp16 = concat(axis = var_11618, interleave = out_19_interleave_1, values = (var_11896_cast_fp16, var_12017_cast_fp16, var_12138_cast_fp16, var_12259_cast_fp16, var_12380_cast_fp16, var_12501_cast_fp16, var_12622_cast_fp16, o_head_19_cast_fp16))[name = string("out_19_cast_fp16")];
            tensor<int32, [3]> var_12746_perm_1 = const()[name = string("op_12746_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_451_axes_1 = const()[name = string("input_451_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_12746_cast_fp16 = transpose(perm = var_12746_perm_1, x = out_19_cast_fp16)[name = string("transpose_579")];
            tensor<fp16, [1, 1024, 51, 1]> input_451_cast_fp16 = expand_dims(axes = input_451_axes_1, x = var_12746_cast_fp16)[name = string("input_451_cast_fp16")];
            string dense_output_801_pad_type_1 = const()[name = string("dense_output_801_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_801_strides_1 = const()[name = string("dense_output_801_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_801_pad_1 = const()[name = string("dense_output_801_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_801_dilations_1 = const()[name = string("dense_output_801_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_801_groups_1 = const()[name = string("dense_output_801_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283190464))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284239104))))[name = string("layers_9_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_801_cast_fp16 = conv(dilations = dense_output_801_dilations_1, groups = dense_output_801_groups_1, pad = dense_output_801_pad_1, pad_type = dense_output_801_pad_type_1, strides = dense_output_801_strides_1, weight = layers_9_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_451_cast_fp16)[name = string("dense_output_801_cast_fp16")];
            string sparse_output_801_pad_type_1 = const()[name = string("sparse_output_801_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_801_strides_1 = const()[name = string("sparse_output_801_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_801_pad_1 = const()[name = string("sparse_output_801_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_801_dilations_1 = const()[name = string("sparse_output_801_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_801_groups_1 = const()[name = string("sparse_output_801_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284260736))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284239680))))[name = string("layers_9_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_801_cast_fp16 = conv(dilations = sparse_output_801_dilations_1, groups = sparse_output_801_groups_1, pad = sparse_output_801_pad_1, pad_type = sparse_output_801_pad_type_1, strides = sparse_output_801_strides_1, weight = layers_9_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_451_cast_fp16)[name = string("sparse_output_801_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_19_cast_fp16 = add(x = dense_output_801_cast_fp16, y = sparse_output_801_cast_fp16)[name = string("out_conv_19_cast_fp16")];
            tensor<int32, [1]> var_12763_axes_1 = const()[name = string("op_12763_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_12763_cast_fp16 = squeeze(axes = var_12763_axes_1, x = out_conv_19_cast_fp16)[name = string("op_12763_cast_fp16")];
            tensor<int32, [3]> var_12764_perm_1 = const()[name = string("op_12764_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_12764_cast_fp16 = transpose(perm = var_12764_perm_1, x = var_12763_cast_fp16)[name = string("transpose_578")];
            tensor<fp16, [1, 51, 1024]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_12764_cast_fp16)[name = string("input_453_cast_fp16")];
            tensor<int32, [1]> x_841_axes_1 = const()[name = string("x_841_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_9_norm_conv_weight_to_fp16 = const()[name = string("layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284391872)))];
            tensor<fp16, [1024]> layers_9_norm_conv_bias_to_fp16 = const()[name = string("layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284393984)))];
            tensor<fp16, [1, 51, 1024]> x_841_cast_fp16 = layer_norm(axes = x_841_axes_1, beta = layers_9_norm_conv_bias_to_fp16, epsilon = var_11633_to_fp16, gamma = layers_9_norm_conv_weight_to_fp16, x = input_453_cast_fp16)[name = string("x_841_cast_fp16")];
            tensor<int32, [3]> var_12774_perm_1 = const()[name = string("op_12774_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_455_axes_1 = const()[name = string("input_455_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_12774_cast_fp16 = transpose(perm = var_12774_perm_1, x = x_841_cast_fp16)[name = string("transpose_577")];
            tensor<fp16, [1, 1024, 51, 1]> input_455_cast_fp16 = expand_dims(axes = input_455_axes_1, x = var_12774_cast_fp16)[name = string("input_455_cast_fp16")];
            string dense_output_803_pad_type_1 = const()[name = string("dense_output_803_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_803_strides_1 = const()[name = string("dense_output_803_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_803_pad_1 = const()[name = string("dense_output_803_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_803_dilations_1 = const()[name = string("dense_output_803_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_803_groups_1 = const()[name = string("dense_output_803_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_9_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284396096))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286493312))))[name = string("layers_9_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_803_cast_fp16 = conv(dilations = dense_output_803_dilations_1, groups = dense_output_803_groups_1, pad = dense_output_803_pad_1, pad_type = dense_output_803_pad_type_1, strides = dense_output_803_strides_1, weight = layers_9_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("dense_output_803_cast_fp16")];
            string sparse_output_803_pad_type_1 = const()[name = string("sparse_output_803_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_803_strides_1 = const()[name = string("sparse_output_803_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_803_pad_1 = const()[name = string("sparse_output_803_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_803_dilations_1 = const()[name = string("sparse_output_803_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_803_groups_1 = const()[name = string("sparse_output_803_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_9_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286535936))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286493888))))[name = string("layers_9_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_803_cast_fp16 = conv(dilations = sparse_output_803_dilations_1, groups = sparse_output_803_groups_1, pad = sparse_output_803_pad_1, pad_type = sparse_output_803_pad_type_1, strides = sparse_output_803_strides_1, weight = layers_9_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_455_cast_fp16)[name = string("sparse_output_803_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_457_cast_fp16 = add(x = dense_output_803_cast_fp16, y = sparse_output_803_cast_fp16)[name = string("input_457_cast_fp16")];
            int32 input_459_split_num_splits_1 = const()[name = string("input_459_split_num_splits_1"), val = int32(2)];
            int32 input_459_split_axis_1 = const()[name = string("input_459_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_459_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_459_split_cast_fp16_1 = split(axis = input_459_split_axis_1, num_splits = input_459_split_num_splits_1, x = input_457_cast_fp16)[name = string("input_459_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_459_split_1_sigmoid_cast_fp16 = sigmoid(x = input_459_split_cast_fp16_1)[name = string("input_459_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_459_cast_fp16 = mul(x = input_459_split_cast_fp16_0, y = input_459_split_1_sigmoid_cast_fp16)[name = string("input_459_cast_fp16")];
            tensor<int32, [8]> input_461_pad_1 = const()[name = string("input_461_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_461_mode_1 = const()[name = string("input_461_mode_1"), val = string("constant")];
            fp16 const_1839_to_fp16 = const()[name = string("const_1839_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_461_cast_fp16 = pad(constant_val = const_1839_to_fp16, mode = input_461_mode_1, pad = input_461_pad_1, x = input_459_cast_fp16)[name = string("input_461_cast_fp16")];
            string dense_output_805_pad_type_1 = const()[name = string("dense_output_805_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_805_strides_1 = const()[name = string("dense_output_805_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_805_pad_1 = const()[name = string("dense_output_805_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_805_dilations_1 = const()[name = string("dense_output_805_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_805_groups_1 = const()[name = string("dense_output_805_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_805_cast_fp16 = conv(dilations = dense_output_805_dilations_1, groups = dense_output_805_groups_1, pad = dense_output_805_pad_1, pad_type = dense_output_805_pad_type_1, strides = dense_output_805_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_461_cast_fp16)[name = string("dense_output_805_cast_fp16")];
            string sparse_output_805_pad_type_1 = const()[name = string("sparse_output_805_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_805_strides_1 = const()[name = string("sparse_output_805_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_805_pad_1 = const()[name = string("sparse_output_805_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_805_dilations_1 = const()[name = string("sparse_output_805_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_805_groups_1 = const()[name = string("sparse_output_805_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_9_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286798144))))[name = string("layers_9_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_805_cast_fp16 = conv(dilations = sparse_output_805_dilations_1, groups = sparse_output_805_groups_1, pad = sparse_output_805_pad_1, pad_type = sparse_output_805_pad_type_1, strides = sparse_output_805_strides_1, weight = layers_9_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_461_cast_fp16)[name = string("sparse_output_805_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_463_cast_fp16 = add(x = dense_output_805_cast_fp16, y = sparse_output_805_cast_fp16)[name = string("input_463_cast_fp16")];
            tensor<fp16, [1024]> layers_9_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_9_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286816640)))];
            tensor<fp16, [1024]> layers_9_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_9_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286818752)))];
            tensor<fp16, [1024]> layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286820864)))];
            tensor<fp16, [1024]> layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286822976)))];
            tensor<fp16, [1, 1024, 51, 1]> input_465_cast_fp16 = batch_norm(beta = layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_11633_to_fp16, gamma = layers_9_conv_batch_norm_weight_to_fp16, mean = layers_9_conv_batch_norm_running_mean_to_fp16, variance = layers_9_conv_batch_norm_running_var_to_fp16, x = input_463_cast_fp16)[name = string("input_465_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_467_cast_fp16 = silu(x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
            string dense_output_807_pad_type_1 = const()[name = string("dense_output_807_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_807_strides_1 = const()[name = string("dense_output_807_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_807_pad_1 = const()[name = string("dense_output_807_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_807_dilations_1 = const()[name = string("dense_output_807_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_807_groups_1 = const()[name = string("dense_output_807_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286825088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287873728))))[name = string("layers_9_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_807_cast_fp16 = conv(dilations = dense_output_807_dilations_1, groups = dense_output_807_groups_1, pad = dense_output_807_pad_1, pad_type = dense_output_807_pad_type_1, strides = dense_output_807_strides_1, weight = layers_9_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_467_cast_fp16)[name = string("dense_output_807_cast_fp16")];
            string sparse_output_807_pad_type_1 = const()[name = string("sparse_output_807_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_807_strides_1 = const()[name = string("sparse_output_807_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_807_pad_1 = const()[name = string("sparse_output_807_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_807_dilations_1 = const()[name = string("sparse_output_807_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_807_groups_1 = const()[name = string("sparse_output_807_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_9_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287895360))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287874304))))[name = string("layers_9_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_807_cast_fp16 = conv(dilations = sparse_output_807_dilations_1, groups = sparse_output_807_groups_1, pad = sparse_output_807_pad_1, pad_type = sparse_output_807_pad_type_1, strides = sparse_output_807_strides_1, weight = layers_9_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_467_cast_fp16)[name = string("sparse_output_807_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_843_cast_fp16 = add(x = dense_output_807_cast_fp16, y = sparse_output_807_cast_fp16)[name = string("x_843_cast_fp16")];
            tensor<int32, [1]> var_12830_axes_1 = const()[name = string("op_12830_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_12830_cast_fp16 = squeeze(axes = var_12830_axes_1, x = x_843_cast_fp16)[name = string("op_12830_cast_fp16")];
            tensor<int32, [3]> var_12831_perm_1 = const()[name = string("op_12831_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_12831_cast_fp16 = transpose(perm = var_12831_perm_1, x = var_12830_cast_fp16)[name = string("transpose_576")];
            tensor<fp16, [1, 51, 1024]> input_469_cast_fp16 = add(x = input_453_cast_fp16, y = var_12831_cast_fp16)[name = string("input_469_cast_fp16")];
            tensor<int32, [1]> x_845_axes_1 = const()[name = string("x_845_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288026496)))];
            tensor<fp16, [1024]> layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288028608)))];
            tensor<fp16, [1, 51, 1024]> x_845_cast_fp16 = layer_norm(axes = x_845_axes_1, beta = layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_11633_to_fp16, gamma = layers_9_norm_feed_forward2_weight_to_fp16, x = input_469_cast_fp16)[name = string("x_845_cast_fp16")];
            tensor<int32, [4]> var_12841 = const()[name = string("op_12841"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_847_cast_fp16 = reshape(shape = var_12841, x = x_845_cast_fp16)[name = string("x_847_cast_fp16")];
            tensor<int32, [4]> input_471_perm_1 = const()[name = string("input_471_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_809_pad_type_1 = const()[name = string("dense_output_809_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_809_strides_1 = const()[name = string("dense_output_809_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_809_pad_1 = const()[name = string("dense_output_809_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_809_dilations_1 = const()[name = string("dense_output_809_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_809_groups_1 = const()[name = string("dense_output_809_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_9_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288030720))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292225088))))[name = string("layers_9_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_471_cast_fp16 = transpose(perm = input_471_perm_1, x = x_847_cast_fp16)[name = string("transpose_575")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_809_cast_fp16 = conv(dilations = dense_output_809_dilations_1, groups = dense_output_809_groups_1, pad = dense_output_809_pad_1, pad_type = dense_output_809_pad_type_1, strides = dense_output_809_strides_1, weight = layers_9_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_471_cast_fp16)[name = string("dense_output_809_cast_fp16")];
            string sparse_output_809_pad_type_1 = const()[name = string("sparse_output_809_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_809_strides_1 = const()[name = string("sparse_output_809_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_809_pad_1 = const()[name = string("sparse_output_809_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_809_dilations_1 = const()[name = string("sparse_output_809_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_809_groups_1 = const()[name = string("sparse_output_809_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_9_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292309632))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292225664))))[name = string("layers_9_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_809_cast_fp16 = conv(dilations = sparse_output_809_dilations_1, groups = sparse_output_809_groups_1, pad = sparse_output_809_pad_1, pad_type = sparse_output_809_pad_type_1, strides = sparse_output_809_strides_1, weight = layers_9_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_471_cast_fp16)[name = string("sparse_output_809_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_473_cast_fp16 = add(x = dense_output_809_cast_fp16, y = sparse_output_809_cast_fp16)[name = string("input_473_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
            string dense_output_811_pad_type_1 = const()[name = string("dense_output_811_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_811_strides_1 = const()[name = string("dense_output_811_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_811_pad_1 = const()[name = string("dense_output_811_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_811_dilations_1 = const()[name = string("dense_output_811_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_811_groups_1 = const()[name = string("dense_output_811_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_9_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292833984))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297028352))))[name = string("layers_9_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_811_cast_fp16 = conv(dilations = dense_output_811_dilations_1, groups = dense_output_811_groups_1, pad = dense_output_811_pad_1, pad_type = dense_output_811_pad_type_1, strides = dense_output_811_strides_1, weight = layers_9_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_475_cast_fp16)[name = string("dense_output_811_cast_fp16")];
            string sparse_output_811_pad_type_1 = const()[name = string("sparse_output_811_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_811_strides_1 = const()[name = string("sparse_output_811_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_811_pad_1 = const()[name = string("sparse_output_811_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_811_dilations_1 = const()[name = string("sparse_output_811_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_811_groups_1 = const()[name = string("sparse_output_811_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_9_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297112896))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297028928))))[name = string("layers_9_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_811_cast_fp16 = conv(dilations = sparse_output_811_dilations_1, groups = sparse_output_811_groups_1, pad = sparse_output_811_pad_1, pad_type = sparse_output_811_pad_type_1, strides = sparse_output_811_strides_1, weight = layers_9_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_475_cast_fp16)[name = string("sparse_output_811_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_849_cast_fp16 = add(x = dense_output_811_cast_fp16, y = sparse_output_811_cast_fp16)[name = string("x_849_cast_fp16")];
            tensor<int32, [4]> x_851_perm_1 = const()[name = string("x_851_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_12876 = const()[name = string("op_12876"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_851_cast_fp16 = transpose(perm = x_851_perm_1, x = x_849_cast_fp16)[name = string("transpose_574")];
            tensor<fp16, [1, 51, 1024]> var_12877_cast_fp16 = reshape(shape = var_12876, x = x_851_cast_fp16)[name = string("op_12877_cast_fp16")];
            fp16 var_12878_to_fp16 = const()[name = string("op_12878_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_12879_cast_fp16 = mul(x = var_12877_cast_fp16, y = var_12878_to_fp16)[name = string("op_12879_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_477_cast_fp16 = add(x = input_469_cast_fp16, y = var_12879_cast_fp16)[name = string("input_477_cast_fp16")];
            tensor<int32, [1]> input_479_axes_1 = const()[name = string("input_479_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_9_norm_out_weight_to_fp16 = const()[name = string("layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297637248)))];
            tensor<fp16, [1024]> layers_9_norm_out_bias_to_fp16 = const()[name = string("layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297639360)))];
            tensor<fp16, [1, 51, 1024]> input_479_cast_fp16 = layer_norm(axes = input_479_axes_1, beta = layers_9_norm_out_bias_to_fp16, epsilon = var_11633_to_fp16, gamma = layers_9_norm_out_weight_to_fp16, x = input_477_cast_fp16)[name = string("input_479_cast_fp16")];
            int32 var_12887 = const()[name = string("op_12887"), val = int32(-1)];
            tensor<int32, [1]> x_853_axes_1 = const()[name = string("x_853_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297641472)))];
            tensor<fp16, [1024]> layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297643584)))];
            fp16 var_12902_to_fp16 = const()[name = string("op_12902_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_853_cast_fp16 = layer_norm(axes = x_853_axes_1, beta = layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_12902_to_fp16, gamma = layers_10_norm_feed_forward1_weight_to_fp16, x = input_479_cast_fp16)[name = string("x_853_cast_fp16")];
            tensor<int32, [4]> var_12921 = const()[name = string("op_12921"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_855_cast_fp16 = reshape(shape = var_12921, x = x_853_cast_fp16)[name = string("x_855_cast_fp16")];
            tensor<int32, [4]> input_481_perm_1 = const()[name = string("input_481_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_813_pad_type_1 = const()[name = string("dense_output_813_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_813_strides_1 = const()[name = string("dense_output_813_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_813_pad_1 = const()[name = string("dense_output_813_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_813_dilations_1 = const()[name = string("dense_output_813_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_813_groups_1 = const()[name = string("dense_output_813_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_10_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297645696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301840064))))[name = string("layers_10_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_481_cast_fp16 = transpose(perm = input_481_perm_1, x = x_855_cast_fp16)[name = string("transpose_573")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_813_cast_fp16 = conv(dilations = dense_output_813_dilations_1, groups = dense_output_813_groups_1, pad = dense_output_813_pad_1, pad_type = dense_output_813_pad_type_1, strides = dense_output_813_strides_1, weight = layers_10_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("dense_output_813_cast_fp16")];
            string sparse_output_813_pad_type_1 = const()[name = string("sparse_output_813_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_813_strides_1 = const()[name = string("sparse_output_813_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_813_pad_1 = const()[name = string("sparse_output_813_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_813_dilations_1 = const()[name = string("sparse_output_813_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_813_groups_1 = const()[name = string("sparse_output_813_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_10_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301924608))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301840640))))[name = string("layers_10_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_813_cast_fp16 = conv(dilations = sparse_output_813_dilations_1, groups = sparse_output_813_groups_1, pad = sparse_output_813_pad_1, pad_type = sparse_output_813_pad_type_1, strides = sparse_output_813_strides_1, weight = layers_10_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_481_cast_fp16)[name = string("sparse_output_813_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_483_cast_fp16 = add(x = dense_output_813_cast_fp16, y = sparse_output_813_cast_fp16)[name = string("input_483_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_485_cast_fp16 = silu(x = input_483_cast_fp16)[name = string("input_485_cast_fp16")];
            string dense_output_815_pad_type_1 = const()[name = string("dense_output_815_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_815_strides_1 = const()[name = string("dense_output_815_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_815_pad_1 = const()[name = string("dense_output_815_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_815_dilations_1 = const()[name = string("dense_output_815_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_815_groups_1 = const()[name = string("dense_output_815_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_10_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302448960))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306643328))))[name = string("layers_10_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_815_cast_fp16 = conv(dilations = dense_output_815_dilations_1, groups = dense_output_815_groups_1, pad = dense_output_815_pad_1, pad_type = dense_output_815_pad_type_1, strides = dense_output_815_strides_1, weight = layers_10_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("dense_output_815_cast_fp16")];
            string sparse_output_815_pad_type_1 = const()[name = string("sparse_output_815_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_815_strides_1 = const()[name = string("sparse_output_815_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_815_pad_1 = const()[name = string("sparse_output_815_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_815_dilations_1 = const()[name = string("sparse_output_815_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_815_groups_1 = const()[name = string("sparse_output_815_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_10_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306727872))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306643904))))[name = string("layers_10_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_815_cast_fp16 = conv(dilations = sparse_output_815_dilations_1, groups = sparse_output_815_groups_1, pad = sparse_output_815_pad_1, pad_type = sparse_output_815_pad_type_1, strides = sparse_output_815_strides_1, weight = layers_10_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_485_cast_fp16)[name = string("sparse_output_815_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_857_cast_fp16 = add(x = dense_output_815_cast_fp16, y = sparse_output_815_cast_fp16)[name = string("x_857_cast_fp16")];
            tensor<int32, [4]> x_859_perm_1 = const()[name = string("x_859_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_12956 = const()[name = string("op_12956"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_859_cast_fp16 = transpose(perm = x_859_perm_1, x = x_857_cast_fp16)[name = string("transpose_572")];
            tensor<fp16, [1, 51, 1024]> var_12957_cast_fp16 = reshape(shape = var_12956, x = x_859_cast_fp16)[name = string("op_12957_cast_fp16")];
            fp16 var_12958_to_fp16 = const()[name = string("op_12958_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_12959_cast_fp16 = mul(x = var_12957_cast_fp16, y = var_12958_to_fp16)[name = string("op_12959_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_487_cast_fp16 = add(x = input_479_cast_fp16, y = var_12959_cast_fp16)[name = string("input_487_cast_fp16")];
            tensor<int32, [1]> q_21_axes_1 = const()[name = string("q_21_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_10_norm_self_att_weight_to_fp16 = const()[name = string("layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307252224)))];
            tensor<fp16, [1024]> layers_10_norm_self_att_bias_to_fp16 = const()[name = string("layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307254336)))];
            tensor<fp16, [1, 51, 1024]> q_21_cast_fp16 = layer_norm(axes = q_21_axes_1, beta = layers_10_norm_self_att_bias_to_fp16, epsilon = var_12902_to_fp16, gamma = layers_10_norm_self_att_weight_to_fp16, x = input_487_cast_fp16)[name = string("q_21_cast_fp16")];
            tensor<int32, [3]> var_13033 = const()[name = string("op_13033"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_489_axes_1 = const()[name = string("input_489_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_13034_cast_fp16 = transpose(perm = var_13033, x = q_21_cast_fp16)[name = string("transpose_571")];
            tensor<fp16, [1, 1024, 51, 1]> input_489_cast_fp16 = expand_dims(axes = input_489_axes_1, x = var_13034_cast_fp16)[name = string("input_489_cast_fp16")];
            string dense_output_817_pad_type_1 = const()[name = string("dense_output_817_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_817_strides_1 = const()[name = string("dense_output_817_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_817_pad_1 = const()[name = string("dense_output_817_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_817_dilations_1 = const()[name = string("dense_output_817_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_817_groups_1 = const()[name = string("dense_output_817_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307256448))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307387584))))[name = string("layers_10_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_817_cast_fp16 = conv(dilations = dense_output_817_dilations_1, groups = dense_output_817_groups_1, pad = dense_output_817_pad_1, pad_type = dense_output_817_pad_type_1, strides = dense_output_817_strides_1, weight = layers_10_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_817_cast_fp16")];
            string sparse_output_817_pad_type_1 = const()[name = string("sparse_output_817_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_817_strides_1 = const()[name = string("sparse_output_817_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_817_pad_1 = const()[name = string("sparse_output_817_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_817_dilations_1 = const()[name = string("sparse_output_817_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_817_groups_1 = const()[name = string("sparse_output_817_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307390848))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307388160))))[name = string("layers_10_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_817_cast_fp16 = conv(dilations = sparse_output_817_dilations_1, groups = sparse_output_817_groups_1, pad = sparse_output_817_pad_1, pad_type = sparse_output_817_pad_type_1, strides = sparse_output_817_strides_1, weight = layers_10_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_817_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13059_cast_fp16 = add(x = dense_output_817_cast_fp16, y = sparse_output_817_cast_fp16)[name = string("op_13059_cast_fp16")];
            tensor<int32, [4]> var_13060 = const()[name = string("op_13060"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13062 = const()[name = string("op_13062"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13061_cast_fp16 = transpose(perm = var_13060, x = var_13059_cast_fp16)[name = string("transpose_570")];
            tensor<fp16, [1, 51, 128]> q_head_161_cast_fp16 = reshape(shape = var_13062, x = var_13061_cast_fp16)[name = string("q_head_161_cast_fp16")];
            string dense_output_819_pad_type_1 = const()[name = string("dense_output_819_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_819_strides_1 = const()[name = string("dense_output_819_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_819_pad_1 = const()[name = string("dense_output_819_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_819_dilations_1 = const()[name = string("dense_output_819_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_819_groups_1 = const()[name = string("dense_output_819_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307407296))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307538432))))[name = string("layers_10_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_819_cast_fp16 = conv(dilations = dense_output_819_dilations_1, groups = dense_output_819_groups_1, pad = dense_output_819_pad_1, pad_type = dense_output_819_pad_type_1, strides = dense_output_819_strides_1, weight = layers_10_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_819_cast_fp16")];
            string sparse_output_819_pad_type_1 = const()[name = string("sparse_output_819_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_819_strides_1 = const()[name = string("sparse_output_819_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_819_pad_1 = const()[name = string("sparse_output_819_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_819_dilations_1 = const()[name = string("sparse_output_819_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_819_groups_1 = const()[name = string("sparse_output_819_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307541696))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307539008))))[name = string("layers_10_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_819_cast_fp16 = conv(dilations = sparse_output_819_dilations_1, groups = sparse_output_819_groups_1, pad = sparse_output_819_pad_1, pad_type = sparse_output_819_pad_type_1, strides = sparse_output_819_strides_1, weight = layers_10_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_819_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13078_cast_fp16 = add(x = dense_output_819_cast_fp16, y = sparse_output_819_cast_fp16)[name = string("op_13078_cast_fp16")];
            tensor<int32, [4]> var_13079 = const()[name = string("op_13079"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13081 = const()[name = string("op_13081"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13080_cast_fp16 = transpose(perm = var_13079, x = var_13078_cast_fp16)[name = string("transpose_569")];
            tensor<fp16, [1, 51, 128]> k_head_321_cast_fp16 = reshape(shape = var_13081, x = var_13080_cast_fp16)[name = string("k_head_321_cast_fp16")];
            string dense_output_821_pad_type_1 = const()[name = string("dense_output_821_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_821_strides_1 = const()[name = string("dense_output_821_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_821_pad_1 = const()[name = string("dense_output_821_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_821_dilations_1 = const()[name = string("dense_output_821_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_821_groups_1 = const()[name = string("dense_output_821_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307558144))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307689280))))[name = string("layers_10_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_821_cast_fp16 = conv(dilations = dense_output_821_dilations_1, groups = dense_output_821_groups_1, pad = dense_output_821_pad_1, pad_type = dense_output_821_pad_type_1, strides = dense_output_821_strides_1, weight = layers_10_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_821_cast_fp16")];
            string sparse_output_821_pad_type_1 = const()[name = string("sparse_output_821_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_821_strides_1 = const()[name = string("sparse_output_821_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_821_pad_1 = const()[name = string("sparse_output_821_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_821_dilations_1 = const()[name = string("sparse_output_821_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_821_groups_1 = const()[name = string("sparse_output_821_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307692544))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307689856))))[name = string("layers_10_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_821_cast_fp16 = conv(dilations = sparse_output_821_dilations_1, groups = sparse_output_821_groups_1, pad = sparse_output_821_pad_1, pad_type = sparse_output_821_pad_type_1, strides = sparse_output_821_strides_1, weight = layers_10_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_821_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13097_cast_fp16 = add(x = dense_output_821_cast_fp16, y = sparse_output_821_cast_fp16)[name = string("op_13097_cast_fp16")];
            tensor<int32, [4]> var_13098 = const()[name = string("op_13098"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13100 = const()[name = string("op_13100"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13099_cast_fp16 = transpose(perm = var_13098, x = var_13097_cast_fp16)[name = string("transpose_568")];
            tensor<fp16, [1, 51, 128]> v_head_321_cast_fp16 = reshape(shape = var_13100, x = var_13099_cast_fp16)[name = string("v_head_321_cast_fp16")];
            string dense_output_823_pad_type_1 = const()[name = string("dense_output_823_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_823_strides_1 = const()[name = string("dense_output_823_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_823_pad_1 = const()[name = string("dense_output_823_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_823_dilations_1 = const()[name = string("dense_output_823_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_823_groups_1 = const()[name = string("dense_output_823_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307708992))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307840128))))[name = string("layers_10_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_823_cast_fp16 = conv(dilations = dense_output_823_dilations_1, groups = dense_output_823_groups_1, pad = dense_output_823_pad_1, pad_type = dense_output_823_pad_type_1, strides = dense_output_823_strides_1, weight = layers_10_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_823_cast_fp16")];
            string sparse_output_823_pad_type_1 = const()[name = string("sparse_output_823_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_823_strides_1 = const()[name = string("sparse_output_823_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_823_pad_1 = const()[name = string("sparse_output_823_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_823_dilations_1 = const()[name = string("sparse_output_823_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_823_groups_1 = const()[name = string("sparse_output_823_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307843392))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307840704))))[name = string("layers_10_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_823_cast_fp16 = conv(dilations = sparse_output_823_dilations_1, groups = sparse_output_823_groups_1, pad = sparse_output_823_pad_1, pad_type = sparse_output_823_pad_type_1, strides = sparse_output_823_strides_1, weight = layers_10_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_823_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_13116_cast_fp16 = add(x = dense_output_823_cast_fp16, y = sparse_output_823_cast_fp16)[name = string("op_13116_cast_fp16")];
            tensor<int32, [4]> var_13117 = const()[name = string("op_13117"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13119 = const()[name = string("op_13119"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_13118_cast_fp16 = transpose(perm = var_13117, x = var_13116_cast_fp16)[name = string("transpose_567")];
            tensor<fp16, [1, 101, 128]> p_head_321_cast_fp16 = reshape(shape = var_13119, x = var_13118_cast_fp16)[name = string("p_head_321_cast_fp16")];
            tensor<fp16, [128]> var_13121_to_fp16 = const()[name = string("op_13121_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307859840)))];
            tensor<fp16, [1, 51, 128]> var_13122_cast_fp16 = add(x = q_head_161_cast_fp16, y = var_13121_to_fp16)[name = string("op_13122_cast_fp16")];
            tensor<int32, [1]> q_u_161_axes_1 = const()[name = string("q_u_161_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_161_cast_fp16 = expand_dims(axes = q_u_161_axes_1, x = var_13122_cast_fp16)[name = string("q_u_161_cast_fp16")];
            tensor<fp16, [128]> var_13124_to_fp16 = const()[name = string("op_13124_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307860160)))];
            tensor<fp16, [1, 51, 128]> var_13125_cast_fp16 = add(x = q_head_161_cast_fp16, y = var_13124_to_fp16)[name = string("op_13125_cast_fp16")];
            tensor<int32, [1]> q_v_161_axes_1 = const()[name = string("q_v_161_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_161_cast_fp16 = expand_dims(axes = q_v_161_axes_1, x = var_13125_cast_fp16)[name = string("q_v_161_cast_fp16")];
            tensor<int32, [1]> k_head_323_axes_1 = const()[name = string("k_head_323_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_323_cast_fp16 = expand_dims(axes = k_head_323_axes_1, x = k_head_321_cast_fp16)[name = string("k_head_323_cast_fp16")];
            tensor<int32, [1]> v_head_323_axes_1 = const()[name = string("v_head_323_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_323_cast_fp16 = expand_dims(axes = v_head_323_axes_1, x = v_head_321_cast_fp16)[name = string("v_head_323_cast_fp16")];
            tensor<int32, [1]> p_head_323_axes_1 = const()[name = string("p_head_323_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_323_cast_fp16 = expand_dims(axes = p_head_323_axes_1, x = p_head_321_cast_fp16)[name = string("p_head_323_cast_fp16")];
            bool var_13131_transpose_x_3 = const()[name = string("op_13131_transpose_x_3"), val = bool(false)];
            bool var_13131_transpose_y_3 = const()[name = string("op_13131_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_13131_cast_fp16 = matmul(transpose_x = var_13131_transpose_x_3, transpose_y = var_13131_transpose_y_3, x = q_u_161_cast_fp16, y = k_head_323_cast_fp16)[name = string("op_13131_cast_fp16")];
            fp16 var_13132_to_fp16 = const()[name = string("op_13132_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_161_cast_fp16 = mul(x = var_13131_cast_fp16, y = var_13132_to_fp16)[name = string("scores_content_161_cast_fp16")];
            bool x_861_transpose_x_3 = const()[name = string("x_861_transpose_x_3"), val = bool(false)];
            bool x_861_transpose_y_3 = const()[name = string("x_861_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_861_cast_fp16 = matmul(transpose_x = x_861_transpose_x_3, transpose_y = x_861_transpose_y_3, x = q_v_161_cast_fp16, y = p_head_323_cast_fp16)[name = string("x_861_cast_fp16")];
            tensor<int32, [8]> x_863_pad_1 = const()[name = string("x_863_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_863_mode_1 = const()[name = string("x_863_mode_1"), val = string("constant")];
            fp16 const_1849_to_fp16 = const()[name = string("const_1849_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_863_cast_fp16 = pad(constant_val = const_1849_to_fp16, mode = x_863_mode_1, pad = x_863_pad_1, x = x_861_cast_fp16)[name = string("x_863_cast_fp16")];
            tensor<int32, [4]> var_13146 = const()[name = string("op_13146"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_865_cast_fp16 = reshape(shape = var_13146, x = x_863_cast_fp16)[name = string("x_865_cast_fp16")];
            tensor<int32, [4]> var_13150_begin_1 = const()[name = string("op_13150_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13150_end_1 = const()[name = string("op_13150_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_13150_end_mask_1 = const()[name = string("op_13150_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_13150_cast_fp16 = slice_by_index(begin = var_13150_begin_1, end = var_13150_end_1, end_mask = var_13150_end_mask_1, x = x_865_cast_fp16)[name = string("op_13150_cast_fp16")];
            tensor<int32, [4]> var_13152 = const()[name = string("op_13152"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_13153_cast_fp16 = reshape(shape = var_13152, x = var_13150_cast_fp16)[name = string("op_13153_cast_fp16")];
            tensor<int32, [4]> var_13158_begin_1 = const()[name = string("op_13158_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13158_end_1 = const()[name = string("op_13158_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_13158_end_mask_1 = const()[name = string("op_13158_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_13158_cast_fp16 = slice_by_index(begin = var_13158_begin_1, end = var_13158_end_1, end_mask = var_13158_end_mask_1, x = var_13153_cast_fp16)[name = string("op_13158_cast_fp16")];
            fp16 var_13159_to_fp16 = const()[name = string("op_13159_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_161_cast_fp16 = mul(x = var_13158_cast_fp16, y = var_13159_to_fp16)[name = string("scores_pos_161_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_161_cast_fp16 = add(x = scores_content_161_cast_fp16, y = scores_pos_161_cast_fp16)[name = string("logits_161_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_13162_cast_fp16 = softmax(axis = var_12887, x = logits_161_cast_fp16)[name = string("op_13162_cast_fp16")];
            bool var_13164_transpose_x_1 = const()[name = string("op_13164_transpose_x_1"), val = bool(false)];
            bool var_13164_transpose_y_1 = const()[name = string("op_13164_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_13164_cast_fp16 = matmul(transpose_x = var_13164_transpose_x_1, transpose_y = var_13164_transpose_y_1, x = var_13162_cast_fp16, y = v_head_323_cast_fp16)[name = string("op_13164_cast_fp16")];
            tensor<int32, [1]> var_13165_axes_1 = const()[name = string("op_13165_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_13165_cast_fp16 = squeeze(axes = var_13165_axes_1, x = var_13164_cast_fp16)[name = string("op_13165_cast_fp16")];
            string dense_output_825_pad_type_1 = const()[name = string("dense_output_825_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_825_strides_1 = const()[name = string("dense_output_825_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_825_pad_1 = const()[name = string("dense_output_825_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_825_dilations_1 = const()[name = string("dense_output_825_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_825_groups_1 = const()[name = string("dense_output_825_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307860480))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307991616))))[name = string("layers_10_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_825_cast_fp16 = conv(dilations = dense_output_825_dilations_1, groups = dense_output_825_groups_1, pad = dense_output_825_pad_1, pad_type = dense_output_825_pad_type_1, strides = dense_output_825_strides_1, weight = layers_10_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_825_cast_fp16")];
            string sparse_output_825_pad_type_1 = const()[name = string("sparse_output_825_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_825_strides_1 = const()[name = string("sparse_output_825_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_825_pad_1 = const()[name = string("sparse_output_825_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_825_dilations_1 = const()[name = string("sparse_output_825_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_825_groups_1 = const()[name = string("sparse_output_825_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307994880))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307992192))))[name = string("layers_10_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_825_cast_fp16 = conv(dilations = sparse_output_825_dilations_1, groups = sparse_output_825_groups_1, pad = sparse_output_825_pad_1, pad_type = sparse_output_825_pad_type_1, strides = sparse_output_825_strides_1, weight = layers_10_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_825_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13180_cast_fp16 = add(x = dense_output_825_cast_fp16, y = sparse_output_825_cast_fp16)[name = string("op_13180_cast_fp16")];
            tensor<int32, [4]> var_13181 = const()[name = string("op_13181"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13183 = const()[name = string("op_13183"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13182_cast_fp16 = transpose(perm = var_13181, x = var_13180_cast_fp16)[name = string("transpose_566")];
            tensor<fp16, [1, 51, 128]> q_head_163_cast_fp16 = reshape(shape = var_13183, x = var_13182_cast_fp16)[name = string("q_head_163_cast_fp16")];
            string dense_output_827_pad_type_1 = const()[name = string("dense_output_827_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_827_strides_1 = const()[name = string("dense_output_827_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_827_pad_1 = const()[name = string("dense_output_827_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_827_dilations_1 = const()[name = string("dense_output_827_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_827_groups_1 = const()[name = string("dense_output_827_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308011328))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308142464))))[name = string("layers_10_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_827_cast_fp16 = conv(dilations = dense_output_827_dilations_1, groups = dense_output_827_groups_1, pad = dense_output_827_pad_1, pad_type = dense_output_827_pad_type_1, strides = dense_output_827_strides_1, weight = layers_10_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_827_cast_fp16")];
            string sparse_output_827_pad_type_1 = const()[name = string("sparse_output_827_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_827_strides_1 = const()[name = string("sparse_output_827_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_827_pad_1 = const()[name = string("sparse_output_827_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_827_dilations_1 = const()[name = string("sparse_output_827_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_827_groups_1 = const()[name = string("sparse_output_827_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308145728))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308143040))))[name = string("layers_10_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_827_cast_fp16 = conv(dilations = sparse_output_827_dilations_1, groups = sparse_output_827_groups_1, pad = sparse_output_827_pad_1, pad_type = sparse_output_827_pad_type_1, strides = sparse_output_827_strides_1, weight = layers_10_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_827_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13199_cast_fp16 = add(x = dense_output_827_cast_fp16, y = sparse_output_827_cast_fp16)[name = string("op_13199_cast_fp16")];
            tensor<int32, [4]> var_13200 = const()[name = string("op_13200"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13202 = const()[name = string("op_13202"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13201_cast_fp16 = transpose(perm = var_13200, x = var_13199_cast_fp16)[name = string("transpose_565")];
            tensor<fp16, [1, 51, 128]> k_head_325_cast_fp16 = reshape(shape = var_13202, x = var_13201_cast_fp16)[name = string("k_head_325_cast_fp16")];
            string dense_output_829_pad_type_1 = const()[name = string("dense_output_829_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_829_strides_1 = const()[name = string("dense_output_829_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_829_pad_1 = const()[name = string("dense_output_829_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_829_dilations_1 = const()[name = string("dense_output_829_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_829_groups_1 = const()[name = string("dense_output_829_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308162176))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308293312))))[name = string("layers_10_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_829_cast_fp16 = conv(dilations = dense_output_829_dilations_1, groups = dense_output_829_groups_1, pad = dense_output_829_pad_1, pad_type = dense_output_829_pad_type_1, strides = dense_output_829_strides_1, weight = layers_10_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_829_cast_fp16")];
            string sparse_output_829_pad_type_1 = const()[name = string("sparse_output_829_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_829_strides_1 = const()[name = string("sparse_output_829_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_829_pad_1 = const()[name = string("sparse_output_829_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_829_dilations_1 = const()[name = string("sparse_output_829_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_829_groups_1 = const()[name = string("sparse_output_829_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308296576))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308293888))))[name = string("layers_10_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_829_cast_fp16 = conv(dilations = sparse_output_829_dilations_1, groups = sparse_output_829_groups_1, pad = sparse_output_829_pad_1, pad_type = sparse_output_829_pad_type_1, strides = sparse_output_829_strides_1, weight = layers_10_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_829_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13218_cast_fp16 = add(x = dense_output_829_cast_fp16, y = sparse_output_829_cast_fp16)[name = string("op_13218_cast_fp16")];
            tensor<int32, [4]> var_13219 = const()[name = string("op_13219"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13221 = const()[name = string("op_13221"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13220_cast_fp16 = transpose(perm = var_13219, x = var_13218_cast_fp16)[name = string("transpose_564")];
            tensor<fp16, [1, 51, 128]> v_head_325_cast_fp16 = reshape(shape = var_13221, x = var_13220_cast_fp16)[name = string("v_head_325_cast_fp16")];
            string dense_output_831_pad_type_1 = const()[name = string("dense_output_831_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_831_strides_1 = const()[name = string("dense_output_831_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_831_pad_1 = const()[name = string("dense_output_831_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_831_dilations_1 = const()[name = string("dense_output_831_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_831_groups_1 = const()[name = string("dense_output_831_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308313024))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308444160))))[name = string("layers_10_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_831_cast_fp16 = conv(dilations = dense_output_831_dilations_1, groups = dense_output_831_groups_1, pad = dense_output_831_pad_1, pad_type = dense_output_831_pad_type_1, strides = dense_output_831_strides_1, weight = layers_10_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_831_cast_fp16")];
            string sparse_output_831_pad_type_1 = const()[name = string("sparse_output_831_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_831_strides_1 = const()[name = string("sparse_output_831_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_831_pad_1 = const()[name = string("sparse_output_831_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_831_dilations_1 = const()[name = string("sparse_output_831_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_831_groups_1 = const()[name = string("sparse_output_831_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308447424))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308444736))))[name = string("layers_10_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_831_cast_fp16 = conv(dilations = sparse_output_831_dilations_1, groups = sparse_output_831_groups_1, pad = sparse_output_831_pad_1, pad_type = sparse_output_831_pad_type_1, strides = sparse_output_831_strides_1, weight = layers_10_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_831_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_13237_cast_fp16 = add(x = dense_output_831_cast_fp16, y = sparse_output_831_cast_fp16)[name = string("op_13237_cast_fp16")];
            tensor<int32, [4]> var_13238 = const()[name = string("op_13238"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13240 = const()[name = string("op_13240"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_13239_cast_fp16 = transpose(perm = var_13238, x = var_13237_cast_fp16)[name = string("transpose_563")];
            tensor<fp16, [1, 101, 128]> p_head_325_cast_fp16 = reshape(shape = var_13240, x = var_13239_cast_fp16)[name = string("p_head_325_cast_fp16")];
            tensor<fp16, [128]> var_13242_to_fp16 = const()[name = string("op_13242_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308463872)))];
            tensor<fp16, [1, 51, 128]> var_13243_cast_fp16 = add(x = q_head_163_cast_fp16, y = var_13242_to_fp16)[name = string("op_13243_cast_fp16")];
            tensor<int32, [1]> q_u_163_axes_1 = const()[name = string("q_u_163_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_163_cast_fp16 = expand_dims(axes = q_u_163_axes_1, x = var_13243_cast_fp16)[name = string("q_u_163_cast_fp16")];
            tensor<fp16, [128]> var_13245_to_fp16 = const()[name = string("op_13245_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308464192)))];
            tensor<fp16, [1, 51, 128]> var_13246_cast_fp16 = add(x = q_head_163_cast_fp16, y = var_13245_to_fp16)[name = string("op_13246_cast_fp16")];
            tensor<int32, [1]> q_v_163_axes_1 = const()[name = string("q_v_163_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_163_cast_fp16 = expand_dims(axes = q_v_163_axes_1, x = var_13246_cast_fp16)[name = string("q_v_163_cast_fp16")];
            tensor<int32, [1]> k_head_327_axes_1 = const()[name = string("k_head_327_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_327_cast_fp16 = expand_dims(axes = k_head_327_axes_1, x = k_head_325_cast_fp16)[name = string("k_head_327_cast_fp16")];
            tensor<int32, [1]> v_head_327_axes_1 = const()[name = string("v_head_327_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_327_cast_fp16 = expand_dims(axes = v_head_327_axes_1, x = v_head_325_cast_fp16)[name = string("v_head_327_cast_fp16")];
            tensor<int32, [1]> p_head_327_axes_1 = const()[name = string("p_head_327_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_327_cast_fp16 = expand_dims(axes = p_head_327_axes_1, x = p_head_325_cast_fp16)[name = string("p_head_327_cast_fp16")];
            bool var_13252_transpose_x_3 = const()[name = string("op_13252_transpose_x_3"), val = bool(false)];
            bool var_13252_transpose_y_3 = const()[name = string("op_13252_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_13252_cast_fp16 = matmul(transpose_x = var_13252_transpose_x_3, transpose_y = var_13252_transpose_y_3, x = q_u_163_cast_fp16, y = k_head_327_cast_fp16)[name = string("op_13252_cast_fp16")];
            fp16 var_13253_to_fp16 = const()[name = string("op_13253_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_163_cast_fp16 = mul(x = var_13252_cast_fp16, y = var_13253_to_fp16)[name = string("scores_content_163_cast_fp16")];
            bool x_869_transpose_x_3 = const()[name = string("x_869_transpose_x_3"), val = bool(false)];
            bool x_869_transpose_y_3 = const()[name = string("x_869_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_869_cast_fp16 = matmul(transpose_x = x_869_transpose_x_3, transpose_y = x_869_transpose_y_3, x = q_v_163_cast_fp16, y = p_head_327_cast_fp16)[name = string("x_869_cast_fp16")];
            tensor<int32, [8]> x_871_pad_1 = const()[name = string("x_871_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_871_mode_1 = const()[name = string("x_871_mode_1"), val = string("constant")];
            fp16 const_1855_to_fp16 = const()[name = string("const_1855_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_871_cast_fp16 = pad(constant_val = const_1855_to_fp16, mode = x_871_mode_1, pad = x_871_pad_1, x = x_869_cast_fp16)[name = string("x_871_cast_fp16")];
            tensor<int32, [4]> var_13267 = const()[name = string("op_13267"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_873_cast_fp16 = reshape(shape = var_13267, x = x_871_cast_fp16)[name = string("x_873_cast_fp16")];
            tensor<int32, [4]> var_13271_begin_1 = const()[name = string("op_13271_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13271_end_1 = const()[name = string("op_13271_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_13271_end_mask_1 = const()[name = string("op_13271_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_13271_cast_fp16 = slice_by_index(begin = var_13271_begin_1, end = var_13271_end_1, end_mask = var_13271_end_mask_1, x = x_873_cast_fp16)[name = string("op_13271_cast_fp16")];
            tensor<int32, [4]> var_13273 = const()[name = string("op_13273"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_13274_cast_fp16 = reshape(shape = var_13273, x = var_13271_cast_fp16)[name = string("op_13274_cast_fp16")];
            tensor<int32, [4]> var_13279_begin_1 = const()[name = string("op_13279_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13279_end_1 = const()[name = string("op_13279_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_13279_end_mask_1 = const()[name = string("op_13279_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_13279_cast_fp16 = slice_by_index(begin = var_13279_begin_1, end = var_13279_end_1, end_mask = var_13279_end_mask_1, x = var_13274_cast_fp16)[name = string("op_13279_cast_fp16")];
            fp16 var_13280_to_fp16 = const()[name = string("op_13280_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_163_cast_fp16 = mul(x = var_13279_cast_fp16, y = var_13280_to_fp16)[name = string("scores_pos_163_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_163_cast_fp16 = add(x = scores_content_163_cast_fp16, y = scores_pos_163_cast_fp16)[name = string("logits_163_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_13283_cast_fp16 = softmax(axis = var_12887, x = logits_163_cast_fp16)[name = string("op_13283_cast_fp16")];
            bool var_13285_transpose_x_1 = const()[name = string("op_13285_transpose_x_1"), val = bool(false)];
            bool var_13285_transpose_y_1 = const()[name = string("op_13285_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_13285_cast_fp16 = matmul(transpose_x = var_13285_transpose_x_1, transpose_y = var_13285_transpose_y_1, x = var_13283_cast_fp16, y = v_head_327_cast_fp16)[name = string("op_13285_cast_fp16")];
            tensor<int32, [1]> var_13286_axes_1 = const()[name = string("op_13286_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_13286_cast_fp16 = squeeze(axes = var_13286_axes_1, x = var_13285_cast_fp16)[name = string("op_13286_cast_fp16")];
            string dense_output_833_pad_type_1 = const()[name = string("dense_output_833_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_833_strides_1 = const()[name = string("dense_output_833_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_833_pad_1 = const()[name = string("dense_output_833_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_833_dilations_1 = const()[name = string("dense_output_833_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_833_groups_1 = const()[name = string("dense_output_833_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308464512))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308595648))))[name = string("layers_10_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_833_cast_fp16 = conv(dilations = dense_output_833_dilations_1, groups = dense_output_833_groups_1, pad = dense_output_833_pad_1, pad_type = dense_output_833_pad_type_1, strides = dense_output_833_strides_1, weight = layers_10_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_833_cast_fp16")];
            string sparse_output_833_pad_type_1 = const()[name = string("sparse_output_833_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_833_strides_1 = const()[name = string("sparse_output_833_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_833_pad_1 = const()[name = string("sparse_output_833_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_833_dilations_1 = const()[name = string("sparse_output_833_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_833_groups_1 = const()[name = string("sparse_output_833_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308598912))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308596224))))[name = string("layers_10_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_833_cast_fp16 = conv(dilations = sparse_output_833_dilations_1, groups = sparse_output_833_groups_1, pad = sparse_output_833_pad_1, pad_type = sparse_output_833_pad_type_1, strides = sparse_output_833_strides_1, weight = layers_10_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_833_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13301_cast_fp16 = add(x = dense_output_833_cast_fp16, y = sparse_output_833_cast_fp16)[name = string("op_13301_cast_fp16")];
            tensor<int32, [4]> var_13302 = const()[name = string("op_13302"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13304 = const()[name = string("op_13304"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13303_cast_fp16 = transpose(perm = var_13302, x = var_13301_cast_fp16)[name = string("transpose_562")];
            tensor<fp16, [1, 51, 128]> q_head_165_cast_fp16 = reshape(shape = var_13304, x = var_13303_cast_fp16)[name = string("q_head_165_cast_fp16")];
            string dense_output_835_pad_type_1 = const()[name = string("dense_output_835_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_835_strides_1 = const()[name = string("dense_output_835_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_835_pad_1 = const()[name = string("dense_output_835_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_835_dilations_1 = const()[name = string("dense_output_835_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_835_groups_1 = const()[name = string("dense_output_835_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308615360))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308746496))))[name = string("layers_10_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_835_cast_fp16 = conv(dilations = dense_output_835_dilations_1, groups = dense_output_835_groups_1, pad = dense_output_835_pad_1, pad_type = dense_output_835_pad_type_1, strides = dense_output_835_strides_1, weight = layers_10_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_835_cast_fp16")];
            string sparse_output_835_pad_type_1 = const()[name = string("sparse_output_835_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_835_strides_1 = const()[name = string("sparse_output_835_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_835_pad_1 = const()[name = string("sparse_output_835_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_835_dilations_1 = const()[name = string("sparse_output_835_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_835_groups_1 = const()[name = string("sparse_output_835_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308749760))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308747072))))[name = string("layers_10_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_835_cast_fp16 = conv(dilations = sparse_output_835_dilations_1, groups = sparse_output_835_groups_1, pad = sparse_output_835_pad_1, pad_type = sparse_output_835_pad_type_1, strides = sparse_output_835_strides_1, weight = layers_10_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_835_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13320_cast_fp16 = add(x = dense_output_835_cast_fp16, y = sparse_output_835_cast_fp16)[name = string("op_13320_cast_fp16")];
            tensor<int32, [4]> var_13321 = const()[name = string("op_13321"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13323 = const()[name = string("op_13323"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13322_cast_fp16 = transpose(perm = var_13321, x = var_13320_cast_fp16)[name = string("transpose_561")];
            tensor<fp16, [1, 51, 128]> k_head_329_cast_fp16 = reshape(shape = var_13323, x = var_13322_cast_fp16)[name = string("k_head_329_cast_fp16")];
            string dense_output_837_pad_type_1 = const()[name = string("dense_output_837_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_837_strides_1 = const()[name = string("dense_output_837_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_837_pad_1 = const()[name = string("dense_output_837_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_837_dilations_1 = const()[name = string("dense_output_837_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_837_groups_1 = const()[name = string("dense_output_837_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308766208))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308897344))))[name = string("layers_10_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_837_cast_fp16 = conv(dilations = dense_output_837_dilations_1, groups = dense_output_837_groups_1, pad = dense_output_837_pad_1, pad_type = dense_output_837_pad_type_1, strides = dense_output_837_strides_1, weight = layers_10_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_837_cast_fp16")];
            string sparse_output_837_pad_type_1 = const()[name = string("sparse_output_837_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_837_strides_1 = const()[name = string("sparse_output_837_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_837_pad_1 = const()[name = string("sparse_output_837_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_837_dilations_1 = const()[name = string("sparse_output_837_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_837_groups_1 = const()[name = string("sparse_output_837_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308900608))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308897920))))[name = string("layers_10_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_837_cast_fp16 = conv(dilations = sparse_output_837_dilations_1, groups = sparse_output_837_groups_1, pad = sparse_output_837_pad_1, pad_type = sparse_output_837_pad_type_1, strides = sparse_output_837_strides_1, weight = layers_10_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_837_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13339_cast_fp16 = add(x = dense_output_837_cast_fp16, y = sparse_output_837_cast_fp16)[name = string("op_13339_cast_fp16")];
            tensor<int32, [4]> var_13340 = const()[name = string("op_13340"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13342 = const()[name = string("op_13342"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13341_cast_fp16 = transpose(perm = var_13340, x = var_13339_cast_fp16)[name = string("transpose_560")];
            tensor<fp16, [1, 51, 128]> v_head_329_cast_fp16 = reshape(shape = var_13342, x = var_13341_cast_fp16)[name = string("v_head_329_cast_fp16")];
            string dense_output_839_pad_type_1 = const()[name = string("dense_output_839_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_839_strides_1 = const()[name = string("dense_output_839_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_839_pad_1 = const()[name = string("dense_output_839_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_839_dilations_1 = const()[name = string("dense_output_839_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_839_groups_1 = const()[name = string("dense_output_839_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308917056))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309048192))))[name = string("layers_10_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_839_cast_fp16 = conv(dilations = dense_output_839_dilations_1, groups = dense_output_839_groups_1, pad = dense_output_839_pad_1, pad_type = dense_output_839_pad_type_1, strides = dense_output_839_strides_1, weight = layers_10_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_839_cast_fp16")];
            string sparse_output_839_pad_type_1 = const()[name = string("sparse_output_839_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_839_strides_1 = const()[name = string("sparse_output_839_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_839_pad_1 = const()[name = string("sparse_output_839_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_839_dilations_1 = const()[name = string("sparse_output_839_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_839_groups_1 = const()[name = string("sparse_output_839_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309051456))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309048768))))[name = string("layers_10_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_839_cast_fp16 = conv(dilations = sparse_output_839_dilations_1, groups = sparse_output_839_groups_1, pad = sparse_output_839_pad_1, pad_type = sparse_output_839_pad_type_1, strides = sparse_output_839_strides_1, weight = layers_10_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_839_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_13358_cast_fp16 = add(x = dense_output_839_cast_fp16, y = sparse_output_839_cast_fp16)[name = string("op_13358_cast_fp16")];
            tensor<int32, [4]> var_13359 = const()[name = string("op_13359"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13361 = const()[name = string("op_13361"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_13360_cast_fp16 = transpose(perm = var_13359, x = var_13358_cast_fp16)[name = string("transpose_559")];
            tensor<fp16, [1, 101, 128]> p_head_329_cast_fp16 = reshape(shape = var_13361, x = var_13360_cast_fp16)[name = string("p_head_329_cast_fp16")];
            tensor<fp16, [128]> var_13363_to_fp16 = const()[name = string("op_13363_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309067904)))];
            tensor<fp16, [1, 51, 128]> var_13364_cast_fp16 = add(x = q_head_165_cast_fp16, y = var_13363_to_fp16)[name = string("op_13364_cast_fp16")];
            tensor<int32, [1]> q_u_165_axes_1 = const()[name = string("q_u_165_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_165_cast_fp16 = expand_dims(axes = q_u_165_axes_1, x = var_13364_cast_fp16)[name = string("q_u_165_cast_fp16")];
            tensor<fp16, [128]> var_13366_to_fp16 = const()[name = string("op_13366_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309068224)))];
            tensor<fp16, [1, 51, 128]> var_13367_cast_fp16 = add(x = q_head_165_cast_fp16, y = var_13366_to_fp16)[name = string("op_13367_cast_fp16")];
            tensor<int32, [1]> q_v_165_axes_1 = const()[name = string("q_v_165_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_165_cast_fp16 = expand_dims(axes = q_v_165_axes_1, x = var_13367_cast_fp16)[name = string("q_v_165_cast_fp16")];
            tensor<int32, [1]> k_head_331_axes_1 = const()[name = string("k_head_331_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_331_cast_fp16 = expand_dims(axes = k_head_331_axes_1, x = k_head_329_cast_fp16)[name = string("k_head_331_cast_fp16")];
            tensor<int32, [1]> v_head_331_axes_1 = const()[name = string("v_head_331_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_331_cast_fp16 = expand_dims(axes = v_head_331_axes_1, x = v_head_329_cast_fp16)[name = string("v_head_331_cast_fp16")];
            tensor<int32, [1]> p_head_331_axes_1 = const()[name = string("p_head_331_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_331_cast_fp16 = expand_dims(axes = p_head_331_axes_1, x = p_head_329_cast_fp16)[name = string("p_head_331_cast_fp16")];
            bool var_13373_transpose_x_3 = const()[name = string("op_13373_transpose_x_3"), val = bool(false)];
            bool var_13373_transpose_y_3 = const()[name = string("op_13373_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_13373_cast_fp16 = matmul(transpose_x = var_13373_transpose_x_3, transpose_y = var_13373_transpose_y_3, x = q_u_165_cast_fp16, y = k_head_331_cast_fp16)[name = string("op_13373_cast_fp16")];
            fp16 var_13374_to_fp16 = const()[name = string("op_13374_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_165_cast_fp16 = mul(x = var_13373_cast_fp16, y = var_13374_to_fp16)[name = string("scores_content_165_cast_fp16")];
            bool x_877_transpose_x_3 = const()[name = string("x_877_transpose_x_3"), val = bool(false)];
            bool x_877_transpose_y_3 = const()[name = string("x_877_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_877_cast_fp16 = matmul(transpose_x = x_877_transpose_x_3, transpose_y = x_877_transpose_y_3, x = q_v_165_cast_fp16, y = p_head_331_cast_fp16)[name = string("x_877_cast_fp16")];
            tensor<int32, [8]> x_879_pad_1 = const()[name = string("x_879_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_879_mode_1 = const()[name = string("x_879_mode_1"), val = string("constant")];
            fp16 const_1861_to_fp16 = const()[name = string("const_1861_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_879_cast_fp16 = pad(constant_val = const_1861_to_fp16, mode = x_879_mode_1, pad = x_879_pad_1, x = x_877_cast_fp16)[name = string("x_879_cast_fp16")];
            tensor<int32, [4]> var_13388 = const()[name = string("op_13388"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_881_cast_fp16 = reshape(shape = var_13388, x = x_879_cast_fp16)[name = string("x_881_cast_fp16")];
            tensor<int32, [4]> var_13392_begin_1 = const()[name = string("op_13392_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13392_end_1 = const()[name = string("op_13392_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_13392_end_mask_1 = const()[name = string("op_13392_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_13392_cast_fp16 = slice_by_index(begin = var_13392_begin_1, end = var_13392_end_1, end_mask = var_13392_end_mask_1, x = x_881_cast_fp16)[name = string("op_13392_cast_fp16")];
            tensor<int32, [4]> var_13394 = const()[name = string("op_13394"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_13395_cast_fp16 = reshape(shape = var_13394, x = var_13392_cast_fp16)[name = string("op_13395_cast_fp16")];
            tensor<int32, [4]> var_13400_begin_1 = const()[name = string("op_13400_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13400_end_1 = const()[name = string("op_13400_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_13400_end_mask_1 = const()[name = string("op_13400_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_13400_cast_fp16 = slice_by_index(begin = var_13400_begin_1, end = var_13400_end_1, end_mask = var_13400_end_mask_1, x = var_13395_cast_fp16)[name = string("op_13400_cast_fp16")];
            fp16 var_13401_to_fp16 = const()[name = string("op_13401_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_165_cast_fp16 = mul(x = var_13400_cast_fp16, y = var_13401_to_fp16)[name = string("scores_pos_165_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_165_cast_fp16 = add(x = scores_content_165_cast_fp16, y = scores_pos_165_cast_fp16)[name = string("logits_165_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_13404_cast_fp16 = softmax(axis = var_12887, x = logits_165_cast_fp16)[name = string("op_13404_cast_fp16")];
            bool var_13406_transpose_x_1 = const()[name = string("op_13406_transpose_x_1"), val = bool(false)];
            bool var_13406_transpose_y_1 = const()[name = string("op_13406_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_13406_cast_fp16 = matmul(transpose_x = var_13406_transpose_x_1, transpose_y = var_13406_transpose_y_1, x = var_13404_cast_fp16, y = v_head_331_cast_fp16)[name = string("op_13406_cast_fp16")];
            tensor<int32, [1]> var_13407_axes_1 = const()[name = string("op_13407_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_13407_cast_fp16 = squeeze(axes = var_13407_axes_1, x = var_13406_cast_fp16)[name = string("op_13407_cast_fp16")];
            string dense_output_841_pad_type_1 = const()[name = string("dense_output_841_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_841_strides_1 = const()[name = string("dense_output_841_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_841_pad_1 = const()[name = string("dense_output_841_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_841_dilations_1 = const()[name = string("dense_output_841_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_841_groups_1 = const()[name = string("dense_output_841_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309068544))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309199680))))[name = string("layers_10_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_841_cast_fp16 = conv(dilations = dense_output_841_dilations_1, groups = dense_output_841_groups_1, pad = dense_output_841_pad_1, pad_type = dense_output_841_pad_type_1, strides = dense_output_841_strides_1, weight = layers_10_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_841_cast_fp16")];
            string sparse_output_841_pad_type_1 = const()[name = string("sparse_output_841_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_841_strides_1 = const()[name = string("sparse_output_841_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_841_pad_1 = const()[name = string("sparse_output_841_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_841_dilations_1 = const()[name = string("sparse_output_841_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_841_groups_1 = const()[name = string("sparse_output_841_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309202944))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309200256))))[name = string("layers_10_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_841_cast_fp16 = conv(dilations = sparse_output_841_dilations_1, groups = sparse_output_841_groups_1, pad = sparse_output_841_pad_1, pad_type = sparse_output_841_pad_type_1, strides = sparse_output_841_strides_1, weight = layers_10_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_841_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13422_cast_fp16 = add(x = dense_output_841_cast_fp16, y = sparse_output_841_cast_fp16)[name = string("op_13422_cast_fp16")];
            tensor<int32, [4]> var_13423 = const()[name = string("op_13423"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13425 = const()[name = string("op_13425"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13424_cast_fp16 = transpose(perm = var_13423, x = var_13422_cast_fp16)[name = string("transpose_558")];
            tensor<fp16, [1, 51, 128]> q_head_167_cast_fp16 = reshape(shape = var_13425, x = var_13424_cast_fp16)[name = string("q_head_167_cast_fp16")];
            string dense_output_843_pad_type_1 = const()[name = string("dense_output_843_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_843_strides_1 = const()[name = string("dense_output_843_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_843_pad_1 = const()[name = string("dense_output_843_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_843_dilations_1 = const()[name = string("dense_output_843_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_843_groups_1 = const()[name = string("dense_output_843_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309219392))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309350528))))[name = string("layers_10_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_843_cast_fp16 = conv(dilations = dense_output_843_dilations_1, groups = dense_output_843_groups_1, pad = dense_output_843_pad_1, pad_type = dense_output_843_pad_type_1, strides = dense_output_843_strides_1, weight = layers_10_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_843_cast_fp16")];
            string sparse_output_843_pad_type_1 = const()[name = string("sparse_output_843_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_843_strides_1 = const()[name = string("sparse_output_843_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_843_pad_1 = const()[name = string("sparse_output_843_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_843_dilations_1 = const()[name = string("sparse_output_843_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_843_groups_1 = const()[name = string("sparse_output_843_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309353792))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309351104))))[name = string("layers_10_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_843_cast_fp16 = conv(dilations = sparse_output_843_dilations_1, groups = sparse_output_843_groups_1, pad = sparse_output_843_pad_1, pad_type = sparse_output_843_pad_type_1, strides = sparse_output_843_strides_1, weight = layers_10_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_843_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13441_cast_fp16 = add(x = dense_output_843_cast_fp16, y = sparse_output_843_cast_fp16)[name = string("op_13441_cast_fp16")];
            tensor<int32, [4]> var_13442 = const()[name = string("op_13442"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13444 = const()[name = string("op_13444"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13443_cast_fp16 = transpose(perm = var_13442, x = var_13441_cast_fp16)[name = string("transpose_557")];
            tensor<fp16, [1, 51, 128]> k_head_333_cast_fp16 = reshape(shape = var_13444, x = var_13443_cast_fp16)[name = string("k_head_333_cast_fp16")];
            string dense_output_845_pad_type_1 = const()[name = string("dense_output_845_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_845_strides_1 = const()[name = string("dense_output_845_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_845_pad_1 = const()[name = string("dense_output_845_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_845_dilations_1 = const()[name = string("dense_output_845_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_845_groups_1 = const()[name = string("dense_output_845_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309370240))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309501376))))[name = string("layers_10_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_845_cast_fp16 = conv(dilations = dense_output_845_dilations_1, groups = dense_output_845_groups_1, pad = dense_output_845_pad_1, pad_type = dense_output_845_pad_type_1, strides = dense_output_845_strides_1, weight = layers_10_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_845_cast_fp16")];
            string sparse_output_845_pad_type_1 = const()[name = string("sparse_output_845_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_845_strides_1 = const()[name = string("sparse_output_845_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_845_pad_1 = const()[name = string("sparse_output_845_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_845_dilations_1 = const()[name = string("sparse_output_845_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_845_groups_1 = const()[name = string("sparse_output_845_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309504640))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309501952))))[name = string("layers_10_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_845_cast_fp16 = conv(dilations = sparse_output_845_dilations_1, groups = sparse_output_845_groups_1, pad = sparse_output_845_pad_1, pad_type = sparse_output_845_pad_type_1, strides = sparse_output_845_strides_1, weight = layers_10_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_845_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13460_cast_fp16 = add(x = dense_output_845_cast_fp16, y = sparse_output_845_cast_fp16)[name = string("op_13460_cast_fp16")];
            tensor<int32, [4]> var_13461 = const()[name = string("op_13461"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13463 = const()[name = string("op_13463"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13462_cast_fp16 = transpose(perm = var_13461, x = var_13460_cast_fp16)[name = string("transpose_556")];
            tensor<fp16, [1, 51, 128]> v_head_333_cast_fp16 = reshape(shape = var_13463, x = var_13462_cast_fp16)[name = string("v_head_333_cast_fp16")];
            string dense_output_847_pad_type_1 = const()[name = string("dense_output_847_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_847_strides_1 = const()[name = string("dense_output_847_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_847_pad_1 = const()[name = string("dense_output_847_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_847_dilations_1 = const()[name = string("dense_output_847_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_847_groups_1 = const()[name = string("dense_output_847_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309521088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309652224))))[name = string("layers_10_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_847_cast_fp16 = conv(dilations = dense_output_847_dilations_1, groups = dense_output_847_groups_1, pad = dense_output_847_pad_1, pad_type = dense_output_847_pad_type_1, strides = dense_output_847_strides_1, weight = layers_10_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_847_cast_fp16")];
            string sparse_output_847_pad_type_1 = const()[name = string("sparse_output_847_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_847_strides_1 = const()[name = string("sparse_output_847_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_847_pad_1 = const()[name = string("sparse_output_847_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_847_dilations_1 = const()[name = string("sparse_output_847_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_847_groups_1 = const()[name = string("sparse_output_847_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309655488))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309652800))))[name = string("layers_10_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_847_cast_fp16 = conv(dilations = sparse_output_847_dilations_1, groups = sparse_output_847_groups_1, pad = sparse_output_847_pad_1, pad_type = sparse_output_847_pad_type_1, strides = sparse_output_847_strides_1, weight = layers_10_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_847_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_13479_cast_fp16 = add(x = dense_output_847_cast_fp16, y = sparse_output_847_cast_fp16)[name = string("op_13479_cast_fp16")];
            tensor<int32, [4]> var_13480 = const()[name = string("op_13480"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13482 = const()[name = string("op_13482"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_13481_cast_fp16 = transpose(perm = var_13480, x = var_13479_cast_fp16)[name = string("transpose_555")];
            tensor<fp16, [1, 101, 128]> p_head_333_cast_fp16 = reshape(shape = var_13482, x = var_13481_cast_fp16)[name = string("p_head_333_cast_fp16")];
            tensor<fp16, [128]> var_13484_to_fp16 = const()[name = string("op_13484_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309671936)))];
            tensor<fp16, [1, 51, 128]> var_13485_cast_fp16 = add(x = q_head_167_cast_fp16, y = var_13484_to_fp16)[name = string("op_13485_cast_fp16")];
            tensor<int32, [1]> q_u_167_axes_1 = const()[name = string("q_u_167_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_167_cast_fp16 = expand_dims(axes = q_u_167_axes_1, x = var_13485_cast_fp16)[name = string("q_u_167_cast_fp16")];
            tensor<fp16, [128]> var_13487_to_fp16 = const()[name = string("op_13487_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309672256)))];
            tensor<fp16, [1, 51, 128]> var_13488_cast_fp16 = add(x = q_head_167_cast_fp16, y = var_13487_to_fp16)[name = string("op_13488_cast_fp16")];
            tensor<int32, [1]> q_v_167_axes_1 = const()[name = string("q_v_167_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_167_cast_fp16 = expand_dims(axes = q_v_167_axes_1, x = var_13488_cast_fp16)[name = string("q_v_167_cast_fp16")];
            tensor<int32, [1]> k_head_335_axes_1 = const()[name = string("k_head_335_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_335_cast_fp16 = expand_dims(axes = k_head_335_axes_1, x = k_head_333_cast_fp16)[name = string("k_head_335_cast_fp16")];
            tensor<int32, [1]> v_head_335_axes_1 = const()[name = string("v_head_335_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_335_cast_fp16 = expand_dims(axes = v_head_335_axes_1, x = v_head_333_cast_fp16)[name = string("v_head_335_cast_fp16")];
            tensor<int32, [1]> p_head_335_axes_1 = const()[name = string("p_head_335_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_335_cast_fp16 = expand_dims(axes = p_head_335_axes_1, x = p_head_333_cast_fp16)[name = string("p_head_335_cast_fp16")];
            bool var_13494_transpose_x_3 = const()[name = string("op_13494_transpose_x_3"), val = bool(false)];
            bool var_13494_transpose_y_3 = const()[name = string("op_13494_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_13494_cast_fp16 = matmul(transpose_x = var_13494_transpose_x_3, transpose_y = var_13494_transpose_y_3, x = q_u_167_cast_fp16, y = k_head_335_cast_fp16)[name = string("op_13494_cast_fp16")];
            fp16 var_13495_to_fp16 = const()[name = string("op_13495_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_167_cast_fp16 = mul(x = var_13494_cast_fp16, y = var_13495_to_fp16)[name = string("scores_content_167_cast_fp16")];
            bool x_885_transpose_x_3 = const()[name = string("x_885_transpose_x_3"), val = bool(false)];
            bool x_885_transpose_y_3 = const()[name = string("x_885_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_885_cast_fp16 = matmul(transpose_x = x_885_transpose_x_3, transpose_y = x_885_transpose_y_3, x = q_v_167_cast_fp16, y = p_head_335_cast_fp16)[name = string("x_885_cast_fp16")];
            tensor<int32, [8]> x_887_pad_1 = const()[name = string("x_887_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_887_mode_1 = const()[name = string("x_887_mode_1"), val = string("constant")];
            fp16 const_1867_to_fp16 = const()[name = string("const_1867_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_887_cast_fp16 = pad(constant_val = const_1867_to_fp16, mode = x_887_mode_1, pad = x_887_pad_1, x = x_885_cast_fp16)[name = string("x_887_cast_fp16")];
            tensor<int32, [4]> var_13509 = const()[name = string("op_13509"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_889_cast_fp16 = reshape(shape = var_13509, x = x_887_cast_fp16)[name = string("x_889_cast_fp16")];
            tensor<int32, [4]> var_13513_begin_1 = const()[name = string("op_13513_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13513_end_1 = const()[name = string("op_13513_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_13513_end_mask_1 = const()[name = string("op_13513_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_13513_cast_fp16 = slice_by_index(begin = var_13513_begin_1, end = var_13513_end_1, end_mask = var_13513_end_mask_1, x = x_889_cast_fp16)[name = string("op_13513_cast_fp16")];
            tensor<int32, [4]> var_13515 = const()[name = string("op_13515"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_13516_cast_fp16 = reshape(shape = var_13515, x = var_13513_cast_fp16)[name = string("op_13516_cast_fp16")];
            tensor<int32, [4]> var_13521_begin_1 = const()[name = string("op_13521_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13521_end_1 = const()[name = string("op_13521_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_13521_end_mask_1 = const()[name = string("op_13521_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_13521_cast_fp16 = slice_by_index(begin = var_13521_begin_1, end = var_13521_end_1, end_mask = var_13521_end_mask_1, x = var_13516_cast_fp16)[name = string("op_13521_cast_fp16")];
            fp16 var_13522_to_fp16 = const()[name = string("op_13522_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_167_cast_fp16 = mul(x = var_13521_cast_fp16, y = var_13522_to_fp16)[name = string("scores_pos_167_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_167_cast_fp16 = add(x = scores_content_167_cast_fp16, y = scores_pos_167_cast_fp16)[name = string("logits_167_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_13525_cast_fp16 = softmax(axis = var_12887, x = logits_167_cast_fp16)[name = string("op_13525_cast_fp16")];
            bool var_13527_transpose_x_1 = const()[name = string("op_13527_transpose_x_1"), val = bool(false)];
            bool var_13527_transpose_y_1 = const()[name = string("op_13527_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_13527_cast_fp16 = matmul(transpose_x = var_13527_transpose_x_1, transpose_y = var_13527_transpose_y_1, x = var_13525_cast_fp16, y = v_head_335_cast_fp16)[name = string("op_13527_cast_fp16")];
            tensor<int32, [1]> var_13528_axes_1 = const()[name = string("op_13528_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_13528_cast_fp16 = squeeze(axes = var_13528_axes_1, x = var_13527_cast_fp16)[name = string("op_13528_cast_fp16")];
            string dense_output_849_pad_type_1 = const()[name = string("dense_output_849_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_849_strides_1 = const()[name = string("dense_output_849_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_849_pad_1 = const()[name = string("dense_output_849_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_849_dilations_1 = const()[name = string("dense_output_849_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_849_groups_1 = const()[name = string("dense_output_849_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309672576))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309803712))))[name = string("layers_10_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_849_cast_fp16 = conv(dilations = dense_output_849_dilations_1, groups = dense_output_849_groups_1, pad = dense_output_849_pad_1, pad_type = dense_output_849_pad_type_1, strides = dense_output_849_strides_1, weight = layers_10_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_849_cast_fp16")];
            string sparse_output_849_pad_type_1 = const()[name = string("sparse_output_849_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_849_strides_1 = const()[name = string("sparse_output_849_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_849_pad_1 = const()[name = string("sparse_output_849_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_849_dilations_1 = const()[name = string("sparse_output_849_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_849_groups_1 = const()[name = string("sparse_output_849_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309806976))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309804288))))[name = string("layers_10_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_849_cast_fp16 = conv(dilations = sparse_output_849_dilations_1, groups = sparse_output_849_groups_1, pad = sparse_output_849_pad_1, pad_type = sparse_output_849_pad_type_1, strides = sparse_output_849_strides_1, weight = layers_10_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_849_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13543_cast_fp16 = add(x = dense_output_849_cast_fp16, y = sparse_output_849_cast_fp16)[name = string("op_13543_cast_fp16")];
            tensor<int32, [4]> var_13544 = const()[name = string("op_13544"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13546 = const()[name = string("op_13546"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13545_cast_fp16 = transpose(perm = var_13544, x = var_13543_cast_fp16)[name = string("transpose_554")];
            tensor<fp16, [1, 51, 128]> q_head_169_cast_fp16 = reshape(shape = var_13546, x = var_13545_cast_fp16)[name = string("q_head_169_cast_fp16")];
            string dense_output_851_pad_type_1 = const()[name = string("dense_output_851_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_851_strides_1 = const()[name = string("dense_output_851_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_851_pad_1 = const()[name = string("dense_output_851_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_851_dilations_1 = const()[name = string("dense_output_851_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_851_groups_1 = const()[name = string("dense_output_851_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309823424))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309954560))))[name = string("layers_10_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_851_cast_fp16 = conv(dilations = dense_output_851_dilations_1, groups = dense_output_851_groups_1, pad = dense_output_851_pad_1, pad_type = dense_output_851_pad_type_1, strides = dense_output_851_strides_1, weight = layers_10_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_851_cast_fp16")];
            string sparse_output_851_pad_type_1 = const()[name = string("sparse_output_851_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_851_strides_1 = const()[name = string("sparse_output_851_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_851_pad_1 = const()[name = string("sparse_output_851_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_851_dilations_1 = const()[name = string("sparse_output_851_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_851_groups_1 = const()[name = string("sparse_output_851_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309957824))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309955136))))[name = string("layers_10_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_851_cast_fp16 = conv(dilations = sparse_output_851_dilations_1, groups = sparse_output_851_groups_1, pad = sparse_output_851_pad_1, pad_type = sparse_output_851_pad_type_1, strides = sparse_output_851_strides_1, weight = layers_10_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_851_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13562_cast_fp16 = add(x = dense_output_851_cast_fp16, y = sparse_output_851_cast_fp16)[name = string("op_13562_cast_fp16")];
            tensor<int32, [4]> var_13563 = const()[name = string("op_13563"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13565 = const()[name = string("op_13565"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13564_cast_fp16 = transpose(perm = var_13563, x = var_13562_cast_fp16)[name = string("transpose_553")];
            tensor<fp16, [1, 51, 128]> k_head_337_cast_fp16 = reshape(shape = var_13565, x = var_13564_cast_fp16)[name = string("k_head_337_cast_fp16")];
            string dense_output_853_pad_type_1 = const()[name = string("dense_output_853_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_853_strides_1 = const()[name = string("dense_output_853_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_853_pad_1 = const()[name = string("dense_output_853_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_853_dilations_1 = const()[name = string("dense_output_853_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_853_groups_1 = const()[name = string("dense_output_853_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309974272))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310105408))))[name = string("layers_10_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_853_cast_fp16 = conv(dilations = dense_output_853_dilations_1, groups = dense_output_853_groups_1, pad = dense_output_853_pad_1, pad_type = dense_output_853_pad_type_1, strides = dense_output_853_strides_1, weight = layers_10_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_853_cast_fp16")];
            string sparse_output_853_pad_type_1 = const()[name = string("sparse_output_853_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_853_strides_1 = const()[name = string("sparse_output_853_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_853_pad_1 = const()[name = string("sparse_output_853_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_853_dilations_1 = const()[name = string("sparse_output_853_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_853_groups_1 = const()[name = string("sparse_output_853_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310108672))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310105984))))[name = string("layers_10_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_853_cast_fp16 = conv(dilations = sparse_output_853_dilations_1, groups = sparse_output_853_groups_1, pad = sparse_output_853_pad_1, pad_type = sparse_output_853_pad_type_1, strides = sparse_output_853_strides_1, weight = layers_10_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_853_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13581_cast_fp16 = add(x = dense_output_853_cast_fp16, y = sparse_output_853_cast_fp16)[name = string("op_13581_cast_fp16")];
            tensor<int32, [4]> var_13582 = const()[name = string("op_13582"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13584 = const()[name = string("op_13584"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13583_cast_fp16 = transpose(perm = var_13582, x = var_13581_cast_fp16)[name = string("transpose_552")];
            tensor<fp16, [1, 51, 128]> v_head_337_cast_fp16 = reshape(shape = var_13584, x = var_13583_cast_fp16)[name = string("v_head_337_cast_fp16")];
            string dense_output_855_pad_type_1 = const()[name = string("dense_output_855_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_855_strides_1 = const()[name = string("dense_output_855_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_855_pad_1 = const()[name = string("dense_output_855_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_855_dilations_1 = const()[name = string("dense_output_855_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_855_groups_1 = const()[name = string("dense_output_855_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310125120))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310256256))))[name = string("layers_10_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_855_cast_fp16 = conv(dilations = dense_output_855_dilations_1, groups = dense_output_855_groups_1, pad = dense_output_855_pad_1, pad_type = dense_output_855_pad_type_1, strides = dense_output_855_strides_1, weight = layers_10_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_855_cast_fp16")];
            string sparse_output_855_pad_type_1 = const()[name = string("sparse_output_855_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_855_strides_1 = const()[name = string("sparse_output_855_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_855_pad_1 = const()[name = string("sparse_output_855_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_855_dilations_1 = const()[name = string("sparse_output_855_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_855_groups_1 = const()[name = string("sparse_output_855_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310259520))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310256832))))[name = string("layers_10_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_855_cast_fp16 = conv(dilations = sparse_output_855_dilations_1, groups = sparse_output_855_groups_1, pad = sparse_output_855_pad_1, pad_type = sparse_output_855_pad_type_1, strides = sparse_output_855_strides_1, weight = layers_10_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_855_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_13600_cast_fp16 = add(x = dense_output_855_cast_fp16, y = sparse_output_855_cast_fp16)[name = string("op_13600_cast_fp16")];
            tensor<int32, [4]> var_13601 = const()[name = string("op_13601"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13603 = const()[name = string("op_13603"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_13602_cast_fp16 = transpose(perm = var_13601, x = var_13600_cast_fp16)[name = string("transpose_551")];
            tensor<fp16, [1, 101, 128]> p_head_337_cast_fp16 = reshape(shape = var_13603, x = var_13602_cast_fp16)[name = string("p_head_337_cast_fp16")];
            tensor<fp16, [128]> var_13605_to_fp16 = const()[name = string("op_13605_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310275968)))];
            tensor<fp16, [1, 51, 128]> var_13606_cast_fp16 = add(x = q_head_169_cast_fp16, y = var_13605_to_fp16)[name = string("op_13606_cast_fp16")];
            tensor<int32, [1]> q_u_169_axes_1 = const()[name = string("q_u_169_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_169_cast_fp16 = expand_dims(axes = q_u_169_axes_1, x = var_13606_cast_fp16)[name = string("q_u_169_cast_fp16")];
            tensor<fp16, [128]> var_13608_to_fp16 = const()[name = string("op_13608_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310276288)))];
            tensor<fp16, [1, 51, 128]> var_13609_cast_fp16 = add(x = q_head_169_cast_fp16, y = var_13608_to_fp16)[name = string("op_13609_cast_fp16")];
            tensor<int32, [1]> q_v_169_axes_1 = const()[name = string("q_v_169_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_169_cast_fp16 = expand_dims(axes = q_v_169_axes_1, x = var_13609_cast_fp16)[name = string("q_v_169_cast_fp16")];
            tensor<int32, [1]> k_head_339_axes_1 = const()[name = string("k_head_339_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_339_cast_fp16 = expand_dims(axes = k_head_339_axes_1, x = k_head_337_cast_fp16)[name = string("k_head_339_cast_fp16")];
            tensor<int32, [1]> v_head_339_axes_1 = const()[name = string("v_head_339_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_339_cast_fp16 = expand_dims(axes = v_head_339_axes_1, x = v_head_337_cast_fp16)[name = string("v_head_339_cast_fp16")];
            tensor<int32, [1]> p_head_339_axes_1 = const()[name = string("p_head_339_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_339_cast_fp16 = expand_dims(axes = p_head_339_axes_1, x = p_head_337_cast_fp16)[name = string("p_head_339_cast_fp16")];
            bool var_13615_transpose_x_3 = const()[name = string("op_13615_transpose_x_3"), val = bool(false)];
            bool var_13615_transpose_y_3 = const()[name = string("op_13615_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_13615_cast_fp16 = matmul(transpose_x = var_13615_transpose_x_3, transpose_y = var_13615_transpose_y_3, x = q_u_169_cast_fp16, y = k_head_339_cast_fp16)[name = string("op_13615_cast_fp16")];
            fp16 var_13616_to_fp16 = const()[name = string("op_13616_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_169_cast_fp16 = mul(x = var_13615_cast_fp16, y = var_13616_to_fp16)[name = string("scores_content_169_cast_fp16")];
            bool x_893_transpose_x_3 = const()[name = string("x_893_transpose_x_3"), val = bool(false)];
            bool x_893_transpose_y_3 = const()[name = string("x_893_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_893_cast_fp16 = matmul(transpose_x = x_893_transpose_x_3, transpose_y = x_893_transpose_y_3, x = q_v_169_cast_fp16, y = p_head_339_cast_fp16)[name = string("x_893_cast_fp16")];
            tensor<int32, [8]> x_895_pad_1 = const()[name = string("x_895_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_895_mode_1 = const()[name = string("x_895_mode_1"), val = string("constant")];
            fp16 const_1873_to_fp16 = const()[name = string("const_1873_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_895_cast_fp16 = pad(constant_val = const_1873_to_fp16, mode = x_895_mode_1, pad = x_895_pad_1, x = x_893_cast_fp16)[name = string("x_895_cast_fp16")];
            tensor<int32, [4]> var_13630 = const()[name = string("op_13630"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_897_cast_fp16 = reshape(shape = var_13630, x = x_895_cast_fp16)[name = string("x_897_cast_fp16")];
            tensor<int32, [4]> var_13634_begin_1 = const()[name = string("op_13634_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13634_end_1 = const()[name = string("op_13634_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_13634_end_mask_1 = const()[name = string("op_13634_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_13634_cast_fp16 = slice_by_index(begin = var_13634_begin_1, end = var_13634_end_1, end_mask = var_13634_end_mask_1, x = x_897_cast_fp16)[name = string("op_13634_cast_fp16")];
            tensor<int32, [4]> var_13636 = const()[name = string("op_13636"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_13637_cast_fp16 = reshape(shape = var_13636, x = var_13634_cast_fp16)[name = string("op_13637_cast_fp16")];
            tensor<int32, [4]> var_13642_begin_1 = const()[name = string("op_13642_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13642_end_1 = const()[name = string("op_13642_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_13642_end_mask_1 = const()[name = string("op_13642_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_13642_cast_fp16 = slice_by_index(begin = var_13642_begin_1, end = var_13642_end_1, end_mask = var_13642_end_mask_1, x = var_13637_cast_fp16)[name = string("op_13642_cast_fp16")];
            fp16 var_13643_to_fp16 = const()[name = string("op_13643_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_169_cast_fp16 = mul(x = var_13642_cast_fp16, y = var_13643_to_fp16)[name = string("scores_pos_169_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_169_cast_fp16 = add(x = scores_content_169_cast_fp16, y = scores_pos_169_cast_fp16)[name = string("logits_169_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_13646_cast_fp16 = softmax(axis = var_12887, x = logits_169_cast_fp16)[name = string("op_13646_cast_fp16")];
            bool var_13648_transpose_x_1 = const()[name = string("op_13648_transpose_x_1"), val = bool(false)];
            bool var_13648_transpose_y_1 = const()[name = string("op_13648_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_13648_cast_fp16 = matmul(transpose_x = var_13648_transpose_x_1, transpose_y = var_13648_transpose_y_1, x = var_13646_cast_fp16, y = v_head_339_cast_fp16)[name = string("op_13648_cast_fp16")];
            tensor<int32, [1]> var_13649_axes_1 = const()[name = string("op_13649_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_13649_cast_fp16 = squeeze(axes = var_13649_axes_1, x = var_13648_cast_fp16)[name = string("op_13649_cast_fp16")];
            string dense_output_857_pad_type_1 = const()[name = string("dense_output_857_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_857_strides_1 = const()[name = string("dense_output_857_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_857_pad_1 = const()[name = string("dense_output_857_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_857_dilations_1 = const()[name = string("dense_output_857_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_857_groups_1 = const()[name = string("dense_output_857_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310276608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310407744))))[name = string("layers_10_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_857_cast_fp16 = conv(dilations = dense_output_857_dilations_1, groups = dense_output_857_groups_1, pad = dense_output_857_pad_1, pad_type = dense_output_857_pad_type_1, strides = dense_output_857_strides_1, weight = layers_10_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_857_cast_fp16")];
            string sparse_output_857_pad_type_1 = const()[name = string("sparse_output_857_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_857_strides_1 = const()[name = string("sparse_output_857_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_857_pad_1 = const()[name = string("sparse_output_857_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_857_dilations_1 = const()[name = string("sparse_output_857_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_857_groups_1 = const()[name = string("sparse_output_857_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310411008))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310408320))))[name = string("layers_10_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_857_cast_fp16 = conv(dilations = sparse_output_857_dilations_1, groups = sparse_output_857_groups_1, pad = sparse_output_857_pad_1, pad_type = sparse_output_857_pad_type_1, strides = sparse_output_857_strides_1, weight = layers_10_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_857_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13664_cast_fp16 = add(x = dense_output_857_cast_fp16, y = sparse_output_857_cast_fp16)[name = string("op_13664_cast_fp16")];
            tensor<int32, [4]> var_13665 = const()[name = string("op_13665"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13667 = const()[name = string("op_13667"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13666_cast_fp16 = transpose(perm = var_13665, x = var_13664_cast_fp16)[name = string("transpose_550")];
            tensor<fp16, [1, 51, 128]> q_head_171_cast_fp16 = reshape(shape = var_13667, x = var_13666_cast_fp16)[name = string("q_head_171_cast_fp16")];
            string dense_output_859_pad_type_1 = const()[name = string("dense_output_859_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_859_strides_1 = const()[name = string("dense_output_859_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_859_pad_1 = const()[name = string("dense_output_859_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_859_dilations_1 = const()[name = string("dense_output_859_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_859_groups_1 = const()[name = string("dense_output_859_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310427456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310558592))))[name = string("layers_10_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_859_cast_fp16 = conv(dilations = dense_output_859_dilations_1, groups = dense_output_859_groups_1, pad = dense_output_859_pad_1, pad_type = dense_output_859_pad_type_1, strides = dense_output_859_strides_1, weight = layers_10_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_859_cast_fp16")];
            string sparse_output_859_pad_type_1 = const()[name = string("sparse_output_859_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_859_strides_1 = const()[name = string("sparse_output_859_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_859_pad_1 = const()[name = string("sparse_output_859_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_859_dilations_1 = const()[name = string("sparse_output_859_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_859_groups_1 = const()[name = string("sparse_output_859_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310561856))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310559168))))[name = string("layers_10_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_859_cast_fp16 = conv(dilations = sparse_output_859_dilations_1, groups = sparse_output_859_groups_1, pad = sparse_output_859_pad_1, pad_type = sparse_output_859_pad_type_1, strides = sparse_output_859_strides_1, weight = layers_10_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_859_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13683_cast_fp16 = add(x = dense_output_859_cast_fp16, y = sparse_output_859_cast_fp16)[name = string("op_13683_cast_fp16")];
            tensor<int32, [4]> var_13684 = const()[name = string("op_13684"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13686 = const()[name = string("op_13686"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13685_cast_fp16 = transpose(perm = var_13684, x = var_13683_cast_fp16)[name = string("transpose_549")];
            tensor<fp16, [1, 51, 128]> k_head_341_cast_fp16 = reshape(shape = var_13686, x = var_13685_cast_fp16)[name = string("k_head_341_cast_fp16")];
            string dense_output_861_pad_type_1 = const()[name = string("dense_output_861_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_861_strides_1 = const()[name = string("dense_output_861_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_861_pad_1 = const()[name = string("dense_output_861_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_861_dilations_1 = const()[name = string("dense_output_861_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_861_groups_1 = const()[name = string("dense_output_861_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310578304))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310709440))))[name = string("layers_10_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_861_cast_fp16 = conv(dilations = dense_output_861_dilations_1, groups = dense_output_861_groups_1, pad = dense_output_861_pad_1, pad_type = dense_output_861_pad_type_1, strides = dense_output_861_strides_1, weight = layers_10_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_861_cast_fp16")];
            string sparse_output_861_pad_type_1 = const()[name = string("sparse_output_861_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_861_strides_1 = const()[name = string("sparse_output_861_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_861_pad_1 = const()[name = string("sparse_output_861_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_861_dilations_1 = const()[name = string("sparse_output_861_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_861_groups_1 = const()[name = string("sparse_output_861_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310712704))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310710016))))[name = string("layers_10_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_861_cast_fp16 = conv(dilations = sparse_output_861_dilations_1, groups = sparse_output_861_groups_1, pad = sparse_output_861_pad_1, pad_type = sparse_output_861_pad_type_1, strides = sparse_output_861_strides_1, weight = layers_10_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_861_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13702_cast_fp16 = add(x = dense_output_861_cast_fp16, y = sparse_output_861_cast_fp16)[name = string("op_13702_cast_fp16")];
            tensor<int32, [4]> var_13703 = const()[name = string("op_13703"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13705 = const()[name = string("op_13705"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13704_cast_fp16 = transpose(perm = var_13703, x = var_13702_cast_fp16)[name = string("transpose_548")];
            tensor<fp16, [1, 51, 128]> v_head_341_cast_fp16 = reshape(shape = var_13705, x = var_13704_cast_fp16)[name = string("v_head_341_cast_fp16")];
            string dense_output_863_pad_type_1 = const()[name = string("dense_output_863_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_863_strides_1 = const()[name = string("dense_output_863_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_863_pad_1 = const()[name = string("dense_output_863_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_863_dilations_1 = const()[name = string("dense_output_863_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_863_groups_1 = const()[name = string("dense_output_863_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310729152))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310860288))))[name = string("layers_10_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_863_cast_fp16 = conv(dilations = dense_output_863_dilations_1, groups = dense_output_863_groups_1, pad = dense_output_863_pad_1, pad_type = dense_output_863_pad_type_1, strides = dense_output_863_strides_1, weight = layers_10_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_863_cast_fp16")];
            string sparse_output_863_pad_type_1 = const()[name = string("sparse_output_863_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_863_strides_1 = const()[name = string("sparse_output_863_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_863_pad_1 = const()[name = string("sparse_output_863_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_863_dilations_1 = const()[name = string("sparse_output_863_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_863_groups_1 = const()[name = string("sparse_output_863_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310863552))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310860864))))[name = string("layers_10_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_863_cast_fp16 = conv(dilations = sparse_output_863_dilations_1, groups = sparse_output_863_groups_1, pad = sparse_output_863_pad_1, pad_type = sparse_output_863_pad_type_1, strides = sparse_output_863_strides_1, weight = layers_10_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_863_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_13721_cast_fp16 = add(x = dense_output_863_cast_fp16, y = sparse_output_863_cast_fp16)[name = string("op_13721_cast_fp16")];
            tensor<int32, [4]> var_13722 = const()[name = string("op_13722"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13724 = const()[name = string("op_13724"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_13723_cast_fp16 = transpose(perm = var_13722, x = var_13721_cast_fp16)[name = string("transpose_547")];
            tensor<fp16, [1, 101, 128]> p_head_341_cast_fp16 = reshape(shape = var_13724, x = var_13723_cast_fp16)[name = string("p_head_341_cast_fp16")];
            tensor<fp16, [128]> var_13726_to_fp16 = const()[name = string("op_13726_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310880000)))];
            tensor<fp16, [1, 51, 128]> var_13727_cast_fp16 = add(x = q_head_171_cast_fp16, y = var_13726_to_fp16)[name = string("op_13727_cast_fp16")];
            tensor<int32, [1]> q_u_171_axes_1 = const()[name = string("q_u_171_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_171_cast_fp16 = expand_dims(axes = q_u_171_axes_1, x = var_13727_cast_fp16)[name = string("q_u_171_cast_fp16")];
            tensor<fp16, [128]> var_13729_to_fp16 = const()[name = string("op_13729_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310880320)))];
            tensor<fp16, [1, 51, 128]> var_13730_cast_fp16 = add(x = q_head_171_cast_fp16, y = var_13729_to_fp16)[name = string("op_13730_cast_fp16")];
            tensor<int32, [1]> q_v_171_axes_1 = const()[name = string("q_v_171_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_171_cast_fp16 = expand_dims(axes = q_v_171_axes_1, x = var_13730_cast_fp16)[name = string("q_v_171_cast_fp16")];
            tensor<int32, [1]> k_head_343_axes_1 = const()[name = string("k_head_343_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_343_cast_fp16 = expand_dims(axes = k_head_343_axes_1, x = k_head_341_cast_fp16)[name = string("k_head_343_cast_fp16")];
            tensor<int32, [1]> v_head_343_axes_1 = const()[name = string("v_head_343_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_343_cast_fp16 = expand_dims(axes = v_head_343_axes_1, x = v_head_341_cast_fp16)[name = string("v_head_343_cast_fp16")];
            tensor<int32, [1]> p_head_343_axes_1 = const()[name = string("p_head_343_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_343_cast_fp16 = expand_dims(axes = p_head_343_axes_1, x = p_head_341_cast_fp16)[name = string("p_head_343_cast_fp16")];
            bool var_13736_transpose_x_3 = const()[name = string("op_13736_transpose_x_3"), val = bool(false)];
            bool var_13736_transpose_y_3 = const()[name = string("op_13736_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_13736_cast_fp16 = matmul(transpose_x = var_13736_transpose_x_3, transpose_y = var_13736_transpose_y_3, x = q_u_171_cast_fp16, y = k_head_343_cast_fp16)[name = string("op_13736_cast_fp16")];
            fp16 var_13737_to_fp16 = const()[name = string("op_13737_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_171_cast_fp16 = mul(x = var_13736_cast_fp16, y = var_13737_to_fp16)[name = string("scores_content_171_cast_fp16")];
            bool x_901_transpose_x_3 = const()[name = string("x_901_transpose_x_3"), val = bool(false)];
            bool x_901_transpose_y_3 = const()[name = string("x_901_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_901_cast_fp16 = matmul(transpose_x = x_901_transpose_x_3, transpose_y = x_901_transpose_y_3, x = q_v_171_cast_fp16, y = p_head_343_cast_fp16)[name = string("x_901_cast_fp16")];
            tensor<int32, [8]> x_903_pad_1 = const()[name = string("x_903_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_903_mode_1 = const()[name = string("x_903_mode_1"), val = string("constant")];
            fp16 const_1879_to_fp16 = const()[name = string("const_1879_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_903_cast_fp16 = pad(constant_val = const_1879_to_fp16, mode = x_903_mode_1, pad = x_903_pad_1, x = x_901_cast_fp16)[name = string("x_903_cast_fp16")];
            tensor<int32, [4]> var_13751 = const()[name = string("op_13751"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_905_cast_fp16 = reshape(shape = var_13751, x = x_903_cast_fp16)[name = string("x_905_cast_fp16")];
            tensor<int32, [4]> var_13755_begin_1 = const()[name = string("op_13755_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13755_end_1 = const()[name = string("op_13755_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_13755_end_mask_1 = const()[name = string("op_13755_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_13755_cast_fp16 = slice_by_index(begin = var_13755_begin_1, end = var_13755_end_1, end_mask = var_13755_end_mask_1, x = x_905_cast_fp16)[name = string("op_13755_cast_fp16")];
            tensor<int32, [4]> var_13757 = const()[name = string("op_13757"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_13758_cast_fp16 = reshape(shape = var_13757, x = var_13755_cast_fp16)[name = string("op_13758_cast_fp16")];
            tensor<int32, [4]> var_13763_begin_1 = const()[name = string("op_13763_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13763_end_1 = const()[name = string("op_13763_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_13763_end_mask_1 = const()[name = string("op_13763_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_13763_cast_fp16 = slice_by_index(begin = var_13763_begin_1, end = var_13763_end_1, end_mask = var_13763_end_mask_1, x = var_13758_cast_fp16)[name = string("op_13763_cast_fp16")];
            fp16 var_13764_to_fp16 = const()[name = string("op_13764_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_171_cast_fp16 = mul(x = var_13763_cast_fp16, y = var_13764_to_fp16)[name = string("scores_pos_171_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_171_cast_fp16 = add(x = scores_content_171_cast_fp16, y = scores_pos_171_cast_fp16)[name = string("logits_171_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_13767_cast_fp16 = softmax(axis = var_12887, x = logits_171_cast_fp16)[name = string("op_13767_cast_fp16")];
            bool var_13769_transpose_x_1 = const()[name = string("op_13769_transpose_x_1"), val = bool(false)];
            bool var_13769_transpose_y_1 = const()[name = string("op_13769_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_13769_cast_fp16 = matmul(transpose_x = var_13769_transpose_x_1, transpose_y = var_13769_transpose_y_1, x = var_13767_cast_fp16, y = v_head_343_cast_fp16)[name = string("op_13769_cast_fp16")];
            tensor<int32, [1]> var_13770_axes_1 = const()[name = string("op_13770_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_13770_cast_fp16 = squeeze(axes = var_13770_axes_1, x = var_13769_cast_fp16)[name = string("op_13770_cast_fp16")];
            string dense_output_865_pad_type_1 = const()[name = string("dense_output_865_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_865_strides_1 = const()[name = string("dense_output_865_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_865_pad_1 = const()[name = string("dense_output_865_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_865_dilations_1 = const()[name = string("dense_output_865_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_865_groups_1 = const()[name = string("dense_output_865_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310880640))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311011776))))[name = string("layers_10_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_865_cast_fp16 = conv(dilations = dense_output_865_dilations_1, groups = dense_output_865_groups_1, pad = dense_output_865_pad_1, pad_type = dense_output_865_pad_type_1, strides = dense_output_865_strides_1, weight = layers_10_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_865_cast_fp16")];
            string sparse_output_865_pad_type_1 = const()[name = string("sparse_output_865_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_865_strides_1 = const()[name = string("sparse_output_865_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_865_pad_1 = const()[name = string("sparse_output_865_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_865_dilations_1 = const()[name = string("sparse_output_865_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_865_groups_1 = const()[name = string("sparse_output_865_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311015040))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311012352))))[name = string("layers_10_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_865_cast_fp16 = conv(dilations = sparse_output_865_dilations_1, groups = sparse_output_865_groups_1, pad = sparse_output_865_pad_1, pad_type = sparse_output_865_pad_type_1, strides = sparse_output_865_strides_1, weight = layers_10_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_865_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13785_cast_fp16 = add(x = dense_output_865_cast_fp16, y = sparse_output_865_cast_fp16)[name = string("op_13785_cast_fp16")];
            tensor<int32, [4]> var_13786 = const()[name = string("op_13786"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13788 = const()[name = string("op_13788"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13787_cast_fp16 = transpose(perm = var_13786, x = var_13785_cast_fp16)[name = string("transpose_546")];
            tensor<fp16, [1, 51, 128]> q_head_173_cast_fp16 = reshape(shape = var_13788, x = var_13787_cast_fp16)[name = string("q_head_173_cast_fp16")];
            string dense_output_867_pad_type_1 = const()[name = string("dense_output_867_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_867_strides_1 = const()[name = string("dense_output_867_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_867_pad_1 = const()[name = string("dense_output_867_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_867_dilations_1 = const()[name = string("dense_output_867_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_867_groups_1 = const()[name = string("dense_output_867_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311031488))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311162624))))[name = string("layers_10_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_867_cast_fp16 = conv(dilations = dense_output_867_dilations_1, groups = dense_output_867_groups_1, pad = dense_output_867_pad_1, pad_type = dense_output_867_pad_type_1, strides = dense_output_867_strides_1, weight = layers_10_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_867_cast_fp16")];
            string sparse_output_867_pad_type_1 = const()[name = string("sparse_output_867_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_867_strides_1 = const()[name = string("sparse_output_867_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_867_pad_1 = const()[name = string("sparse_output_867_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_867_dilations_1 = const()[name = string("sparse_output_867_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_867_groups_1 = const()[name = string("sparse_output_867_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311165888))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311163200))))[name = string("layers_10_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_867_cast_fp16 = conv(dilations = sparse_output_867_dilations_1, groups = sparse_output_867_groups_1, pad = sparse_output_867_pad_1, pad_type = sparse_output_867_pad_type_1, strides = sparse_output_867_strides_1, weight = layers_10_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_867_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13804_cast_fp16 = add(x = dense_output_867_cast_fp16, y = sparse_output_867_cast_fp16)[name = string("op_13804_cast_fp16")];
            tensor<int32, [4]> var_13805 = const()[name = string("op_13805"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13807 = const()[name = string("op_13807"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13806_cast_fp16 = transpose(perm = var_13805, x = var_13804_cast_fp16)[name = string("transpose_545")];
            tensor<fp16, [1, 51, 128]> k_head_345_cast_fp16 = reshape(shape = var_13807, x = var_13806_cast_fp16)[name = string("k_head_345_cast_fp16")];
            string dense_output_869_pad_type_1 = const()[name = string("dense_output_869_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_869_strides_1 = const()[name = string("dense_output_869_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_869_pad_1 = const()[name = string("dense_output_869_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_869_dilations_1 = const()[name = string("dense_output_869_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_869_groups_1 = const()[name = string("dense_output_869_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311182336))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311313472))))[name = string("layers_10_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_869_cast_fp16 = conv(dilations = dense_output_869_dilations_1, groups = dense_output_869_groups_1, pad = dense_output_869_pad_1, pad_type = dense_output_869_pad_type_1, strides = dense_output_869_strides_1, weight = layers_10_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_869_cast_fp16")];
            string sparse_output_869_pad_type_1 = const()[name = string("sparse_output_869_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_869_strides_1 = const()[name = string("sparse_output_869_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_869_pad_1 = const()[name = string("sparse_output_869_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_869_dilations_1 = const()[name = string("sparse_output_869_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_869_groups_1 = const()[name = string("sparse_output_869_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311316736))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311314048))))[name = string("layers_10_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_869_cast_fp16 = conv(dilations = sparse_output_869_dilations_1, groups = sparse_output_869_groups_1, pad = sparse_output_869_pad_1, pad_type = sparse_output_869_pad_type_1, strides = sparse_output_869_strides_1, weight = layers_10_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_869_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13823_cast_fp16 = add(x = dense_output_869_cast_fp16, y = sparse_output_869_cast_fp16)[name = string("op_13823_cast_fp16")];
            tensor<int32, [4]> var_13824 = const()[name = string("op_13824"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13826 = const()[name = string("op_13826"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13825_cast_fp16 = transpose(perm = var_13824, x = var_13823_cast_fp16)[name = string("transpose_544")];
            tensor<fp16, [1, 51, 128]> v_head_345_cast_fp16 = reshape(shape = var_13826, x = var_13825_cast_fp16)[name = string("v_head_345_cast_fp16")];
            string dense_output_871_pad_type_1 = const()[name = string("dense_output_871_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_871_strides_1 = const()[name = string("dense_output_871_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_871_pad_1 = const()[name = string("dense_output_871_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_871_dilations_1 = const()[name = string("dense_output_871_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_871_groups_1 = const()[name = string("dense_output_871_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311333184))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311464320))))[name = string("layers_10_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_871_cast_fp16 = conv(dilations = dense_output_871_dilations_1, groups = dense_output_871_groups_1, pad = dense_output_871_pad_1, pad_type = dense_output_871_pad_type_1, strides = dense_output_871_strides_1, weight = layers_10_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_871_cast_fp16")];
            string sparse_output_871_pad_type_1 = const()[name = string("sparse_output_871_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_871_strides_1 = const()[name = string("sparse_output_871_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_871_pad_1 = const()[name = string("sparse_output_871_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_871_dilations_1 = const()[name = string("sparse_output_871_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_871_groups_1 = const()[name = string("sparse_output_871_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311467584))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311464896))))[name = string("layers_10_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_871_cast_fp16 = conv(dilations = sparse_output_871_dilations_1, groups = sparse_output_871_groups_1, pad = sparse_output_871_pad_1, pad_type = sparse_output_871_pad_type_1, strides = sparse_output_871_strides_1, weight = layers_10_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_871_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_13842_cast_fp16 = add(x = dense_output_871_cast_fp16, y = sparse_output_871_cast_fp16)[name = string("op_13842_cast_fp16")];
            tensor<int32, [4]> var_13843 = const()[name = string("op_13843"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13845 = const()[name = string("op_13845"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_13844_cast_fp16 = transpose(perm = var_13843, x = var_13842_cast_fp16)[name = string("transpose_543")];
            tensor<fp16, [1, 101, 128]> p_head_345_cast_fp16 = reshape(shape = var_13845, x = var_13844_cast_fp16)[name = string("p_head_345_cast_fp16")];
            tensor<fp16, [128]> var_13847_to_fp16 = const()[name = string("op_13847_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311484032)))];
            tensor<fp16, [1, 51, 128]> var_13848_cast_fp16 = add(x = q_head_173_cast_fp16, y = var_13847_to_fp16)[name = string("op_13848_cast_fp16")];
            tensor<int32, [1]> q_u_173_axes_1 = const()[name = string("q_u_173_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_173_cast_fp16 = expand_dims(axes = q_u_173_axes_1, x = var_13848_cast_fp16)[name = string("q_u_173_cast_fp16")];
            tensor<fp16, [128]> var_13850_to_fp16 = const()[name = string("op_13850_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311484352)))];
            tensor<fp16, [1, 51, 128]> var_13851_cast_fp16 = add(x = q_head_173_cast_fp16, y = var_13850_to_fp16)[name = string("op_13851_cast_fp16")];
            tensor<int32, [1]> q_v_173_axes_1 = const()[name = string("q_v_173_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_173_cast_fp16 = expand_dims(axes = q_v_173_axes_1, x = var_13851_cast_fp16)[name = string("q_v_173_cast_fp16")];
            tensor<int32, [1]> k_head_347_axes_1 = const()[name = string("k_head_347_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_347_cast_fp16 = expand_dims(axes = k_head_347_axes_1, x = k_head_345_cast_fp16)[name = string("k_head_347_cast_fp16")];
            tensor<int32, [1]> v_head_347_axes_1 = const()[name = string("v_head_347_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_347_cast_fp16 = expand_dims(axes = v_head_347_axes_1, x = v_head_345_cast_fp16)[name = string("v_head_347_cast_fp16")];
            tensor<int32, [1]> p_head_347_axes_1 = const()[name = string("p_head_347_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_347_cast_fp16 = expand_dims(axes = p_head_347_axes_1, x = p_head_345_cast_fp16)[name = string("p_head_347_cast_fp16")];
            bool var_13857_transpose_x_3 = const()[name = string("op_13857_transpose_x_3"), val = bool(false)];
            bool var_13857_transpose_y_3 = const()[name = string("op_13857_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_13857_cast_fp16 = matmul(transpose_x = var_13857_transpose_x_3, transpose_y = var_13857_transpose_y_3, x = q_u_173_cast_fp16, y = k_head_347_cast_fp16)[name = string("op_13857_cast_fp16")];
            fp16 var_13858_to_fp16 = const()[name = string("op_13858_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_173_cast_fp16 = mul(x = var_13857_cast_fp16, y = var_13858_to_fp16)[name = string("scores_content_173_cast_fp16")];
            bool x_909_transpose_x_3 = const()[name = string("x_909_transpose_x_3"), val = bool(false)];
            bool x_909_transpose_y_3 = const()[name = string("x_909_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_909_cast_fp16 = matmul(transpose_x = x_909_transpose_x_3, transpose_y = x_909_transpose_y_3, x = q_v_173_cast_fp16, y = p_head_347_cast_fp16)[name = string("x_909_cast_fp16")];
            tensor<int32, [8]> x_911_pad_1 = const()[name = string("x_911_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_911_mode_1 = const()[name = string("x_911_mode_1"), val = string("constant")];
            fp16 const_1885_to_fp16 = const()[name = string("const_1885_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_911_cast_fp16 = pad(constant_val = const_1885_to_fp16, mode = x_911_mode_1, pad = x_911_pad_1, x = x_909_cast_fp16)[name = string("x_911_cast_fp16")];
            tensor<int32, [4]> var_13872 = const()[name = string("op_13872"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_913_cast_fp16 = reshape(shape = var_13872, x = x_911_cast_fp16)[name = string("x_913_cast_fp16")];
            tensor<int32, [4]> var_13876_begin_1 = const()[name = string("op_13876_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13876_end_1 = const()[name = string("op_13876_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_13876_end_mask_1 = const()[name = string("op_13876_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_13876_cast_fp16 = slice_by_index(begin = var_13876_begin_1, end = var_13876_end_1, end_mask = var_13876_end_mask_1, x = x_913_cast_fp16)[name = string("op_13876_cast_fp16")];
            tensor<int32, [4]> var_13878 = const()[name = string("op_13878"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_13879_cast_fp16 = reshape(shape = var_13878, x = var_13876_cast_fp16)[name = string("op_13879_cast_fp16")];
            tensor<int32, [4]> var_13884_begin_1 = const()[name = string("op_13884_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_13884_end_1 = const()[name = string("op_13884_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_13884_end_mask_1 = const()[name = string("op_13884_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_13884_cast_fp16 = slice_by_index(begin = var_13884_begin_1, end = var_13884_end_1, end_mask = var_13884_end_mask_1, x = var_13879_cast_fp16)[name = string("op_13884_cast_fp16")];
            fp16 var_13885_to_fp16 = const()[name = string("op_13885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_173_cast_fp16 = mul(x = var_13884_cast_fp16, y = var_13885_to_fp16)[name = string("scores_pos_173_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_173_cast_fp16 = add(x = scores_content_173_cast_fp16, y = scores_pos_173_cast_fp16)[name = string("logits_173_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_13888_cast_fp16 = softmax(axis = var_12887, x = logits_173_cast_fp16)[name = string("op_13888_cast_fp16")];
            bool var_13890_transpose_x_1 = const()[name = string("op_13890_transpose_x_1"), val = bool(false)];
            bool var_13890_transpose_y_1 = const()[name = string("op_13890_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_13890_cast_fp16 = matmul(transpose_x = var_13890_transpose_x_1, transpose_y = var_13890_transpose_y_1, x = var_13888_cast_fp16, y = v_head_347_cast_fp16)[name = string("op_13890_cast_fp16")];
            tensor<int32, [1]> var_13891_axes_1 = const()[name = string("op_13891_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_13891_cast_fp16 = squeeze(axes = var_13891_axes_1, x = var_13890_cast_fp16)[name = string("op_13891_cast_fp16")];
            string dense_output_873_pad_type_1 = const()[name = string("dense_output_873_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_873_strides_1 = const()[name = string("dense_output_873_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_873_pad_1 = const()[name = string("dense_output_873_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_873_dilations_1 = const()[name = string("dense_output_873_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_873_groups_1 = const()[name = string("dense_output_873_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311484672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311615808))))[name = string("layers_10_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_873_cast_fp16 = conv(dilations = dense_output_873_dilations_1, groups = dense_output_873_groups_1, pad = dense_output_873_pad_1, pad_type = dense_output_873_pad_type_1, strides = dense_output_873_strides_1, weight = layers_10_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_873_cast_fp16")];
            string sparse_output_873_pad_type_1 = const()[name = string("sparse_output_873_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_873_strides_1 = const()[name = string("sparse_output_873_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_873_pad_1 = const()[name = string("sparse_output_873_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_873_dilations_1 = const()[name = string("sparse_output_873_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_873_groups_1 = const()[name = string("sparse_output_873_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311619072))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311616384))))[name = string("layers_10_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_873_cast_fp16 = conv(dilations = sparse_output_873_dilations_1, groups = sparse_output_873_groups_1, pad = sparse_output_873_pad_1, pad_type = sparse_output_873_pad_type_1, strides = sparse_output_873_strides_1, weight = layers_10_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_873_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13906_cast_fp16 = add(x = dense_output_873_cast_fp16, y = sparse_output_873_cast_fp16)[name = string("op_13906_cast_fp16")];
            tensor<int32, [4]> var_13907 = const()[name = string("op_13907"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13909 = const()[name = string("op_13909"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13908_cast_fp16 = transpose(perm = var_13907, x = var_13906_cast_fp16)[name = string("transpose_542")];
            tensor<fp16, [1, 51, 128]> q_head_175_cast_fp16 = reshape(shape = var_13909, x = var_13908_cast_fp16)[name = string("q_head_175_cast_fp16")];
            string dense_output_875_pad_type_1 = const()[name = string("dense_output_875_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_875_strides_1 = const()[name = string("dense_output_875_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_875_pad_1 = const()[name = string("dense_output_875_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_875_dilations_1 = const()[name = string("dense_output_875_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_875_groups_1 = const()[name = string("dense_output_875_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311635520))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311766656))))[name = string("layers_10_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_875_cast_fp16 = conv(dilations = dense_output_875_dilations_1, groups = dense_output_875_groups_1, pad = dense_output_875_pad_1, pad_type = dense_output_875_pad_type_1, strides = dense_output_875_strides_1, weight = layers_10_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_875_cast_fp16")];
            string sparse_output_875_pad_type_1 = const()[name = string("sparse_output_875_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_875_strides_1 = const()[name = string("sparse_output_875_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_875_pad_1 = const()[name = string("sparse_output_875_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_875_dilations_1 = const()[name = string("sparse_output_875_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_875_groups_1 = const()[name = string("sparse_output_875_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311769920))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311767232))))[name = string("layers_10_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_875_cast_fp16 = conv(dilations = sparse_output_875_dilations_1, groups = sparse_output_875_groups_1, pad = sparse_output_875_pad_1, pad_type = sparse_output_875_pad_type_1, strides = sparse_output_875_strides_1, weight = layers_10_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_875_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13925_cast_fp16 = add(x = dense_output_875_cast_fp16, y = sparse_output_875_cast_fp16)[name = string("op_13925_cast_fp16")];
            tensor<int32, [4]> var_13926 = const()[name = string("op_13926"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13928 = const()[name = string("op_13928"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13927_cast_fp16 = transpose(perm = var_13926, x = var_13925_cast_fp16)[name = string("transpose_541")];
            tensor<fp16, [1, 51, 128]> k_head_349_cast_fp16 = reshape(shape = var_13928, x = var_13927_cast_fp16)[name = string("k_head_349_cast_fp16")];
            string dense_output_877_pad_type_1 = const()[name = string("dense_output_877_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_877_strides_1 = const()[name = string("dense_output_877_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_877_pad_1 = const()[name = string("dense_output_877_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_877_dilations_1 = const()[name = string("dense_output_877_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_877_groups_1 = const()[name = string("dense_output_877_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311786368))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311917504))))[name = string("layers_10_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_877_cast_fp16 = conv(dilations = dense_output_877_dilations_1, groups = dense_output_877_groups_1, pad = dense_output_877_pad_1, pad_type = dense_output_877_pad_type_1, strides = dense_output_877_strides_1, weight = layers_10_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_489_cast_fp16)[name = string("dense_output_877_cast_fp16")];
            string sparse_output_877_pad_type_1 = const()[name = string("sparse_output_877_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_877_strides_1 = const()[name = string("sparse_output_877_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_877_pad_1 = const()[name = string("sparse_output_877_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_877_dilations_1 = const()[name = string("sparse_output_877_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_877_groups_1 = const()[name = string("sparse_output_877_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311920768))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311918080))))[name = string("layers_10_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_877_cast_fp16 = conv(dilations = sparse_output_877_dilations_1, groups = sparse_output_877_groups_1, pad = sparse_output_877_pad_1, pad_type = sparse_output_877_pad_type_1, strides = sparse_output_877_strides_1, weight = layers_10_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_489_cast_fp16)[name = string("sparse_output_877_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_13944_cast_fp16 = add(x = dense_output_877_cast_fp16, y = sparse_output_877_cast_fp16)[name = string("op_13944_cast_fp16")];
            tensor<int32, [4]> var_13945 = const()[name = string("op_13945"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13947 = const()[name = string("op_13947"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_13946_cast_fp16 = transpose(perm = var_13945, x = var_13944_cast_fp16)[name = string("transpose_540")];
            tensor<fp16, [1, 51, 128]> v_head_349_cast_fp16 = reshape(shape = var_13947, x = var_13946_cast_fp16)[name = string("v_head_349_cast_fp16")];
            string dense_output_879_pad_type_1 = const()[name = string("dense_output_879_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_879_strides_1 = const()[name = string("dense_output_879_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_879_pad_1 = const()[name = string("dense_output_879_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_879_dilations_1 = const()[name = string("dense_output_879_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_879_groups_1 = const()[name = string("dense_output_879_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311937216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312068352))))[name = string("layers_10_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_879_cast_fp16 = conv(dilations = dense_output_879_dilations_1, groups = dense_output_879_groups_1, pad = dense_output_879_pad_1, pad_type = dense_output_879_pad_type_1, strides = dense_output_879_strides_1, weight = layers_10_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_879_cast_fp16")];
            string sparse_output_879_pad_type_1 = const()[name = string("sparse_output_879_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_879_strides_1 = const()[name = string("sparse_output_879_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_879_pad_1 = const()[name = string("sparse_output_879_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_879_dilations_1 = const()[name = string("sparse_output_879_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_879_groups_1 = const()[name = string("sparse_output_879_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_10_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312071616))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312068928))))[name = string("layers_10_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_879_cast_fp16 = conv(dilations = sparse_output_879_dilations_1, groups = sparse_output_879_groups_1, pad = sparse_output_879_pad_1, pad_type = sparse_output_879_pad_type_1, strides = sparse_output_879_strides_1, weight = layers_10_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_879_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_13963_cast_fp16 = add(x = dense_output_879_cast_fp16, y = sparse_output_879_cast_fp16)[name = string("op_13963_cast_fp16")];
            tensor<int32, [4]> var_13964 = const()[name = string("op_13964"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_13966 = const()[name = string("op_13966"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_13965_cast_fp16 = transpose(perm = var_13964, x = var_13963_cast_fp16)[name = string("transpose_539")];
            tensor<fp16, [1, 101, 128]> p_head_349_cast_fp16 = reshape(shape = var_13966, x = var_13965_cast_fp16)[name = string("p_head_349_cast_fp16")];
            tensor<fp16, [128]> var_13968_to_fp16 = const()[name = string("op_13968_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312088064)))];
            tensor<fp16, [1, 51, 128]> var_13969_cast_fp16 = add(x = q_head_175_cast_fp16, y = var_13968_to_fp16)[name = string("op_13969_cast_fp16")];
            tensor<int32, [1]> q_u_175_axes_1 = const()[name = string("q_u_175_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_175_cast_fp16 = expand_dims(axes = q_u_175_axes_1, x = var_13969_cast_fp16)[name = string("q_u_175_cast_fp16")];
            tensor<fp16, [128]> var_13971_to_fp16 = const()[name = string("op_13971_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312088384)))];
            tensor<fp16, [1, 51, 128]> var_13972_cast_fp16 = add(x = q_head_175_cast_fp16, y = var_13971_to_fp16)[name = string("op_13972_cast_fp16")];
            tensor<int32, [1]> q_v_175_axes_1 = const()[name = string("q_v_175_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_175_cast_fp16 = expand_dims(axes = q_v_175_axes_1, x = var_13972_cast_fp16)[name = string("q_v_175_cast_fp16")];
            tensor<int32, [1]> k_head_351_axes_1 = const()[name = string("k_head_351_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_351_cast_fp16 = expand_dims(axes = k_head_351_axes_1, x = k_head_349_cast_fp16)[name = string("k_head_351_cast_fp16")];
            tensor<int32, [1]> v_head_351_axes_1 = const()[name = string("v_head_351_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_351_cast_fp16 = expand_dims(axes = v_head_351_axes_1, x = v_head_349_cast_fp16)[name = string("v_head_351_cast_fp16")];
            tensor<int32, [1]> p_head_351_axes_1 = const()[name = string("p_head_351_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_351_cast_fp16 = expand_dims(axes = p_head_351_axes_1, x = p_head_349_cast_fp16)[name = string("p_head_351_cast_fp16")];
            bool var_13978_transpose_x_3 = const()[name = string("op_13978_transpose_x_3"), val = bool(false)];
            bool var_13978_transpose_y_3 = const()[name = string("op_13978_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_13978_cast_fp16 = matmul(transpose_x = var_13978_transpose_x_3, transpose_y = var_13978_transpose_y_3, x = q_u_175_cast_fp16, y = k_head_351_cast_fp16)[name = string("op_13978_cast_fp16")];
            fp16 var_13979_to_fp16 = const()[name = string("op_13979_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_175_cast_fp16 = mul(x = var_13978_cast_fp16, y = var_13979_to_fp16)[name = string("scores_content_175_cast_fp16")];
            bool x_917_transpose_x_3 = const()[name = string("x_917_transpose_x_3"), val = bool(false)];
            bool x_917_transpose_y_3 = const()[name = string("x_917_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_917_cast_fp16 = matmul(transpose_x = x_917_transpose_x_3, transpose_y = x_917_transpose_y_3, x = q_v_175_cast_fp16, y = p_head_351_cast_fp16)[name = string("x_917_cast_fp16")];
            tensor<int32, [8]> x_919_pad_1 = const()[name = string("x_919_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_919_mode_1 = const()[name = string("x_919_mode_1"), val = string("constant")];
            fp16 const_1891_to_fp16 = const()[name = string("const_1891_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_919_cast_fp16 = pad(constant_val = const_1891_to_fp16, mode = x_919_mode_1, pad = x_919_pad_1, x = x_917_cast_fp16)[name = string("x_919_cast_fp16")];
            tensor<int32, [4]> var_13993 = const()[name = string("op_13993"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_921_cast_fp16 = reshape(shape = var_13993, x = x_919_cast_fp16)[name = string("x_921_cast_fp16")];
            tensor<int32, [4]> var_13997_begin_1 = const()[name = string("op_13997_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_13997_end_1 = const()[name = string("op_13997_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_13997_end_mask_1 = const()[name = string("op_13997_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_13997_cast_fp16 = slice_by_index(begin = var_13997_begin_1, end = var_13997_end_1, end_mask = var_13997_end_mask_1, x = x_921_cast_fp16)[name = string("op_13997_cast_fp16")];
            tensor<int32, [4]> var_13999 = const()[name = string("op_13999"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_14000_cast_fp16 = reshape(shape = var_13999, x = var_13997_cast_fp16)[name = string("op_14000_cast_fp16")];
            tensor<int32, [4]> var_14005_begin_1 = const()[name = string("op_14005_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14005_end_1 = const()[name = string("op_14005_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_14005_end_mask_1 = const()[name = string("op_14005_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_14005_cast_fp16 = slice_by_index(begin = var_14005_begin_1, end = var_14005_end_1, end_mask = var_14005_end_mask_1, x = var_14000_cast_fp16)[name = string("op_14005_cast_fp16")];
            fp16 var_14006_to_fp16 = const()[name = string("op_14006_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_175_cast_fp16 = mul(x = var_14005_cast_fp16, y = var_14006_to_fp16)[name = string("scores_pos_175_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_175_cast_fp16 = add(x = scores_content_175_cast_fp16, y = scores_pos_175_cast_fp16)[name = string("logits_175_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_14009_cast_fp16 = softmax(axis = var_12887, x = logits_175_cast_fp16)[name = string("op_14009_cast_fp16")];
            bool var_14011_transpose_x_1 = const()[name = string("op_14011_transpose_x_1"), val = bool(false)];
            bool var_14011_transpose_y_1 = const()[name = string("op_14011_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_14011_cast_fp16 = matmul(transpose_x = var_14011_transpose_x_1, transpose_y = var_14011_transpose_y_1, x = var_14009_cast_fp16, y = v_head_351_cast_fp16)[name = string("op_14011_cast_fp16")];
            tensor<int32, [1]> o_head_21_axes_1 = const()[name = string("o_head_21_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_21_cast_fp16 = squeeze(axes = o_head_21_axes_1, x = var_14011_cast_fp16)[name = string("o_head_21_cast_fp16")];
            bool out_21_interleave_1 = const()[name = string("out_21_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_21_cast_fp16 = concat(axis = var_12887, interleave = out_21_interleave_1, values = (var_13165_cast_fp16, var_13286_cast_fp16, var_13407_cast_fp16, var_13528_cast_fp16, var_13649_cast_fp16, var_13770_cast_fp16, var_13891_cast_fp16, o_head_21_cast_fp16))[name = string("out_21_cast_fp16")];
            tensor<int32, [3]> var_14015_perm_1 = const()[name = string("op_14015_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_497_axes_1 = const()[name = string("input_497_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_14015_cast_fp16 = transpose(perm = var_14015_perm_1, x = out_21_cast_fp16)[name = string("transpose_538")];
            tensor<fp16, [1, 1024, 51, 1]> input_497_cast_fp16 = expand_dims(axes = input_497_axes_1, x = var_14015_cast_fp16)[name = string("input_497_cast_fp16")];
            string dense_output_881_pad_type_1 = const()[name = string("dense_output_881_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_881_strides_1 = const()[name = string("dense_output_881_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_881_pad_1 = const()[name = string("dense_output_881_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_881_dilations_1 = const()[name = string("dense_output_881_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_881_groups_1 = const()[name = string("dense_output_881_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312088704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313137344))))[name = string("layers_10_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_881_cast_fp16 = conv(dilations = dense_output_881_dilations_1, groups = dense_output_881_groups_1, pad = dense_output_881_pad_1, pad_type = dense_output_881_pad_type_1, strides = dense_output_881_strides_1, weight = layers_10_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_497_cast_fp16)[name = string("dense_output_881_cast_fp16")];
            string sparse_output_881_pad_type_1 = const()[name = string("sparse_output_881_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_881_strides_1 = const()[name = string("sparse_output_881_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_881_pad_1 = const()[name = string("sparse_output_881_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_881_dilations_1 = const()[name = string("sparse_output_881_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_881_groups_1 = const()[name = string("sparse_output_881_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313158976))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313137920))))[name = string("layers_10_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_881_cast_fp16 = conv(dilations = sparse_output_881_dilations_1, groups = sparse_output_881_groups_1, pad = sparse_output_881_pad_1, pad_type = sparse_output_881_pad_type_1, strides = sparse_output_881_strides_1, weight = layers_10_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_497_cast_fp16)[name = string("sparse_output_881_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_21_cast_fp16 = add(x = dense_output_881_cast_fp16, y = sparse_output_881_cast_fp16)[name = string("out_conv_21_cast_fp16")];
            tensor<int32, [1]> var_14032_axes_1 = const()[name = string("op_14032_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_14032_cast_fp16 = squeeze(axes = var_14032_axes_1, x = out_conv_21_cast_fp16)[name = string("op_14032_cast_fp16")];
            tensor<int32, [3]> var_14033_perm_1 = const()[name = string("op_14033_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_14033_cast_fp16 = transpose(perm = var_14033_perm_1, x = var_14032_cast_fp16)[name = string("transpose_537")];
            tensor<fp16, [1, 51, 1024]> input_499_cast_fp16 = add(x = input_487_cast_fp16, y = var_14033_cast_fp16)[name = string("input_499_cast_fp16")];
            tensor<int32, [1]> x_925_axes_1 = const()[name = string("x_925_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_10_norm_conv_weight_to_fp16 = const()[name = string("layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313290112)))];
            tensor<fp16, [1024]> layers_10_norm_conv_bias_to_fp16 = const()[name = string("layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313292224)))];
            tensor<fp16, [1, 51, 1024]> x_925_cast_fp16 = layer_norm(axes = x_925_axes_1, beta = layers_10_norm_conv_bias_to_fp16, epsilon = var_12902_to_fp16, gamma = layers_10_norm_conv_weight_to_fp16, x = input_499_cast_fp16)[name = string("x_925_cast_fp16")];
            tensor<int32, [3]> var_14043_perm_1 = const()[name = string("op_14043_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_501_axes_1 = const()[name = string("input_501_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_14043_cast_fp16 = transpose(perm = var_14043_perm_1, x = x_925_cast_fp16)[name = string("transpose_536")];
            tensor<fp16, [1, 1024, 51, 1]> input_501_cast_fp16 = expand_dims(axes = input_501_axes_1, x = var_14043_cast_fp16)[name = string("input_501_cast_fp16")];
            string dense_output_883_pad_type_1 = const()[name = string("dense_output_883_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_883_strides_1 = const()[name = string("dense_output_883_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_883_pad_1 = const()[name = string("dense_output_883_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_883_dilations_1 = const()[name = string("dense_output_883_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_883_groups_1 = const()[name = string("dense_output_883_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_10_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313294336))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315391552))))[name = string("layers_10_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_883_cast_fp16 = conv(dilations = dense_output_883_dilations_1, groups = dense_output_883_groups_1, pad = dense_output_883_pad_1, pad_type = dense_output_883_pad_type_1, strides = dense_output_883_strides_1, weight = layers_10_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_501_cast_fp16)[name = string("dense_output_883_cast_fp16")];
            string sparse_output_883_pad_type_1 = const()[name = string("sparse_output_883_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_883_strides_1 = const()[name = string("sparse_output_883_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_883_pad_1 = const()[name = string("sparse_output_883_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_883_dilations_1 = const()[name = string("sparse_output_883_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_883_groups_1 = const()[name = string("sparse_output_883_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_10_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315434176))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315392128))))[name = string("layers_10_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_883_cast_fp16 = conv(dilations = sparse_output_883_dilations_1, groups = sparse_output_883_groups_1, pad = sparse_output_883_pad_1, pad_type = sparse_output_883_pad_type_1, strides = sparse_output_883_strides_1, weight = layers_10_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_501_cast_fp16)[name = string("sparse_output_883_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_503_cast_fp16 = add(x = dense_output_883_cast_fp16, y = sparse_output_883_cast_fp16)[name = string("input_503_cast_fp16")];
            int32 input_505_split_num_splits_1 = const()[name = string("input_505_split_num_splits_1"), val = int32(2)];
            int32 input_505_split_axis_1 = const()[name = string("input_505_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_505_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_505_split_cast_fp16_1 = split(axis = input_505_split_axis_1, num_splits = input_505_split_num_splits_1, x = input_503_cast_fp16)[name = string("input_505_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_505_split_1_sigmoid_cast_fp16 = sigmoid(x = input_505_split_cast_fp16_1)[name = string("input_505_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_505_cast_fp16 = mul(x = input_505_split_cast_fp16_0, y = input_505_split_1_sigmoid_cast_fp16)[name = string("input_505_cast_fp16")];
            tensor<int32, [8]> input_507_pad_1 = const()[name = string("input_507_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_507_mode_1 = const()[name = string("input_507_mode_1"), val = string("constant")];
            fp16 const_1893_to_fp16 = const()[name = string("const_1893_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_507_cast_fp16 = pad(constant_val = const_1893_to_fp16, mode = input_507_mode_1, pad = input_507_pad_1, x = input_505_cast_fp16)[name = string("input_507_cast_fp16")];
            string dense_output_885_pad_type_1 = const()[name = string("dense_output_885_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_885_strides_1 = const()[name = string("dense_output_885_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_885_pad_1 = const()[name = string("dense_output_885_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_885_dilations_1 = const()[name = string("dense_output_885_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_885_groups_1 = const()[name = string("dense_output_885_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_885_cast_fp16 = conv(dilations = dense_output_885_dilations_1, groups = dense_output_885_groups_1, pad = dense_output_885_pad_1, pad_type = dense_output_885_pad_type_1, strides = dense_output_885_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_507_cast_fp16)[name = string("dense_output_885_cast_fp16")];
            string sparse_output_885_pad_type_1 = const()[name = string("sparse_output_885_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_885_strides_1 = const()[name = string("sparse_output_885_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_885_pad_1 = const()[name = string("sparse_output_885_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_885_dilations_1 = const()[name = string("sparse_output_885_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_885_groups_1 = const()[name = string("sparse_output_885_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_10_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315696384))))[name = string("layers_10_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_885_cast_fp16 = conv(dilations = sparse_output_885_dilations_1, groups = sparse_output_885_groups_1, pad = sparse_output_885_pad_1, pad_type = sparse_output_885_pad_type_1, strides = sparse_output_885_strides_1, weight = layers_10_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_507_cast_fp16)[name = string("sparse_output_885_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_509_cast_fp16 = add(x = dense_output_885_cast_fp16, y = sparse_output_885_cast_fp16)[name = string("input_509_cast_fp16")];
            tensor<fp16, [1024]> layers_10_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_10_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315714880)))];
            tensor<fp16, [1024]> layers_10_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_10_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315716992)))];
            tensor<fp16, [1024]> layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315719104)))];
            tensor<fp16, [1024]> layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315721216)))];
            tensor<fp16, [1, 1024, 51, 1]> input_511_cast_fp16 = batch_norm(beta = layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_12902_to_fp16, gamma = layers_10_conv_batch_norm_weight_to_fp16, mean = layers_10_conv_batch_norm_running_mean_to_fp16, variance = layers_10_conv_batch_norm_running_var_to_fp16, x = input_509_cast_fp16)[name = string("input_511_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_513_cast_fp16 = silu(x = input_511_cast_fp16)[name = string("input_513_cast_fp16")];
            string dense_output_887_pad_type_1 = const()[name = string("dense_output_887_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_887_strides_1 = const()[name = string("dense_output_887_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_887_pad_1 = const()[name = string("dense_output_887_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_887_dilations_1 = const()[name = string("dense_output_887_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_887_groups_1 = const()[name = string("dense_output_887_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315723328))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316771968))))[name = string("layers_10_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_887_cast_fp16 = conv(dilations = dense_output_887_dilations_1, groups = dense_output_887_groups_1, pad = dense_output_887_pad_1, pad_type = dense_output_887_pad_type_1, strides = dense_output_887_strides_1, weight = layers_10_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_513_cast_fp16)[name = string("dense_output_887_cast_fp16")];
            string sparse_output_887_pad_type_1 = const()[name = string("sparse_output_887_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_887_strides_1 = const()[name = string("sparse_output_887_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_887_pad_1 = const()[name = string("sparse_output_887_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_887_dilations_1 = const()[name = string("sparse_output_887_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_887_groups_1 = const()[name = string("sparse_output_887_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_10_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793600))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316772544))))[name = string("layers_10_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_887_cast_fp16 = conv(dilations = sparse_output_887_dilations_1, groups = sparse_output_887_groups_1, pad = sparse_output_887_pad_1, pad_type = sparse_output_887_pad_type_1, strides = sparse_output_887_strides_1, weight = layers_10_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_513_cast_fp16)[name = string("sparse_output_887_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_927_cast_fp16 = add(x = dense_output_887_cast_fp16, y = sparse_output_887_cast_fp16)[name = string("x_927_cast_fp16")];
            tensor<int32, [1]> var_14099_axes_1 = const()[name = string("op_14099_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_14099_cast_fp16 = squeeze(axes = var_14099_axes_1, x = x_927_cast_fp16)[name = string("op_14099_cast_fp16")];
            tensor<int32, [3]> var_14100_perm_1 = const()[name = string("op_14100_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_14100_cast_fp16 = transpose(perm = var_14100_perm_1, x = var_14099_cast_fp16)[name = string("transpose_535")];
            tensor<fp16, [1, 51, 1024]> input_515_cast_fp16 = add(x = input_499_cast_fp16, y = var_14100_cast_fp16)[name = string("input_515_cast_fp16")];
            tensor<int32, [1]> x_929_axes_1 = const()[name = string("x_929_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316924736)))];
            tensor<fp16, [1024]> layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316926848)))];
            tensor<fp16, [1, 51, 1024]> x_929_cast_fp16 = layer_norm(axes = x_929_axes_1, beta = layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_12902_to_fp16, gamma = layers_10_norm_feed_forward2_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_929_cast_fp16")];
            tensor<int32, [4]> var_14110 = const()[name = string("op_14110"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_931_cast_fp16 = reshape(shape = var_14110, x = x_929_cast_fp16)[name = string("x_931_cast_fp16")];
            tensor<int32, [4]> input_517_perm_1 = const()[name = string("input_517_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_889_pad_type_1 = const()[name = string("dense_output_889_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_889_strides_1 = const()[name = string("dense_output_889_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_889_pad_1 = const()[name = string("dense_output_889_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_889_dilations_1 = const()[name = string("dense_output_889_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_889_groups_1 = const()[name = string("dense_output_889_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_10_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316928960))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321123328))))[name = string("layers_10_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_517_cast_fp16 = transpose(perm = input_517_perm_1, x = x_931_cast_fp16)[name = string("transpose_534")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_889_cast_fp16 = conv(dilations = dense_output_889_dilations_1, groups = dense_output_889_groups_1, pad = dense_output_889_pad_1, pad_type = dense_output_889_pad_type_1, strides = dense_output_889_strides_1, weight = layers_10_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_517_cast_fp16)[name = string("dense_output_889_cast_fp16")];
            string sparse_output_889_pad_type_1 = const()[name = string("sparse_output_889_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_889_strides_1 = const()[name = string("sparse_output_889_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_889_pad_1 = const()[name = string("sparse_output_889_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_889_dilations_1 = const()[name = string("sparse_output_889_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_889_groups_1 = const()[name = string("sparse_output_889_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_10_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321207872))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321123904))))[name = string("layers_10_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_889_cast_fp16 = conv(dilations = sparse_output_889_dilations_1, groups = sparse_output_889_groups_1, pad = sparse_output_889_pad_1, pad_type = sparse_output_889_pad_type_1, strides = sparse_output_889_strides_1, weight = layers_10_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_517_cast_fp16)[name = string("sparse_output_889_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_519_cast_fp16 = add(x = dense_output_889_cast_fp16, y = sparse_output_889_cast_fp16)[name = string("input_519_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_521_cast_fp16 = silu(x = input_519_cast_fp16)[name = string("input_521_cast_fp16")];
            string dense_output_891_pad_type_1 = const()[name = string("dense_output_891_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_891_strides_1 = const()[name = string("dense_output_891_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_891_pad_1 = const()[name = string("dense_output_891_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_891_dilations_1 = const()[name = string("dense_output_891_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_891_groups_1 = const()[name = string("dense_output_891_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_10_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321732224))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325926592))))[name = string("layers_10_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_891_cast_fp16 = conv(dilations = dense_output_891_dilations_1, groups = dense_output_891_groups_1, pad = dense_output_891_pad_1, pad_type = dense_output_891_pad_type_1, strides = dense_output_891_strides_1, weight = layers_10_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_521_cast_fp16)[name = string("dense_output_891_cast_fp16")];
            string sparse_output_891_pad_type_1 = const()[name = string("sparse_output_891_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_891_strides_1 = const()[name = string("sparse_output_891_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_891_pad_1 = const()[name = string("sparse_output_891_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_891_dilations_1 = const()[name = string("sparse_output_891_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_891_groups_1 = const()[name = string("sparse_output_891_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_10_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326011136))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325927168))))[name = string("layers_10_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_891_cast_fp16 = conv(dilations = sparse_output_891_dilations_1, groups = sparse_output_891_groups_1, pad = sparse_output_891_pad_1, pad_type = sparse_output_891_pad_type_1, strides = sparse_output_891_strides_1, weight = layers_10_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_521_cast_fp16)[name = string("sparse_output_891_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_933_cast_fp16 = add(x = dense_output_891_cast_fp16, y = sparse_output_891_cast_fp16)[name = string("x_933_cast_fp16")];
            tensor<int32, [4]> x_935_perm_1 = const()[name = string("x_935_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_14145 = const()[name = string("op_14145"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_935_cast_fp16 = transpose(perm = x_935_perm_1, x = x_933_cast_fp16)[name = string("transpose_533")];
            tensor<fp16, [1, 51, 1024]> var_14146_cast_fp16 = reshape(shape = var_14145, x = x_935_cast_fp16)[name = string("op_14146_cast_fp16")];
            fp16 var_14147_to_fp16 = const()[name = string("op_14147_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_14148_cast_fp16 = mul(x = var_14146_cast_fp16, y = var_14147_to_fp16)[name = string("op_14148_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_523_cast_fp16 = add(x = input_515_cast_fp16, y = var_14148_cast_fp16)[name = string("input_523_cast_fp16")];
            tensor<int32, [1]> input_525_axes_1 = const()[name = string("input_525_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_10_norm_out_weight_to_fp16 = const()[name = string("layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326535488)))];
            tensor<fp16, [1024]> layers_10_norm_out_bias_to_fp16 = const()[name = string("layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326537600)))];
            tensor<fp16, [1, 51, 1024]> input_525_cast_fp16 = layer_norm(axes = input_525_axes_1, beta = layers_10_norm_out_bias_to_fp16, epsilon = var_12902_to_fp16, gamma = layers_10_norm_out_weight_to_fp16, x = input_523_cast_fp16)[name = string("input_525_cast_fp16")];
            int32 var_14156 = const()[name = string("op_14156"), val = int32(-1)];
            tensor<int32, [1]> x_937_axes_1 = const()[name = string("x_937_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326539712)))];
            tensor<fp16, [1024]> layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326541824)))];
            fp16 var_14171_to_fp16 = const()[name = string("op_14171_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_937_cast_fp16 = layer_norm(axes = x_937_axes_1, beta = layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_14171_to_fp16, gamma = layers_11_norm_feed_forward1_weight_to_fp16, x = input_525_cast_fp16)[name = string("x_937_cast_fp16")];
            tensor<int32, [4]> var_14190 = const()[name = string("op_14190"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_939_cast_fp16 = reshape(shape = var_14190, x = x_937_cast_fp16)[name = string("x_939_cast_fp16")];
            tensor<int32, [4]> input_527_perm_1 = const()[name = string("input_527_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_893_pad_type_1 = const()[name = string("dense_output_893_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_893_strides_1 = const()[name = string("dense_output_893_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_893_pad_1 = const()[name = string("dense_output_893_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_893_dilations_1 = const()[name = string("dense_output_893_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_893_groups_1 = const()[name = string("dense_output_893_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_11_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326543936))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330738304))))[name = string("layers_11_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_527_cast_fp16 = transpose(perm = input_527_perm_1, x = x_939_cast_fp16)[name = string("transpose_532")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_893_cast_fp16 = conv(dilations = dense_output_893_dilations_1, groups = dense_output_893_groups_1, pad = dense_output_893_pad_1, pad_type = dense_output_893_pad_type_1, strides = dense_output_893_strides_1, weight = layers_11_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_527_cast_fp16)[name = string("dense_output_893_cast_fp16")];
            string sparse_output_893_pad_type_1 = const()[name = string("sparse_output_893_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_893_strides_1 = const()[name = string("sparse_output_893_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_893_pad_1 = const()[name = string("sparse_output_893_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_893_dilations_1 = const()[name = string("sparse_output_893_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_893_groups_1 = const()[name = string("sparse_output_893_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_11_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330822848))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330738880))))[name = string("layers_11_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_893_cast_fp16 = conv(dilations = sparse_output_893_dilations_1, groups = sparse_output_893_groups_1, pad = sparse_output_893_pad_1, pad_type = sparse_output_893_pad_type_1, strides = sparse_output_893_strides_1, weight = layers_11_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_527_cast_fp16)[name = string("sparse_output_893_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_529_cast_fp16 = add(x = dense_output_893_cast_fp16, y = sparse_output_893_cast_fp16)[name = string("input_529_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_531_cast_fp16 = silu(x = input_529_cast_fp16)[name = string("input_531_cast_fp16")];
            string dense_output_895_pad_type_1 = const()[name = string("dense_output_895_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_895_strides_1 = const()[name = string("dense_output_895_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_895_pad_1 = const()[name = string("dense_output_895_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_895_dilations_1 = const()[name = string("dense_output_895_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_895_groups_1 = const()[name = string("dense_output_895_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_11_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331347200))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335541568))))[name = string("layers_11_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_895_cast_fp16 = conv(dilations = dense_output_895_dilations_1, groups = dense_output_895_groups_1, pad = dense_output_895_pad_1, pad_type = dense_output_895_pad_type_1, strides = dense_output_895_strides_1, weight = layers_11_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_531_cast_fp16)[name = string("dense_output_895_cast_fp16")];
            string sparse_output_895_pad_type_1 = const()[name = string("sparse_output_895_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_895_strides_1 = const()[name = string("sparse_output_895_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_895_pad_1 = const()[name = string("sparse_output_895_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_895_dilations_1 = const()[name = string("sparse_output_895_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_895_groups_1 = const()[name = string("sparse_output_895_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_11_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335626112))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335542144))))[name = string("layers_11_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_895_cast_fp16 = conv(dilations = sparse_output_895_dilations_1, groups = sparse_output_895_groups_1, pad = sparse_output_895_pad_1, pad_type = sparse_output_895_pad_type_1, strides = sparse_output_895_strides_1, weight = layers_11_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_531_cast_fp16)[name = string("sparse_output_895_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_941_cast_fp16 = add(x = dense_output_895_cast_fp16, y = sparse_output_895_cast_fp16)[name = string("x_941_cast_fp16")];
            tensor<int32, [4]> x_943_perm_1 = const()[name = string("x_943_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_14225 = const()[name = string("op_14225"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_943_cast_fp16 = transpose(perm = x_943_perm_1, x = x_941_cast_fp16)[name = string("transpose_531")];
            tensor<fp16, [1, 51, 1024]> var_14226_cast_fp16 = reshape(shape = var_14225, x = x_943_cast_fp16)[name = string("op_14226_cast_fp16")];
            fp16 var_14227_to_fp16 = const()[name = string("op_14227_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_14228_cast_fp16 = mul(x = var_14226_cast_fp16, y = var_14227_to_fp16)[name = string("op_14228_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_533_cast_fp16 = add(x = input_525_cast_fp16, y = var_14228_cast_fp16)[name = string("input_533_cast_fp16")];
            tensor<int32, [1]> q_23_axes_1 = const()[name = string("q_23_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_11_norm_self_att_weight_to_fp16 = const()[name = string("layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336150464)))];
            tensor<fp16, [1024]> layers_11_norm_self_att_bias_to_fp16 = const()[name = string("layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336152576)))];
            tensor<fp16, [1, 51, 1024]> q_23_cast_fp16 = layer_norm(axes = q_23_axes_1, beta = layers_11_norm_self_att_bias_to_fp16, epsilon = var_14171_to_fp16, gamma = layers_11_norm_self_att_weight_to_fp16, x = input_533_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<int32, [3]> var_14302 = const()[name = string("op_14302"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_535_axes_1 = const()[name = string("input_535_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_14303_cast_fp16 = transpose(perm = var_14302, x = q_23_cast_fp16)[name = string("transpose_530")];
            tensor<fp16, [1, 1024, 51, 1]> input_535_cast_fp16 = expand_dims(axes = input_535_axes_1, x = var_14303_cast_fp16)[name = string("input_535_cast_fp16")];
            string dense_output_897_pad_type_1 = const()[name = string("dense_output_897_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_897_strides_1 = const()[name = string("dense_output_897_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_897_pad_1 = const()[name = string("dense_output_897_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_897_dilations_1 = const()[name = string("dense_output_897_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_897_groups_1 = const()[name = string("dense_output_897_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336154688))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336285824))))[name = string("layers_11_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_897_cast_fp16 = conv(dilations = dense_output_897_dilations_1, groups = dense_output_897_groups_1, pad = dense_output_897_pad_1, pad_type = dense_output_897_pad_type_1, strides = dense_output_897_strides_1, weight = layers_11_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_897_cast_fp16")];
            string sparse_output_897_pad_type_1 = const()[name = string("sparse_output_897_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_897_strides_1 = const()[name = string("sparse_output_897_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_897_pad_1 = const()[name = string("sparse_output_897_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_897_dilations_1 = const()[name = string("sparse_output_897_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_897_groups_1 = const()[name = string("sparse_output_897_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336289088))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336286400))))[name = string("layers_11_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_897_cast_fp16 = conv(dilations = sparse_output_897_dilations_1, groups = sparse_output_897_groups_1, pad = sparse_output_897_pad_1, pad_type = sparse_output_897_pad_type_1, strides = sparse_output_897_strides_1, weight = layers_11_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_897_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14328_cast_fp16 = add(x = dense_output_897_cast_fp16, y = sparse_output_897_cast_fp16)[name = string("op_14328_cast_fp16")];
            tensor<int32, [4]> var_14329 = const()[name = string("op_14329"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14331 = const()[name = string("op_14331"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14330_cast_fp16 = transpose(perm = var_14329, x = var_14328_cast_fp16)[name = string("transpose_529")];
            tensor<fp16, [1, 51, 128]> q_head_177_cast_fp16 = reshape(shape = var_14331, x = var_14330_cast_fp16)[name = string("q_head_177_cast_fp16")];
            string dense_output_899_pad_type_1 = const()[name = string("dense_output_899_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_899_strides_1 = const()[name = string("dense_output_899_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_899_pad_1 = const()[name = string("dense_output_899_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_899_dilations_1 = const()[name = string("dense_output_899_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_899_groups_1 = const()[name = string("dense_output_899_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336305536))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336436672))))[name = string("layers_11_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_899_cast_fp16 = conv(dilations = dense_output_899_dilations_1, groups = dense_output_899_groups_1, pad = dense_output_899_pad_1, pad_type = dense_output_899_pad_type_1, strides = dense_output_899_strides_1, weight = layers_11_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_899_cast_fp16")];
            string sparse_output_899_pad_type_1 = const()[name = string("sparse_output_899_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_899_strides_1 = const()[name = string("sparse_output_899_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_899_pad_1 = const()[name = string("sparse_output_899_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_899_dilations_1 = const()[name = string("sparse_output_899_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_899_groups_1 = const()[name = string("sparse_output_899_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336439936))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336437248))))[name = string("layers_11_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_899_cast_fp16 = conv(dilations = sparse_output_899_dilations_1, groups = sparse_output_899_groups_1, pad = sparse_output_899_pad_1, pad_type = sparse_output_899_pad_type_1, strides = sparse_output_899_strides_1, weight = layers_11_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_899_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14347_cast_fp16 = add(x = dense_output_899_cast_fp16, y = sparse_output_899_cast_fp16)[name = string("op_14347_cast_fp16")];
            tensor<int32, [4]> var_14348 = const()[name = string("op_14348"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14350 = const()[name = string("op_14350"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14349_cast_fp16 = transpose(perm = var_14348, x = var_14347_cast_fp16)[name = string("transpose_528")];
            tensor<fp16, [1, 51, 128]> k_head_353_cast_fp16 = reshape(shape = var_14350, x = var_14349_cast_fp16)[name = string("k_head_353_cast_fp16")];
            string dense_output_901_pad_type_1 = const()[name = string("dense_output_901_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_901_strides_1 = const()[name = string("dense_output_901_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_901_pad_1 = const()[name = string("dense_output_901_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_901_dilations_1 = const()[name = string("dense_output_901_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_901_groups_1 = const()[name = string("dense_output_901_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336456384))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336587520))))[name = string("layers_11_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_901_cast_fp16 = conv(dilations = dense_output_901_dilations_1, groups = dense_output_901_groups_1, pad = dense_output_901_pad_1, pad_type = dense_output_901_pad_type_1, strides = dense_output_901_strides_1, weight = layers_11_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_901_cast_fp16")];
            string sparse_output_901_pad_type_1 = const()[name = string("sparse_output_901_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_901_strides_1 = const()[name = string("sparse_output_901_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_901_pad_1 = const()[name = string("sparse_output_901_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_901_dilations_1 = const()[name = string("sparse_output_901_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_901_groups_1 = const()[name = string("sparse_output_901_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336590784))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336588096))))[name = string("layers_11_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_901_cast_fp16 = conv(dilations = sparse_output_901_dilations_1, groups = sparse_output_901_groups_1, pad = sparse_output_901_pad_1, pad_type = sparse_output_901_pad_type_1, strides = sparse_output_901_strides_1, weight = layers_11_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_901_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14366_cast_fp16 = add(x = dense_output_901_cast_fp16, y = sparse_output_901_cast_fp16)[name = string("op_14366_cast_fp16")];
            tensor<int32, [4]> var_14367 = const()[name = string("op_14367"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14369 = const()[name = string("op_14369"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14368_cast_fp16 = transpose(perm = var_14367, x = var_14366_cast_fp16)[name = string("transpose_527")];
            tensor<fp16, [1, 51, 128]> v_head_353_cast_fp16 = reshape(shape = var_14369, x = var_14368_cast_fp16)[name = string("v_head_353_cast_fp16")];
            string dense_output_903_pad_type_1 = const()[name = string("dense_output_903_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_903_strides_1 = const()[name = string("dense_output_903_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_903_pad_1 = const()[name = string("dense_output_903_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_903_dilations_1 = const()[name = string("dense_output_903_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_903_groups_1 = const()[name = string("dense_output_903_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336607232))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336738368))))[name = string("layers_11_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_903_cast_fp16 = conv(dilations = dense_output_903_dilations_1, groups = dense_output_903_groups_1, pad = dense_output_903_pad_1, pad_type = dense_output_903_pad_type_1, strides = dense_output_903_strides_1, weight = layers_11_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_903_cast_fp16")];
            string sparse_output_903_pad_type_1 = const()[name = string("sparse_output_903_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_903_strides_1 = const()[name = string("sparse_output_903_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_903_pad_1 = const()[name = string("sparse_output_903_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_903_dilations_1 = const()[name = string("sparse_output_903_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_903_groups_1 = const()[name = string("sparse_output_903_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336741632))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336738944))))[name = string("layers_11_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_903_cast_fp16 = conv(dilations = sparse_output_903_dilations_1, groups = sparse_output_903_groups_1, pad = sparse_output_903_pad_1, pad_type = sparse_output_903_pad_type_1, strides = sparse_output_903_strides_1, weight = layers_11_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_903_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_14385_cast_fp16 = add(x = dense_output_903_cast_fp16, y = sparse_output_903_cast_fp16)[name = string("op_14385_cast_fp16")];
            tensor<int32, [4]> var_14386 = const()[name = string("op_14386"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14388 = const()[name = string("op_14388"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_14387_cast_fp16 = transpose(perm = var_14386, x = var_14385_cast_fp16)[name = string("transpose_526")];
            tensor<fp16, [1, 101, 128]> p_head_353_cast_fp16 = reshape(shape = var_14388, x = var_14387_cast_fp16)[name = string("p_head_353_cast_fp16")];
            tensor<fp16, [128]> var_14390_to_fp16 = const()[name = string("op_14390_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336758080)))];
            tensor<fp16, [1, 51, 128]> var_14391_cast_fp16 = add(x = q_head_177_cast_fp16, y = var_14390_to_fp16)[name = string("op_14391_cast_fp16")];
            tensor<int32, [1]> q_u_177_axes_1 = const()[name = string("q_u_177_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_177_cast_fp16 = expand_dims(axes = q_u_177_axes_1, x = var_14391_cast_fp16)[name = string("q_u_177_cast_fp16")];
            tensor<fp16, [128]> var_14393_to_fp16 = const()[name = string("op_14393_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336758400)))];
            tensor<fp16, [1, 51, 128]> var_14394_cast_fp16 = add(x = q_head_177_cast_fp16, y = var_14393_to_fp16)[name = string("op_14394_cast_fp16")];
            tensor<int32, [1]> q_v_177_axes_1 = const()[name = string("q_v_177_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_177_cast_fp16 = expand_dims(axes = q_v_177_axes_1, x = var_14394_cast_fp16)[name = string("q_v_177_cast_fp16")];
            tensor<int32, [1]> k_head_355_axes_1 = const()[name = string("k_head_355_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_355_cast_fp16 = expand_dims(axes = k_head_355_axes_1, x = k_head_353_cast_fp16)[name = string("k_head_355_cast_fp16")];
            tensor<int32, [1]> v_head_355_axes_1 = const()[name = string("v_head_355_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_355_cast_fp16 = expand_dims(axes = v_head_355_axes_1, x = v_head_353_cast_fp16)[name = string("v_head_355_cast_fp16")];
            tensor<int32, [1]> p_head_355_axes_1 = const()[name = string("p_head_355_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_355_cast_fp16 = expand_dims(axes = p_head_355_axes_1, x = p_head_353_cast_fp16)[name = string("p_head_355_cast_fp16")];
            bool var_14400_transpose_x_3 = const()[name = string("op_14400_transpose_x_3"), val = bool(false)];
            bool var_14400_transpose_y_3 = const()[name = string("op_14400_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_14400_cast_fp16 = matmul(transpose_x = var_14400_transpose_x_3, transpose_y = var_14400_transpose_y_3, x = q_u_177_cast_fp16, y = k_head_355_cast_fp16)[name = string("op_14400_cast_fp16")];
            fp16 var_14401_to_fp16 = const()[name = string("op_14401_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_177_cast_fp16 = mul(x = var_14400_cast_fp16, y = var_14401_to_fp16)[name = string("scores_content_177_cast_fp16")];
            bool x_945_transpose_x_3 = const()[name = string("x_945_transpose_x_3"), val = bool(false)];
            bool x_945_transpose_y_3 = const()[name = string("x_945_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_945_cast_fp16 = matmul(transpose_x = x_945_transpose_x_3, transpose_y = x_945_transpose_y_3, x = q_v_177_cast_fp16, y = p_head_355_cast_fp16)[name = string("x_945_cast_fp16")];
            tensor<int32, [8]> x_947_pad_1 = const()[name = string("x_947_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_947_mode_1 = const()[name = string("x_947_mode_1"), val = string("constant")];
            fp16 const_1903_to_fp16 = const()[name = string("const_1903_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_947_cast_fp16 = pad(constant_val = const_1903_to_fp16, mode = x_947_mode_1, pad = x_947_pad_1, x = x_945_cast_fp16)[name = string("x_947_cast_fp16")];
            tensor<int32, [4]> var_14415 = const()[name = string("op_14415"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_949_cast_fp16 = reshape(shape = var_14415, x = x_947_cast_fp16)[name = string("x_949_cast_fp16")];
            tensor<int32, [4]> var_14419_begin_1 = const()[name = string("op_14419_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14419_end_1 = const()[name = string("op_14419_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_14419_end_mask_1 = const()[name = string("op_14419_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_14419_cast_fp16 = slice_by_index(begin = var_14419_begin_1, end = var_14419_end_1, end_mask = var_14419_end_mask_1, x = x_949_cast_fp16)[name = string("op_14419_cast_fp16")];
            tensor<int32, [4]> var_14421 = const()[name = string("op_14421"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_14422_cast_fp16 = reshape(shape = var_14421, x = var_14419_cast_fp16)[name = string("op_14422_cast_fp16")];
            tensor<int32, [4]> var_14427_begin_1 = const()[name = string("op_14427_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14427_end_1 = const()[name = string("op_14427_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_14427_end_mask_1 = const()[name = string("op_14427_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_14427_cast_fp16 = slice_by_index(begin = var_14427_begin_1, end = var_14427_end_1, end_mask = var_14427_end_mask_1, x = var_14422_cast_fp16)[name = string("op_14427_cast_fp16")];
            fp16 var_14428_to_fp16 = const()[name = string("op_14428_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_177_cast_fp16 = mul(x = var_14427_cast_fp16, y = var_14428_to_fp16)[name = string("scores_pos_177_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_177_cast_fp16 = add(x = scores_content_177_cast_fp16, y = scores_pos_177_cast_fp16)[name = string("logits_177_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_14431_cast_fp16 = softmax(axis = var_14156, x = logits_177_cast_fp16)[name = string("op_14431_cast_fp16")];
            bool var_14433_transpose_x_1 = const()[name = string("op_14433_transpose_x_1"), val = bool(false)];
            bool var_14433_transpose_y_1 = const()[name = string("op_14433_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_14433_cast_fp16 = matmul(transpose_x = var_14433_transpose_x_1, transpose_y = var_14433_transpose_y_1, x = var_14431_cast_fp16, y = v_head_355_cast_fp16)[name = string("op_14433_cast_fp16")];
            tensor<int32, [1]> var_14434_axes_1 = const()[name = string("op_14434_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_14434_cast_fp16 = squeeze(axes = var_14434_axes_1, x = var_14433_cast_fp16)[name = string("op_14434_cast_fp16")];
            string dense_output_905_pad_type_1 = const()[name = string("dense_output_905_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_905_strides_1 = const()[name = string("dense_output_905_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_905_pad_1 = const()[name = string("dense_output_905_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_905_dilations_1 = const()[name = string("dense_output_905_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_905_groups_1 = const()[name = string("dense_output_905_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336758720))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336889856))))[name = string("layers_11_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_905_cast_fp16 = conv(dilations = dense_output_905_dilations_1, groups = dense_output_905_groups_1, pad = dense_output_905_pad_1, pad_type = dense_output_905_pad_type_1, strides = dense_output_905_strides_1, weight = layers_11_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_905_cast_fp16")];
            string sparse_output_905_pad_type_1 = const()[name = string("sparse_output_905_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_905_strides_1 = const()[name = string("sparse_output_905_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_905_pad_1 = const()[name = string("sparse_output_905_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_905_dilations_1 = const()[name = string("sparse_output_905_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_905_groups_1 = const()[name = string("sparse_output_905_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336893120))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336890432))))[name = string("layers_11_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_905_cast_fp16 = conv(dilations = sparse_output_905_dilations_1, groups = sparse_output_905_groups_1, pad = sparse_output_905_pad_1, pad_type = sparse_output_905_pad_type_1, strides = sparse_output_905_strides_1, weight = layers_11_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_905_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14449_cast_fp16 = add(x = dense_output_905_cast_fp16, y = sparse_output_905_cast_fp16)[name = string("op_14449_cast_fp16")];
            tensor<int32, [4]> var_14450 = const()[name = string("op_14450"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14452 = const()[name = string("op_14452"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14451_cast_fp16 = transpose(perm = var_14450, x = var_14449_cast_fp16)[name = string("transpose_525")];
            tensor<fp16, [1, 51, 128]> q_head_179_cast_fp16 = reshape(shape = var_14452, x = var_14451_cast_fp16)[name = string("q_head_179_cast_fp16")];
            string dense_output_907_pad_type_1 = const()[name = string("dense_output_907_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_907_strides_1 = const()[name = string("dense_output_907_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_907_pad_1 = const()[name = string("dense_output_907_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_907_dilations_1 = const()[name = string("dense_output_907_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_907_groups_1 = const()[name = string("dense_output_907_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336909568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337040704))))[name = string("layers_11_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_907_cast_fp16 = conv(dilations = dense_output_907_dilations_1, groups = dense_output_907_groups_1, pad = dense_output_907_pad_1, pad_type = dense_output_907_pad_type_1, strides = dense_output_907_strides_1, weight = layers_11_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_907_cast_fp16")];
            string sparse_output_907_pad_type_1 = const()[name = string("sparse_output_907_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_907_strides_1 = const()[name = string("sparse_output_907_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_907_pad_1 = const()[name = string("sparse_output_907_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_907_dilations_1 = const()[name = string("sparse_output_907_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_907_groups_1 = const()[name = string("sparse_output_907_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337043968))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337041280))))[name = string("layers_11_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_907_cast_fp16 = conv(dilations = sparse_output_907_dilations_1, groups = sparse_output_907_groups_1, pad = sparse_output_907_pad_1, pad_type = sparse_output_907_pad_type_1, strides = sparse_output_907_strides_1, weight = layers_11_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_907_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14468_cast_fp16 = add(x = dense_output_907_cast_fp16, y = sparse_output_907_cast_fp16)[name = string("op_14468_cast_fp16")];
            tensor<int32, [4]> var_14469 = const()[name = string("op_14469"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14471 = const()[name = string("op_14471"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14470_cast_fp16 = transpose(perm = var_14469, x = var_14468_cast_fp16)[name = string("transpose_524")];
            tensor<fp16, [1, 51, 128]> k_head_357_cast_fp16 = reshape(shape = var_14471, x = var_14470_cast_fp16)[name = string("k_head_357_cast_fp16")];
            string dense_output_909_pad_type_1 = const()[name = string("dense_output_909_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_909_strides_1 = const()[name = string("dense_output_909_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_909_pad_1 = const()[name = string("dense_output_909_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_909_dilations_1 = const()[name = string("dense_output_909_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_909_groups_1 = const()[name = string("dense_output_909_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337060416))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337191552))))[name = string("layers_11_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_909_cast_fp16 = conv(dilations = dense_output_909_dilations_1, groups = dense_output_909_groups_1, pad = dense_output_909_pad_1, pad_type = dense_output_909_pad_type_1, strides = dense_output_909_strides_1, weight = layers_11_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_909_cast_fp16")];
            string sparse_output_909_pad_type_1 = const()[name = string("sparse_output_909_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_909_strides_1 = const()[name = string("sparse_output_909_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_909_pad_1 = const()[name = string("sparse_output_909_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_909_dilations_1 = const()[name = string("sparse_output_909_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_909_groups_1 = const()[name = string("sparse_output_909_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337194816))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337192128))))[name = string("layers_11_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_909_cast_fp16 = conv(dilations = sparse_output_909_dilations_1, groups = sparse_output_909_groups_1, pad = sparse_output_909_pad_1, pad_type = sparse_output_909_pad_type_1, strides = sparse_output_909_strides_1, weight = layers_11_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_909_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14487_cast_fp16 = add(x = dense_output_909_cast_fp16, y = sparse_output_909_cast_fp16)[name = string("op_14487_cast_fp16")];
            tensor<int32, [4]> var_14488 = const()[name = string("op_14488"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14490 = const()[name = string("op_14490"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14489_cast_fp16 = transpose(perm = var_14488, x = var_14487_cast_fp16)[name = string("transpose_523")];
            tensor<fp16, [1, 51, 128]> v_head_357_cast_fp16 = reshape(shape = var_14490, x = var_14489_cast_fp16)[name = string("v_head_357_cast_fp16")];
            string dense_output_911_pad_type_1 = const()[name = string("dense_output_911_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_911_strides_1 = const()[name = string("dense_output_911_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_911_pad_1 = const()[name = string("dense_output_911_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_911_dilations_1 = const()[name = string("dense_output_911_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_911_groups_1 = const()[name = string("dense_output_911_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337211264))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337342400))))[name = string("layers_11_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_911_cast_fp16 = conv(dilations = dense_output_911_dilations_1, groups = dense_output_911_groups_1, pad = dense_output_911_pad_1, pad_type = dense_output_911_pad_type_1, strides = dense_output_911_strides_1, weight = layers_11_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_911_cast_fp16")];
            string sparse_output_911_pad_type_1 = const()[name = string("sparse_output_911_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_911_strides_1 = const()[name = string("sparse_output_911_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_911_pad_1 = const()[name = string("sparse_output_911_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_911_dilations_1 = const()[name = string("sparse_output_911_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_911_groups_1 = const()[name = string("sparse_output_911_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337345664))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337342976))))[name = string("layers_11_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_911_cast_fp16 = conv(dilations = sparse_output_911_dilations_1, groups = sparse_output_911_groups_1, pad = sparse_output_911_pad_1, pad_type = sparse_output_911_pad_type_1, strides = sparse_output_911_strides_1, weight = layers_11_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_911_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_14506_cast_fp16 = add(x = dense_output_911_cast_fp16, y = sparse_output_911_cast_fp16)[name = string("op_14506_cast_fp16")];
            tensor<int32, [4]> var_14507 = const()[name = string("op_14507"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14509 = const()[name = string("op_14509"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_14508_cast_fp16 = transpose(perm = var_14507, x = var_14506_cast_fp16)[name = string("transpose_522")];
            tensor<fp16, [1, 101, 128]> p_head_357_cast_fp16 = reshape(shape = var_14509, x = var_14508_cast_fp16)[name = string("p_head_357_cast_fp16")];
            tensor<fp16, [128]> var_14511_to_fp16 = const()[name = string("op_14511_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337362112)))];
            tensor<fp16, [1, 51, 128]> var_14512_cast_fp16 = add(x = q_head_179_cast_fp16, y = var_14511_to_fp16)[name = string("op_14512_cast_fp16")];
            tensor<int32, [1]> q_u_179_axes_1 = const()[name = string("q_u_179_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_179_cast_fp16 = expand_dims(axes = q_u_179_axes_1, x = var_14512_cast_fp16)[name = string("q_u_179_cast_fp16")];
            tensor<fp16, [128]> var_14514_to_fp16 = const()[name = string("op_14514_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337362432)))];
            tensor<fp16, [1, 51, 128]> var_14515_cast_fp16 = add(x = q_head_179_cast_fp16, y = var_14514_to_fp16)[name = string("op_14515_cast_fp16")];
            tensor<int32, [1]> q_v_179_axes_1 = const()[name = string("q_v_179_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_179_cast_fp16 = expand_dims(axes = q_v_179_axes_1, x = var_14515_cast_fp16)[name = string("q_v_179_cast_fp16")];
            tensor<int32, [1]> k_head_359_axes_1 = const()[name = string("k_head_359_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_359_cast_fp16 = expand_dims(axes = k_head_359_axes_1, x = k_head_357_cast_fp16)[name = string("k_head_359_cast_fp16")];
            tensor<int32, [1]> v_head_359_axes_1 = const()[name = string("v_head_359_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_359_cast_fp16 = expand_dims(axes = v_head_359_axes_1, x = v_head_357_cast_fp16)[name = string("v_head_359_cast_fp16")];
            tensor<int32, [1]> p_head_359_axes_1 = const()[name = string("p_head_359_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_359_cast_fp16 = expand_dims(axes = p_head_359_axes_1, x = p_head_357_cast_fp16)[name = string("p_head_359_cast_fp16")];
            bool var_14521_transpose_x_3 = const()[name = string("op_14521_transpose_x_3"), val = bool(false)];
            bool var_14521_transpose_y_3 = const()[name = string("op_14521_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_14521_cast_fp16 = matmul(transpose_x = var_14521_transpose_x_3, transpose_y = var_14521_transpose_y_3, x = q_u_179_cast_fp16, y = k_head_359_cast_fp16)[name = string("op_14521_cast_fp16")];
            fp16 var_14522_to_fp16 = const()[name = string("op_14522_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_179_cast_fp16 = mul(x = var_14521_cast_fp16, y = var_14522_to_fp16)[name = string("scores_content_179_cast_fp16")];
            bool x_953_transpose_x_3 = const()[name = string("x_953_transpose_x_3"), val = bool(false)];
            bool x_953_transpose_y_3 = const()[name = string("x_953_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_953_cast_fp16 = matmul(transpose_x = x_953_transpose_x_3, transpose_y = x_953_transpose_y_3, x = q_v_179_cast_fp16, y = p_head_359_cast_fp16)[name = string("x_953_cast_fp16")];
            tensor<int32, [8]> x_955_pad_1 = const()[name = string("x_955_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_955_mode_1 = const()[name = string("x_955_mode_1"), val = string("constant")];
            fp16 const_1909_to_fp16 = const()[name = string("const_1909_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_955_cast_fp16 = pad(constant_val = const_1909_to_fp16, mode = x_955_mode_1, pad = x_955_pad_1, x = x_953_cast_fp16)[name = string("x_955_cast_fp16")];
            tensor<int32, [4]> var_14536 = const()[name = string("op_14536"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_957_cast_fp16 = reshape(shape = var_14536, x = x_955_cast_fp16)[name = string("x_957_cast_fp16")];
            tensor<int32, [4]> var_14540_begin_1 = const()[name = string("op_14540_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14540_end_1 = const()[name = string("op_14540_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_14540_end_mask_1 = const()[name = string("op_14540_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_14540_cast_fp16 = slice_by_index(begin = var_14540_begin_1, end = var_14540_end_1, end_mask = var_14540_end_mask_1, x = x_957_cast_fp16)[name = string("op_14540_cast_fp16")];
            tensor<int32, [4]> var_14542 = const()[name = string("op_14542"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_14543_cast_fp16 = reshape(shape = var_14542, x = var_14540_cast_fp16)[name = string("op_14543_cast_fp16")];
            tensor<int32, [4]> var_14548_begin_1 = const()[name = string("op_14548_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14548_end_1 = const()[name = string("op_14548_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_14548_end_mask_1 = const()[name = string("op_14548_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_14548_cast_fp16 = slice_by_index(begin = var_14548_begin_1, end = var_14548_end_1, end_mask = var_14548_end_mask_1, x = var_14543_cast_fp16)[name = string("op_14548_cast_fp16")];
            fp16 var_14549_to_fp16 = const()[name = string("op_14549_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_179_cast_fp16 = mul(x = var_14548_cast_fp16, y = var_14549_to_fp16)[name = string("scores_pos_179_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_179_cast_fp16 = add(x = scores_content_179_cast_fp16, y = scores_pos_179_cast_fp16)[name = string("logits_179_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_14552_cast_fp16 = softmax(axis = var_14156, x = logits_179_cast_fp16)[name = string("op_14552_cast_fp16")];
            bool var_14554_transpose_x_1 = const()[name = string("op_14554_transpose_x_1"), val = bool(false)];
            bool var_14554_transpose_y_1 = const()[name = string("op_14554_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_14554_cast_fp16 = matmul(transpose_x = var_14554_transpose_x_1, transpose_y = var_14554_transpose_y_1, x = var_14552_cast_fp16, y = v_head_359_cast_fp16)[name = string("op_14554_cast_fp16")];
            tensor<int32, [1]> var_14555_axes_1 = const()[name = string("op_14555_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_14555_cast_fp16 = squeeze(axes = var_14555_axes_1, x = var_14554_cast_fp16)[name = string("op_14555_cast_fp16")];
            string dense_output_913_pad_type_1 = const()[name = string("dense_output_913_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_913_strides_1 = const()[name = string("dense_output_913_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_913_pad_1 = const()[name = string("dense_output_913_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_913_dilations_1 = const()[name = string("dense_output_913_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_913_groups_1 = const()[name = string("dense_output_913_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337362752))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337493888))))[name = string("layers_11_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_913_cast_fp16 = conv(dilations = dense_output_913_dilations_1, groups = dense_output_913_groups_1, pad = dense_output_913_pad_1, pad_type = dense_output_913_pad_type_1, strides = dense_output_913_strides_1, weight = layers_11_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_913_cast_fp16")];
            string sparse_output_913_pad_type_1 = const()[name = string("sparse_output_913_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_913_strides_1 = const()[name = string("sparse_output_913_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_913_pad_1 = const()[name = string("sparse_output_913_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_913_dilations_1 = const()[name = string("sparse_output_913_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_913_groups_1 = const()[name = string("sparse_output_913_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337497152))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337494464))))[name = string("layers_11_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_913_cast_fp16 = conv(dilations = sparse_output_913_dilations_1, groups = sparse_output_913_groups_1, pad = sparse_output_913_pad_1, pad_type = sparse_output_913_pad_type_1, strides = sparse_output_913_strides_1, weight = layers_11_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_913_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14570_cast_fp16 = add(x = dense_output_913_cast_fp16, y = sparse_output_913_cast_fp16)[name = string("op_14570_cast_fp16")];
            tensor<int32, [4]> var_14571 = const()[name = string("op_14571"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14573 = const()[name = string("op_14573"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14572_cast_fp16 = transpose(perm = var_14571, x = var_14570_cast_fp16)[name = string("transpose_521")];
            tensor<fp16, [1, 51, 128]> q_head_181_cast_fp16 = reshape(shape = var_14573, x = var_14572_cast_fp16)[name = string("q_head_181_cast_fp16")];
            string dense_output_915_pad_type_1 = const()[name = string("dense_output_915_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_915_strides_1 = const()[name = string("dense_output_915_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_915_pad_1 = const()[name = string("dense_output_915_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_915_dilations_1 = const()[name = string("dense_output_915_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_915_groups_1 = const()[name = string("dense_output_915_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337513600))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337644736))))[name = string("layers_11_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_915_cast_fp16 = conv(dilations = dense_output_915_dilations_1, groups = dense_output_915_groups_1, pad = dense_output_915_pad_1, pad_type = dense_output_915_pad_type_1, strides = dense_output_915_strides_1, weight = layers_11_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_915_cast_fp16")];
            string sparse_output_915_pad_type_1 = const()[name = string("sparse_output_915_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_915_strides_1 = const()[name = string("sparse_output_915_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_915_pad_1 = const()[name = string("sparse_output_915_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_915_dilations_1 = const()[name = string("sparse_output_915_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_915_groups_1 = const()[name = string("sparse_output_915_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337648000))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337645312))))[name = string("layers_11_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_915_cast_fp16 = conv(dilations = sparse_output_915_dilations_1, groups = sparse_output_915_groups_1, pad = sparse_output_915_pad_1, pad_type = sparse_output_915_pad_type_1, strides = sparse_output_915_strides_1, weight = layers_11_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_915_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14589_cast_fp16 = add(x = dense_output_915_cast_fp16, y = sparse_output_915_cast_fp16)[name = string("op_14589_cast_fp16")];
            tensor<int32, [4]> var_14590 = const()[name = string("op_14590"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14592 = const()[name = string("op_14592"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14591_cast_fp16 = transpose(perm = var_14590, x = var_14589_cast_fp16)[name = string("transpose_520")];
            tensor<fp16, [1, 51, 128]> k_head_361_cast_fp16 = reshape(shape = var_14592, x = var_14591_cast_fp16)[name = string("k_head_361_cast_fp16")];
            string dense_output_917_pad_type_1 = const()[name = string("dense_output_917_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_917_strides_1 = const()[name = string("dense_output_917_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_917_pad_1 = const()[name = string("dense_output_917_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_917_dilations_1 = const()[name = string("dense_output_917_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_917_groups_1 = const()[name = string("dense_output_917_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337664448))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337795584))))[name = string("layers_11_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_917_cast_fp16 = conv(dilations = dense_output_917_dilations_1, groups = dense_output_917_groups_1, pad = dense_output_917_pad_1, pad_type = dense_output_917_pad_type_1, strides = dense_output_917_strides_1, weight = layers_11_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_917_cast_fp16")];
            string sparse_output_917_pad_type_1 = const()[name = string("sparse_output_917_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_917_strides_1 = const()[name = string("sparse_output_917_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_917_pad_1 = const()[name = string("sparse_output_917_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_917_dilations_1 = const()[name = string("sparse_output_917_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_917_groups_1 = const()[name = string("sparse_output_917_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337798848))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337796160))))[name = string("layers_11_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_917_cast_fp16 = conv(dilations = sparse_output_917_dilations_1, groups = sparse_output_917_groups_1, pad = sparse_output_917_pad_1, pad_type = sparse_output_917_pad_type_1, strides = sparse_output_917_strides_1, weight = layers_11_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_917_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14608_cast_fp16 = add(x = dense_output_917_cast_fp16, y = sparse_output_917_cast_fp16)[name = string("op_14608_cast_fp16")];
            tensor<int32, [4]> var_14609 = const()[name = string("op_14609"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14611 = const()[name = string("op_14611"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14610_cast_fp16 = transpose(perm = var_14609, x = var_14608_cast_fp16)[name = string("transpose_519")];
            tensor<fp16, [1, 51, 128]> v_head_361_cast_fp16 = reshape(shape = var_14611, x = var_14610_cast_fp16)[name = string("v_head_361_cast_fp16")];
            string dense_output_919_pad_type_1 = const()[name = string("dense_output_919_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_919_strides_1 = const()[name = string("dense_output_919_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_919_pad_1 = const()[name = string("dense_output_919_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_919_dilations_1 = const()[name = string("dense_output_919_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_919_groups_1 = const()[name = string("dense_output_919_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337815296))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337946432))))[name = string("layers_11_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_919_cast_fp16 = conv(dilations = dense_output_919_dilations_1, groups = dense_output_919_groups_1, pad = dense_output_919_pad_1, pad_type = dense_output_919_pad_type_1, strides = dense_output_919_strides_1, weight = layers_11_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_919_cast_fp16")];
            string sparse_output_919_pad_type_1 = const()[name = string("sparse_output_919_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_919_strides_1 = const()[name = string("sparse_output_919_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_919_pad_1 = const()[name = string("sparse_output_919_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_919_dilations_1 = const()[name = string("sparse_output_919_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_919_groups_1 = const()[name = string("sparse_output_919_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337949696))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337947008))))[name = string("layers_11_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_919_cast_fp16 = conv(dilations = sparse_output_919_dilations_1, groups = sparse_output_919_groups_1, pad = sparse_output_919_pad_1, pad_type = sparse_output_919_pad_type_1, strides = sparse_output_919_strides_1, weight = layers_11_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_919_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_14627_cast_fp16 = add(x = dense_output_919_cast_fp16, y = sparse_output_919_cast_fp16)[name = string("op_14627_cast_fp16")];
            tensor<int32, [4]> var_14628 = const()[name = string("op_14628"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14630 = const()[name = string("op_14630"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_14629_cast_fp16 = transpose(perm = var_14628, x = var_14627_cast_fp16)[name = string("transpose_518")];
            tensor<fp16, [1, 101, 128]> p_head_361_cast_fp16 = reshape(shape = var_14630, x = var_14629_cast_fp16)[name = string("p_head_361_cast_fp16")];
            tensor<fp16, [128]> var_14632_to_fp16 = const()[name = string("op_14632_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337966144)))];
            tensor<fp16, [1, 51, 128]> var_14633_cast_fp16 = add(x = q_head_181_cast_fp16, y = var_14632_to_fp16)[name = string("op_14633_cast_fp16")];
            tensor<int32, [1]> q_u_181_axes_1 = const()[name = string("q_u_181_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_181_cast_fp16 = expand_dims(axes = q_u_181_axes_1, x = var_14633_cast_fp16)[name = string("q_u_181_cast_fp16")];
            tensor<fp16, [128]> var_14635_to_fp16 = const()[name = string("op_14635_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337966464)))];
            tensor<fp16, [1, 51, 128]> var_14636_cast_fp16 = add(x = q_head_181_cast_fp16, y = var_14635_to_fp16)[name = string("op_14636_cast_fp16")];
            tensor<int32, [1]> q_v_181_axes_1 = const()[name = string("q_v_181_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_181_cast_fp16 = expand_dims(axes = q_v_181_axes_1, x = var_14636_cast_fp16)[name = string("q_v_181_cast_fp16")];
            tensor<int32, [1]> k_head_363_axes_1 = const()[name = string("k_head_363_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_363_cast_fp16 = expand_dims(axes = k_head_363_axes_1, x = k_head_361_cast_fp16)[name = string("k_head_363_cast_fp16")];
            tensor<int32, [1]> v_head_363_axes_1 = const()[name = string("v_head_363_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_363_cast_fp16 = expand_dims(axes = v_head_363_axes_1, x = v_head_361_cast_fp16)[name = string("v_head_363_cast_fp16")];
            tensor<int32, [1]> p_head_363_axes_1 = const()[name = string("p_head_363_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_363_cast_fp16 = expand_dims(axes = p_head_363_axes_1, x = p_head_361_cast_fp16)[name = string("p_head_363_cast_fp16")];
            bool var_14642_transpose_x_3 = const()[name = string("op_14642_transpose_x_3"), val = bool(false)];
            bool var_14642_transpose_y_3 = const()[name = string("op_14642_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_14642_cast_fp16 = matmul(transpose_x = var_14642_transpose_x_3, transpose_y = var_14642_transpose_y_3, x = q_u_181_cast_fp16, y = k_head_363_cast_fp16)[name = string("op_14642_cast_fp16")];
            fp16 var_14643_to_fp16 = const()[name = string("op_14643_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_181_cast_fp16 = mul(x = var_14642_cast_fp16, y = var_14643_to_fp16)[name = string("scores_content_181_cast_fp16")];
            bool x_961_transpose_x_3 = const()[name = string("x_961_transpose_x_3"), val = bool(false)];
            bool x_961_transpose_y_3 = const()[name = string("x_961_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_961_cast_fp16 = matmul(transpose_x = x_961_transpose_x_3, transpose_y = x_961_transpose_y_3, x = q_v_181_cast_fp16, y = p_head_363_cast_fp16)[name = string("x_961_cast_fp16")];
            tensor<int32, [8]> x_963_pad_1 = const()[name = string("x_963_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_963_mode_1 = const()[name = string("x_963_mode_1"), val = string("constant")];
            fp16 const_1915_to_fp16 = const()[name = string("const_1915_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_963_cast_fp16 = pad(constant_val = const_1915_to_fp16, mode = x_963_mode_1, pad = x_963_pad_1, x = x_961_cast_fp16)[name = string("x_963_cast_fp16")];
            tensor<int32, [4]> var_14657 = const()[name = string("op_14657"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_965_cast_fp16 = reshape(shape = var_14657, x = x_963_cast_fp16)[name = string("x_965_cast_fp16")];
            tensor<int32, [4]> var_14661_begin_1 = const()[name = string("op_14661_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14661_end_1 = const()[name = string("op_14661_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_14661_end_mask_1 = const()[name = string("op_14661_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_14661_cast_fp16 = slice_by_index(begin = var_14661_begin_1, end = var_14661_end_1, end_mask = var_14661_end_mask_1, x = x_965_cast_fp16)[name = string("op_14661_cast_fp16")];
            tensor<int32, [4]> var_14663 = const()[name = string("op_14663"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_14664_cast_fp16 = reshape(shape = var_14663, x = var_14661_cast_fp16)[name = string("op_14664_cast_fp16")];
            tensor<int32, [4]> var_14669_begin_1 = const()[name = string("op_14669_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14669_end_1 = const()[name = string("op_14669_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_14669_end_mask_1 = const()[name = string("op_14669_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_14669_cast_fp16 = slice_by_index(begin = var_14669_begin_1, end = var_14669_end_1, end_mask = var_14669_end_mask_1, x = var_14664_cast_fp16)[name = string("op_14669_cast_fp16")];
            fp16 var_14670_to_fp16 = const()[name = string("op_14670_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_181_cast_fp16 = mul(x = var_14669_cast_fp16, y = var_14670_to_fp16)[name = string("scores_pos_181_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_181_cast_fp16 = add(x = scores_content_181_cast_fp16, y = scores_pos_181_cast_fp16)[name = string("logits_181_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_14673_cast_fp16 = softmax(axis = var_14156, x = logits_181_cast_fp16)[name = string("op_14673_cast_fp16")];
            bool var_14675_transpose_x_1 = const()[name = string("op_14675_transpose_x_1"), val = bool(false)];
            bool var_14675_transpose_y_1 = const()[name = string("op_14675_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_14675_cast_fp16 = matmul(transpose_x = var_14675_transpose_x_1, transpose_y = var_14675_transpose_y_1, x = var_14673_cast_fp16, y = v_head_363_cast_fp16)[name = string("op_14675_cast_fp16")];
            tensor<int32, [1]> var_14676_axes_1 = const()[name = string("op_14676_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_14676_cast_fp16 = squeeze(axes = var_14676_axes_1, x = var_14675_cast_fp16)[name = string("op_14676_cast_fp16")];
            string dense_output_921_pad_type_1 = const()[name = string("dense_output_921_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_921_strides_1 = const()[name = string("dense_output_921_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_921_pad_1 = const()[name = string("dense_output_921_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_921_dilations_1 = const()[name = string("dense_output_921_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_921_groups_1 = const()[name = string("dense_output_921_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337966784))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338097920))))[name = string("layers_11_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_921_cast_fp16 = conv(dilations = dense_output_921_dilations_1, groups = dense_output_921_groups_1, pad = dense_output_921_pad_1, pad_type = dense_output_921_pad_type_1, strides = dense_output_921_strides_1, weight = layers_11_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_921_cast_fp16")];
            string sparse_output_921_pad_type_1 = const()[name = string("sparse_output_921_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_921_strides_1 = const()[name = string("sparse_output_921_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_921_pad_1 = const()[name = string("sparse_output_921_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_921_dilations_1 = const()[name = string("sparse_output_921_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_921_groups_1 = const()[name = string("sparse_output_921_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338101184))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338098496))))[name = string("layers_11_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_921_cast_fp16 = conv(dilations = sparse_output_921_dilations_1, groups = sparse_output_921_groups_1, pad = sparse_output_921_pad_1, pad_type = sparse_output_921_pad_type_1, strides = sparse_output_921_strides_1, weight = layers_11_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_921_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14691_cast_fp16 = add(x = dense_output_921_cast_fp16, y = sparse_output_921_cast_fp16)[name = string("op_14691_cast_fp16")];
            tensor<int32, [4]> var_14692 = const()[name = string("op_14692"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14694 = const()[name = string("op_14694"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14693_cast_fp16 = transpose(perm = var_14692, x = var_14691_cast_fp16)[name = string("transpose_517")];
            tensor<fp16, [1, 51, 128]> q_head_183_cast_fp16 = reshape(shape = var_14694, x = var_14693_cast_fp16)[name = string("q_head_183_cast_fp16")];
            string dense_output_923_pad_type_1 = const()[name = string("dense_output_923_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_923_strides_1 = const()[name = string("dense_output_923_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_923_pad_1 = const()[name = string("dense_output_923_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_923_dilations_1 = const()[name = string("dense_output_923_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_923_groups_1 = const()[name = string("dense_output_923_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338117632))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338248768))))[name = string("layers_11_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_923_cast_fp16 = conv(dilations = dense_output_923_dilations_1, groups = dense_output_923_groups_1, pad = dense_output_923_pad_1, pad_type = dense_output_923_pad_type_1, strides = dense_output_923_strides_1, weight = layers_11_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_923_cast_fp16")];
            string sparse_output_923_pad_type_1 = const()[name = string("sparse_output_923_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_923_strides_1 = const()[name = string("sparse_output_923_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_923_pad_1 = const()[name = string("sparse_output_923_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_923_dilations_1 = const()[name = string("sparse_output_923_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_923_groups_1 = const()[name = string("sparse_output_923_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338252032))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338249344))))[name = string("layers_11_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_923_cast_fp16 = conv(dilations = sparse_output_923_dilations_1, groups = sparse_output_923_groups_1, pad = sparse_output_923_pad_1, pad_type = sparse_output_923_pad_type_1, strides = sparse_output_923_strides_1, weight = layers_11_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_923_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14710_cast_fp16 = add(x = dense_output_923_cast_fp16, y = sparse_output_923_cast_fp16)[name = string("op_14710_cast_fp16")];
            tensor<int32, [4]> var_14711 = const()[name = string("op_14711"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14713 = const()[name = string("op_14713"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14712_cast_fp16 = transpose(perm = var_14711, x = var_14710_cast_fp16)[name = string("transpose_516")];
            tensor<fp16, [1, 51, 128]> k_head_365_cast_fp16 = reshape(shape = var_14713, x = var_14712_cast_fp16)[name = string("k_head_365_cast_fp16")];
            string dense_output_925_pad_type_1 = const()[name = string("dense_output_925_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_925_strides_1 = const()[name = string("dense_output_925_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_925_pad_1 = const()[name = string("dense_output_925_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_925_dilations_1 = const()[name = string("dense_output_925_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_925_groups_1 = const()[name = string("dense_output_925_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338268480))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338399616))))[name = string("layers_11_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_925_cast_fp16 = conv(dilations = dense_output_925_dilations_1, groups = dense_output_925_groups_1, pad = dense_output_925_pad_1, pad_type = dense_output_925_pad_type_1, strides = dense_output_925_strides_1, weight = layers_11_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_925_cast_fp16")];
            string sparse_output_925_pad_type_1 = const()[name = string("sparse_output_925_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_925_strides_1 = const()[name = string("sparse_output_925_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_925_pad_1 = const()[name = string("sparse_output_925_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_925_dilations_1 = const()[name = string("sparse_output_925_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_925_groups_1 = const()[name = string("sparse_output_925_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338402880))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338400192))))[name = string("layers_11_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_925_cast_fp16 = conv(dilations = sparse_output_925_dilations_1, groups = sparse_output_925_groups_1, pad = sparse_output_925_pad_1, pad_type = sparse_output_925_pad_type_1, strides = sparse_output_925_strides_1, weight = layers_11_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_925_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14729_cast_fp16 = add(x = dense_output_925_cast_fp16, y = sparse_output_925_cast_fp16)[name = string("op_14729_cast_fp16")];
            tensor<int32, [4]> var_14730 = const()[name = string("op_14730"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14732 = const()[name = string("op_14732"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14731_cast_fp16 = transpose(perm = var_14730, x = var_14729_cast_fp16)[name = string("transpose_515")];
            tensor<fp16, [1, 51, 128]> v_head_365_cast_fp16 = reshape(shape = var_14732, x = var_14731_cast_fp16)[name = string("v_head_365_cast_fp16")];
            string dense_output_927_pad_type_1 = const()[name = string("dense_output_927_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_927_strides_1 = const()[name = string("dense_output_927_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_927_pad_1 = const()[name = string("dense_output_927_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_927_dilations_1 = const()[name = string("dense_output_927_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_927_groups_1 = const()[name = string("dense_output_927_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338419328))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338550464))))[name = string("layers_11_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_927_cast_fp16 = conv(dilations = dense_output_927_dilations_1, groups = dense_output_927_groups_1, pad = dense_output_927_pad_1, pad_type = dense_output_927_pad_type_1, strides = dense_output_927_strides_1, weight = layers_11_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_927_cast_fp16")];
            string sparse_output_927_pad_type_1 = const()[name = string("sparse_output_927_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_927_strides_1 = const()[name = string("sparse_output_927_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_927_pad_1 = const()[name = string("sparse_output_927_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_927_dilations_1 = const()[name = string("sparse_output_927_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_927_groups_1 = const()[name = string("sparse_output_927_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338553728))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338551040))))[name = string("layers_11_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_927_cast_fp16 = conv(dilations = sparse_output_927_dilations_1, groups = sparse_output_927_groups_1, pad = sparse_output_927_pad_1, pad_type = sparse_output_927_pad_type_1, strides = sparse_output_927_strides_1, weight = layers_11_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_927_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_14748_cast_fp16 = add(x = dense_output_927_cast_fp16, y = sparse_output_927_cast_fp16)[name = string("op_14748_cast_fp16")];
            tensor<int32, [4]> var_14749 = const()[name = string("op_14749"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14751 = const()[name = string("op_14751"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_14750_cast_fp16 = transpose(perm = var_14749, x = var_14748_cast_fp16)[name = string("transpose_514")];
            tensor<fp16, [1, 101, 128]> p_head_365_cast_fp16 = reshape(shape = var_14751, x = var_14750_cast_fp16)[name = string("p_head_365_cast_fp16")];
            tensor<fp16, [128]> var_14753_to_fp16 = const()[name = string("op_14753_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338570176)))];
            tensor<fp16, [1, 51, 128]> var_14754_cast_fp16 = add(x = q_head_183_cast_fp16, y = var_14753_to_fp16)[name = string("op_14754_cast_fp16")];
            tensor<int32, [1]> q_u_183_axes_1 = const()[name = string("q_u_183_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_183_cast_fp16 = expand_dims(axes = q_u_183_axes_1, x = var_14754_cast_fp16)[name = string("q_u_183_cast_fp16")];
            tensor<fp16, [128]> var_14756_to_fp16 = const()[name = string("op_14756_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338570496)))];
            tensor<fp16, [1, 51, 128]> var_14757_cast_fp16 = add(x = q_head_183_cast_fp16, y = var_14756_to_fp16)[name = string("op_14757_cast_fp16")];
            tensor<int32, [1]> q_v_183_axes_1 = const()[name = string("q_v_183_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_183_cast_fp16 = expand_dims(axes = q_v_183_axes_1, x = var_14757_cast_fp16)[name = string("q_v_183_cast_fp16")];
            tensor<int32, [1]> k_head_367_axes_1 = const()[name = string("k_head_367_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_367_cast_fp16 = expand_dims(axes = k_head_367_axes_1, x = k_head_365_cast_fp16)[name = string("k_head_367_cast_fp16")];
            tensor<int32, [1]> v_head_367_axes_1 = const()[name = string("v_head_367_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_367_cast_fp16 = expand_dims(axes = v_head_367_axes_1, x = v_head_365_cast_fp16)[name = string("v_head_367_cast_fp16")];
            tensor<int32, [1]> p_head_367_axes_1 = const()[name = string("p_head_367_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_367_cast_fp16 = expand_dims(axes = p_head_367_axes_1, x = p_head_365_cast_fp16)[name = string("p_head_367_cast_fp16")];
            bool var_14763_transpose_x_3 = const()[name = string("op_14763_transpose_x_3"), val = bool(false)];
            bool var_14763_transpose_y_3 = const()[name = string("op_14763_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_14763_cast_fp16 = matmul(transpose_x = var_14763_transpose_x_3, transpose_y = var_14763_transpose_y_3, x = q_u_183_cast_fp16, y = k_head_367_cast_fp16)[name = string("op_14763_cast_fp16")];
            fp16 var_14764_to_fp16 = const()[name = string("op_14764_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_183_cast_fp16 = mul(x = var_14763_cast_fp16, y = var_14764_to_fp16)[name = string("scores_content_183_cast_fp16")];
            bool x_969_transpose_x_3 = const()[name = string("x_969_transpose_x_3"), val = bool(false)];
            bool x_969_transpose_y_3 = const()[name = string("x_969_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_969_cast_fp16 = matmul(transpose_x = x_969_transpose_x_3, transpose_y = x_969_transpose_y_3, x = q_v_183_cast_fp16, y = p_head_367_cast_fp16)[name = string("x_969_cast_fp16")];
            tensor<int32, [8]> x_971_pad_1 = const()[name = string("x_971_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_971_mode_1 = const()[name = string("x_971_mode_1"), val = string("constant")];
            fp16 const_1921_to_fp16 = const()[name = string("const_1921_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_971_cast_fp16 = pad(constant_val = const_1921_to_fp16, mode = x_971_mode_1, pad = x_971_pad_1, x = x_969_cast_fp16)[name = string("x_971_cast_fp16")];
            tensor<int32, [4]> var_14778 = const()[name = string("op_14778"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_973_cast_fp16 = reshape(shape = var_14778, x = x_971_cast_fp16)[name = string("x_973_cast_fp16")];
            tensor<int32, [4]> var_14782_begin_1 = const()[name = string("op_14782_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14782_end_1 = const()[name = string("op_14782_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_14782_end_mask_1 = const()[name = string("op_14782_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_14782_cast_fp16 = slice_by_index(begin = var_14782_begin_1, end = var_14782_end_1, end_mask = var_14782_end_mask_1, x = x_973_cast_fp16)[name = string("op_14782_cast_fp16")];
            tensor<int32, [4]> var_14784 = const()[name = string("op_14784"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_14785_cast_fp16 = reshape(shape = var_14784, x = var_14782_cast_fp16)[name = string("op_14785_cast_fp16")];
            tensor<int32, [4]> var_14790_begin_1 = const()[name = string("op_14790_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14790_end_1 = const()[name = string("op_14790_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_14790_end_mask_1 = const()[name = string("op_14790_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_14790_cast_fp16 = slice_by_index(begin = var_14790_begin_1, end = var_14790_end_1, end_mask = var_14790_end_mask_1, x = var_14785_cast_fp16)[name = string("op_14790_cast_fp16")];
            fp16 var_14791_to_fp16 = const()[name = string("op_14791_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_183_cast_fp16 = mul(x = var_14790_cast_fp16, y = var_14791_to_fp16)[name = string("scores_pos_183_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_183_cast_fp16 = add(x = scores_content_183_cast_fp16, y = scores_pos_183_cast_fp16)[name = string("logits_183_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_14794_cast_fp16 = softmax(axis = var_14156, x = logits_183_cast_fp16)[name = string("op_14794_cast_fp16")];
            bool var_14796_transpose_x_1 = const()[name = string("op_14796_transpose_x_1"), val = bool(false)];
            bool var_14796_transpose_y_1 = const()[name = string("op_14796_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_14796_cast_fp16 = matmul(transpose_x = var_14796_transpose_x_1, transpose_y = var_14796_transpose_y_1, x = var_14794_cast_fp16, y = v_head_367_cast_fp16)[name = string("op_14796_cast_fp16")];
            tensor<int32, [1]> var_14797_axes_1 = const()[name = string("op_14797_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_14797_cast_fp16 = squeeze(axes = var_14797_axes_1, x = var_14796_cast_fp16)[name = string("op_14797_cast_fp16")];
            string dense_output_929_pad_type_1 = const()[name = string("dense_output_929_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_929_strides_1 = const()[name = string("dense_output_929_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_929_pad_1 = const()[name = string("dense_output_929_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_929_dilations_1 = const()[name = string("dense_output_929_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_929_groups_1 = const()[name = string("dense_output_929_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338570816))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338701952))))[name = string("layers_11_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_929_cast_fp16 = conv(dilations = dense_output_929_dilations_1, groups = dense_output_929_groups_1, pad = dense_output_929_pad_1, pad_type = dense_output_929_pad_type_1, strides = dense_output_929_strides_1, weight = layers_11_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_929_cast_fp16")];
            string sparse_output_929_pad_type_1 = const()[name = string("sparse_output_929_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_929_strides_1 = const()[name = string("sparse_output_929_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_929_pad_1 = const()[name = string("sparse_output_929_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_929_dilations_1 = const()[name = string("sparse_output_929_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_929_groups_1 = const()[name = string("sparse_output_929_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338705216))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338702528))))[name = string("layers_11_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_929_cast_fp16 = conv(dilations = sparse_output_929_dilations_1, groups = sparse_output_929_groups_1, pad = sparse_output_929_pad_1, pad_type = sparse_output_929_pad_type_1, strides = sparse_output_929_strides_1, weight = layers_11_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_929_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14812_cast_fp16 = add(x = dense_output_929_cast_fp16, y = sparse_output_929_cast_fp16)[name = string("op_14812_cast_fp16")];
            tensor<int32, [4]> var_14813 = const()[name = string("op_14813"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14815 = const()[name = string("op_14815"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14814_cast_fp16 = transpose(perm = var_14813, x = var_14812_cast_fp16)[name = string("transpose_513")];
            tensor<fp16, [1, 51, 128]> q_head_185_cast_fp16 = reshape(shape = var_14815, x = var_14814_cast_fp16)[name = string("q_head_185_cast_fp16")];
            string dense_output_931_pad_type_1 = const()[name = string("dense_output_931_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_931_strides_1 = const()[name = string("dense_output_931_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_931_pad_1 = const()[name = string("dense_output_931_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_931_dilations_1 = const()[name = string("dense_output_931_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_931_groups_1 = const()[name = string("dense_output_931_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338721664))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338852800))))[name = string("layers_11_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_931_cast_fp16 = conv(dilations = dense_output_931_dilations_1, groups = dense_output_931_groups_1, pad = dense_output_931_pad_1, pad_type = dense_output_931_pad_type_1, strides = dense_output_931_strides_1, weight = layers_11_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_931_cast_fp16")];
            string sparse_output_931_pad_type_1 = const()[name = string("sparse_output_931_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_931_strides_1 = const()[name = string("sparse_output_931_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_931_pad_1 = const()[name = string("sparse_output_931_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_931_dilations_1 = const()[name = string("sparse_output_931_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_931_groups_1 = const()[name = string("sparse_output_931_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338856064))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338853376))))[name = string("layers_11_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_931_cast_fp16 = conv(dilations = sparse_output_931_dilations_1, groups = sparse_output_931_groups_1, pad = sparse_output_931_pad_1, pad_type = sparse_output_931_pad_type_1, strides = sparse_output_931_strides_1, weight = layers_11_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_931_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14831_cast_fp16 = add(x = dense_output_931_cast_fp16, y = sparse_output_931_cast_fp16)[name = string("op_14831_cast_fp16")];
            tensor<int32, [4]> var_14832 = const()[name = string("op_14832"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14834 = const()[name = string("op_14834"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14833_cast_fp16 = transpose(perm = var_14832, x = var_14831_cast_fp16)[name = string("transpose_512")];
            tensor<fp16, [1, 51, 128]> k_head_369_cast_fp16 = reshape(shape = var_14834, x = var_14833_cast_fp16)[name = string("k_head_369_cast_fp16")];
            string dense_output_933_pad_type_1 = const()[name = string("dense_output_933_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_933_strides_1 = const()[name = string("dense_output_933_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_933_pad_1 = const()[name = string("dense_output_933_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_933_dilations_1 = const()[name = string("dense_output_933_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_933_groups_1 = const()[name = string("dense_output_933_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338872512))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339003648))))[name = string("layers_11_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_933_cast_fp16 = conv(dilations = dense_output_933_dilations_1, groups = dense_output_933_groups_1, pad = dense_output_933_pad_1, pad_type = dense_output_933_pad_type_1, strides = dense_output_933_strides_1, weight = layers_11_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_933_cast_fp16")];
            string sparse_output_933_pad_type_1 = const()[name = string("sparse_output_933_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_933_strides_1 = const()[name = string("sparse_output_933_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_933_pad_1 = const()[name = string("sparse_output_933_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_933_dilations_1 = const()[name = string("sparse_output_933_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_933_groups_1 = const()[name = string("sparse_output_933_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339006912))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339004224))))[name = string("layers_11_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_933_cast_fp16 = conv(dilations = sparse_output_933_dilations_1, groups = sparse_output_933_groups_1, pad = sparse_output_933_pad_1, pad_type = sparse_output_933_pad_type_1, strides = sparse_output_933_strides_1, weight = layers_11_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_933_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14850_cast_fp16 = add(x = dense_output_933_cast_fp16, y = sparse_output_933_cast_fp16)[name = string("op_14850_cast_fp16")];
            tensor<int32, [4]> var_14851 = const()[name = string("op_14851"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14853 = const()[name = string("op_14853"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14852_cast_fp16 = transpose(perm = var_14851, x = var_14850_cast_fp16)[name = string("transpose_511")];
            tensor<fp16, [1, 51, 128]> v_head_369_cast_fp16 = reshape(shape = var_14853, x = var_14852_cast_fp16)[name = string("v_head_369_cast_fp16")];
            string dense_output_935_pad_type_1 = const()[name = string("dense_output_935_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_935_strides_1 = const()[name = string("dense_output_935_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_935_pad_1 = const()[name = string("dense_output_935_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_935_dilations_1 = const()[name = string("dense_output_935_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_935_groups_1 = const()[name = string("dense_output_935_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339023360))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339154496))))[name = string("layers_11_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_935_cast_fp16 = conv(dilations = dense_output_935_dilations_1, groups = dense_output_935_groups_1, pad = dense_output_935_pad_1, pad_type = dense_output_935_pad_type_1, strides = dense_output_935_strides_1, weight = layers_11_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_935_cast_fp16")];
            string sparse_output_935_pad_type_1 = const()[name = string("sparse_output_935_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_935_strides_1 = const()[name = string("sparse_output_935_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_935_pad_1 = const()[name = string("sparse_output_935_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_935_dilations_1 = const()[name = string("sparse_output_935_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_935_groups_1 = const()[name = string("sparse_output_935_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339157760))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339155072))))[name = string("layers_11_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_935_cast_fp16 = conv(dilations = sparse_output_935_dilations_1, groups = sparse_output_935_groups_1, pad = sparse_output_935_pad_1, pad_type = sparse_output_935_pad_type_1, strides = sparse_output_935_strides_1, weight = layers_11_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_935_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_14869_cast_fp16 = add(x = dense_output_935_cast_fp16, y = sparse_output_935_cast_fp16)[name = string("op_14869_cast_fp16")];
            tensor<int32, [4]> var_14870 = const()[name = string("op_14870"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14872 = const()[name = string("op_14872"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_14871_cast_fp16 = transpose(perm = var_14870, x = var_14869_cast_fp16)[name = string("transpose_510")];
            tensor<fp16, [1, 101, 128]> p_head_369_cast_fp16 = reshape(shape = var_14872, x = var_14871_cast_fp16)[name = string("p_head_369_cast_fp16")];
            tensor<fp16, [128]> var_14874_to_fp16 = const()[name = string("op_14874_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339174208)))];
            tensor<fp16, [1, 51, 128]> var_14875_cast_fp16 = add(x = q_head_185_cast_fp16, y = var_14874_to_fp16)[name = string("op_14875_cast_fp16")];
            tensor<int32, [1]> q_u_185_axes_1 = const()[name = string("q_u_185_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_185_cast_fp16 = expand_dims(axes = q_u_185_axes_1, x = var_14875_cast_fp16)[name = string("q_u_185_cast_fp16")];
            tensor<fp16, [128]> var_14877_to_fp16 = const()[name = string("op_14877_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339174528)))];
            tensor<fp16, [1, 51, 128]> var_14878_cast_fp16 = add(x = q_head_185_cast_fp16, y = var_14877_to_fp16)[name = string("op_14878_cast_fp16")];
            tensor<int32, [1]> q_v_185_axes_1 = const()[name = string("q_v_185_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_185_cast_fp16 = expand_dims(axes = q_v_185_axes_1, x = var_14878_cast_fp16)[name = string("q_v_185_cast_fp16")];
            tensor<int32, [1]> k_head_371_axes_1 = const()[name = string("k_head_371_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_371_cast_fp16 = expand_dims(axes = k_head_371_axes_1, x = k_head_369_cast_fp16)[name = string("k_head_371_cast_fp16")];
            tensor<int32, [1]> v_head_371_axes_1 = const()[name = string("v_head_371_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_371_cast_fp16 = expand_dims(axes = v_head_371_axes_1, x = v_head_369_cast_fp16)[name = string("v_head_371_cast_fp16")];
            tensor<int32, [1]> p_head_371_axes_1 = const()[name = string("p_head_371_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_371_cast_fp16 = expand_dims(axes = p_head_371_axes_1, x = p_head_369_cast_fp16)[name = string("p_head_371_cast_fp16")];
            bool var_14884_transpose_x_3 = const()[name = string("op_14884_transpose_x_3"), val = bool(false)];
            bool var_14884_transpose_y_3 = const()[name = string("op_14884_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_14884_cast_fp16 = matmul(transpose_x = var_14884_transpose_x_3, transpose_y = var_14884_transpose_y_3, x = q_u_185_cast_fp16, y = k_head_371_cast_fp16)[name = string("op_14884_cast_fp16")];
            fp16 var_14885_to_fp16 = const()[name = string("op_14885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_185_cast_fp16 = mul(x = var_14884_cast_fp16, y = var_14885_to_fp16)[name = string("scores_content_185_cast_fp16")];
            bool x_977_transpose_x_3 = const()[name = string("x_977_transpose_x_3"), val = bool(false)];
            bool x_977_transpose_y_3 = const()[name = string("x_977_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_977_cast_fp16 = matmul(transpose_x = x_977_transpose_x_3, transpose_y = x_977_transpose_y_3, x = q_v_185_cast_fp16, y = p_head_371_cast_fp16)[name = string("x_977_cast_fp16")];
            tensor<int32, [8]> x_979_pad_1 = const()[name = string("x_979_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_979_mode_1 = const()[name = string("x_979_mode_1"), val = string("constant")];
            fp16 const_1927_to_fp16 = const()[name = string("const_1927_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_979_cast_fp16 = pad(constant_val = const_1927_to_fp16, mode = x_979_mode_1, pad = x_979_pad_1, x = x_977_cast_fp16)[name = string("x_979_cast_fp16")];
            tensor<int32, [4]> var_14899 = const()[name = string("op_14899"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_981_cast_fp16 = reshape(shape = var_14899, x = x_979_cast_fp16)[name = string("x_981_cast_fp16")];
            tensor<int32, [4]> var_14903_begin_1 = const()[name = string("op_14903_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_14903_end_1 = const()[name = string("op_14903_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_14903_end_mask_1 = const()[name = string("op_14903_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_14903_cast_fp16 = slice_by_index(begin = var_14903_begin_1, end = var_14903_end_1, end_mask = var_14903_end_mask_1, x = x_981_cast_fp16)[name = string("op_14903_cast_fp16")];
            tensor<int32, [4]> var_14905 = const()[name = string("op_14905"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_14906_cast_fp16 = reshape(shape = var_14905, x = var_14903_cast_fp16)[name = string("op_14906_cast_fp16")];
            tensor<int32, [4]> var_14911_begin_1 = const()[name = string("op_14911_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_14911_end_1 = const()[name = string("op_14911_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_14911_end_mask_1 = const()[name = string("op_14911_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_14911_cast_fp16 = slice_by_index(begin = var_14911_begin_1, end = var_14911_end_1, end_mask = var_14911_end_mask_1, x = var_14906_cast_fp16)[name = string("op_14911_cast_fp16")];
            fp16 var_14912_to_fp16 = const()[name = string("op_14912_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_185_cast_fp16 = mul(x = var_14911_cast_fp16, y = var_14912_to_fp16)[name = string("scores_pos_185_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_185_cast_fp16 = add(x = scores_content_185_cast_fp16, y = scores_pos_185_cast_fp16)[name = string("logits_185_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_14915_cast_fp16 = softmax(axis = var_14156, x = logits_185_cast_fp16)[name = string("op_14915_cast_fp16")];
            bool var_14917_transpose_x_1 = const()[name = string("op_14917_transpose_x_1"), val = bool(false)];
            bool var_14917_transpose_y_1 = const()[name = string("op_14917_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_14917_cast_fp16 = matmul(transpose_x = var_14917_transpose_x_1, transpose_y = var_14917_transpose_y_1, x = var_14915_cast_fp16, y = v_head_371_cast_fp16)[name = string("op_14917_cast_fp16")];
            tensor<int32, [1]> var_14918_axes_1 = const()[name = string("op_14918_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_14918_cast_fp16 = squeeze(axes = var_14918_axes_1, x = var_14917_cast_fp16)[name = string("op_14918_cast_fp16")];
            string dense_output_937_pad_type_1 = const()[name = string("dense_output_937_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_937_strides_1 = const()[name = string("dense_output_937_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_937_pad_1 = const()[name = string("dense_output_937_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_937_dilations_1 = const()[name = string("dense_output_937_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_937_groups_1 = const()[name = string("dense_output_937_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339174848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339305984))))[name = string("layers_11_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_937_cast_fp16 = conv(dilations = dense_output_937_dilations_1, groups = dense_output_937_groups_1, pad = dense_output_937_pad_1, pad_type = dense_output_937_pad_type_1, strides = dense_output_937_strides_1, weight = layers_11_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_937_cast_fp16")];
            string sparse_output_937_pad_type_1 = const()[name = string("sparse_output_937_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_937_strides_1 = const()[name = string("sparse_output_937_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_937_pad_1 = const()[name = string("sparse_output_937_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_937_dilations_1 = const()[name = string("sparse_output_937_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_937_groups_1 = const()[name = string("sparse_output_937_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339309248))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339306560))))[name = string("layers_11_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_937_cast_fp16 = conv(dilations = sparse_output_937_dilations_1, groups = sparse_output_937_groups_1, pad = sparse_output_937_pad_1, pad_type = sparse_output_937_pad_type_1, strides = sparse_output_937_strides_1, weight = layers_11_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_937_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14933_cast_fp16 = add(x = dense_output_937_cast_fp16, y = sparse_output_937_cast_fp16)[name = string("op_14933_cast_fp16")];
            tensor<int32, [4]> var_14934 = const()[name = string("op_14934"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14936 = const()[name = string("op_14936"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14935_cast_fp16 = transpose(perm = var_14934, x = var_14933_cast_fp16)[name = string("transpose_509")];
            tensor<fp16, [1, 51, 128]> q_head_187_cast_fp16 = reshape(shape = var_14936, x = var_14935_cast_fp16)[name = string("q_head_187_cast_fp16")];
            string dense_output_939_pad_type_1 = const()[name = string("dense_output_939_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_939_strides_1 = const()[name = string("dense_output_939_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_939_pad_1 = const()[name = string("dense_output_939_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_939_dilations_1 = const()[name = string("dense_output_939_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_939_groups_1 = const()[name = string("dense_output_939_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339325696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339456832))))[name = string("layers_11_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_939_cast_fp16 = conv(dilations = dense_output_939_dilations_1, groups = dense_output_939_groups_1, pad = dense_output_939_pad_1, pad_type = dense_output_939_pad_type_1, strides = dense_output_939_strides_1, weight = layers_11_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_939_cast_fp16")];
            string sparse_output_939_pad_type_1 = const()[name = string("sparse_output_939_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_939_strides_1 = const()[name = string("sparse_output_939_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_939_pad_1 = const()[name = string("sparse_output_939_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_939_dilations_1 = const()[name = string("sparse_output_939_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_939_groups_1 = const()[name = string("sparse_output_939_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339460096))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339457408))))[name = string("layers_11_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_939_cast_fp16 = conv(dilations = sparse_output_939_dilations_1, groups = sparse_output_939_groups_1, pad = sparse_output_939_pad_1, pad_type = sparse_output_939_pad_type_1, strides = sparse_output_939_strides_1, weight = layers_11_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_939_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14952_cast_fp16 = add(x = dense_output_939_cast_fp16, y = sparse_output_939_cast_fp16)[name = string("op_14952_cast_fp16")];
            tensor<int32, [4]> var_14953 = const()[name = string("op_14953"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14955 = const()[name = string("op_14955"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14954_cast_fp16 = transpose(perm = var_14953, x = var_14952_cast_fp16)[name = string("transpose_508")];
            tensor<fp16, [1, 51, 128]> k_head_373_cast_fp16 = reshape(shape = var_14955, x = var_14954_cast_fp16)[name = string("k_head_373_cast_fp16")];
            string dense_output_941_pad_type_1 = const()[name = string("dense_output_941_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_941_strides_1 = const()[name = string("dense_output_941_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_941_pad_1 = const()[name = string("dense_output_941_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_941_dilations_1 = const()[name = string("dense_output_941_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_941_groups_1 = const()[name = string("dense_output_941_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339476544))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339607680))))[name = string("layers_11_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_941_cast_fp16 = conv(dilations = dense_output_941_dilations_1, groups = dense_output_941_groups_1, pad = dense_output_941_pad_1, pad_type = dense_output_941_pad_type_1, strides = dense_output_941_strides_1, weight = layers_11_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_941_cast_fp16")];
            string sparse_output_941_pad_type_1 = const()[name = string("sparse_output_941_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_941_strides_1 = const()[name = string("sparse_output_941_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_941_pad_1 = const()[name = string("sparse_output_941_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_941_dilations_1 = const()[name = string("sparse_output_941_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_941_groups_1 = const()[name = string("sparse_output_941_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339610944))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339608256))))[name = string("layers_11_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_941_cast_fp16 = conv(dilations = sparse_output_941_dilations_1, groups = sparse_output_941_groups_1, pad = sparse_output_941_pad_1, pad_type = sparse_output_941_pad_type_1, strides = sparse_output_941_strides_1, weight = layers_11_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_941_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_14971_cast_fp16 = add(x = dense_output_941_cast_fp16, y = sparse_output_941_cast_fp16)[name = string("op_14971_cast_fp16")];
            tensor<int32, [4]> var_14972 = const()[name = string("op_14972"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14974 = const()[name = string("op_14974"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_14973_cast_fp16 = transpose(perm = var_14972, x = var_14971_cast_fp16)[name = string("transpose_507")];
            tensor<fp16, [1, 51, 128]> v_head_373_cast_fp16 = reshape(shape = var_14974, x = var_14973_cast_fp16)[name = string("v_head_373_cast_fp16")];
            string dense_output_943_pad_type_1 = const()[name = string("dense_output_943_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_943_strides_1 = const()[name = string("dense_output_943_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_943_pad_1 = const()[name = string("dense_output_943_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_943_dilations_1 = const()[name = string("dense_output_943_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_943_groups_1 = const()[name = string("dense_output_943_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339627392))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339758528))))[name = string("layers_11_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_943_cast_fp16 = conv(dilations = dense_output_943_dilations_1, groups = dense_output_943_groups_1, pad = dense_output_943_pad_1, pad_type = dense_output_943_pad_type_1, strides = dense_output_943_strides_1, weight = layers_11_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_943_cast_fp16")];
            string sparse_output_943_pad_type_1 = const()[name = string("sparse_output_943_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_943_strides_1 = const()[name = string("sparse_output_943_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_943_pad_1 = const()[name = string("sparse_output_943_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_943_dilations_1 = const()[name = string("sparse_output_943_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_943_groups_1 = const()[name = string("sparse_output_943_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339761792))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339759104))))[name = string("layers_11_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_943_cast_fp16 = conv(dilations = sparse_output_943_dilations_1, groups = sparse_output_943_groups_1, pad = sparse_output_943_pad_1, pad_type = sparse_output_943_pad_type_1, strides = sparse_output_943_strides_1, weight = layers_11_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_943_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_14990_cast_fp16 = add(x = dense_output_943_cast_fp16, y = sparse_output_943_cast_fp16)[name = string("op_14990_cast_fp16")];
            tensor<int32, [4]> var_14991 = const()[name = string("op_14991"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_14993 = const()[name = string("op_14993"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_14992_cast_fp16 = transpose(perm = var_14991, x = var_14990_cast_fp16)[name = string("transpose_506")];
            tensor<fp16, [1, 101, 128]> p_head_373_cast_fp16 = reshape(shape = var_14993, x = var_14992_cast_fp16)[name = string("p_head_373_cast_fp16")];
            tensor<fp16, [128]> var_14995_to_fp16 = const()[name = string("op_14995_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339778240)))];
            tensor<fp16, [1, 51, 128]> var_14996_cast_fp16 = add(x = q_head_187_cast_fp16, y = var_14995_to_fp16)[name = string("op_14996_cast_fp16")];
            tensor<int32, [1]> q_u_187_axes_1 = const()[name = string("q_u_187_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_187_cast_fp16 = expand_dims(axes = q_u_187_axes_1, x = var_14996_cast_fp16)[name = string("q_u_187_cast_fp16")];
            tensor<fp16, [128]> var_14998_to_fp16 = const()[name = string("op_14998_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339778560)))];
            tensor<fp16, [1, 51, 128]> var_14999_cast_fp16 = add(x = q_head_187_cast_fp16, y = var_14998_to_fp16)[name = string("op_14999_cast_fp16")];
            tensor<int32, [1]> q_v_187_axes_1 = const()[name = string("q_v_187_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_187_cast_fp16 = expand_dims(axes = q_v_187_axes_1, x = var_14999_cast_fp16)[name = string("q_v_187_cast_fp16")];
            tensor<int32, [1]> k_head_375_axes_1 = const()[name = string("k_head_375_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_375_cast_fp16 = expand_dims(axes = k_head_375_axes_1, x = k_head_373_cast_fp16)[name = string("k_head_375_cast_fp16")];
            tensor<int32, [1]> v_head_375_axes_1 = const()[name = string("v_head_375_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_375_cast_fp16 = expand_dims(axes = v_head_375_axes_1, x = v_head_373_cast_fp16)[name = string("v_head_375_cast_fp16")];
            tensor<int32, [1]> p_head_375_axes_1 = const()[name = string("p_head_375_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_375_cast_fp16 = expand_dims(axes = p_head_375_axes_1, x = p_head_373_cast_fp16)[name = string("p_head_375_cast_fp16")];
            bool var_15005_transpose_x_3 = const()[name = string("op_15005_transpose_x_3"), val = bool(false)];
            bool var_15005_transpose_y_3 = const()[name = string("op_15005_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_15005_cast_fp16 = matmul(transpose_x = var_15005_transpose_x_3, transpose_y = var_15005_transpose_y_3, x = q_u_187_cast_fp16, y = k_head_375_cast_fp16)[name = string("op_15005_cast_fp16")];
            fp16 var_15006_to_fp16 = const()[name = string("op_15006_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_187_cast_fp16 = mul(x = var_15005_cast_fp16, y = var_15006_to_fp16)[name = string("scores_content_187_cast_fp16")];
            bool x_985_transpose_x_3 = const()[name = string("x_985_transpose_x_3"), val = bool(false)];
            bool x_985_transpose_y_3 = const()[name = string("x_985_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_985_cast_fp16 = matmul(transpose_x = x_985_transpose_x_3, transpose_y = x_985_transpose_y_3, x = q_v_187_cast_fp16, y = p_head_375_cast_fp16)[name = string("x_985_cast_fp16")];
            tensor<int32, [8]> x_987_pad_1 = const()[name = string("x_987_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_987_mode_1 = const()[name = string("x_987_mode_1"), val = string("constant")];
            fp16 const_1933_to_fp16 = const()[name = string("const_1933_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_987_cast_fp16 = pad(constant_val = const_1933_to_fp16, mode = x_987_mode_1, pad = x_987_pad_1, x = x_985_cast_fp16)[name = string("x_987_cast_fp16")];
            tensor<int32, [4]> var_15020 = const()[name = string("op_15020"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_989_cast_fp16 = reshape(shape = var_15020, x = x_987_cast_fp16)[name = string("x_989_cast_fp16")];
            tensor<int32, [4]> var_15024_begin_1 = const()[name = string("op_15024_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15024_end_1 = const()[name = string("op_15024_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_15024_end_mask_1 = const()[name = string("op_15024_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_15024_cast_fp16 = slice_by_index(begin = var_15024_begin_1, end = var_15024_end_1, end_mask = var_15024_end_mask_1, x = x_989_cast_fp16)[name = string("op_15024_cast_fp16")];
            tensor<int32, [4]> var_15026 = const()[name = string("op_15026"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_15027_cast_fp16 = reshape(shape = var_15026, x = var_15024_cast_fp16)[name = string("op_15027_cast_fp16")];
            tensor<int32, [4]> var_15032_begin_1 = const()[name = string("op_15032_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15032_end_1 = const()[name = string("op_15032_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_15032_end_mask_1 = const()[name = string("op_15032_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_15032_cast_fp16 = slice_by_index(begin = var_15032_begin_1, end = var_15032_end_1, end_mask = var_15032_end_mask_1, x = var_15027_cast_fp16)[name = string("op_15032_cast_fp16")];
            fp16 var_15033_to_fp16 = const()[name = string("op_15033_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_187_cast_fp16 = mul(x = var_15032_cast_fp16, y = var_15033_to_fp16)[name = string("scores_pos_187_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_187_cast_fp16 = add(x = scores_content_187_cast_fp16, y = scores_pos_187_cast_fp16)[name = string("logits_187_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_15036_cast_fp16 = softmax(axis = var_14156, x = logits_187_cast_fp16)[name = string("op_15036_cast_fp16")];
            bool var_15038_transpose_x_1 = const()[name = string("op_15038_transpose_x_1"), val = bool(false)];
            bool var_15038_transpose_y_1 = const()[name = string("op_15038_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_15038_cast_fp16 = matmul(transpose_x = var_15038_transpose_x_1, transpose_y = var_15038_transpose_y_1, x = var_15036_cast_fp16, y = v_head_375_cast_fp16)[name = string("op_15038_cast_fp16")];
            tensor<int32, [1]> var_15039_axes_1 = const()[name = string("op_15039_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_15039_cast_fp16 = squeeze(axes = var_15039_axes_1, x = var_15038_cast_fp16)[name = string("op_15039_cast_fp16")];
            string dense_output_945_pad_type_1 = const()[name = string("dense_output_945_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_945_strides_1 = const()[name = string("dense_output_945_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_945_pad_1 = const()[name = string("dense_output_945_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_945_dilations_1 = const()[name = string("dense_output_945_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_945_groups_1 = const()[name = string("dense_output_945_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339778880))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339910016))))[name = string("layers_11_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_945_cast_fp16 = conv(dilations = dense_output_945_dilations_1, groups = dense_output_945_groups_1, pad = dense_output_945_pad_1, pad_type = dense_output_945_pad_type_1, strides = dense_output_945_strides_1, weight = layers_11_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_945_cast_fp16")];
            string sparse_output_945_pad_type_1 = const()[name = string("sparse_output_945_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_945_strides_1 = const()[name = string("sparse_output_945_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_945_pad_1 = const()[name = string("sparse_output_945_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_945_dilations_1 = const()[name = string("sparse_output_945_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_945_groups_1 = const()[name = string("sparse_output_945_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339913280))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339910592))))[name = string("layers_11_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_945_cast_fp16 = conv(dilations = sparse_output_945_dilations_1, groups = sparse_output_945_groups_1, pad = sparse_output_945_pad_1, pad_type = sparse_output_945_pad_type_1, strides = sparse_output_945_strides_1, weight = layers_11_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_945_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15054_cast_fp16 = add(x = dense_output_945_cast_fp16, y = sparse_output_945_cast_fp16)[name = string("op_15054_cast_fp16")];
            tensor<int32, [4]> var_15055 = const()[name = string("op_15055"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15057 = const()[name = string("op_15057"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15056_cast_fp16 = transpose(perm = var_15055, x = var_15054_cast_fp16)[name = string("transpose_505")];
            tensor<fp16, [1, 51, 128]> q_head_189_cast_fp16 = reshape(shape = var_15057, x = var_15056_cast_fp16)[name = string("q_head_189_cast_fp16")];
            string dense_output_947_pad_type_1 = const()[name = string("dense_output_947_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_947_strides_1 = const()[name = string("dense_output_947_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_947_pad_1 = const()[name = string("dense_output_947_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_947_dilations_1 = const()[name = string("dense_output_947_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_947_groups_1 = const()[name = string("dense_output_947_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339929728))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340060864))))[name = string("layers_11_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_947_cast_fp16 = conv(dilations = dense_output_947_dilations_1, groups = dense_output_947_groups_1, pad = dense_output_947_pad_1, pad_type = dense_output_947_pad_type_1, strides = dense_output_947_strides_1, weight = layers_11_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_947_cast_fp16")];
            string sparse_output_947_pad_type_1 = const()[name = string("sparse_output_947_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_947_strides_1 = const()[name = string("sparse_output_947_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_947_pad_1 = const()[name = string("sparse_output_947_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_947_dilations_1 = const()[name = string("sparse_output_947_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_947_groups_1 = const()[name = string("sparse_output_947_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340064128))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340061440))))[name = string("layers_11_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_947_cast_fp16 = conv(dilations = sparse_output_947_dilations_1, groups = sparse_output_947_groups_1, pad = sparse_output_947_pad_1, pad_type = sparse_output_947_pad_type_1, strides = sparse_output_947_strides_1, weight = layers_11_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_947_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15073_cast_fp16 = add(x = dense_output_947_cast_fp16, y = sparse_output_947_cast_fp16)[name = string("op_15073_cast_fp16")];
            tensor<int32, [4]> var_15074 = const()[name = string("op_15074"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15076 = const()[name = string("op_15076"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15075_cast_fp16 = transpose(perm = var_15074, x = var_15073_cast_fp16)[name = string("transpose_504")];
            tensor<fp16, [1, 51, 128]> k_head_377_cast_fp16 = reshape(shape = var_15076, x = var_15075_cast_fp16)[name = string("k_head_377_cast_fp16")];
            string dense_output_949_pad_type_1 = const()[name = string("dense_output_949_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_949_strides_1 = const()[name = string("dense_output_949_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_949_pad_1 = const()[name = string("dense_output_949_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_949_dilations_1 = const()[name = string("dense_output_949_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_949_groups_1 = const()[name = string("dense_output_949_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340080576))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340211712))))[name = string("layers_11_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_949_cast_fp16 = conv(dilations = dense_output_949_dilations_1, groups = dense_output_949_groups_1, pad = dense_output_949_pad_1, pad_type = dense_output_949_pad_type_1, strides = dense_output_949_strides_1, weight = layers_11_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_949_cast_fp16")];
            string sparse_output_949_pad_type_1 = const()[name = string("sparse_output_949_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_949_strides_1 = const()[name = string("sparse_output_949_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_949_pad_1 = const()[name = string("sparse_output_949_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_949_dilations_1 = const()[name = string("sparse_output_949_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_949_groups_1 = const()[name = string("sparse_output_949_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340214976))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340212288))))[name = string("layers_11_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_949_cast_fp16 = conv(dilations = sparse_output_949_dilations_1, groups = sparse_output_949_groups_1, pad = sparse_output_949_pad_1, pad_type = sparse_output_949_pad_type_1, strides = sparse_output_949_strides_1, weight = layers_11_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_949_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15092_cast_fp16 = add(x = dense_output_949_cast_fp16, y = sparse_output_949_cast_fp16)[name = string("op_15092_cast_fp16")];
            tensor<int32, [4]> var_15093 = const()[name = string("op_15093"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15095 = const()[name = string("op_15095"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15094_cast_fp16 = transpose(perm = var_15093, x = var_15092_cast_fp16)[name = string("transpose_503")];
            tensor<fp16, [1, 51, 128]> v_head_377_cast_fp16 = reshape(shape = var_15095, x = var_15094_cast_fp16)[name = string("v_head_377_cast_fp16")];
            string dense_output_951_pad_type_1 = const()[name = string("dense_output_951_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_951_strides_1 = const()[name = string("dense_output_951_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_951_pad_1 = const()[name = string("dense_output_951_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_951_dilations_1 = const()[name = string("dense_output_951_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_951_groups_1 = const()[name = string("dense_output_951_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340231424))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340362560))))[name = string("layers_11_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_951_cast_fp16 = conv(dilations = dense_output_951_dilations_1, groups = dense_output_951_groups_1, pad = dense_output_951_pad_1, pad_type = dense_output_951_pad_type_1, strides = dense_output_951_strides_1, weight = layers_11_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_951_cast_fp16")];
            string sparse_output_951_pad_type_1 = const()[name = string("sparse_output_951_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_951_strides_1 = const()[name = string("sparse_output_951_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_951_pad_1 = const()[name = string("sparse_output_951_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_951_dilations_1 = const()[name = string("sparse_output_951_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_951_groups_1 = const()[name = string("sparse_output_951_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340365824))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340363136))))[name = string("layers_11_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_951_cast_fp16 = conv(dilations = sparse_output_951_dilations_1, groups = sparse_output_951_groups_1, pad = sparse_output_951_pad_1, pad_type = sparse_output_951_pad_type_1, strides = sparse_output_951_strides_1, weight = layers_11_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_951_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_15111_cast_fp16 = add(x = dense_output_951_cast_fp16, y = sparse_output_951_cast_fp16)[name = string("op_15111_cast_fp16")];
            tensor<int32, [4]> var_15112 = const()[name = string("op_15112"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15114 = const()[name = string("op_15114"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_15113_cast_fp16 = transpose(perm = var_15112, x = var_15111_cast_fp16)[name = string("transpose_502")];
            tensor<fp16, [1, 101, 128]> p_head_377_cast_fp16 = reshape(shape = var_15114, x = var_15113_cast_fp16)[name = string("p_head_377_cast_fp16")];
            tensor<fp16, [128]> var_15116_to_fp16 = const()[name = string("op_15116_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340382272)))];
            tensor<fp16, [1, 51, 128]> var_15117_cast_fp16 = add(x = q_head_189_cast_fp16, y = var_15116_to_fp16)[name = string("op_15117_cast_fp16")];
            tensor<int32, [1]> q_u_189_axes_1 = const()[name = string("q_u_189_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_189_cast_fp16 = expand_dims(axes = q_u_189_axes_1, x = var_15117_cast_fp16)[name = string("q_u_189_cast_fp16")];
            tensor<fp16, [128]> var_15119_to_fp16 = const()[name = string("op_15119_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340382592)))];
            tensor<fp16, [1, 51, 128]> var_15120_cast_fp16 = add(x = q_head_189_cast_fp16, y = var_15119_to_fp16)[name = string("op_15120_cast_fp16")];
            tensor<int32, [1]> q_v_189_axes_1 = const()[name = string("q_v_189_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_189_cast_fp16 = expand_dims(axes = q_v_189_axes_1, x = var_15120_cast_fp16)[name = string("q_v_189_cast_fp16")];
            tensor<int32, [1]> k_head_379_axes_1 = const()[name = string("k_head_379_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_379_cast_fp16 = expand_dims(axes = k_head_379_axes_1, x = k_head_377_cast_fp16)[name = string("k_head_379_cast_fp16")];
            tensor<int32, [1]> v_head_379_axes_1 = const()[name = string("v_head_379_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_379_cast_fp16 = expand_dims(axes = v_head_379_axes_1, x = v_head_377_cast_fp16)[name = string("v_head_379_cast_fp16")];
            tensor<int32, [1]> p_head_379_axes_1 = const()[name = string("p_head_379_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_379_cast_fp16 = expand_dims(axes = p_head_379_axes_1, x = p_head_377_cast_fp16)[name = string("p_head_379_cast_fp16")];
            bool var_15126_transpose_x_3 = const()[name = string("op_15126_transpose_x_3"), val = bool(false)];
            bool var_15126_transpose_y_3 = const()[name = string("op_15126_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_15126_cast_fp16 = matmul(transpose_x = var_15126_transpose_x_3, transpose_y = var_15126_transpose_y_3, x = q_u_189_cast_fp16, y = k_head_379_cast_fp16)[name = string("op_15126_cast_fp16")];
            fp16 var_15127_to_fp16 = const()[name = string("op_15127_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_189_cast_fp16 = mul(x = var_15126_cast_fp16, y = var_15127_to_fp16)[name = string("scores_content_189_cast_fp16")];
            bool x_993_transpose_x_3 = const()[name = string("x_993_transpose_x_3"), val = bool(false)];
            bool x_993_transpose_y_3 = const()[name = string("x_993_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_993_cast_fp16 = matmul(transpose_x = x_993_transpose_x_3, transpose_y = x_993_transpose_y_3, x = q_v_189_cast_fp16, y = p_head_379_cast_fp16)[name = string("x_993_cast_fp16")];
            tensor<int32, [8]> x_995_pad_1 = const()[name = string("x_995_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_995_mode_1 = const()[name = string("x_995_mode_1"), val = string("constant")];
            fp16 const_1939_to_fp16 = const()[name = string("const_1939_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_995_cast_fp16 = pad(constant_val = const_1939_to_fp16, mode = x_995_mode_1, pad = x_995_pad_1, x = x_993_cast_fp16)[name = string("x_995_cast_fp16")];
            tensor<int32, [4]> var_15141 = const()[name = string("op_15141"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_997_cast_fp16 = reshape(shape = var_15141, x = x_995_cast_fp16)[name = string("x_997_cast_fp16")];
            tensor<int32, [4]> var_15145_begin_1 = const()[name = string("op_15145_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15145_end_1 = const()[name = string("op_15145_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_15145_end_mask_1 = const()[name = string("op_15145_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_15145_cast_fp16 = slice_by_index(begin = var_15145_begin_1, end = var_15145_end_1, end_mask = var_15145_end_mask_1, x = x_997_cast_fp16)[name = string("op_15145_cast_fp16")];
            tensor<int32, [4]> var_15147 = const()[name = string("op_15147"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_15148_cast_fp16 = reshape(shape = var_15147, x = var_15145_cast_fp16)[name = string("op_15148_cast_fp16")];
            tensor<int32, [4]> var_15153_begin_1 = const()[name = string("op_15153_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15153_end_1 = const()[name = string("op_15153_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_15153_end_mask_1 = const()[name = string("op_15153_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_15153_cast_fp16 = slice_by_index(begin = var_15153_begin_1, end = var_15153_end_1, end_mask = var_15153_end_mask_1, x = var_15148_cast_fp16)[name = string("op_15153_cast_fp16")];
            fp16 var_15154_to_fp16 = const()[name = string("op_15154_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_189_cast_fp16 = mul(x = var_15153_cast_fp16, y = var_15154_to_fp16)[name = string("scores_pos_189_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_189_cast_fp16 = add(x = scores_content_189_cast_fp16, y = scores_pos_189_cast_fp16)[name = string("logits_189_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_15157_cast_fp16 = softmax(axis = var_14156, x = logits_189_cast_fp16)[name = string("op_15157_cast_fp16")];
            bool var_15159_transpose_x_1 = const()[name = string("op_15159_transpose_x_1"), val = bool(false)];
            bool var_15159_transpose_y_1 = const()[name = string("op_15159_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_15159_cast_fp16 = matmul(transpose_x = var_15159_transpose_x_1, transpose_y = var_15159_transpose_y_1, x = var_15157_cast_fp16, y = v_head_379_cast_fp16)[name = string("op_15159_cast_fp16")];
            tensor<int32, [1]> var_15160_axes_1 = const()[name = string("op_15160_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_15160_cast_fp16 = squeeze(axes = var_15160_axes_1, x = var_15159_cast_fp16)[name = string("op_15160_cast_fp16")];
            string dense_output_953_pad_type_1 = const()[name = string("dense_output_953_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_953_strides_1 = const()[name = string("dense_output_953_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_953_pad_1 = const()[name = string("dense_output_953_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_953_dilations_1 = const()[name = string("dense_output_953_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_953_groups_1 = const()[name = string("dense_output_953_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340382912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340514048))))[name = string("layers_11_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_953_cast_fp16 = conv(dilations = dense_output_953_dilations_1, groups = dense_output_953_groups_1, pad = dense_output_953_pad_1, pad_type = dense_output_953_pad_type_1, strides = dense_output_953_strides_1, weight = layers_11_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_953_cast_fp16")];
            string sparse_output_953_pad_type_1 = const()[name = string("sparse_output_953_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_953_strides_1 = const()[name = string("sparse_output_953_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_953_pad_1 = const()[name = string("sparse_output_953_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_953_dilations_1 = const()[name = string("sparse_output_953_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_953_groups_1 = const()[name = string("sparse_output_953_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340517312))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340514624))))[name = string("layers_11_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_953_cast_fp16 = conv(dilations = sparse_output_953_dilations_1, groups = sparse_output_953_groups_1, pad = sparse_output_953_pad_1, pad_type = sparse_output_953_pad_type_1, strides = sparse_output_953_strides_1, weight = layers_11_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_953_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15175_cast_fp16 = add(x = dense_output_953_cast_fp16, y = sparse_output_953_cast_fp16)[name = string("op_15175_cast_fp16")];
            tensor<int32, [4]> var_15176 = const()[name = string("op_15176"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15178 = const()[name = string("op_15178"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15177_cast_fp16 = transpose(perm = var_15176, x = var_15175_cast_fp16)[name = string("transpose_501")];
            tensor<fp16, [1, 51, 128]> q_head_191_cast_fp16 = reshape(shape = var_15178, x = var_15177_cast_fp16)[name = string("q_head_191_cast_fp16")];
            string dense_output_955_pad_type_1 = const()[name = string("dense_output_955_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_955_strides_1 = const()[name = string("dense_output_955_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_955_pad_1 = const()[name = string("dense_output_955_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_955_dilations_1 = const()[name = string("dense_output_955_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_955_groups_1 = const()[name = string("dense_output_955_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340533760))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340664896))))[name = string("layers_11_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_955_cast_fp16 = conv(dilations = dense_output_955_dilations_1, groups = dense_output_955_groups_1, pad = dense_output_955_pad_1, pad_type = dense_output_955_pad_type_1, strides = dense_output_955_strides_1, weight = layers_11_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_955_cast_fp16")];
            string sparse_output_955_pad_type_1 = const()[name = string("sparse_output_955_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_955_strides_1 = const()[name = string("sparse_output_955_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_955_pad_1 = const()[name = string("sparse_output_955_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_955_dilations_1 = const()[name = string("sparse_output_955_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_955_groups_1 = const()[name = string("sparse_output_955_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340668160))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340665472))))[name = string("layers_11_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_955_cast_fp16 = conv(dilations = sparse_output_955_dilations_1, groups = sparse_output_955_groups_1, pad = sparse_output_955_pad_1, pad_type = sparse_output_955_pad_type_1, strides = sparse_output_955_strides_1, weight = layers_11_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_955_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15194_cast_fp16 = add(x = dense_output_955_cast_fp16, y = sparse_output_955_cast_fp16)[name = string("op_15194_cast_fp16")];
            tensor<int32, [4]> var_15195 = const()[name = string("op_15195"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15197 = const()[name = string("op_15197"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15196_cast_fp16 = transpose(perm = var_15195, x = var_15194_cast_fp16)[name = string("transpose_500")];
            tensor<fp16, [1, 51, 128]> k_head_381_cast_fp16 = reshape(shape = var_15197, x = var_15196_cast_fp16)[name = string("k_head_381_cast_fp16")];
            string dense_output_957_pad_type_1 = const()[name = string("dense_output_957_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_957_strides_1 = const()[name = string("dense_output_957_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_957_pad_1 = const()[name = string("dense_output_957_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_957_dilations_1 = const()[name = string("dense_output_957_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_957_groups_1 = const()[name = string("dense_output_957_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340684608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340815744))))[name = string("layers_11_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_957_cast_fp16 = conv(dilations = dense_output_957_dilations_1, groups = dense_output_957_groups_1, pad = dense_output_957_pad_1, pad_type = dense_output_957_pad_type_1, strides = dense_output_957_strides_1, weight = layers_11_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_535_cast_fp16)[name = string("dense_output_957_cast_fp16")];
            string sparse_output_957_pad_type_1 = const()[name = string("sparse_output_957_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_957_strides_1 = const()[name = string("sparse_output_957_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_957_pad_1 = const()[name = string("sparse_output_957_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_957_dilations_1 = const()[name = string("sparse_output_957_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_957_groups_1 = const()[name = string("sparse_output_957_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340819008))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340816320))))[name = string("layers_11_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_957_cast_fp16 = conv(dilations = sparse_output_957_dilations_1, groups = sparse_output_957_groups_1, pad = sparse_output_957_pad_1, pad_type = sparse_output_957_pad_type_1, strides = sparse_output_957_strides_1, weight = layers_11_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_535_cast_fp16)[name = string("sparse_output_957_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15213_cast_fp16 = add(x = dense_output_957_cast_fp16, y = sparse_output_957_cast_fp16)[name = string("op_15213_cast_fp16")];
            tensor<int32, [4]> var_15214 = const()[name = string("op_15214"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15216 = const()[name = string("op_15216"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15215_cast_fp16 = transpose(perm = var_15214, x = var_15213_cast_fp16)[name = string("transpose_499")];
            tensor<fp16, [1, 51, 128]> v_head_381_cast_fp16 = reshape(shape = var_15216, x = var_15215_cast_fp16)[name = string("v_head_381_cast_fp16")];
            string dense_output_959_pad_type_1 = const()[name = string("dense_output_959_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_959_strides_1 = const()[name = string("dense_output_959_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_959_pad_1 = const()[name = string("dense_output_959_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_959_dilations_1 = const()[name = string("dense_output_959_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_959_groups_1 = const()[name = string("dense_output_959_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340835456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340966592))))[name = string("layers_11_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_959_cast_fp16 = conv(dilations = dense_output_959_dilations_1, groups = dense_output_959_groups_1, pad = dense_output_959_pad_1, pad_type = dense_output_959_pad_type_1, strides = dense_output_959_strides_1, weight = layers_11_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_959_cast_fp16")];
            string sparse_output_959_pad_type_1 = const()[name = string("sparse_output_959_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_959_strides_1 = const()[name = string("sparse_output_959_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_959_pad_1 = const()[name = string("sparse_output_959_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_959_dilations_1 = const()[name = string("sparse_output_959_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_959_groups_1 = const()[name = string("sparse_output_959_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_11_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340969856))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340967168))))[name = string("layers_11_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_959_cast_fp16 = conv(dilations = sparse_output_959_dilations_1, groups = sparse_output_959_groups_1, pad = sparse_output_959_pad_1, pad_type = sparse_output_959_pad_type_1, strides = sparse_output_959_strides_1, weight = layers_11_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_959_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_15232_cast_fp16 = add(x = dense_output_959_cast_fp16, y = sparse_output_959_cast_fp16)[name = string("op_15232_cast_fp16")];
            tensor<int32, [4]> var_15233 = const()[name = string("op_15233"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15235 = const()[name = string("op_15235"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_15234_cast_fp16 = transpose(perm = var_15233, x = var_15232_cast_fp16)[name = string("transpose_498")];
            tensor<fp16, [1, 101, 128]> p_head_381_cast_fp16 = reshape(shape = var_15235, x = var_15234_cast_fp16)[name = string("p_head_381_cast_fp16")];
            tensor<fp16, [128]> var_15237_to_fp16 = const()[name = string("op_15237_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340986304)))];
            tensor<fp16, [1, 51, 128]> var_15238_cast_fp16 = add(x = q_head_191_cast_fp16, y = var_15237_to_fp16)[name = string("op_15238_cast_fp16")];
            tensor<int32, [1]> q_u_191_axes_1 = const()[name = string("q_u_191_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_191_cast_fp16 = expand_dims(axes = q_u_191_axes_1, x = var_15238_cast_fp16)[name = string("q_u_191_cast_fp16")];
            tensor<fp16, [128]> var_15240_to_fp16 = const()[name = string("op_15240_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340986624)))];
            tensor<fp16, [1, 51, 128]> var_15241_cast_fp16 = add(x = q_head_191_cast_fp16, y = var_15240_to_fp16)[name = string("op_15241_cast_fp16")];
            tensor<int32, [1]> q_v_191_axes_1 = const()[name = string("q_v_191_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_191_cast_fp16 = expand_dims(axes = q_v_191_axes_1, x = var_15241_cast_fp16)[name = string("q_v_191_cast_fp16")];
            tensor<int32, [1]> k_head_383_axes_1 = const()[name = string("k_head_383_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_383_cast_fp16 = expand_dims(axes = k_head_383_axes_1, x = k_head_381_cast_fp16)[name = string("k_head_383_cast_fp16")];
            tensor<int32, [1]> v_head_383_axes_1 = const()[name = string("v_head_383_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_383_cast_fp16 = expand_dims(axes = v_head_383_axes_1, x = v_head_381_cast_fp16)[name = string("v_head_383_cast_fp16")];
            tensor<int32, [1]> p_head_383_axes_1 = const()[name = string("p_head_383_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_383_cast_fp16 = expand_dims(axes = p_head_383_axes_1, x = p_head_381_cast_fp16)[name = string("p_head_383_cast_fp16")];
            bool var_15247_transpose_x_3 = const()[name = string("op_15247_transpose_x_3"), val = bool(false)];
            bool var_15247_transpose_y_3 = const()[name = string("op_15247_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_15247_cast_fp16 = matmul(transpose_x = var_15247_transpose_x_3, transpose_y = var_15247_transpose_y_3, x = q_u_191_cast_fp16, y = k_head_383_cast_fp16)[name = string("op_15247_cast_fp16")];
            fp16 var_15248_to_fp16 = const()[name = string("op_15248_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_191_cast_fp16 = mul(x = var_15247_cast_fp16, y = var_15248_to_fp16)[name = string("scores_content_191_cast_fp16")];
            bool x_1001_transpose_x_3 = const()[name = string("x_1001_transpose_x_3"), val = bool(false)];
            bool x_1001_transpose_y_3 = const()[name = string("x_1001_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1001_cast_fp16 = matmul(transpose_x = x_1001_transpose_x_3, transpose_y = x_1001_transpose_y_3, x = q_v_191_cast_fp16, y = p_head_383_cast_fp16)[name = string("x_1001_cast_fp16")];
            tensor<int32, [8]> x_1003_pad_1 = const()[name = string("x_1003_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1003_mode_1 = const()[name = string("x_1003_mode_1"), val = string("constant")];
            fp16 const_1945_to_fp16 = const()[name = string("const_1945_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1003_cast_fp16 = pad(constant_val = const_1945_to_fp16, mode = x_1003_mode_1, pad = x_1003_pad_1, x = x_1001_cast_fp16)[name = string("x_1003_cast_fp16")];
            tensor<int32, [4]> var_15262 = const()[name = string("op_15262"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1005_cast_fp16 = reshape(shape = var_15262, x = x_1003_cast_fp16)[name = string("x_1005_cast_fp16")];
            tensor<int32, [4]> var_15266_begin_1 = const()[name = string("op_15266_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15266_end_1 = const()[name = string("op_15266_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_15266_end_mask_1 = const()[name = string("op_15266_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_15266_cast_fp16 = slice_by_index(begin = var_15266_begin_1, end = var_15266_end_1, end_mask = var_15266_end_mask_1, x = x_1005_cast_fp16)[name = string("op_15266_cast_fp16")];
            tensor<int32, [4]> var_15268 = const()[name = string("op_15268"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_15269_cast_fp16 = reshape(shape = var_15268, x = var_15266_cast_fp16)[name = string("op_15269_cast_fp16")];
            tensor<int32, [4]> var_15274_begin_1 = const()[name = string("op_15274_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15274_end_1 = const()[name = string("op_15274_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_15274_end_mask_1 = const()[name = string("op_15274_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_15274_cast_fp16 = slice_by_index(begin = var_15274_begin_1, end = var_15274_end_1, end_mask = var_15274_end_mask_1, x = var_15269_cast_fp16)[name = string("op_15274_cast_fp16")];
            fp16 var_15275_to_fp16 = const()[name = string("op_15275_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_191_cast_fp16 = mul(x = var_15274_cast_fp16, y = var_15275_to_fp16)[name = string("scores_pos_191_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_191_cast_fp16 = add(x = scores_content_191_cast_fp16, y = scores_pos_191_cast_fp16)[name = string("logits_191_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_15278_cast_fp16 = softmax(axis = var_14156, x = logits_191_cast_fp16)[name = string("op_15278_cast_fp16")];
            bool var_15280_transpose_x_1 = const()[name = string("op_15280_transpose_x_1"), val = bool(false)];
            bool var_15280_transpose_y_1 = const()[name = string("op_15280_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_15280_cast_fp16 = matmul(transpose_x = var_15280_transpose_x_1, transpose_y = var_15280_transpose_y_1, x = var_15278_cast_fp16, y = v_head_383_cast_fp16)[name = string("op_15280_cast_fp16")];
            tensor<int32, [1]> o_head_23_axes_1 = const()[name = string("o_head_23_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_23_cast_fp16 = squeeze(axes = o_head_23_axes_1, x = var_15280_cast_fp16)[name = string("o_head_23_cast_fp16")];
            bool out_23_interleave_1 = const()[name = string("out_23_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_23_cast_fp16 = concat(axis = var_14156, interleave = out_23_interleave_1, values = (var_14434_cast_fp16, var_14555_cast_fp16, var_14676_cast_fp16, var_14797_cast_fp16, var_14918_cast_fp16, var_15039_cast_fp16, var_15160_cast_fp16, o_head_23_cast_fp16))[name = string("out_23_cast_fp16")];
            tensor<int32, [3]> var_15284_perm_1 = const()[name = string("op_15284_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_543_axes_1 = const()[name = string("input_543_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_15284_cast_fp16 = transpose(perm = var_15284_perm_1, x = out_23_cast_fp16)[name = string("transpose_497")];
            tensor<fp16, [1, 1024, 51, 1]> input_543_cast_fp16 = expand_dims(axes = input_543_axes_1, x = var_15284_cast_fp16)[name = string("input_543_cast_fp16")];
            string dense_output_961_pad_type_1 = const()[name = string("dense_output_961_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_961_strides_1 = const()[name = string("dense_output_961_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_961_pad_1 = const()[name = string("dense_output_961_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_961_dilations_1 = const()[name = string("dense_output_961_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_961_groups_1 = const()[name = string("dense_output_961_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340986944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342035584))))[name = string("layers_11_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_961_cast_fp16 = conv(dilations = dense_output_961_dilations_1, groups = dense_output_961_groups_1, pad = dense_output_961_pad_1, pad_type = dense_output_961_pad_type_1, strides = dense_output_961_strides_1, weight = layers_11_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_543_cast_fp16)[name = string("dense_output_961_cast_fp16")];
            string sparse_output_961_pad_type_1 = const()[name = string("sparse_output_961_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_961_strides_1 = const()[name = string("sparse_output_961_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_961_pad_1 = const()[name = string("sparse_output_961_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_961_dilations_1 = const()[name = string("sparse_output_961_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_961_groups_1 = const()[name = string("sparse_output_961_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342057216))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342036160))))[name = string("layers_11_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_961_cast_fp16 = conv(dilations = sparse_output_961_dilations_1, groups = sparse_output_961_groups_1, pad = sparse_output_961_pad_1, pad_type = sparse_output_961_pad_type_1, strides = sparse_output_961_strides_1, weight = layers_11_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_543_cast_fp16)[name = string("sparse_output_961_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_23_cast_fp16 = add(x = dense_output_961_cast_fp16, y = sparse_output_961_cast_fp16)[name = string("out_conv_23_cast_fp16")];
            tensor<int32, [1]> var_15301_axes_1 = const()[name = string("op_15301_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_15301_cast_fp16 = squeeze(axes = var_15301_axes_1, x = out_conv_23_cast_fp16)[name = string("op_15301_cast_fp16")];
            tensor<int32, [3]> var_15302_perm_1 = const()[name = string("op_15302_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_15302_cast_fp16 = transpose(perm = var_15302_perm_1, x = var_15301_cast_fp16)[name = string("transpose_496")];
            tensor<fp16, [1, 51, 1024]> input_545_cast_fp16 = add(x = input_533_cast_fp16, y = var_15302_cast_fp16)[name = string("input_545_cast_fp16")];
            tensor<int32, [1]> x_1009_axes_1 = const()[name = string("x_1009_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_11_norm_conv_weight_to_fp16 = const()[name = string("layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342188352)))];
            tensor<fp16, [1024]> layers_11_norm_conv_bias_to_fp16 = const()[name = string("layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342190464)))];
            tensor<fp16, [1, 51, 1024]> x_1009_cast_fp16 = layer_norm(axes = x_1009_axes_1, beta = layers_11_norm_conv_bias_to_fp16, epsilon = var_14171_to_fp16, gamma = layers_11_norm_conv_weight_to_fp16, x = input_545_cast_fp16)[name = string("x_1009_cast_fp16")];
            tensor<int32, [3]> var_15312_perm_1 = const()[name = string("op_15312_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_547_axes_1 = const()[name = string("input_547_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_15312_cast_fp16 = transpose(perm = var_15312_perm_1, x = x_1009_cast_fp16)[name = string("transpose_495")];
            tensor<fp16, [1, 1024, 51, 1]> input_547_cast_fp16 = expand_dims(axes = input_547_axes_1, x = var_15312_cast_fp16)[name = string("input_547_cast_fp16")];
            string dense_output_963_pad_type_1 = const()[name = string("dense_output_963_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_963_strides_1 = const()[name = string("dense_output_963_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_963_pad_1 = const()[name = string("dense_output_963_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_963_dilations_1 = const()[name = string("dense_output_963_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_963_groups_1 = const()[name = string("dense_output_963_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_11_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342192576))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344289792))))[name = string("layers_11_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_963_cast_fp16 = conv(dilations = dense_output_963_dilations_1, groups = dense_output_963_groups_1, pad = dense_output_963_pad_1, pad_type = dense_output_963_pad_type_1, strides = dense_output_963_strides_1, weight = layers_11_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("dense_output_963_cast_fp16")];
            string sparse_output_963_pad_type_1 = const()[name = string("sparse_output_963_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_963_strides_1 = const()[name = string("sparse_output_963_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_963_pad_1 = const()[name = string("sparse_output_963_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_963_dilations_1 = const()[name = string("sparse_output_963_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_963_groups_1 = const()[name = string("sparse_output_963_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_11_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344332416))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344290368))))[name = string("layers_11_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_963_cast_fp16 = conv(dilations = sparse_output_963_dilations_1, groups = sparse_output_963_groups_1, pad = sparse_output_963_pad_1, pad_type = sparse_output_963_pad_type_1, strides = sparse_output_963_strides_1, weight = layers_11_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_547_cast_fp16)[name = string("sparse_output_963_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_549_cast_fp16 = add(x = dense_output_963_cast_fp16, y = sparse_output_963_cast_fp16)[name = string("input_549_cast_fp16")];
            int32 input_551_split_num_splits_1 = const()[name = string("input_551_split_num_splits_1"), val = int32(2)];
            int32 input_551_split_axis_1 = const()[name = string("input_551_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_551_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_551_split_cast_fp16_1 = split(axis = input_551_split_axis_1, num_splits = input_551_split_num_splits_1, x = input_549_cast_fp16)[name = string("input_551_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_551_split_1_sigmoid_cast_fp16 = sigmoid(x = input_551_split_cast_fp16_1)[name = string("input_551_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_551_cast_fp16 = mul(x = input_551_split_cast_fp16_0, y = input_551_split_1_sigmoid_cast_fp16)[name = string("input_551_cast_fp16")];
            tensor<int32, [8]> input_553_pad_1 = const()[name = string("input_553_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_553_mode_1 = const()[name = string("input_553_mode_1"), val = string("constant")];
            fp16 const_1947_to_fp16 = const()[name = string("const_1947_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_553_cast_fp16 = pad(constant_val = const_1947_to_fp16, mode = input_553_mode_1, pad = input_553_pad_1, x = input_551_cast_fp16)[name = string("input_553_cast_fp16")];
            string dense_output_965_pad_type_1 = const()[name = string("dense_output_965_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_965_strides_1 = const()[name = string("dense_output_965_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_965_pad_1 = const()[name = string("dense_output_965_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_965_dilations_1 = const()[name = string("dense_output_965_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_965_groups_1 = const()[name = string("dense_output_965_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_965_cast_fp16 = conv(dilations = dense_output_965_dilations_1, groups = dense_output_965_groups_1, pad = dense_output_965_pad_1, pad_type = dense_output_965_pad_type_1, strides = dense_output_965_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_553_cast_fp16)[name = string("dense_output_965_cast_fp16")];
            string sparse_output_965_pad_type_1 = const()[name = string("sparse_output_965_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_965_strides_1 = const()[name = string("sparse_output_965_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_965_pad_1 = const()[name = string("sparse_output_965_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_965_dilations_1 = const()[name = string("sparse_output_965_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_965_groups_1 = const()[name = string("sparse_output_965_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_11_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344594624))))[name = string("layers_11_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_965_cast_fp16 = conv(dilations = sparse_output_965_dilations_1, groups = sparse_output_965_groups_1, pad = sparse_output_965_pad_1, pad_type = sparse_output_965_pad_type_1, strides = sparse_output_965_strides_1, weight = layers_11_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_553_cast_fp16)[name = string("sparse_output_965_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_555_cast_fp16 = add(x = dense_output_965_cast_fp16, y = sparse_output_965_cast_fp16)[name = string("input_555_cast_fp16")];
            tensor<fp16, [1024]> layers_11_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_11_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344613120)))];
            tensor<fp16, [1024]> layers_11_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_11_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344615232)))];
            tensor<fp16, [1024]> layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344617344)))];
            tensor<fp16, [1024]> layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344619456)))];
            tensor<fp16, [1, 1024, 51, 1]> input_557_cast_fp16 = batch_norm(beta = layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_14171_to_fp16, gamma = layers_11_conv_batch_norm_weight_to_fp16, mean = layers_11_conv_batch_norm_running_mean_to_fp16, variance = layers_11_conv_batch_norm_running_var_to_fp16, x = input_555_cast_fp16)[name = string("input_557_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_559_cast_fp16 = silu(x = input_557_cast_fp16)[name = string("input_559_cast_fp16")];
            string dense_output_967_pad_type_1 = const()[name = string("dense_output_967_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_967_strides_1 = const()[name = string("dense_output_967_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_967_pad_1 = const()[name = string("dense_output_967_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_967_dilations_1 = const()[name = string("dense_output_967_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_967_groups_1 = const()[name = string("dense_output_967_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344621568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345670208))))[name = string("layers_11_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_967_cast_fp16 = conv(dilations = dense_output_967_dilations_1, groups = dense_output_967_groups_1, pad = dense_output_967_pad_1, pad_type = dense_output_967_pad_type_1, strides = dense_output_967_strides_1, weight = layers_11_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("dense_output_967_cast_fp16")];
            string sparse_output_967_pad_type_1 = const()[name = string("sparse_output_967_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_967_strides_1 = const()[name = string("sparse_output_967_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_967_pad_1 = const()[name = string("sparse_output_967_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_967_dilations_1 = const()[name = string("sparse_output_967_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_967_groups_1 = const()[name = string("sparse_output_967_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_11_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345691840))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345670784))))[name = string("layers_11_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_967_cast_fp16 = conv(dilations = sparse_output_967_dilations_1, groups = sparse_output_967_groups_1, pad = sparse_output_967_pad_1, pad_type = sparse_output_967_pad_type_1, strides = sparse_output_967_strides_1, weight = layers_11_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_559_cast_fp16)[name = string("sparse_output_967_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1011_cast_fp16 = add(x = dense_output_967_cast_fp16, y = sparse_output_967_cast_fp16)[name = string("x_1011_cast_fp16")];
            tensor<int32, [1]> var_15368_axes_1 = const()[name = string("op_15368_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_15368_cast_fp16 = squeeze(axes = var_15368_axes_1, x = x_1011_cast_fp16)[name = string("op_15368_cast_fp16")];
            tensor<int32, [3]> var_15369_perm_1 = const()[name = string("op_15369_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_15369_cast_fp16 = transpose(perm = var_15369_perm_1, x = var_15368_cast_fp16)[name = string("transpose_494")];
            tensor<fp16, [1, 51, 1024]> input_561_cast_fp16 = add(x = input_545_cast_fp16, y = var_15369_cast_fp16)[name = string("input_561_cast_fp16")];
            tensor<int32, [1]> x_1013_axes_1 = const()[name = string("x_1013_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345822976)))];
            tensor<fp16, [1024]> layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345825088)))];
            tensor<fp16, [1, 51, 1024]> x_1013_cast_fp16 = layer_norm(axes = x_1013_axes_1, beta = layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_14171_to_fp16, gamma = layers_11_norm_feed_forward2_weight_to_fp16, x = input_561_cast_fp16)[name = string("x_1013_cast_fp16")];
            tensor<int32, [4]> var_15379 = const()[name = string("op_15379"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1015_cast_fp16 = reshape(shape = var_15379, x = x_1013_cast_fp16)[name = string("x_1015_cast_fp16")];
            tensor<int32, [4]> input_563_perm_1 = const()[name = string("input_563_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_969_pad_type_1 = const()[name = string("dense_output_969_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_969_strides_1 = const()[name = string("dense_output_969_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_969_pad_1 = const()[name = string("dense_output_969_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_969_dilations_1 = const()[name = string("dense_output_969_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_969_groups_1 = const()[name = string("dense_output_969_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_11_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345827200))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350021568))))[name = string("layers_11_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_563_cast_fp16 = transpose(perm = input_563_perm_1, x = x_1015_cast_fp16)[name = string("transpose_493")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_969_cast_fp16 = conv(dilations = dense_output_969_dilations_1, groups = dense_output_969_groups_1, pad = dense_output_969_pad_1, pad_type = dense_output_969_pad_type_1, strides = dense_output_969_strides_1, weight = layers_11_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("dense_output_969_cast_fp16")];
            string sparse_output_969_pad_type_1 = const()[name = string("sparse_output_969_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_969_strides_1 = const()[name = string("sparse_output_969_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_969_pad_1 = const()[name = string("sparse_output_969_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_969_dilations_1 = const()[name = string("sparse_output_969_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_969_groups_1 = const()[name = string("sparse_output_969_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_11_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350106112))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350022144))))[name = string("layers_11_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_969_cast_fp16 = conv(dilations = sparse_output_969_dilations_1, groups = sparse_output_969_groups_1, pad = sparse_output_969_pad_1, pad_type = sparse_output_969_pad_type_1, strides = sparse_output_969_strides_1, weight = layers_11_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_563_cast_fp16)[name = string("sparse_output_969_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_565_cast_fp16 = add(x = dense_output_969_cast_fp16, y = sparse_output_969_cast_fp16)[name = string("input_565_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_567_cast_fp16 = silu(x = input_565_cast_fp16)[name = string("input_567_cast_fp16")];
            string dense_output_971_pad_type_1 = const()[name = string("dense_output_971_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_971_strides_1 = const()[name = string("dense_output_971_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_971_pad_1 = const()[name = string("dense_output_971_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_971_dilations_1 = const()[name = string("dense_output_971_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_971_groups_1 = const()[name = string("dense_output_971_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_11_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350630464))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354824832))))[name = string("layers_11_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_971_cast_fp16 = conv(dilations = dense_output_971_dilations_1, groups = dense_output_971_groups_1, pad = dense_output_971_pad_1, pad_type = dense_output_971_pad_type_1, strides = dense_output_971_strides_1, weight = layers_11_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_567_cast_fp16)[name = string("dense_output_971_cast_fp16")];
            string sparse_output_971_pad_type_1 = const()[name = string("sparse_output_971_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_971_strides_1 = const()[name = string("sparse_output_971_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_971_pad_1 = const()[name = string("sparse_output_971_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_971_dilations_1 = const()[name = string("sparse_output_971_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_971_groups_1 = const()[name = string("sparse_output_971_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_11_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354909376))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354825408))))[name = string("layers_11_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_971_cast_fp16 = conv(dilations = sparse_output_971_dilations_1, groups = sparse_output_971_groups_1, pad = sparse_output_971_pad_1, pad_type = sparse_output_971_pad_type_1, strides = sparse_output_971_strides_1, weight = layers_11_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_567_cast_fp16)[name = string("sparse_output_971_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1017_cast_fp16 = add(x = dense_output_971_cast_fp16, y = sparse_output_971_cast_fp16)[name = string("x_1017_cast_fp16")];
            tensor<int32, [4]> x_1019_perm_1 = const()[name = string("x_1019_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_15414 = const()[name = string("op_15414"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1019_cast_fp16 = transpose(perm = x_1019_perm_1, x = x_1017_cast_fp16)[name = string("transpose_492")];
            tensor<fp16, [1, 51, 1024]> var_15415_cast_fp16 = reshape(shape = var_15414, x = x_1019_cast_fp16)[name = string("op_15415_cast_fp16")];
            fp16 var_15416_to_fp16 = const()[name = string("op_15416_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_15417_cast_fp16 = mul(x = var_15415_cast_fp16, y = var_15416_to_fp16)[name = string("op_15417_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_569_cast_fp16 = add(x = input_561_cast_fp16, y = var_15417_cast_fp16)[name = string("input_569_cast_fp16")];
            tensor<int32, [1]> input_571_axes_1 = const()[name = string("input_571_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_11_norm_out_weight_to_fp16 = const()[name = string("layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355433728)))];
            tensor<fp16, [1024]> layers_11_norm_out_bias_to_fp16 = const()[name = string("layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355435840)))];
            tensor<fp16, [1, 51, 1024]> input_571_cast_fp16 = layer_norm(axes = input_571_axes_1, beta = layers_11_norm_out_bias_to_fp16, epsilon = var_14171_to_fp16, gamma = layers_11_norm_out_weight_to_fp16, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
            int32 var_15425 = const()[name = string("op_15425"), val = int32(-1)];
            tensor<int32, [1]> x_1021_axes_1 = const()[name = string("x_1021_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355437952)))];
            tensor<fp16, [1024]> layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355440064)))];
            fp16 var_15440_to_fp16 = const()[name = string("op_15440_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1021_cast_fp16 = layer_norm(axes = x_1021_axes_1, beta = layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_15440_to_fp16, gamma = layers_12_norm_feed_forward1_weight_to_fp16, x = input_571_cast_fp16)[name = string("x_1021_cast_fp16")];
            tensor<int32, [4]> var_15459 = const()[name = string("op_15459"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1023_cast_fp16 = reshape(shape = var_15459, x = x_1021_cast_fp16)[name = string("x_1023_cast_fp16")];
            tensor<int32, [4]> input_573_perm_1 = const()[name = string("input_573_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_973_pad_type_1 = const()[name = string("dense_output_973_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_973_strides_1 = const()[name = string("dense_output_973_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_973_pad_1 = const()[name = string("dense_output_973_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_973_dilations_1 = const()[name = string("dense_output_973_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_973_groups_1 = const()[name = string("dense_output_973_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_12_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355442176))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359636544))))[name = string("layers_12_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_573_cast_fp16 = transpose(perm = input_573_perm_1, x = x_1023_cast_fp16)[name = string("transpose_491")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_973_cast_fp16 = conv(dilations = dense_output_973_dilations_1, groups = dense_output_973_groups_1, pad = dense_output_973_pad_1, pad_type = dense_output_973_pad_type_1, strides = dense_output_973_strides_1, weight = layers_12_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_573_cast_fp16)[name = string("dense_output_973_cast_fp16")];
            string sparse_output_973_pad_type_1 = const()[name = string("sparse_output_973_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_973_strides_1 = const()[name = string("sparse_output_973_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_973_pad_1 = const()[name = string("sparse_output_973_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_973_dilations_1 = const()[name = string("sparse_output_973_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_973_groups_1 = const()[name = string("sparse_output_973_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_12_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359721088))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359637120))))[name = string("layers_12_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_973_cast_fp16 = conv(dilations = sparse_output_973_dilations_1, groups = sparse_output_973_groups_1, pad = sparse_output_973_pad_1, pad_type = sparse_output_973_pad_type_1, strides = sparse_output_973_strides_1, weight = layers_12_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_573_cast_fp16)[name = string("sparse_output_973_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_575_cast_fp16 = add(x = dense_output_973_cast_fp16, y = sparse_output_973_cast_fp16)[name = string("input_575_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_577_cast_fp16 = silu(x = input_575_cast_fp16)[name = string("input_577_cast_fp16")];
            string dense_output_975_pad_type_1 = const()[name = string("dense_output_975_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_975_strides_1 = const()[name = string("dense_output_975_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_975_pad_1 = const()[name = string("dense_output_975_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_975_dilations_1 = const()[name = string("dense_output_975_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_975_groups_1 = const()[name = string("dense_output_975_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_12_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360245440))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364439808))))[name = string("layers_12_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_975_cast_fp16 = conv(dilations = dense_output_975_dilations_1, groups = dense_output_975_groups_1, pad = dense_output_975_pad_1, pad_type = dense_output_975_pad_type_1, strides = dense_output_975_strides_1, weight = layers_12_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_577_cast_fp16)[name = string("dense_output_975_cast_fp16")];
            string sparse_output_975_pad_type_1 = const()[name = string("sparse_output_975_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_975_strides_1 = const()[name = string("sparse_output_975_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_975_pad_1 = const()[name = string("sparse_output_975_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_975_dilations_1 = const()[name = string("sparse_output_975_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_975_groups_1 = const()[name = string("sparse_output_975_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_12_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364524352))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364440384))))[name = string("layers_12_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_975_cast_fp16 = conv(dilations = sparse_output_975_dilations_1, groups = sparse_output_975_groups_1, pad = sparse_output_975_pad_1, pad_type = sparse_output_975_pad_type_1, strides = sparse_output_975_strides_1, weight = layers_12_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_577_cast_fp16)[name = string("sparse_output_975_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1025_cast_fp16 = add(x = dense_output_975_cast_fp16, y = sparse_output_975_cast_fp16)[name = string("x_1025_cast_fp16")];
            tensor<int32, [4]> x_1027_perm_1 = const()[name = string("x_1027_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_15494 = const()[name = string("op_15494"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1027_cast_fp16 = transpose(perm = x_1027_perm_1, x = x_1025_cast_fp16)[name = string("transpose_490")];
            tensor<fp16, [1, 51, 1024]> var_15495_cast_fp16 = reshape(shape = var_15494, x = x_1027_cast_fp16)[name = string("op_15495_cast_fp16")];
            fp16 var_15496_to_fp16 = const()[name = string("op_15496_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_15497_cast_fp16 = mul(x = var_15495_cast_fp16, y = var_15496_to_fp16)[name = string("op_15497_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_579_cast_fp16 = add(x = input_571_cast_fp16, y = var_15497_cast_fp16)[name = string("input_579_cast_fp16")];
            tensor<int32, [1]> q_25_axes_1 = const()[name = string("q_25_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_12_norm_self_att_weight_to_fp16 = const()[name = string("layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365048704)))];
            tensor<fp16, [1024]> layers_12_norm_self_att_bias_to_fp16 = const()[name = string("layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365050816)))];
            tensor<fp16, [1, 51, 1024]> q_25_cast_fp16 = layer_norm(axes = q_25_axes_1, beta = layers_12_norm_self_att_bias_to_fp16, epsilon = var_15440_to_fp16, gamma = layers_12_norm_self_att_weight_to_fp16, x = input_579_cast_fp16)[name = string("q_25_cast_fp16")];
            tensor<int32, [3]> var_15571 = const()[name = string("op_15571"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_581_axes_1 = const()[name = string("input_581_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_15572_cast_fp16 = transpose(perm = var_15571, x = q_25_cast_fp16)[name = string("transpose_489")];
            tensor<fp16, [1, 1024, 51, 1]> input_581_cast_fp16 = expand_dims(axes = input_581_axes_1, x = var_15572_cast_fp16)[name = string("input_581_cast_fp16")];
            string dense_output_977_pad_type_1 = const()[name = string("dense_output_977_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_977_strides_1 = const()[name = string("dense_output_977_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_977_pad_1 = const()[name = string("dense_output_977_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_977_dilations_1 = const()[name = string("dense_output_977_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_977_groups_1 = const()[name = string("dense_output_977_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365052928))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365184064))))[name = string("layers_12_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_977_cast_fp16 = conv(dilations = dense_output_977_dilations_1, groups = dense_output_977_groups_1, pad = dense_output_977_pad_1, pad_type = dense_output_977_pad_type_1, strides = dense_output_977_strides_1, weight = layers_12_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_977_cast_fp16")];
            string sparse_output_977_pad_type_1 = const()[name = string("sparse_output_977_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_977_strides_1 = const()[name = string("sparse_output_977_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_977_pad_1 = const()[name = string("sparse_output_977_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_977_dilations_1 = const()[name = string("sparse_output_977_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_977_groups_1 = const()[name = string("sparse_output_977_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365187328))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365184640))))[name = string("layers_12_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_977_cast_fp16 = conv(dilations = sparse_output_977_dilations_1, groups = sparse_output_977_groups_1, pad = sparse_output_977_pad_1, pad_type = sparse_output_977_pad_type_1, strides = sparse_output_977_strides_1, weight = layers_12_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_977_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15597_cast_fp16 = add(x = dense_output_977_cast_fp16, y = sparse_output_977_cast_fp16)[name = string("op_15597_cast_fp16")];
            tensor<int32, [4]> var_15598 = const()[name = string("op_15598"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15600 = const()[name = string("op_15600"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15599_cast_fp16 = transpose(perm = var_15598, x = var_15597_cast_fp16)[name = string("transpose_488")];
            tensor<fp16, [1, 51, 128]> q_head_193_cast_fp16 = reshape(shape = var_15600, x = var_15599_cast_fp16)[name = string("q_head_193_cast_fp16")];
            string dense_output_979_pad_type_1 = const()[name = string("dense_output_979_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_979_strides_1 = const()[name = string("dense_output_979_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_979_pad_1 = const()[name = string("dense_output_979_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_979_dilations_1 = const()[name = string("dense_output_979_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_979_groups_1 = const()[name = string("dense_output_979_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365203776))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365334912))))[name = string("layers_12_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_979_cast_fp16 = conv(dilations = dense_output_979_dilations_1, groups = dense_output_979_groups_1, pad = dense_output_979_pad_1, pad_type = dense_output_979_pad_type_1, strides = dense_output_979_strides_1, weight = layers_12_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_979_cast_fp16")];
            string sparse_output_979_pad_type_1 = const()[name = string("sparse_output_979_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_979_strides_1 = const()[name = string("sparse_output_979_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_979_pad_1 = const()[name = string("sparse_output_979_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_979_dilations_1 = const()[name = string("sparse_output_979_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_979_groups_1 = const()[name = string("sparse_output_979_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365338176))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365335488))))[name = string("layers_12_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_979_cast_fp16 = conv(dilations = sparse_output_979_dilations_1, groups = sparse_output_979_groups_1, pad = sparse_output_979_pad_1, pad_type = sparse_output_979_pad_type_1, strides = sparse_output_979_strides_1, weight = layers_12_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_979_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15616_cast_fp16 = add(x = dense_output_979_cast_fp16, y = sparse_output_979_cast_fp16)[name = string("op_15616_cast_fp16")];
            tensor<int32, [4]> var_15617 = const()[name = string("op_15617"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15619 = const()[name = string("op_15619"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15618_cast_fp16 = transpose(perm = var_15617, x = var_15616_cast_fp16)[name = string("transpose_487")];
            tensor<fp16, [1, 51, 128]> k_head_385_cast_fp16 = reshape(shape = var_15619, x = var_15618_cast_fp16)[name = string("k_head_385_cast_fp16")];
            string dense_output_981_pad_type_1 = const()[name = string("dense_output_981_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_981_strides_1 = const()[name = string("dense_output_981_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_981_pad_1 = const()[name = string("dense_output_981_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_981_dilations_1 = const()[name = string("dense_output_981_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_981_groups_1 = const()[name = string("dense_output_981_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365354624))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365485760))))[name = string("layers_12_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_981_cast_fp16 = conv(dilations = dense_output_981_dilations_1, groups = dense_output_981_groups_1, pad = dense_output_981_pad_1, pad_type = dense_output_981_pad_type_1, strides = dense_output_981_strides_1, weight = layers_12_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_981_cast_fp16")];
            string sparse_output_981_pad_type_1 = const()[name = string("sparse_output_981_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_981_strides_1 = const()[name = string("sparse_output_981_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_981_pad_1 = const()[name = string("sparse_output_981_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_981_dilations_1 = const()[name = string("sparse_output_981_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_981_groups_1 = const()[name = string("sparse_output_981_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365489024))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365486336))))[name = string("layers_12_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_981_cast_fp16 = conv(dilations = sparse_output_981_dilations_1, groups = sparse_output_981_groups_1, pad = sparse_output_981_pad_1, pad_type = sparse_output_981_pad_type_1, strides = sparse_output_981_strides_1, weight = layers_12_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_981_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15635_cast_fp16 = add(x = dense_output_981_cast_fp16, y = sparse_output_981_cast_fp16)[name = string("op_15635_cast_fp16")];
            tensor<int32, [4]> var_15636 = const()[name = string("op_15636"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15638 = const()[name = string("op_15638"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15637_cast_fp16 = transpose(perm = var_15636, x = var_15635_cast_fp16)[name = string("transpose_486")];
            tensor<fp16, [1, 51, 128]> v_head_385_cast_fp16 = reshape(shape = var_15638, x = var_15637_cast_fp16)[name = string("v_head_385_cast_fp16")];
            string dense_output_983_pad_type_1 = const()[name = string("dense_output_983_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_983_strides_1 = const()[name = string("dense_output_983_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_983_pad_1 = const()[name = string("dense_output_983_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_983_dilations_1 = const()[name = string("dense_output_983_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_983_groups_1 = const()[name = string("dense_output_983_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365505472))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365636608))))[name = string("layers_12_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_983_cast_fp16 = conv(dilations = dense_output_983_dilations_1, groups = dense_output_983_groups_1, pad = dense_output_983_pad_1, pad_type = dense_output_983_pad_type_1, strides = dense_output_983_strides_1, weight = layers_12_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_983_cast_fp16")];
            string sparse_output_983_pad_type_1 = const()[name = string("sparse_output_983_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_983_strides_1 = const()[name = string("sparse_output_983_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_983_pad_1 = const()[name = string("sparse_output_983_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_983_dilations_1 = const()[name = string("sparse_output_983_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_983_groups_1 = const()[name = string("sparse_output_983_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365639872))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365637184))))[name = string("layers_12_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_983_cast_fp16 = conv(dilations = sparse_output_983_dilations_1, groups = sparse_output_983_groups_1, pad = sparse_output_983_pad_1, pad_type = sparse_output_983_pad_type_1, strides = sparse_output_983_strides_1, weight = layers_12_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_983_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_15654_cast_fp16 = add(x = dense_output_983_cast_fp16, y = sparse_output_983_cast_fp16)[name = string("op_15654_cast_fp16")];
            tensor<int32, [4]> var_15655 = const()[name = string("op_15655"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15657 = const()[name = string("op_15657"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_15656_cast_fp16 = transpose(perm = var_15655, x = var_15654_cast_fp16)[name = string("transpose_485")];
            tensor<fp16, [1, 101, 128]> p_head_385_cast_fp16 = reshape(shape = var_15657, x = var_15656_cast_fp16)[name = string("p_head_385_cast_fp16")];
            tensor<fp16, [128]> var_15659_to_fp16 = const()[name = string("op_15659_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365656320)))];
            tensor<fp16, [1, 51, 128]> var_15660_cast_fp16 = add(x = q_head_193_cast_fp16, y = var_15659_to_fp16)[name = string("op_15660_cast_fp16")];
            tensor<int32, [1]> q_u_193_axes_1 = const()[name = string("q_u_193_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_193_cast_fp16 = expand_dims(axes = q_u_193_axes_1, x = var_15660_cast_fp16)[name = string("q_u_193_cast_fp16")];
            tensor<fp16, [128]> var_15662_to_fp16 = const()[name = string("op_15662_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365656640)))];
            tensor<fp16, [1, 51, 128]> var_15663_cast_fp16 = add(x = q_head_193_cast_fp16, y = var_15662_to_fp16)[name = string("op_15663_cast_fp16")];
            tensor<int32, [1]> q_v_193_axes_1 = const()[name = string("q_v_193_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_193_cast_fp16 = expand_dims(axes = q_v_193_axes_1, x = var_15663_cast_fp16)[name = string("q_v_193_cast_fp16")];
            tensor<int32, [1]> k_head_387_axes_1 = const()[name = string("k_head_387_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_387_cast_fp16 = expand_dims(axes = k_head_387_axes_1, x = k_head_385_cast_fp16)[name = string("k_head_387_cast_fp16")];
            tensor<int32, [1]> v_head_387_axes_1 = const()[name = string("v_head_387_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_387_cast_fp16 = expand_dims(axes = v_head_387_axes_1, x = v_head_385_cast_fp16)[name = string("v_head_387_cast_fp16")];
            tensor<int32, [1]> p_head_387_axes_1 = const()[name = string("p_head_387_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_387_cast_fp16 = expand_dims(axes = p_head_387_axes_1, x = p_head_385_cast_fp16)[name = string("p_head_387_cast_fp16")];
            bool var_15669_transpose_x_3 = const()[name = string("op_15669_transpose_x_3"), val = bool(false)];
            bool var_15669_transpose_y_3 = const()[name = string("op_15669_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_15669_cast_fp16 = matmul(transpose_x = var_15669_transpose_x_3, transpose_y = var_15669_transpose_y_3, x = q_u_193_cast_fp16, y = k_head_387_cast_fp16)[name = string("op_15669_cast_fp16")];
            fp16 var_15670_to_fp16 = const()[name = string("op_15670_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_193_cast_fp16 = mul(x = var_15669_cast_fp16, y = var_15670_to_fp16)[name = string("scores_content_193_cast_fp16")];
            bool x_1029_transpose_x_3 = const()[name = string("x_1029_transpose_x_3"), val = bool(false)];
            bool x_1029_transpose_y_3 = const()[name = string("x_1029_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1029_cast_fp16 = matmul(transpose_x = x_1029_transpose_x_3, transpose_y = x_1029_transpose_y_3, x = q_v_193_cast_fp16, y = p_head_387_cast_fp16)[name = string("x_1029_cast_fp16")];
            tensor<int32, [8]> x_1031_pad_1 = const()[name = string("x_1031_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1031_mode_1 = const()[name = string("x_1031_mode_1"), val = string("constant")];
            fp16 const_1957_to_fp16 = const()[name = string("const_1957_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1031_cast_fp16 = pad(constant_val = const_1957_to_fp16, mode = x_1031_mode_1, pad = x_1031_pad_1, x = x_1029_cast_fp16)[name = string("x_1031_cast_fp16")];
            tensor<int32, [4]> var_15684 = const()[name = string("op_15684"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1033_cast_fp16 = reshape(shape = var_15684, x = x_1031_cast_fp16)[name = string("x_1033_cast_fp16")];
            tensor<int32, [4]> var_15688_begin_1 = const()[name = string("op_15688_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15688_end_1 = const()[name = string("op_15688_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_15688_end_mask_1 = const()[name = string("op_15688_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_15688_cast_fp16 = slice_by_index(begin = var_15688_begin_1, end = var_15688_end_1, end_mask = var_15688_end_mask_1, x = x_1033_cast_fp16)[name = string("op_15688_cast_fp16")];
            tensor<int32, [4]> var_15690 = const()[name = string("op_15690"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_15691_cast_fp16 = reshape(shape = var_15690, x = var_15688_cast_fp16)[name = string("op_15691_cast_fp16")];
            tensor<int32, [4]> var_15696_begin_1 = const()[name = string("op_15696_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15696_end_1 = const()[name = string("op_15696_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_15696_end_mask_1 = const()[name = string("op_15696_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_15696_cast_fp16 = slice_by_index(begin = var_15696_begin_1, end = var_15696_end_1, end_mask = var_15696_end_mask_1, x = var_15691_cast_fp16)[name = string("op_15696_cast_fp16")];
            fp16 var_15697_to_fp16 = const()[name = string("op_15697_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_193_cast_fp16 = mul(x = var_15696_cast_fp16, y = var_15697_to_fp16)[name = string("scores_pos_193_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_193_cast_fp16 = add(x = scores_content_193_cast_fp16, y = scores_pos_193_cast_fp16)[name = string("logits_193_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_15700_cast_fp16 = softmax(axis = var_15425, x = logits_193_cast_fp16)[name = string("op_15700_cast_fp16")];
            bool var_15702_transpose_x_1 = const()[name = string("op_15702_transpose_x_1"), val = bool(false)];
            bool var_15702_transpose_y_1 = const()[name = string("op_15702_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_15702_cast_fp16 = matmul(transpose_x = var_15702_transpose_x_1, transpose_y = var_15702_transpose_y_1, x = var_15700_cast_fp16, y = v_head_387_cast_fp16)[name = string("op_15702_cast_fp16")];
            tensor<int32, [1]> var_15703_axes_1 = const()[name = string("op_15703_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_15703_cast_fp16 = squeeze(axes = var_15703_axes_1, x = var_15702_cast_fp16)[name = string("op_15703_cast_fp16")];
            string dense_output_985_pad_type_1 = const()[name = string("dense_output_985_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_985_strides_1 = const()[name = string("dense_output_985_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_985_pad_1 = const()[name = string("dense_output_985_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_985_dilations_1 = const()[name = string("dense_output_985_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_985_groups_1 = const()[name = string("dense_output_985_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365656960))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365788096))))[name = string("layers_12_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_985_cast_fp16 = conv(dilations = dense_output_985_dilations_1, groups = dense_output_985_groups_1, pad = dense_output_985_pad_1, pad_type = dense_output_985_pad_type_1, strides = dense_output_985_strides_1, weight = layers_12_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_985_cast_fp16")];
            string sparse_output_985_pad_type_1 = const()[name = string("sparse_output_985_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_985_strides_1 = const()[name = string("sparse_output_985_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_985_pad_1 = const()[name = string("sparse_output_985_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_985_dilations_1 = const()[name = string("sparse_output_985_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_985_groups_1 = const()[name = string("sparse_output_985_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365791360))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365788672))))[name = string("layers_12_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_985_cast_fp16 = conv(dilations = sparse_output_985_dilations_1, groups = sparse_output_985_groups_1, pad = sparse_output_985_pad_1, pad_type = sparse_output_985_pad_type_1, strides = sparse_output_985_strides_1, weight = layers_12_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_985_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15718_cast_fp16 = add(x = dense_output_985_cast_fp16, y = sparse_output_985_cast_fp16)[name = string("op_15718_cast_fp16")];
            tensor<int32, [4]> var_15719 = const()[name = string("op_15719"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15721 = const()[name = string("op_15721"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15720_cast_fp16 = transpose(perm = var_15719, x = var_15718_cast_fp16)[name = string("transpose_484")];
            tensor<fp16, [1, 51, 128]> q_head_195_cast_fp16 = reshape(shape = var_15721, x = var_15720_cast_fp16)[name = string("q_head_195_cast_fp16")];
            string dense_output_987_pad_type_1 = const()[name = string("dense_output_987_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_987_strides_1 = const()[name = string("dense_output_987_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_987_pad_1 = const()[name = string("dense_output_987_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_987_dilations_1 = const()[name = string("dense_output_987_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_987_groups_1 = const()[name = string("dense_output_987_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365807808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365938944))))[name = string("layers_12_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_987_cast_fp16 = conv(dilations = dense_output_987_dilations_1, groups = dense_output_987_groups_1, pad = dense_output_987_pad_1, pad_type = dense_output_987_pad_type_1, strides = dense_output_987_strides_1, weight = layers_12_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_987_cast_fp16")];
            string sparse_output_987_pad_type_1 = const()[name = string("sparse_output_987_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_987_strides_1 = const()[name = string("sparse_output_987_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_987_pad_1 = const()[name = string("sparse_output_987_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_987_dilations_1 = const()[name = string("sparse_output_987_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_987_groups_1 = const()[name = string("sparse_output_987_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365942208))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365939520))))[name = string("layers_12_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_987_cast_fp16 = conv(dilations = sparse_output_987_dilations_1, groups = sparse_output_987_groups_1, pad = sparse_output_987_pad_1, pad_type = sparse_output_987_pad_type_1, strides = sparse_output_987_strides_1, weight = layers_12_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_987_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15737_cast_fp16 = add(x = dense_output_987_cast_fp16, y = sparse_output_987_cast_fp16)[name = string("op_15737_cast_fp16")];
            tensor<int32, [4]> var_15738 = const()[name = string("op_15738"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15740 = const()[name = string("op_15740"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15739_cast_fp16 = transpose(perm = var_15738, x = var_15737_cast_fp16)[name = string("transpose_483")];
            tensor<fp16, [1, 51, 128]> k_head_389_cast_fp16 = reshape(shape = var_15740, x = var_15739_cast_fp16)[name = string("k_head_389_cast_fp16")];
            string dense_output_989_pad_type_1 = const()[name = string("dense_output_989_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_989_strides_1 = const()[name = string("dense_output_989_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_989_pad_1 = const()[name = string("dense_output_989_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_989_dilations_1 = const()[name = string("dense_output_989_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_989_groups_1 = const()[name = string("dense_output_989_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365958656))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366089792))))[name = string("layers_12_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_989_cast_fp16 = conv(dilations = dense_output_989_dilations_1, groups = dense_output_989_groups_1, pad = dense_output_989_pad_1, pad_type = dense_output_989_pad_type_1, strides = dense_output_989_strides_1, weight = layers_12_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_989_cast_fp16")];
            string sparse_output_989_pad_type_1 = const()[name = string("sparse_output_989_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_989_strides_1 = const()[name = string("sparse_output_989_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_989_pad_1 = const()[name = string("sparse_output_989_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_989_dilations_1 = const()[name = string("sparse_output_989_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_989_groups_1 = const()[name = string("sparse_output_989_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366093056))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366090368))))[name = string("layers_12_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_989_cast_fp16 = conv(dilations = sparse_output_989_dilations_1, groups = sparse_output_989_groups_1, pad = sparse_output_989_pad_1, pad_type = sparse_output_989_pad_type_1, strides = sparse_output_989_strides_1, weight = layers_12_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_989_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15756_cast_fp16 = add(x = dense_output_989_cast_fp16, y = sparse_output_989_cast_fp16)[name = string("op_15756_cast_fp16")];
            tensor<int32, [4]> var_15757 = const()[name = string("op_15757"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15759 = const()[name = string("op_15759"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15758_cast_fp16 = transpose(perm = var_15757, x = var_15756_cast_fp16)[name = string("transpose_482")];
            tensor<fp16, [1, 51, 128]> v_head_389_cast_fp16 = reshape(shape = var_15759, x = var_15758_cast_fp16)[name = string("v_head_389_cast_fp16")];
            string dense_output_991_pad_type_1 = const()[name = string("dense_output_991_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_991_strides_1 = const()[name = string("dense_output_991_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_991_pad_1 = const()[name = string("dense_output_991_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_991_dilations_1 = const()[name = string("dense_output_991_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_991_groups_1 = const()[name = string("dense_output_991_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366109504))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366240640))))[name = string("layers_12_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_991_cast_fp16 = conv(dilations = dense_output_991_dilations_1, groups = dense_output_991_groups_1, pad = dense_output_991_pad_1, pad_type = dense_output_991_pad_type_1, strides = dense_output_991_strides_1, weight = layers_12_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_991_cast_fp16")];
            string sparse_output_991_pad_type_1 = const()[name = string("sparse_output_991_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_991_strides_1 = const()[name = string("sparse_output_991_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_991_pad_1 = const()[name = string("sparse_output_991_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_991_dilations_1 = const()[name = string("sparse_output_991_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_991_groups_1 = const()[name = string("sparse_output_991_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366243904))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366241216))))[name = string("layers_12_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_991_cast_fp16 = conv(dilations = sparse_output_991_dilations_1, groups = sparse_output_991_groups_1, pad = sparse_output_991_pad_1, pad_type = sparse_output_991_pad_type_1, strides = sparse_output_991_strides_1, weight = layers_12_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_991_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_15775_cast_fp16 = add(x = dense_output_991_cast_fp16, y = sparse_output_991_cast_fp16)[name = string("op_15775_cast_fp16")];
            tensor<int32, [4]> var_15776 = const()[name = string("op_15776"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15778 = const()[name = string("op_15778"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_15777_cast_fp16 = transpose(perm = var_15776, x = var_15775_cast_fp16)[name = string("transpose_481")];
            tensor<fp16, [1, 101, 128]> p_head_389_cast_fp16 = reshape(shape = var_15778, x = var_15777_cast_fp16)[name = string("p_head_389_cast_fp16")];
            tensor<fp16, [128]> var_15780_to_fp16 = const()[name = string("op_15780_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366260352)))];
            tensor<fp16, [1, 51, 128]> var_15781_cast_fp16 = add(x = q_head_195_cast_fp16, y = var_15780_to_fp16)[name = string("op_15781_cast_fp16")];
            tensor<int32, [1]> q_u_195_axes_1 = const()[name = string("q_u_195_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_195_cast_fp16 = expand_dims(axes = q_u_195_axes_1, x = var_15781_cast_fp16)[name = string("q_u_195_cast_fp16")];
            tensor<fp16, [128]> var_15783_to_fp16 = const()[name = string("op_15783_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366260672)))];
            tensor<fp16, [1, 51, 128]> var_15784_cast_fp16 = add(x = q_head_195_cast_fp16, y = var_15783_to_fp16)[name = string("op_15784_cast_fp16")];
            tensor<int32, [1]> q_v_195_axes_1 = const()[name = string("q_v_195_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_195_cast_fp16 = expand_dims(axes = q_v_195_axes_1, x = var_15784_cast_fp16)[name = string("q_v_195_cast_fp16")];
            tensor<int32, [1]> k_head_391_axes_1 = const()[name = string("k_head_391_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_391_cast_fp16 = expand_dims(axes = k_head_391_axes_1, x = k_head_389_cast_fp16)[name = string("k_head_391_cast_fp16")];
            tensor<int32, [1]> v_head_391_axes_1 = const()[name = string("v_head_391_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_391_cast_fp16 = expand_dims(axes = v_head_391_axes_1, x = v_head_389_cast_fp16)[name = string("v_head_391_cast_fp16")];
            tensor<int32, [1]> p_head_391_axes_1 = const()[name = string("p_head_391_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_391_cast_fp16 = expand_dims(axes = p_head_391_axes_1, x = p_head_389_cast_fp16)[name = string("p_head_391_cast_fp16")];
            bool var_15790_transpose_x_3 = const()[name = string("op_15790_transpose_x_3"), val = bool(false)];
            bool var_15790_transpose_y_3 = const()[name = string("op_15790_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_15790_cast_fp16 = matmul(transpose_x = var_15790_transpose_x_3, transpose_y = var_15790_transpose_y_3, x = q_u_195_cast_fp16, y = k_head_391_cast_fp16)[name = string("op_15790_cast_fp16")];
            fp16 var_15791_to_fp16 = const()[name = string("op_15791_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_195_cast_fp16 = mul(x = var_15790_cast_fp16, y = var_15791_to_fp16)[name = string("scores_content_195_cast_fp16")];
            bool x_1037_transpose_x_3 = const()[name = string("x_1037_transpose_x_3"), val = bool(false)];
            bool x_1037_transpose_y_3 = const()[name = string("x_1037_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1037_cast_fp16 = matmul(transpose_x = x_1037_transpose_x_3, transpose_y = x_1037_transpose_y_3, x = q_v_195_cast_fp16, y = p_head_391_cast_fp16)[name = string("x_1037_cast_fp16")];
            tensor<int32, [8]> x_1039_pad_1 = const()[name = string("x_1039_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1039_mode_1 = const()[name = string("x_1039_mode_1"), val = string("constant")];
            fp16 const_1963_to_fp16 = const()[name = string("const_1963_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1039_cast_fp16 = pad(constant_val = const_1963_to_fp16, mode = x_1039_mode_1, pad = x_1039_pad_1, x = x_1037_cast_fp16)[name = string("x_1039_cast_fp16")];
            tensor<int32, [4]> var_15805 = const()[name = string("op_15805"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1041_cast_fp16 = reshape(shape = var_15805, x = x_1039_cast_fp16)[name = string("x_1041_cast_fp16")];
            tensor<int32, [4]> var_15809_begin_1 = const()[name = string("op_15809_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15809_end_1 = const()[name = string("op_15809_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_15809_end_mask_1 = const()[name = string("op_15809_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_15809_cast_fp16 = slice_by_index(begin = var_15809_begin_1, end = var_15809_end_1, end_mask = var_15809_end_mask_1, x = x_1041_cast_fp16)[name = string("op_15809_cast_fp16")];
            tensor<int32, [4]> var_15811 = const()[name = string("op_15811"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_15812_cast_fp16 = reshape(shape = var_15811, x = var_15809_cast_fp16)[name = string("op_15812_cast_fp16")];
            tensor<int32, [4]> var_15817_begin_1 = const()[name = string("op_15817_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15817_end_1 = const()[name = string("op_15817_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_15817_end_mask_1 = const()[name = string("op_15817_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_15817_cast_fp16 = slice_by_index(begin = var_15817_begin_1, end = var_15817_end_1, end_mask = var_15817_end_mask_1, x = var_15812_cast_fp16)[name = string("op_15817_cast_fp16")];
            fp16 var_15818_to_fp16 = const()[name = string("op_15818_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_195_cast_fp16 = mul(x = var_15817_cast_fp16, y = var_15818_to_fp16)[name = string("scores_pos_195_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_195_cast_fp16 = add(x = scores_content_195_cast_fp16, y = scores_pos_195_cast_fp16)[name = string("logits_195_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_15821_cast_fp16 = softmax(axis = var_15425, x = logits_195_cast_fp16)[name = string("op_15821_cast_fp16")];
            bool var_15823_transpose_x_1 = const()[name = string("op_15823_transpose_x_1"), val = bool(false)];
            bool var_15823_transpose_y_1 = const()[name = string("op_15823_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_15823_cast_fp16 = matmul(transpose_x = var_15823_transpose_x_1, transpose_y = var_15823_transpose_y_1, x = var_15821_cast_fp16, y = v_head_391_cast_fp16)[name = string("op_15823_cast_fp16")];
            tensor<int32, [1]> var_15824_axes_1 = const()[name = string("op_15824_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_15824_cast_fp16 = squeeze(axes = var_15824_axes_1, x = var_15823_cast_fp16)[name = string("op_15824_cast_fp16")];
            string dense_output_993_pad_type_1 = const()[name = string("dense_output_993_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_993_strides_1 = const()[name = string("dense_output_993_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_993_pad_1 = const()[name = string("dense_output_993_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_993_dilations_1 = const()[name = string("dense_output_993_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_993_groups_1 = const()[name = string("dense_output_993_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366260992))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366392128))))[name = string("layers_12_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_993_cast_fp16 = conv(dilations = dense_output_993_dilations_1, groups = dense_output_993_groups_1, pad = dense_output_993_pad_1, pad_type = dense_output_993_pad_type_1, strides = dense_output_993_strides_1, weight = layers_12_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_993_cast_fp16")];
            string sparse_output_993_pad_type_1 = const()[name = string("sparse_output_993_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_993_strides_1 = const()[name = string("sparse_output_993_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_993_pad_1 = const()[name = string("sparse_output_993_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_993_dilations_1 = const()[name = string("sparse_output_993_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_993_groups_1 = const()[name = string("sparse_output_993_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366395392))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366392704))))[name = string("layers_12_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_993_cast_fp16 = conv(dilations = sparse_output_993_dilations_1, groups = sparse_output_993_groups_1, pad = sparse_output_993_pad_1, pad_type = sparse_output_993_pad_type_1, strides = sparse_output_993_strides_1, weight = layers_12_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_993_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15839_cast_fp16 = add(x = dense_output_993_cast_fp16, y = sparse_output_993_cast_fp16)[name = string("op_15839_cast_fp16")];
            tensor<int32, [4]> var_15840 = const()[name = string("op_15840"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15842 = const()[name = string("op_15842"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15841_cast_fp16 = transpose(perm = var_15840, x = var_15839_cast_fp16)[name = string("transpose_480")];
            tensor<fp16, [1, 51, 128]> q_head_197_cast_fp16 = reshape(shape = var_15842, x = var_15841_cast_fp16)[name = string("q_head_197_cast_fp16")];
            string dense_output_995_pad_type_1 = const()[name = string("dense_output_995_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_995_strides_1 = const()[name = string("dense_output_995_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_995_pad_1 = const()[name = string("dense_output_995_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_995_dilations_1 = const()[name = string("dense_output_995_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_995_groups_1 = const()[name = string("dense_output_995_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366411840))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366542976))))[name = string("layers_12_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_995_cast_fp16 = conv(dilations = dense_output_995_dilations_1, groups = dense_output_995_groups_1, pad = dense_output_995_pad_1, pad_type = dense_output_995_pad_type_1, strides = dense_output_995_strides_1, weight = layers_12_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_995_cast_fp16")];
            string sparse_output_995_pad_type_1 = const()[name = string("sparse_output_995_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_995_strides_1 = const()[name = string("sparse_output_995_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_995_pad_1 = const()[name = string("sparse_output_995_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_995_dilations_1 = const()[name = string("sparse_output_995_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_995_groups_1 = const()[name = string("sparse_output_995_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366546240))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366543552))))[name = string("layers_12_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_995_cast_fp16 = conv(dilations = sparse_output_995_dilations_1, groups = sparse_output_995_groups_1, pad = sparse_output_995_pad_1, pad_type = sparse_output_995_pad_type_1, strides = sparse_output_995_strides_1, weight = layers_12_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_995_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15858_cast_fp16 = add(x = dense_output_995_cast_fp16, y = sparse_output_995_cast_fp16)[name = string("op_15858_cast_fp16")];
            tensor<int32, [4]> var_15859 = const()[name = string("op_15859"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15861 = const()[name = string("op_15861"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15860_cast_fp16 = transpose(perm = var_15859, x = var_15858_cast_fp16)[name = string("transpose_479")];
            tensor<fp16, [1, 51, 128]> k_head_393_cast_fp16 = reshape(shape = var_15861, x = var_15860_cast_fp16)[name = string("k_head_393_cast_fp16")];
            string dense_output_997_pad_type_1 = const()[name = string("dense_output_997_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_997_strides_1 = const()[name = string("dense_output_997_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_997_pad_1 = const()[name = string("dense_output_997_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_997_dilations_1 = const()[name = string("dense_output_997_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_997_groups_1 = const()[name = string("dense_output_997_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366562688))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366693824))))[name = string("layers_12_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_997_cast_fp16 = conv(dilations = dense_output_997_dilations_1, groups = dense_output_997_groups_1, pad = dense_output_997_pad_1, pad_type = dense_output_997_pad_type_1, strides = dense_output_997_strides_1, weight = layers_12_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_997_cast_fp16")];
            string sparse_output_997_pad_type_1 = const()[name = string("sparse_output_997_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_997_strides_1 = const()[name = string("sparse_output_997_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_997_pad_1 = const()[name = string("sparse_output_997_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_997_dilations_1 = const()[name = string("sparse_output_997_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_997_groups_1 = const()[name = string("sparse_output_997_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366697088))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366694400))))[name = string("layers_12_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_997_cast_fp16 = conv(dilations = sparse_output_997_dilations_1, groups = sparse_output_997_groups_1, pad = sparse_output_997_pad_1, pad_type = sparse_output_997_pad_type_1, strides = sparse_output_997_strides_1, weight = layers_12_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_997_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15877_cast_fp16 = add(x = dense_output_997_cast_fp16, y = sparse_output_997_cast_fp16)[name = string("op_15877_cast_fp16")];
            tensor<int32, [4]> var_15878 = const()[name = string("op_15878"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15880 = const()[name = string("op_15880"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15879_cast_fp16 = transpose(perm = var_15878, x = var_15877_cast_fp16)[name = string("transpose_478")];
            tensor<fp16, [1, 51, 128]> v_head_393_cast_fp16 = reshape(shape = var_15880, x = var_15879_cast_fp16)[name = string("v_head_393_cast_fp16")];
            string dense_output_999_pad_type_1 = const()[name = string("dense_output_999_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_999_strides_1 = const()[name = string("dense_output_999_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_999_pad_1 = const()[name = string("dense_output_999_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_999_dilations_1 = const()[name = string("dense_output_999_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_999_groups_1 = const()[name = string("dense_output_999_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366713536))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366844672))))[name = string("layers_12_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_999_cast_fp16 = conv(dilations = dense_output_999_dilations_1, groups = dense_output_999_groups_1, pad = dense_output_999_pad_1, pad_type = dense_output_999_pad_type_1, strides = dense_output_999_strides_1, weight = layers_12_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_999_cast_fp16")];
            string sparse_output_999_pad_type_1 = const()[name = string("sparse_output_999_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_999_strides_1 = const()[name = string("sparse_output_999_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_999_pad_1 = const()[name = string("sparse_output_999_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_999_dilations_1 = const()[name = string("sparse_output_999_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_999_groups_1 = const()[name = string("sparse_output_999_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366847936))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366845248))))[name = string("layers_12_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_999_cast_fp16 = conv(dilations = sparse_output_999_dilations_1, groups = sparse_output_999_groups_1, pad = sparse_output_999_pad_1, pad_type = sparse_output_999_pad_type_1, strides = sparse_output_999_strides_1, weight = layers_12_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_999_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_15896_cast_fp16 = add(x = dense_output_999_cast_fp16, y = sparse_output_999_cast_fp16)[name = string("op_15896_cast_fp16")];
            tensor<int32, [4]> var_15897 = const()[name = string("op_15897"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15899 = const()[name = string("op_15899"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_15898_cast_fp16 = transpose(perm = var_15897, x = var_15896_cast_fp16)[name = string("transpose_477")];
            tensor<fp16, [1, 101, 128]> p_head_393_cast_fp16 = reshape(shape = var_15899, x = var_15898_cast_fp16)[name = string("p_head_393_cast_fp16")];
            tensor<fp16, [128]> var_15901_to_fp16 = const()[name = string("op_15901_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366864384)))];
            tensor<fp16, [1, 51, 128]> var_15902_cast_fp16 = add(x = q_head_197_cast_fp16, y = var_15901_to_fp16)[name = string("op_15902_cast_fp16")];
            tensor<int32, [1]> q_u_197_axes_1 = const()[name = string("q_u_197_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_197_cast_fp16 = expand_dims(axes = q_u_197_axes_1, x = var_15902_cast_fp16)[name = string("q_u_197_cast_fp16")];
            tensor<fp16, [128]> var_15904_to_fp16 = const()[name = string("op_15904_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366864704)))];
            tensor<fp16, [1, 51, 128]> var_15905_cast_fp16 = add(x = q_head_197_cast_fp16, y = var_15904_to_fp16)[name = string("op_15905_cast_fp16")];
            tensor<int32, [1]> q_v_197_axes_1 = const()[name = string("q_v_197_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_197_cast_fp16 = expand_dims(axes = q_v_197_axes_1, x = var_15905_cast_fp16)[name = string("q_v_197_cast_fp16")];
            tensor<int32, [1]> k_head_395_axes_1 = const()[name = string("k_head_395_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_395_cast_fp16 = expand_dims(axes = k_head_395_axes_1, x = k_head_393_cast_fp16)[name = string("k_head_395_cast_fp16")];
            tensor<int32, [1]> v_head_395_axes_1 = const()[name = string("v_head_395_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_395_cast_fp16 = expand_dims(axes = v_head_395_axes_1, x = v_head_393_cast_fp16)[name = string("v_head_395_cast_fp16")];
            tensor<int32, [1]> p_head_395_axes_1 = const()[name = string("p_head_395_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_395_cast_fp16 = expand_dims(axes = p_head_395_axes_1, x = p_head_393_cast_fp16)[name = string("p_head_395_cast_fp16")];
            bool var_15911_transpose_x_3 = const()[name = string("op_15911_transpose_x_3"), val = bool(false)];
            bool var_15911_transpose_y_3 = const()[name = string("op_15911_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_15911_cast_fp16 = matmul(transpose_x = var_15911_transpose_x_3, transpose_y = var_15911_transpose_y_3, x = q_u_197_cast_fp16, y = k_head_395_cast_fp16)[name = string("op_15911_cast_fp16")];
            fp16 var_15912_to_fp16 = const()[name = string("op_15912_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_197_cast_fp16 = mul(x = var_15911_cast_fp16, y = var_15912_to_fp16)[name = string("scores_content_197_cast_fp16")];
            bool x_1045_transpose_x_3 = const()[name = string("x_1045_transpose_x_3"), val = bool(false)];
            bool x_1045_transpose_y_3 = const()[name = string("x_1045_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1045_cast_fp16 = matmul(transpose_x = x_1045_transpose_x_3, transpose_y = x_1045_transpose_y_3, x = q_v_197_cast_fp16, y = p_head_395_cast_fp16)[name = string("x_1045_cast_fp16")];
            tensor<int32, [8]> x_1047_pad_1 = const()[name = string("x_1047_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1047_mode_1 = const()[name = string("x_1047_mode_1"), val = string("constant")];
            fp16 const_1969_to_fp16 = const()[name = string("const_1969_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1047_cast_fp16 = pad(constant_val = const_1969_to_fp16, mode = x_1047_mode_1, pad = x_1047_pad_1, x = x_1045_cast_fp16)[name = string("x_1047_cast_fp16")];
            tensor<int32, [4]> var_15926 = const()[name = string("op_15926"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1049_cast_fp16 = reshape(shape = var_15926, x = x_1047_cast_fp16)[name = string("x_1049_cast_fp16")];
            tensor<int32, [4]> var_15930_begin_1 = const()[name = string("op_15930_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_15930_end_1 = const()[name = string("op_15930_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_15930_end_mask_1 = const()[name = string("op_15930_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_15930_cast_fp16 = slice_by_index(begin = var_15930_begin_1, end = var_15930_end_1, end_mask = var_15930_end_mask_1, x = x_1049_cast_fp16)[name = string("op_15930_cast_fp16")];
            tensor<int32, [4]> var_15932 = const()[name = string("op_15932"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_15933_cast_fp16 = reshape(shape = var_15932, x = var_15930_cast_fp16)[name = string("op_15933_cast_fp16")];
            tensor<int32, [4]> var_15938_begin_1 = const()[name = string("op_15938_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_15938_end_1 = const()[name = string("op_15938_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_15938_end_mask_1 = const()[name = string("op_15938_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_15938_cast_fp16 = slice_by_index(begin = var_15938_begin_1, end = var_15938_end_1, end_mask = var_15938_end_mask_1, x = var_15933_cast_fp16)[name = string("op_15938_cast_fp16")];
            fp16 var_15939_to_fp16 = const()[name = string("op_15939_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_197_cast_fp16 = mul(x = var_15938_cast_fp16, y = var_15939_to_fp16)[name = string("scores_pos_197_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_197_cast_fp16 = add(x = scores_content_197_cast_fp16, y = scores_pos_197_cast_fp16)[name = string("logits_197_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_15942_cast_fp16 = softmax(axis = var_15425, x = logits_197_cast_fp16)[name = string("op_15942_cast_fp16")];
            bool var_15944_transpose_x_1 = const()[name = string("op_15944_transpose_x_1"), val = bool(false)];
            bool var_15944_transpose_y_1 = const()[name = string("op_15944_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_15944_cast_fp16 = matmul(transpose_x = var_15944_transpose_x_1, transpose_y = var_15944_transpose_y_1, x = var_15942_cast_fp16, y = v_head_395_cast_fp16)[name = string("op_15944_cast_fp16")];
            tensor<int32, [1]> var_15945_axes_1 = const()[name = string("op_15945_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_15945_cast_fp16 = squeeze(axes = var_15945_axes_1, x = var_15944_cast_fp16)[name = string("op_15945_cast_fp16")];
            string dense_output_1001_pad_type_1 = const()[name = string("dense_output_1001_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1001_strides_1 = const()[name = string("dense_output_1001_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1001_pad_1 = const()[name = string("dense_output_1001_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1001_dilations_1 = const()[name = string("dense_output_1001_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1001_groups_1 = const()[name = string("dense_output_1001_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366865024))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366996160))))[name = string("layers_12_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1001_cast_fp16 = conv(dilations = dense_output_1001_dilations_1, groups = dense_output_1001_groups_1, pad = dense_output_1001_pad_1, pad_type = dense_output_1001_pad_type_1, strides = dense_output_1001_strides_1, weight = layers_12_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1001_cast_fp16")];
            string sparse_output_1001_pad_type_1 = const()[name = string("sparse_output_1001_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1001_strides_1 = const()[name = string("sparse_output_1001_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1001_pad_1 = const()[name = string("sparse_output_1001_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1001_dilations_1 = const()[name = string("sparse_output_1001_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1001_groups_1 = const()[name = string("sparse_output_1001_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366999424))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366996736))))[name = string("layers_12_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1001_cast_fp16 = conv(dilations = sparse_output_1001_dilations_1, groups = sparse_output_1001_groups_1, pad = sparse_output_1001_pad_1, pad_type = sparse_output_1001_pad_type_1, strides = sparse_output_1001_strides_1, weight = layers_12_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1001_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15960_cast_fp16 = add(x = dense_output_1001_cast_fp16, y = sparse_output_1001_cast_fp16)[name = string("op_15960_cast_fp16")];
            tensor<int32, [4]> var_15961 = const()[name = string("op_15961"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15963 = const()[name = string("op_15963"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15962_cast_fp16 = transpose(perm = var_15961, x = var_15960_cast_fp16)[name = string("transpose_476")];
            tensor<fp16, [1, 51, 128]> q_head_199_cast_fp16 = reshape(shape = var_15963, x = var_15962_cast_fp16)[name = string("q_head_199_cast_fp16")];
            string dense_output_1003_pad_type_1 = const()[name = string("dense_output_1003_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1003_strides_1 = const()[name = string("dense_output_1003_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1003_pad_1 = const()[name = string("dense_output_1003_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1003_dilations_1 = const()[name = string("dense_output_1003_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1003_groups_1 = const()[name = string("dense_output_1003_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367015872))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367147008))))[name = string("layers_12_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1003_cast_fp16 = conv(dilations = dense_output_1003_dilations_1, groups = dense_output_1003_groups_1, pad = dense_output_1003_pad_1, pad_type = dense_output_1003_pad_type_1, strides = dense_output_1003_strides_1, weight = layers_12_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1003_cast_fp16")];
            string sparse_output_1003_pad_type_1 = const()[name = string("sparse_output_1003_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1003_strides_1 = const()[name = string("sparse_output_1003_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1003_pad_1 = const()[name = string("sparse_output_1003_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1003_dilations_1 = const()[name = string("sparse_output_1003_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1003_groups_1 = const()[name = string("sparse_output_1003_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367150272))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367147584))))[name = string("layers_12_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1003_cast_fp16 = conv(dilations = sparse_output_1003_dilations_1, groups = sparse_output_1003_groups_1, pad = sparse_output_1003_pad_1, pad_type = sparse_output_1003_pad_type_1, strides = sparse_output_1003_strides_1, weight = layers_12_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1003_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15979_cast_fp16 = add(x = dense_output_1003_cast_fp16, y = sparse_output_1003_cast_fp16)[name = string("op_15979_cast_fp16")];
            tensor<int32, [4]> var_15980 = const()[name = string("op_15980"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_15982 = const()[name = string("op_15982"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_15981_cast_fp16 = transpose(perm = var_15980, x = var_15979_cast_fp16)[name = string("transpose_475")];
            tensor<fp16, [1, 51, 128]> k_head_397_cast_fp16 = reshape(shape = var_15982, x = var_15981_cast_fp16)[name = string("k_head_397_cast_fp16")];
            string dense_output_1005_pad_type_1 = const()[name = string("dense_output_1005_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1005_strides_1 = const()[name = string("dense_output_1005_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1005_pad_1 = const()[name = string("dense_output_1005_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1005_dilations_1 = const()[name = string("dense_output_1005_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1005_groups_1 = const()[name = string("dense_output_1005_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367166720))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367297856))))[name = string("layers_12_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1005_cast_fp16 = conv(dilations = dense_output_1005_dilations_1, groups = dense_output_1005_groups_1, pad = dense_output_1005_pad_1, pad_type = dense_output_1005_pad_type_1, strides = dense_output_1005_strides_1, weight = layers_12_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1005_cast_fp16")];
            string sparse_output_1005_pad_type_1 = const()[name = string("sparse_output_1005_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1005_strides_1 = const()[name = string("sparse_output_1005_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1005_pad_1 = const()[name = string("sparse_output_1005_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1005_dilations_1 = const()[name = string("sparse_output_1005_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1005_groups_1 = const()[name = string("sparse_output_1005_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367301120))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367298432))))[name = string("layers_12_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1005_cast_fp16 = conv(dilations = sparse_output_1005_dilations_1, groups = sparse_output_1005_groups_1, pad = sparse_output_1005_pad_1, pad_type = sparse_output_1005_pad_type_1, strides = sparse_output_1005_strides_1, weight = layers_12_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1005_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_15998_cast_fp16 = add(x = dense_output_1005_cast_fp16, y = sparse_output_1005_cast_fp16)[name = string("op_15998_cast_fp16")];
            tensor<int32, [4]> var_15999 = const()[name = string("op_15999"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16001 = const()[name = string("op_16001"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16000_cast_fp16 = transpose(perm = var_15999, x = var_15998_cast_fp16)[name = string("transpose_474")];
            tensor<fp16, [1, 51, 128]> v_head_397_cast_fp16 = reshape(shape = var_16001, x = var_16000_cast_fp16)[name = string("v_head_397_cast_fp16")];
            string dense_output_1007_pad_type_1 = const()[name = string("dense_output_1007_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1007_strides_1 = const()[name = string("dense_output_1007_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1007_pad_1 = const()[name = string("dense_output_1007_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1007_dilations_1 = const()[name = string("dense_output_1007_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1007_groups_1 = const()[name = string("dense_output_1007_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367317568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367448704))))[name = string("layers_12_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1007_cast_fp16 = conv(dilations = dense_output_1007_dilations_1, groups = dense_output_1007_groups_1, pad = dense_output_1007_pad_1, pad_type = dense_output_1007_pad_type_1, strides = dense_output_1007_strides_1, weight = layers_12_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1007_cast_fp16")];
            string sparse_output_1007_pad_type_1 = const()[name = string("sparse_output_1007_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1007_strides_1 = const()[name = string("sparse_output_1007_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1007_pad_1 = const()[name = string("sparse_output_1007_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1007_dilations_1 = const()[name = string("sparse_output_1007_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1007_groups_1 = const()[name = string("sparse_output_1007_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367451968))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367449280))))[name = string("layers_12_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1007_cast_fp16 = conv(dilations = sparse_output_1007_dilations_1, groups = sparse_output_1007_groups_1, pad = sparse_output_1007_pad_1, pad_type = sparse_output_1007_pad_type_1, strides = sparse_output_1007_strides_1, weight = layers_12_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1007_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_16017_cast_fp16 = add(x = dense_output_1007_cast_fp16, y = sparse_output_1007_cast_fp16)[name = string("op_16017_cast_fp16")];
            tensor<int32, [4]> var_16018 = const()[name = string("op_16018"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16020 = const()[name = string("op_16020"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_16019_cast_fp16 = transpose(perm = var_16018, x = var_16017_cast_fp16)[name = string("transpose_473")];
            tensor<fp16, [1, 101, 128]> p_head_397_cast_fp16 = reshape(shape = var_16020, x = var_16019_cast_fp16)[name = string("p_head_397_cast_fp16")];
            tensor<fp16, [128]> var_16022_to_fp16 = const()[name = string("op_16022_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367468416)))];
            tensor<fp16, [1, 51, 128]> var_16023_cast_fp16 = add(x = q_head_199_cast_fp16, y = var_16022_to_fp16)[name = string("op_16023_cast_fp16")];
            tensor<int32, [1]> q_u_199_axes_1 = const()[name = string("q_u_199_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_199_cast_fp16 = expand_dims(axes = q_u_199_axes_1, x = var_16023_cast_fp16)[name = string("q_u_199_cast_fp16")];
            tensor<fp16, [128]> var_16025_to_fp16 = const()[name = string("op_16025_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367468736)))];
            tensor<fp16, [1, 51, 128]> var_16026_cast_fp16 = add(x = q_head_199_cast_fp16, y = var_16025_to_fp16)[name = string("op_16026_cast_fp16")];
            tensor<int32, [1]> q_v_199_axes_1 = const()[name = string("q_v_199_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_199_cast_fp16 = expand_dims(axes = q_v_199_axes_1, x = var_16026_cast_fp16)[name = string("q_v_199_cast_fp16")];
            tensor<int32, [1]> k_head_399_axes_1 = const()[name = string("k_head_399_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_399_cast_fp16 = expand_dims(axes = k_head_399_axes_1, x = k_head_397_cast_fp16)[name = string("k_head_399_cast_fp16")];
            tensor<int32, [1]> v_head_399_axes_1 = const()[name = string("v_head_399_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_399_cast_fp16 = expand_dims(axes = v_head_399_axes_1, x = v_head_397_cast_fp16)[name = string("v_head_399_cast_fp16")];
            tensor<int32, [1]> p_head_399_axes_1 = const()[name = string("p_head_399_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_399_cast_fp16 = expand_dims(axes = p_head_399_axes_1, x = p_head_397_cast_fp16)[name = string("p_head_399_cast_fp16")];
            bool var_16032_transpose_x_3 = const()[name = string("op_16032_transpose_x_3"), val = bool(false)];
            bool var_16032_transpose_y_3 = const()[name = string("op_16032_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_16032_cast_fp16 = matmul(transpose_x = var_16032_transpose_x_3, transpose_y = var_16032_transpose_y_3, x = q_u_199_cast_fp16, y = k_head_399_cast_fp16)[name = string("op_16032_cast_fp16")];
            fp16 var_16033_to_fp16 = const()[name = string("op_16033_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_199_cast_fp16 = mul(x = var_16032_cast_fp16, y = var_16033_to_fp16)[name = string("scores_content_199_cast_fp16")];
            bool x_1053_transpose_x_3 = const()[name = string("x_1053_transpose_x_3"), val = bool(false)];
            bool x_1053_transpose_y_3 = const()[name = string("x_1053_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1053_cast_fp16 = matmul(transpose_x = x_1053_transpose_x_3, transpose_y = x_1053_transpose_y_3, x = q_v_199_cast_fp16, y = p_head_399_cast_fp16)[name = string("x_1053_cast_fp16")];
            tensor<int32, [8]> x_1055_pad_1 = const()[name = string("x_1055_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1055_mode_1 = const()[name = string("x_1055_mode_1"), val = string("constant")];
            fp16 const_1975_to_fp16 = const()[name = string("const_1975_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1055_cast_fp16 = pad(constant_val = const_1975_to_fp16, mode = x_1055_mode_1, pad = x_1055_pad_1, x = x_1053_cast_fp16)[name = string("x_1055_cast_fp16")];
            tensor<int32, [4]> var_16047 = const()[name = string("op_16047"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1057_cast_fp16 = reshape(shape = var_16047, x = x_1055_cast_fp16)[name = string("x_1057_cast_fp16")];
            tensor<int32, [4]> var_16051_begin_1 = const()[name = string("op_16051_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16051_end_1 = const()[name = string("op_16051_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_16051_end_mask_1 = const()[name = string("op_16051_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_16051_cast_fp16 = slice_by_index(begin = var_16051_begin_1, end = var_16051_end_1, end_mask = var_16051_end_mask_1, x = x_1057_cast_fp16)[name = string("op_16051_cast_fp16")];
            tensor<int32, [4]> var_16053 = const()[name = string("op_16053"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_16054_cast_fp16 = reshape(shape = var_16053, x = var_16051_cast_fp16)[name = string("op_16054_cast_fp16")];
            tensor<int32, [4]> var_16059_begin_1 = const()[name = string("op_16059_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16059_end_1 = const()[name = string("op_16059_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_16059_end_mask_1 = const()[name = string("op_16059_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_16059_cast_fp16 = slice_by_index(begin = var_16059_begin_1, end = var_16059_end_1, end_mask = var_16059_end_mask_1, x = var_16054_cast_fp16)[name = string("op_16059_cast_fp16")];
            fp16 var_16060_to_fp16 = const()[name = string("op_16060_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_199_cast_fp16 = mul(x = var_16059_cast_fp16, y = var_16060_to_fp16)[name = string("scores_pos_199_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_199_cast_fp16 = add(x = scores_content_199_cast_fp16, y = scores_pos_199_cast_fp16)[name = string("logits_199_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_16063_cast_fp16 = softmax(axis = var_15425, x = logits_199_cast_fp16)[name = string("op_16063_cast_fp16")];
            bool var_16065_transpose_x_1 = const()[name = string("op_16065_transpose_x_1"), val = bool(false)];
            bool var_16065_transpose_y_1 = const()[name = string("op_16065_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_16065_cast_fp16 = matmul(transpose_x = var_16065_transpose_x_1, transpose_y = var_16065_transpose_y_1, x = var_16063_cast_fp16, y = v_head_399_cast_fp16)[name = string("op_16065_cast_fp16")];
            tensor<int32, [1]> var_16066_axes_1 = const()[name = string("op_16066_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_16066_cast_fp16 = squeeze(axes = var_16066_axes_1, x = var_16065_cast_fp16)[name = string("op_16066_cast_fp16")];
            string dense_output_1009_pad_type_1 = const()[name = string("dense_output_1009_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1009_strides_1 = const()[name = string("dense_output_1009_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1009_pad_1 = const()[name = string("dense_output_1009_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1009_dilations_1 = const()[name = string("dense_output_1009_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1009_groups_1 = const()[name = string("dense_output_1009_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367469056))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367600192))))[name = string("layers_12_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1009_cast_fp16 = conv(dilations = dense_output_1009_dilations_1, groups = dense_output_1009_groups_1, pad = dense_output_1009_pad_1, pad_type = dense_output_1009_pad_type_1, strides = dense_output_1009_strides_1, weight = layers_12_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1009_cast_fp16")];
            string sparse_output_1009_pad_type_1 = const()[name = string("sparse_output_1009_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1009_strides_1 = const()[name = string("sparse_output_1009_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1009_pad_1 = const()[name = string("sparse_output_1009_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1009_dilations_1 = const()[name = string("sparse_output_1009_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1009_groups_1 = const()[name = string("sparse_output_1009_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367603456))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367600768))))[name = string("layers_12_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1009_cast_fp16 = conv(dilations = sparse_output_1009_dilations_1, groups = sparse_output_1009_groups_1, pad = sparse_output_1009_pad_1, pad_type = sparse_output_1009_pad_type_1, strides = sparse_output_1009_strides_1, weight = layers_12_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1009_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16081_cast_fp16 = add(x = dense_output_1009_cast_fp16, y = sparse_output_1009_cast_fp16)[name = string("op_16081_cast_fp16")];
            tensor<int32, [4]> var_16082 = const()[name = string("op_16082"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16084 = const()[name = string("op_16084"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16083_cast_fp16 = transpose(perm = var_16082, x = var_16081_cast_fp16)[name = string("transpose_472")];
            tensor<fp16, [1, 51, 128]> q_head_201_cast_fp16 = reshape(shape = var_16084, x = var_16083_cast_fp16)[name = string("q_head_201_cast_fp16")];
            string dense_output_1011_pad_type_1 = const()[name = string("dense_output_1011_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1011_strides_1 = const()[name = string("dense_output_1011_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1011_pad_1 = const()[name = string("dense_output_1011_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1011_dilations_1 = const()[name = string("dense_output_1011_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1011_groups_1 = const()[name = string("dense_output_1011_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367619904))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367751040))))[name = string("layers_12_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1011_cast_fp16 = conv(dilations = dense_output_1011_dilations_1, groups = dense_output_1011_groups_1, pad = dense_output_1011_pad_1, pad_type = dense_output_1011_pad_type_1, strides = dense_output_1011_strides_1, weight = layers_12_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1011_cast_fp16")];
            string sparse_output_1011_pad_type_1 = const()[name = string("sparse_output_1011_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1011_strides_1 = const()[name = string("sparse_output_1011_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1011_pad_1 = const()[name = string("sparse_output_1011_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1011_dilations_1 = const()[name = string("sparse_output_1011_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1011_groups_1 = const()[name = string("sparse_output_1011_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367754304))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367751616))))[name = string("layers_12_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1011_cast_fp16 = conv(dilations = sparse_output_1011_dilations_1, groups = sparse_output_1011_groups_1, pad = sparse_output_1011_pad_1, pad_type = sparse_output_1011_pad_type_1, strides = sparse_output_1011_strides_1, weight = layers_12_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1011_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16100_cast_fp16 = add(x = dense_output_1011_cast_fp16, y = sparse_output_1011_cast_fp16)[name = string("op_16100_cast_fp16")];
            tensor<int32, [4]> var_16101 = const()[name = string("op_16101"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16103 = const()[name = string("op_16103"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16102_cast_fp16 = transpose(perm = var_16101, x = var_16100_cast_fp16)[name = string("transpose_471")];
            tensor<fp16, [1, 51, 128]> k_head_401_cast_fp16 = reshape(shape = var_16103, x = var_16102_cast_fp16)[name = string("k_head_401_cast_fp16")];
            string dense_output_1013_pad_type_1 = const()[name = string("dense_output_1013_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1013_strides_1 = const()[name = string("dense_output_1013_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1013_pad_1 = const()[name = string("dense_output_1013_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1013_dilations_1 = const()[name = string("dense_output_1013_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1013_groups_1 = const()[name = string("dense_output_1013_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367770752))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367901888))))[name = string("layers_12_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1013_cast_fp16 = conv(dilations = dense_output_1013_dilations_1, groups = dense_output_1013_groups_1, pad = dense_output_1013_pad_1, pad_type = dense_output_1013_pad_type_1, strides = dense_output_1013_strides_1, weight = layers_12_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1013_cast_fp16")];
            string sparse_output_1013_pad_type_1 = const()[name = string("sparse_output_1013_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1013_strides_1 = const()[name = string("sparse_output_1013_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1013_pad_1 = const()[name = string("sparse_output_1013_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1013_dilations_1 = const()[name = string("sparse_output_1013_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1013_groups_1 = const()[name = string("sparse_output_1013_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367905152))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367902464))))[name = string("layers_12_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1013_cast_fp16 = conv(dilations = sparse_output_1013_dilations_1, groups = sparse_output_1013_groups_1, pad = sparse_output_1013_pad_1, pad_type = sparse_output_1013_pad_type_1, strides = sparse_output_1013_strides_1, weight = layers_12_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1013_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16119_cast_fp16 = add(x = dense_output_1013_cast_fp16, y = sparse_output_1013_cast_fp16)[name = string("op_16119_cast_fp16")];
            tensor<int32, [4]> var_16120 = const()[name = string("op_16120"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16122 = const()[name = string("op_16122"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16121_cast_fp16 = transpose(perm = var_16120, x = var_16119_cast_fp16)[name = string("transpose_470")];
            tensor<fp16, [1, 51, 128]> v_head_401_cast_fp16 = reshape(shape = var_16122, x = var_16121_cast_fp16)[name = string("v_head_401_cast_fp16")];
            string dense_output_1015_pad_type_1 = const()[name = string("dense_output_1015_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1015_strides_1 = const()[name = string("dense_output_1015_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1015_pad_1 = const()[name = string("dense_output_1015_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1015_dilations_1 = const()[name = string("dense_output_1015_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1015_groups_1 = const()[name = string("dense_output_1015_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367921600))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368052736))))[name = string("layers_12_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1015_cast_fp16 = conv(dilations = dense_output_1015_dilations_1, groups = dense_output_1015_groups_1, pad = dense_output_1015_pad_1, pad_type = dense_output_1015_pad_type_1, strides = dense_output_1015_strides_1, weight = layers_12_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1015_cast_fp16")];
            string sparse_output_1015_pad_type_1 = const()[name = string("sparse_output_1015_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1015_strides_1 = const()[name = string("sparse_output_1015_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1015_pad_1 = const()[name = string("sparse_output_1015_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1015_dilations_1 = const()[name = string("sparse_output_1015_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1015_groups_1 = const()[name = string("sparse_output_1015_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368056000))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368053312))))[name = string("layers_12_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1015_cast_fp16 = conv(dilations = sparse_output_1015_dilations_1, groups = sparse_output_1015_groups_1, pad = sparse_output_1015_pad_1, pad_type = sparse_output_1015_pad_type_1, strides = sparse_output_1015_strides_1, weight = layers_12_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1015_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_16138_cast_fp16 = add(x = dense_output_1015_cast_fp16, y = sparse_output_1015_cast_fp16)[name = string("op_16138_cast_fp16")];
            tensor<int32, [4]> var_16139 = const()[name = string("op_16139"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16141 = const()[name = string("op_16141"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_16140_cast_fp16 = transpose(perm = var_16139, x = var_16138_cast_fp16)[name = string("transpose_469")];
            tensor<fp16, [1, 101, 128]> p_head_401_cast_fp16 = reshape(shape = var_16141, x = var_16140_cast_fp16)[name = string("p_head_401_cast_fp16")];
            tensor<fp16, [128]> var_16143_to_fp16 = const()[name = string("op_16143_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368072448)))];
            tensor<fp16, [1, 51, 128]> var_16144_cast_fp16 = add(x = q_head_201_cast_fp16, y = var_16143_to_fp16)[name = string("op_16144_cast_fp16")];
            tensor<int32, [1]> q_u_201_axes_1 = const()[name = string("q_u_201_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_201_cast_fp16 = expand_dims(axes = q_u_201_axes_1, x = var_16144_cast_fp16)[name = string("q_u_201_cast_fp16")];
            tensor<fp16, [128]> var_16146_to_fp16 = const()[name = string("op_16146_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368072768)))];
            tensor<fp16, [1, 51, 128]> var_16147_cast_fp16 = add(x = q_head_201_cast_fp16, y = var_16146_to_fp16)[name = string("op_16147_cast_fp16")];
            tensor<int32, [1]> q_v_201_axes_1 = const()[name = string("q_v_201_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_201_cast_fp16 = expand_dims(axes = q_v_201_axes_1, x = var_16147_cast_fp16)[name = string("q_v_201_cast_fp16")];
            tensor<int32, [1]> k_head_403_axes_1 = const()[name = string("k_head_403_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_403_cast_fp16 = expand_dims(axes = k_head_403_axes_1, x = k_head_401_cast_fp16)[name = string("k_head_403_cast_fp16")];
            tensor<int32, [1]> v_head_403_axes_1 = const()[name = string("v_head_403_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_403_cast_fp16 = expand_dims(axes = v_head_403_axes_1, x = v_head_401_cast_fp16)[name = string("v_head_403_cast_fp16")];
            tensor<int32, [1]> p_head_403_axes_1 = const()[name = string("p_head_403_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_403_cast_fp16 = expand_dims(axes = p_head_403_axes_1, x = p_head_401_cast_fp16)[name = string("p_head_403_cast_fp16")];
            bool var_16153_transpose_x_3 = const()[name = string("op_16153_transpose_x_3"), val = bool(false)];
            bool var_16153_transpose_y_3 = const()[name = string("op_16153_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_16153_cast_fp16 = matmul(transpose_x = var_16153_transpose_x_3, transpose_y = var_16153_transpose_y_3, x = q_u_201_cast_fp16, y = k_head_403_cast_fp16)[name = string("op_16153_cast_fp16")];
            fp16 var_16154_to_fp16 = const()[name = string("op_16154_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_201_cast_fp16 = mul(x = var_16153_cast_fp16, y = var_16154_to_fp16)[name = string("scores_content_201_cast_fp16")];
            bool x_1061_transpose_x_3 = const()[name = string("x_1061_transpose_x_3"), val = bool(false)];
            bool x_1061_transpose_y_3 = const()[name = string("x_1061_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1061_cast_fp16 = matmul(transpose_x = x_1061_transpose_x_3, transpose_y = x_1061_transpose_y_3, x = q_v_201_cast_fp16, y = p_head_403_cast_fp16)[name = string("x_1061_cast_fp16")];
            tensor<int32, [8]> x_1063_pad_1 = const()[name = string("x_1063_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1063_mode_1 = const()[name = string("x_1063_mode_1"), val = string("constant")];
            fp16 const_1981_to_fp16 = const()[name = string("const_1981_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1063_cast_fp16 = pad(constant_val = const_1981_to_fp16, mode = x_1063_mode_1, pad = x_1063_pad_1, x = x_1061_cast_fp16)[name = string("x_1063_cast_fp16")];
            tensor<int32, [4]> var_16168 = const()[name = string("op_16168"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1065_cast_fp16 = reshape(shape = var_16168, x = x_1063_cast_fp16)[name = string("x_1065_cast_fp16")];
            tensor<int32, [4]> var_16172_begin_1 = const()[name = string("op_16172_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16172_end_1 = const()[name = string("op_16172_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_16172_end_mask_1 = const()[name = string("op_16172_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_16172_cast_fp16 = slice_by_index(begin = var_16172_begin_1, end = var_16172_end_1, end_mask = var_16172_end_mask_1, x = x_1065_cast_fp16)[name = string("op_16172_cast_fp16")];
            tensor<int32, [4]> var_16174 = const()[name = string("op_16174"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_16175_cast_fp16 = reshape(shape = var_16174, x = var_16172_cast_fp16)[name = string("op_16175_cast_fp16")];
            tensor<int32, [4]> var_16180_begin_1 = const()[name = string("op_16180_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16180_end_1 = const()[name = string("op_16180_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_16180_end_mask_1 = const()[name = string("op_16180_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_16180_cast_fp16 = slice_by_index(begin = var_16180_begin_1, end = var_16180_end_1, end_mask = var_16180_end_mask_1, x = var_16175_cast_fp16)[name = string("op_16180_cast_fp16")];
            fp16 var_16181_to_fp16 = const()[name = string("op_16181_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_201_cast_fp16 = mul(x = var_16180_cast_fp16, y = var_16181_to_fp16)[name = string("scores_pos_201_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_201_cast_fp16 = add(x = scores_content_201_cast_fp16, y = scores_pos_201_cast_fp16)[name = string("logits_201_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_16184_cast_fp16 = softmax(axis = var_15425, x = logits_201_cast_fp16)[name = string("op_16184_cast_fp16")];
            bool var_16186_transpose_x_1 = const()[name = string("op_16186_transpose_x_1"), val = bool(false)];
            bool var_16186_transpose_y_1 = const()[name = string("op_16186_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_16186_cast_fp16 = matmul(transpose_x = var_16186_transpose_x_1, transpose_y = var_16186_transpose_y_1, x = var_16184_cast_fp16, y = v_head_403_cast_fp16)[name = string("op_16186_cast_fp16")];
            tensor<int32, [1]> var_16187_axes_1 = const()[name = string("op_16187_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_16187_cast_fp16 = squeeze(axes = var_16187_axes_1, x = var_16186_cast_fp16)[name = string("op_16187_cast_fp16")];
            string dense_output_1017_pad_type_1 = const()[name = string("dense_output_1017_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1017_strides_1 = const()[name = string("dense_output_1017_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1017_pad_1 = const()[name = string("dense_output_1017_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1017_dilations_1 = const()[name = string("dense_output_1017_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1017_groups_1 = const()[name = string("dense_output_1017_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368073088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368204224))))[name = string("layers_12_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1017_cast_fp16 = conv(dilations = dense_output_1017_dilations_1, groups = dense_output_1017_groups_1, pad = dense_output_1017_pad_1, pad_type = dense_output_1017_pad_type_1, strides = dense_output_1017_strides_1, weight = layers_12_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1017_cast_fp16")];
            string sparse_output_1017_pad_type_1 = const()[name = string("sparse_output_1017_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1017_strides_1 = const()[name = string("sparse_output_1017_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1017_pad_1 = const()[name = string("sparse_output_1017_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1017_dilations_1 = const()[name = string("sparse_output_1017_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1017_groups_1 = const()[name = string("sparse_output_1017_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368207488))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368204800))))[name = string("layers_12_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1017_cast_fp16 = conv(dilations = sparse_output_1017_dilations_1, groups = sparse_output_1017_groups_1, pad = sparse_output_1017_pad_1, pad_type = sparse_output_1017_pad_type_1, strides = sparse_output_1017_strides_1, weight = layers_12_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1017_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16202_cast_fp16 = add(x = dense_output_1017_cast_fp16, y = sparse_output_1017_cast_fp16)[name = string("op_16202_cast_fp16")];
            tensor<int32, [4]> var_16203 = const()[name = string("op_16203"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16205 = const()[name = string("op_16205"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16204_cast_fp16 = transpose(perm = var_16203, x = var_16202_cast_fp16)[name = string("transpose_468")];
            tensor<fp16, [1, 51, 128]> q_head_203_cast_fp16 = reshape(shape = var_16205, x = var_16204_cast_fp16)[name = string("q_head_203_cast_fp16")];
            string dense_output_1019_pad_type_1 = const()[name = string("dense_output_1019_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1019_strides_1 = const()[name = string("dense_output_1019_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1019_pad_1 = const()[name = string("dense_output_1019_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1019_dilations_1 = const()[name = string("dense_output_1019_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1019_groups_1 = const()[name = string("dense_output_1019_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368223936))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368355072))))[name = string("layers_12_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1019_cast_fp16 = conv(dilations = dense_output_1019_dilations_1, groups = dense_output_1019_groups_1, pad = dense_output_1019_pad_1, pad_type = dense_output_1019_pad_type_1, strides = dense_output_1019_strides_1, weight = layers_12_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1019_cast_fp16")];
            string sparse_output_1019_pad_type_1 = const()[name = string("sparse_output_1019_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1019_strides_1 = const()[name = string("sparse_output_1019_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1019_pad_1 = const()[name = string("sparse_output_1019_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1019_dilations_1 = const()[name = string("sparse_output_1019_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1019_groups_1 = const()[name = string("sparse_output_1019_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368358336))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368355648))))[name = string("layers_12_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1019_cast_fp16 = conv(dilations = sparse_output_1019_dilations_1, groups = sparse_output_1019_groups_1, pad = sparse_output_1019_pad_1, pad_type = sparse_output_1019_pad_type_1, strides = sparse_output_1019_strides_1, weight = layers_12_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1019_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16221_cast_fp16 = add(x = dense_output_1019_cast_fp16, y = sparse_output_1019_cast_fp16)[name = string("op_16221_cast_fp16")];
            tensor<int32, [4]> var_16222 = const()[name = string("op_16222"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16224 = const()[name = string("op_16224"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16223_cast_fp16 = transpose(perm = var_16222, x = var_16221_cast_fp16)[name = string("transpose_467")];
            tensor<fp16, [1, 51, 128]> k_head_405_cast_fp16 = reshape(shape = var_16224, x = var_16223_cast_fp16)[name = string("k_head_405_cast_fp16")];
            string dense_output_1021_pad_type_1 = const()[name = string("dense_output_1021_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1021_strides_1 = const()[name = string("dense_output_1021_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1021_pad_1 = const()[name = string("dense_output_1021_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1021_dilations_1 = const()[name = string("dense_output_1021_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1021_groups_1 = const()[name = string("dense_output_1021_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368374784))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368505920))))[name = string("layers_12_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1021_cast_fp16 = conv(dilations = dense_output_1021_dilations_1, groups = dense_output_1021_groups_1, pad = dense_output_1021_pad_1, pad_type = dense_output_1021_pad_type_1, strides = dense_output_1021_strides_1, weight = layers_12_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1021_cast_fp16")];
            string sparse_output_1021_pad_type_1 = const()[name = string("sparse_output_1021_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1021_strides_1 = const()[name = string("sparse_output_1021_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1021_pad_1 = const()[name = string("sparse_output_1021_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1021_dilations_1 = const()[name = string("sparse_output_1021_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1021_groups_1 = const()[name = string("sparse_output_1021_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368509184))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368506496))))[name = string("layers_12_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1021_cast_fp16 = conv(dilations = sparse_output_1021_dilations_1, groups = sparse_output_1021_groups_1, pad = sparse_output_1021_pad_1, pad_type = sparse_output_1021_pad_type_1, strides = sparse_output_1021_strides_1, weight = layers_12_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1021_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16240_cast_fp16 = add(x = dense_output_1021_cast_fp16, y = sparse_output_1021_cast_fp16)[name = string("op_16240_cast_fp16")];
            tensor<int32, [4]> var_16241 = const()[name = string("op_16241"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16243 = const()[name = string("op_16243"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16242_cast_fp16 = transpose(perm = var_16241, x = var_16240_cast_fp16)[name = string("transpose_466")];
            tensor<fp16, [1, 51, 128]> v_head_405_cast_fp16 = reshape(shape = var_16243, x = var_16242_cast_fp16)[name = string("v_head_405_cast_fp16")];
            string dense_output_1023_pad_type_1 = const()[name = string("dense_output_1023_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1023_strides_1 = const()[name = string("dense_output_1023_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1023_pad_1 = const()[name = string("dense_output_1023_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1023_dilations_1 = const()[name = string("dense_output_1023_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1023_groups_1 = const()[name = string("dense_output_1023_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368525632))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368656768))))[name = string("layers_12_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1023_cast_fp16 = conv(dilations = dense_output_1023_dilations_1, groups = dense_output_1023_groups_1, pad = dense_output_1023_pad_1, pad_type = dense_output_1023_pad_type_1, strides = dense_output_1023_strides_1, weight = layers_12_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1023_cast_fp16")];
            string sparse_output_1023_pad_type_1 = const()[name = string("sparse_output_1023_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1023_strides_1 = const()[name = string("sparse_output_1023_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1023_pad_1 = const()[name = string("sparse_output_1023_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1023_dilations_1 = const()[name = string("sparse_output_1023_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1023_groups_1 = const()[name = string("sparse_output_1023_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368660032))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368657344))))[name = string("layers_12_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1023_cast_fp16 = conv(dilations = sparse_output_1023_dilations_1, groups = sparse_output_1023_groups_1, pad = sparse_output_1023_pad_1, pad_type = sparse_output_1023_pad_type_1, strides = sparse_output_1023_strides_1, weight = layers_12_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1023_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_16259_cast_fp16 = add(x = dense_output_1023_cast_fp16, y = sparse_output_1023_cast_fp16)[name = string("op_16259_cast_fp16")];
            tensor<int32, [4]> var_16260 = const()[name = string("op_16260"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16262 = const()[name = string("op_16262"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_16261_cast_fp16 = transpose(perm = var_16260, x = var_16259_cast_fp16)[name = string("transpose_465")];
            tensor<fp16, [1, 101, 128]> p_head_405_cast_fp16 = reshape(shape = var_16262, x = var_16261_cast_fp16)[name = string("p_head_405_cast_fp16")];
            tensor<fp16, [128]> var_16264_to_fp16 = const()[name = string("op_16264_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368676480)))];
            tensor<fp16, [1, 51, 128]> var_16265_cast_fp16 = add(x = q_head_203_cast_fp16, y = var_16264_to_fp16)[name = string("op_16265_cast_fp16")];
            tensor<int32, [1]> q_u_203_axes_1 = const()[name = string("q_u_203_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_203_cast_fp16 = expand_dims(axes = q_u_203_axes_1, x = var_16265_cast_fp16)[name = string("q_u_203_cast_fp16")];
            tensor<fp16, [128]> var_16267_to_fp16 = const()[name = string("op_16267_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368676800)))];
            tensor<fp16, [1, 51, 128]> var_16268_cast_fp16 = add(x = q_head_203_cast_fp16, y = var_16267_to_fp16)[name = string("op_16268_cast_fp16")];
            tensor<int32, [1]> q_v_203_axes_1 = const()[name = string("q_v_203_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_203_cast_fp16 = expand_dims(axes = q_v_203_axes_1, x = var_16268_cast_fp16)[name = string("q_v_203_cast_fp16")];
            tensor<int32, [1]> k_head_407_axes_1 = const()[name = string("k_head_407_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_407_cast_fp16 = expand_dims(axes = k_head_407_axes_1, x = k_head_405_cast_fp16)[name = string("k_head_407_cast_fp16")];
            tensor<int32, [1]> v_head_407_axes_1 = const()[name = string("v_head_407_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_407_cast_fp16 = expand_dims(axes = v_head_407_axes_1, x = v_head_405_cast_fp16)[name = string("v_head_407_cast_fp16")];
            tensor<int32, [1]> p_head_407_axes_1 = const()[name = string("p_head_407_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_407_cast_fp16 = expand_dims(axes = p_head_407_axes_1, x = p_head_405_cast_fp16)[name = string("p_head_407_cast_fp16")];
            bool var_16274_transpose_x_3 = const()[name = string("op_16274_transpose_x_3"), val = bool(false)];
            bool var_16274_transpose_y_3 = const()[name = string("op_16274_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_16274_cast_fp16 = matmul(transpose_x = var_16274_transpose_x_3, transpose_y = var_16274_transpose_y_3, x = q_u_203_cast_fp16, y = k_head_407_cast_fp16)[name = string("op_16274_cast_fp16")];
            fp16 var_16275_to_fp16 = const()[name = string("op_16275_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_203_cast_fp16 = mul(x = var_16274_cast_fp16, y = var_16275_to_fp16)[name = string("scores_content_203_cast_fp16")];
            bool x_1069_transpose_x_3 = const()[name = string("x_1069_transpose_x_3"), val = bool(false)];
            bool x_1069_transpose_y_3 = const()[name = string("x_1069_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1069_cast_fp16 = matmul(transpose_x = x_1069_transpose_x_3, transpose_y = x_1069_transpose_y_3, x = q_v_203_cast_fp16, y = p_head_407_cast_fp16)[name = string("x_1069_cast_fp16")];
            tensor<int32, [8]> x_1071_pad_1 = const()[name = string("x_1071_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1071_mode_1 = const()[name = string("x_1071_mode_1"), val = string("constant")];
            fp16 const_1987_to_fp16 = const()[name = string("const_1987_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1071_cast_fp16 = pad(constant_val = const_1987_to_fp16, mode = x_1071_mode_1, pad = x_1071_pad_1, x = x_1069_cast_fp16)[name = string("x_1071_cast_fp16")];
            tensor<int32, [4]> var_16289 = const()[name = string("op_16289"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1073_cast_fp16 = reshape(shape = var_16289, x = x_1071_cast_fp16)[name = string("x_1073_cast_fp16")];
            tensor<int32, [4]> var_16293_begin_1 = const()[name = string("op_16293_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16293_end_1 = const()[name = string("op_16293_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_16293_end_mask_1 = const()[name = string("op_16293_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_16293_cast_fp16 = slice_by_index(begin = var_16293_begin_1, end = var_16293_end_1, end_mask = var_16293_end_mask_1, x = x_1073_cast_fp16)[name = string("op_16293_cast_fp16")];
            tensor<int32, [4]> var_16295 = const()[name = string("op_16295"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_16296_cast_fp16 = reshape(shape = var_16295, x = var_16293_cast_fp16)[name = string("op_16296_cast_fp16")];
            tensor<int32, [4]> var_16301_begin_1 = const()[name = string("op_16301_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16301_end_1 = const()[name = string("op_16301_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_16301_end_mask_1 = const()[name = string("op_16301_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_16301_cast_fp16 = slice_by_index(begin = var_16301_begin_1, end = var_16301_end_1, end_mask = var_16301_end_mask_1, x = var_16296_cast_fp16)[name = string("op_16301_cast_fp16")];
            fp16 var_16302_to_fp16 = const()[name = string("op_16302_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_203_cast_fp16 = mul(x = var_16301_cast_fp16, y = var_16302_to_fp16)[name = string("scores_pos_203_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_203_cast_fp16 = add(x = scores_content_203_cast_fp16, y = scores_pos_203_cast_fp16)[name = string("logits_203_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_16305_cast_fp16 = softmax(axis = var_15425, x = logits_203_cast_fp16)[name = string("op_16305_cast_fp16")];
            bool var_16307_transpose_x_1 = const()[name = string("op_16307_transpose_x_1"), val = bool(false)];
            bool var_16307_transpose_y_1 = const()[name = string("op_16307_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_16307_cast_fp16 = matmul(transpose_x = var_16307_transpose_x_1, transpose_y = var_16307_transpose_y_1, x = var_16305_cast_fp16, y = v_head_407_cast_fp16)[name = string("op_16307_cast_fp16")];
            tensor<int32, [1]> var_16308_axes_1 = const()[name = string("op_16308_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_16308_cast_fp16 = squeeze(axes = var_16308_axes_1, x = var_16307_cast_fp16)[name = string("op_16308_cast_fp16")];
            string dense_output_1025_pad_type_1 = const()[name = string("dense_output_1025_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1025_strides_1 = const()[name = string("dense_output_1025_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1025_pad_1 = const()[name = string("dense_output_1025_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1025_dilations_1 = const()[name = string("dense_output_1025_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1025_groups_1 = const()[name = string("dense_output_1025_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368677120))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368808256))))[name = string("layers_12_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1025_cast_fp16 = conv(dilations = dense_output_1025_dilations_1, groups = dense_output_1025_groups_1, pad = dense_output_1025_pad_1, pad_type = dense_output_1025_pad_type_1, strides = dense_output_1025_strides_1, weight = layers_12_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1025_cast_fp16")];
            string sparse_output_1025_pad_type_1 = const()[name = string("sparse_output_1025_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1025_strides_1 = const()[name = string("sparse_output_1025_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1025_pad_1 = const()[name = string("sparse_output_1025_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1025_dilations_1 = const()[name = string("sparse_output_1025_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1025_groups_1 = const()[name = string("sparse_output_1025_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368811520))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368808832))))[name = string("layers_12_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1025_cast_fp16 = conv(dilations = sparse_output_1025_dilations_1, groups = sparse_output_1025_groups_1, pad = sparse_output_1025_pad_1, pad_type = sparse_output_1025_pad_type_1, strides = sparse_output_1025_strides_1, weight = layers_12_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1025_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16323_cast_fp16 = add(x = dense_output_1025_cast_fp16, y = sparse_output_1025_cast_fp16)[name = string("op_16323_cast_fp16")];
            tensor<int32, [4]> var_16324 = const()[name = string("op_16324"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16326 = const()[name = string("op_16326"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16325_cast_fp16 = transpose(perm = var_16324, x = var_16323_cast_fp16)[name = string("transpose_464")];
            tensor<fp16, [1, 51, 128]> q_head_205_cast_fp16 = reshape(shape = var_16326, x = var_16325_cast_fp16)[name = string("q_head_205_cast_fp16")];
            string dense_output_1027_pad_type_1 = const()[name = string("dense_output_1027_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1027_strides_1 = const()[name = string("dense_output_1027_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1027_pad_1 = const()[name = string("dense_output_1027_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1027_dilations_1 = const()[name = string("dense_output_1027_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1027_groups_1 = const()[name = string("dense_output_1027_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368827968))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368959104))))[name = string("layers_12_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1027_cast_fp16 = conv(dilations = dense_output_1027_dilations_1, groups = dense_output_1027_groups_1, pad = dense_output_1027_pad_1, pad_type = dense_output_1027_pad_type_1, strides = dense_output_1027_strides_1, weight = layers_12_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1027_cast_fp16")];
            string sparse_output_1027_pad_type_1 = const()[name = string("sparse_output_1027_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1027_strides_1 = const()[name = string("sparse_output_1027_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1027_pad_1 = const()[name = string("sparse_output_1027_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1027_dilations_1 = const()[name = string("sparse_output_1027_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1027_groups_1 = const()[name = string("sparse_output_1027_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368962368))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368959680))))[name = string("layers_12_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1027_cast_fp16 = conv(dilations = sparse_output_1027_dilations_1, groups = sparse_output_1027_groups_1, pad = sparse_output_1027_pad_1, pad_type = sparse_output_1027_pad_type_1, strides = sparse_output_1027_strides_1, weight = layers_12_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1027_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16342_cast_fp16 = add(x = dense_output_1027_cast_fp16, y = sparse_output_1027_cast_fp16)[name = string("op_16342_cast_fp16")];
            tensor<int32, [4]> var_16343 = const()[name = string("op_16343"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16345 = const()[name = string("op_16345"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16344_cast_fp16 = transpose(perm = var_16343, x = var_16342_cast_fp16)[name = string("transpose_463")];
            tensor<fp16, [1, 51, 128]> k_head_409_cast_fp16 = reshape(shape = var_16345, x = var_16344_cast_fp16)[name = string("k_head_409_cast_fp16")];
            string dense_output_1029_pad_type_1 = const()[name = string("dense_output_1029_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1029_strides_1 = const()[name = string("dense_output_1029_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1029_pad_1 = const()[name = string("dense_output_1029_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1029_dilations_1 = const()[name = string("dense_output_1029_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1029_groups_1 = const()[name = string("dense_output_1029_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368978816))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369109952))))[name = string("layers_12_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1029_cast_fp16 = conv(dilations = dense_output_1029_dilations_1, groups = dense_output_1029_groups_1, pad = dense_output_1029_pad_1, pad_type = dense_output_1029_pad_type_1, strides = dense_output_1029_strides_1, weight = layers_12_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1029_cast_fp16")];
            string sparse_output_1029_pad_type_1 = const()[name = string("sparse_output_1029_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1029_strides_1 = const()[name = string("sparse_output_1029_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1029_pad_1 = const()[name = string("sparse_output_1029_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1029_dilations_1 = const()[name = string("sparse_output_1029_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1029_groups_1 = const()[name = string("sparse_output_1029_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369113216))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369110528))))[name = string("layers_12_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1029_cast_fp16 = conv(dilations = sparse_output_1029_dilations_1, groups = sparse_output_1029_groups_1, pad = sparse_output_1029_pad_1, pad_type = sparse_output_1029_pad_type_1, strides = sparse_output_1029_strides_1, weight = layers_12_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1029_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16361_cast_fp16 = add(x = dense_output_1029_cast_fp16, y = sparse_output_1029_cast_fp16)[name = string("op_16361_cast_fp16")];
            tensor<int32, [4]> var_16362 = const()[name = string("op_16362"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16364 = const()[name = string("op_16364"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16363_cast_fp16 = transpose(perm = var_16362, x = var_16361_cast_fp16)[name = string("transpose_462")];
            tensor<fp16, [1, 51, 128]> v_head_409_cast_fp16 = reshape(shape = var_16364, x = var_16363_cast_fp16)[name = string("v_head_409_cast_fp16")];
            string dense_output_1031_pad_type_1 = const()[name = string("dense_output_1031_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1031_strides_1 = const()[name = string("dense_output_1031_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1031_pad_1 = const()[name = string("dense_output_1031_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1031_dilations_1 = const()[name = string("dense_output_1031_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1031_groups_1 = const()[name = string("dense_output_1031_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369129664))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369260800))))[name = string("layers_12_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1031_cast_fp16 = conv(dilations = dense_output_1031_dilations_1, groups = dense_output_1031_groups_1, pad = dense_output_1031_pad_1, pad_type = dense_output_1031_pad_type_1, strides = dense_output_1031_strides_1, weight = layers_12_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1031_cast_fp16")];
            string sparse_output_1031_pad_type_1 = const()[name = string("sparse_output_1031_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1031_strides_1 = const()[name = string("sparse_output_1031_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1031_pad_1 = const()[name = string("sparse_output_1031_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1031_dilations_1 = const()[name = string("sparse_output_1031_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1031_groups_1 = const()[name = string("sparse_output_1031_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369264064))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369261376))))[name = string("layers_12_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1031_cast_fp16 = conv(dilations = sparse_output_1031_dilations_1, groups = sparse_output_1031_groups_1, pad = sparse_output_1031_pad_1, pad_type = sparse_output_1031_pad_type_1, strides = sparse_output_1031_strides_1, weight = layers_12_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1031_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_16380_cast_fp16 = add(x = dense_output_1031_cast_fp16, y = sparse_output_1031_cast_fp16)[name = string("op_16380_cast_fp16")];
            tensor<int32, [4]> var_16381 = const()[name = string("op_16381"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16383 = const()[name = string("op_16383"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_16382_cast_fp16 = transpose(perm = var_16381, x = var_16380_cast_fp16)[name = string("transpose_461")];
            tensor<fp16, [1, 101, 128]> p_head_409_cast_fp16 = reshape(shape = var_16383, x = var_16382_cast_fp16)[name = string("p_head_409_cast_fp16")];
            tensor<fp16, [128]> var_16385_to_fp16 = const()[name = string("op_16385_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369280512)))];
            tensor<fp16, [1, 51, 128]> var_16386_cast_fp16 = add(x = q_head_205_cast_fp16, y = var_16385_to_fp16)[name = string("op_16386_cast_fp16")];
            tensor<int32, [1]> q_u_205_axes_1 = const()[name = string("q_u_205_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_205_cast_fp16 = expand_dims(axes = q_u_205_axes_1, x = var_16386_cast_fp16)[name = string("q_u_205_cast_fp16")];
            tensor<fp16, [128]> var_16388_to_fp16 = const()[name = string("op_16388_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369280832)))];
            tensor<fp16, [1, 51, 128]> var_16389_cast_fp16 = add(x = q_head_205_cast_fp16, y = var_16388_to_fp16)[name = string("op_16389_cast_fp16")];
            tensor<int32, [1]> q_v_205_axes_1 = const()[name = string("q_v_205_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_205_cast_fp16 = expand_dims(axes = q_v_205_axes_1, x = var_16389_cast_fp16)[name = string("q_v_205_cast_fp16")];
            tensor<int32, [1]> k_head_411_axes_1 = const()[name = string("k_head_411_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_411_cast_fp16 = expand_dims(axes = k_head_411_axes_1, x = k_head_409_cast_fp16)[name = string("k_head_411_cast_fp16")];
            tensor<int32, [1]> v_head_411_axes_1 = const()[name = string("v_head_411_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_411_cast_fp16 = expand_dims(axes = v_head_411_axes_1, x = v_head_409_cast_fp16)[name = string("v_head_411_cast_fp16")];
            tensor<int32, [1]> p_head_411_axes_1 = const()[name = string("p_head_411_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_411_cast_fp16 = expand_dims(axes = p_head_411_axes_1, x = p_head_409_cast_fp16)[name = string("p_head_411_cast_fp16")];
            bool var_16395_transpose_x_3 = const()[name = string("op_16395_transpose_x_3"), val = bool(false)];
            bool var_16395_transpose_y_3 = const()[name = string("op_16395_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_16395_cast_fp16 = matmul(transpose_x = var_16395_transpose_x_3, transpose_y = var_16395_transpose_y_3, x = q_u_205_cast_fp16, y = k_head_411_cast_fp16)[name = string("op_16395_cast_fp16")];
            fp16 var_16396_to_fp16 = const()[name = string("op_16396_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_205_cast_fp16 = mul(x = var_16395_cast_fp16, y = var_16396_to_fp16)[name = string("scores_content_205_cast_fp16")];
            bool x_1077_transpose_x_3 = const()[name = string("x_1077_transpose_x_3"), val = bool(false)];
            bool x_1077_transpose_y_3 = const()[name = string("x_1077_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1077_cast_fp16 = matmul(transpose_x = x_1077_transpose_x_3, transpose_y = x_1077_transpose_y_3, x = q_v_205_cast_fp16, y = p_head_411_cast_fp16)[name = string("x_1077_cast_fp16")];
            tensor<int32, [8]> x_1079_pad_1 = const()[name = string("x_1079_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1079_mode_1 = const()[name = string("x_1079_mode_1"), val = string("constant")];
            fp16 const_1993_to_fp16 = const()[name = string("const_1993_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1079_cast_fp16 = pad(constant_val = const_1993_to_fp16, mode = x_1079_mode_1, pad = x_1079_pad_1, x = x_1077_cast_fp16)[name = string("x_1079_cast_fp16")];
            tensor<int32, [4]> var_16410 = const()[name = string("op_16410"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1081_cast_fp16 = reshape(shape = var_16410, x = x_1079_cast_fp16)[name = string("x_1081_cast_fp16")];
            tensor<int32, [4]> var_16414_begin_1 = const()[name = string("op_16414_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16414_end_1 = const()[name = string("op_16414_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_16414_end_mask_1 = const()[name = string("op_16414_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_16414_cast_fp16 = slice_by_index(begin = var_16414_begin_1, end = var_16414_end_1, end_mask = var_16414_end_mask_1, x = x_1081_cast_fp16)[name = string("op_16414_cast_fp16")];
            tensor<int32, [4]> var_16416 = const()[name = string("op_16416"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_16417_cast_fp16 = reshape(shape = var_16416, x = var_16414_cast_fp16)[name = string("op_16417_cast_fp16")];
            tensor<int32, [4]> var_16422_begin_1 = const()[name = string("op_16422_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16422_end_1 = const()[name = string("op_16422_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_16422_end_mask_1 = const()[name = string("op_16422_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_16422_cast_fp16 = slice_by_index(begin = var_16422_begin_1, end = var_16422_end_1, end_mask = var_16422_end_mask_1, x = var_16417_cast_fp16)[name = string("op_16422_cast_fp16")];
            fp16 var_16423_to_fp16 = const()[name = string("op_16423_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_205_cast_fp16 = mul(x = var_16422_cast_fp16, y = var_16423_to_fp16)[name = string("scores_pos_205_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_205_cast_fp16 = add(x = scores_content_205_cast_fp16, y = scores_pos_205_cast_fp16)[name = string("logits_205_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_16426_cast_fp16 = softmax(axis = var_15425, x = logits_205_cast_fp16)[name = string("op_16426_cast_fp16")];
            bool var_16428_transpose_x_1 = const()[name = string("op_16428_transpose_x_1"), val = bool(false)];
            bool var_16428_transpose_y_1 = const()[name = string("op_16428_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_16428_cast_fp16 = matmul(transpose_x = var_16428_transpose_x_1, transpose_y = var_16428_transpose_y_1, x = var_16426_cast_fp16, y = v_head_411_cast_fp16)[name = string("op_16428_cast_fp16")];
            tensor<int32, [1]> var_16429_axes_1 = const()[name = string("op_16429_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_16429_cast_fp16 = squeeze(axes = var_16429_axes_1, x = var_16428_cast_fp16)[name = string("op_16429_cast_fp16")];
            string dense_output_1033_pad_type_1 = const()[name = string("dense_output_1033_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1033_strides_1 = const()[name = string("dense_output_1033_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1033_pad_1 = const()[name = string("dense_output_1033_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1033_dilations_1 = const()[name = string("dense_output_1033_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1033_groups_1 = const()[name = string("dense_output_1033_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369281152))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369412288))))[name = string("layers_12_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1033_cast_fp16 = conv(dilations = dense_output_1033_dilations_1, groups = dense_output_1033_groups_1, pad = dense_output_1033_pad_1, pad_type = dense_output_1033_pad_type_1, strides = dense_output_1033_strides_1, weight = layers_12_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1033_cast_fp16")];
            string sparse_output_1033_pad_type_1 = const()[name = string("sparse_output_1033_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1033_strides_1 = const()[name = string("sparse_output_1033_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1033_pad_1 = const()[name = string("sparse_output_1033_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1033_dilations_1 = const()[name = string("sparse_output_1033_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1033_groups_1 = const()[name = string("sparse_output_1033_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369415552))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369412864))))[name = string("layers_12_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1033_cast_fp16 = conv(dilations = sparse_output_1033_dilations_1, groups = sparse_output_1033_groups_1, pad = sparse_output_1033_pad_1, pad_type = sparse_output_1033_pad_type_1, strides = sparse_output_1033_strides_1, weight = layers_12_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1033_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16444_cast_fp16 = add(x = dense_output_1033_cast_fp16, y = sparse_output_1033_cast_fp16)[name = string("op_16444_cast_fp16")];
            tensor<int32, [4]> var_16445 = const()[name = string("op_16445"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16447 = const()[name = string("op_16447"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16446_cast_fp16 = transpose(perm = var_16445, x = var_16444_cast_fp16)[name = string("transpose_460")];
            tensor<fp16, [1, 51, 128]> q_head_207_cast_fp16 = reshape(shape = var_16447, x = var_16446_cast_fp16)[name = string("q_head_207_cast_fp16")];
            string dense_output_1035_pad_type_1 = const()[name = string("dense_output_1035_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1035_strides_1 = const()[name = string("dense_output_1035_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1035_pad_1 = const()[name = string("dense_output_1035_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1035_dilations_1 = const()[name = string("dense_output_1035_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1035_groups_1 = const()[name = string("dense_output_1035_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369432000))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369563136))))[name = string("layers_12_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1035_cast_fp16 = conv(dilations = dense_output_1035_dilations_1, groups = dense_output_1035_groups_1, pad = dense_output_1035_pad_1, pad_type = dense_output_1035_pad_type_1, strides = dense_output_1035_strides_1, weight = layers_12_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1035_cast_fp16")];
            string sparse_output_1035_pad_type_1 = const()[name = string("sparse_output_1035_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1035_strides_1 = const()[name = string("sparse_output_1035_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1035_pad_1 = const()[name = string("sparse_output_1035_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1035_dilations_1 = const()[name = string("sparse_output_1035_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1035_groups_1 = const()[name = string("sparse_output_1035_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369566400))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369563712))))[name = string("layers_12_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1035_cast_fp16 = conv(dilations = sparse_output_1035_dilations_1, groups = sparse_output_1035_groups_1, pad = sparse_output_1035_pad_1, pad_type = sparse_output_1035_pad_type_1, strides = sparse_output_1035_strides_1, weight = layers_12_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1035_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16463_cast_fp16 = add(x = dense_output_1035_cast_fp16, y = sparse_output_1035_cast_fp16)[name = string("op_16463_cast_fp16")];
            tensor<int32, [4]> var_16464 = const()[name = string("op_16464"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16466 = const()[name = string("op_16466"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16465_cast_fp16 = transpose(perm = var_16464, x = var_16463_cast_fp16)[name = string("transpose_459")];
            tensor<fp16, [1, 51, 128]> k_head_413_cast_fp16 = reshape(shape = var_16466, x = var_16465_cast_fp16)[name = string("k_head_413_cast_fp16")];
            string dense_output_1037_pad_type_1 = const()[name = string("dense_output_1037_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1037_strides_1 = const()[name = string("dense_output_1037_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1037_pad_1 = const()[name = string("dense_output_1037_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1037_dilations_1 = const()[name = string("dense_output_1037_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1037_groups_1 = const()[name = string("dense_output_1037_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369582848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369713984))))[name = string("layers_12_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1037_cast_fp16 = conv(dilations = dense_output_1037_dilations_1, groups = dense_output_1037_groups_1, pad = dense_output_1037_pad_1, pad_type = dense_output_1037_pad_type_1, strides = dense_output_1037_strides_1, weight = layers_12_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = string("dense_output_1037_cast_fp16")];
            string sparse_output_1037_pad_type_1 = const()[name = string("sparse_output_1037_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1037_strides_1 = const()[name = string("sparse_output_1037_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1037_pad_1 = const()[name = string("sparse_output_1037_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1037_dilations_1 = const()[name = string("sparse_output_1037_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1037_groups_1 = const()[name = string("sparse_output_1037_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369717248))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369714560))))[name = string("layers_12_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1037_cast_fp16 = conv(dilations = sparse_output_1037_dilations_1, groups = sparse_output_1037_groups_1, pad = sparse_output_1037_pad_1, pad_type = sparse_output_1037_pad_type_1, strides = sparse_output_1037_strides_1, weight = layers_12_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_581_cast_fp16)[name = string("sparse_output_1037_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16482_cast_fp16 = add(x = dense_output_1037_cast_fp16, y = sparse_output_1037_cast_fp16)[name = string("op_16482_cast_fp16")];
            tensor<int32, [4]> var_16483 = const()[name = string("op_16483"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16485 = const()[name = string("op_16485"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16484_cast_fp16 = transpose(perm = var_16483, x = var_16482_cast_fp16)[name = string("transpose_458")];
            tensor<fp16, [1, 51, 128]> v_head_413_cast_fp16 = reshape(shape = var_16485, x = var_16484_cast_fp16)[name = string("v_head_413_cast_fp16")];
            string dense_output_1039_pad_type_1 = const()[name = string("dense_output_1039_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1039_strides_1 = const()[name = string("dense_output_1039_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1039_pad_1 = const()[name = string("dense_output_1039_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1039_dilations_1 = const()[name = string("dense_output_1039_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1039_groups_1 = const()[name = string("dense_output_1039_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369733696))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369864832))))[name = string("layers_12_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1039_cast_fp16 = conv(dilations = dense_output_1039_dilations_1, groups = dense_output_1039_groups_1, pad = dense_output_1039_pad_1, pad_type = dense_output_1039_pad_type_1, strides = dense_output_1039_strides_1, weight = layers_12_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1039_cast_fp16")];
            string sparse_output_1039_pad_type_1 = const()[name = string("sparse_output_1039_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1039_strides_1 = const()[name = string("sparse_output_1039_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1039_pad_1 = const()[name = string("sparse_output_1039_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1039_dilations_1 = const()[name = string("sparse_output_1039_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1039_groups_1 = const()[name = string("sparse_output_1039_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_12_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369868096))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369865408))))[name = string("layers_12_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1039_cast_fp16 = conv(dilations = sparse_output_1039_dilations_1, groups = sparse_output_1039_groups_1, pad = sparse_output_1039_pad_1, pad_type = sparse_output_1039_pad_type_1, strides = sparse_output_1039_strides_1, weight = layers_12_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1039_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_16501_cast_fp16 = add(x = dense_output_1039_cast_fp16, y = sparse_output_1039_cast_fp16)[name = string("op_16501_cast_fp16")];
            tensor<int32, [4]> var_16502 = const()[name = string("op_16502"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16504 = const()[name = string("op_16504"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_16503_cast_fp16 = transpose(perm = var_16502, x = var_16501_cast_fp16)[name = string("transpose_457")];
            tensor<fp16, [1, 101, 128]> p_head_413_cast_fp16 = reshape(shape = var_16504, x = var_16503_cast_fp16)[name = string("p_head_413_cast_fp16")];
            tensor<fp16, [128]> var_16506_to_fp16 = const()[name = string("op_16506_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369884544)))];
            tensor<fp16, [1, 51, 128]> var_16507_cast_fp16 = add(x = q_head_207_cast_fp16, y = var_16506_to_fp16)[name = string("op_16507_cast_fp16")];
            tensor<int32, [1]> q_u_207_axes_1 = const()[name = string("q_u_207_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_207_cast_fp16 = expand_dims(axes = q_u_207_axes_1, x = var_16507_cast_fp16)[name = string("q_u_207_cast_fp16")];
            tensor<fp16, [128]> var_16509_to_fp16 = const()[name = string("op_16509_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369884864)))];
            tensor<fp16, [1, 51, 128]> var_16510_cast_fp16 = add(x = q_head_207_cast_fp16, y = var_16509_to_fp16)[name = string("op_16510_cast_fp16")];
            tensor<int32, [1]> q_v_207_axes_1 = const()[name = string("q_v_207_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_207_cast_fp16 = expand_dims(axes = q_v_207_axes_1, x = var_16510_cast_fp16)[name = string("q_v_207_cast_fp16")];
            tensor<int32, [1]> k_head_415_axes_1 = const()[name = string("k_head_415_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_415_cast_fp16 = expand_dims(axes = k_head_415_axes_1, x = k_head_413_cast_fp16)[name = string("k_head_415_cast_fp16")];
            tensor<int32, [1]> v_head_415_axes_1 = const()[name = string("v_head_415_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_415_cast_fp16 = expand_dims(axes = v_head_415_axes_1, x = v_head_413_cast_fp16)[name = string("v_head_415_cast_fp16")];
            tensor<int32, [1]> p_head_415_axes_1 = const()[name = string("p_head_415_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_415_cast_fp16 = expand_dims(axes = p_head_415_axes_1, x = p_head_413_cast_fp16)[name = string("p_head_415_cast_fp16")];
            bool var_16516_transpose_x_3 = const()[name = string("op_16516_transpose_x_3"), val = bool(false)];
            bool var_16516_transpose_y_3 = const()[name = string("op_16516_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_16516_cast_fp16 = matmul(transpose_x = var_16516_transpose_x_3, transpose_y = var_16516_transpose_y_3, x = q_u_207_cast_fp16, y = k_head_415_cast_fp16)[name = string("op_16516_cast_fp16")];
            fp16 var_16517_to_fp16 = const()[name = string("op_16517_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_207_cast_fp16 = mul(x = var_16516_cast_fp16, y = var_16517_to_fp16)[name = string("scores_content_207_cast_fp16")];
            bool x_1085_transpose_x_3 = const()[name = string("x_1085_transpose_x_3"), val = bool(false)];
            bool x_1085_transpose_y_3 = const()[name = string("x_1085_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1085_cast_fp16 = matmul(transpose_x = x_1085_transpose_x_3, transpose_y = x_1085_transpose_y_3, x = q_v_207_cast_fp16, y = p_head_415_cast_fp16)[name = string("x_1085_cast_fp16")];
            tensor<int32, [8]> x_1087_pad_1 = const()[name = string("x_1087_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1087_mode_1 = const()[name = string("x_1087_mode_1"), val = string("constant")];
            fp16 const_1999_to_fp16 = const()[name = string("const_1999_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1087_cast_fp16 = pad(constant_val = const_1999_to_fp16, mode = x_1087_mode_1, pad = x_1087_pad_1, x = x_1085_cast_fp16)[name = string("x_1087_cast_fp16")];
            tensor<int32, [4]> var_16531 = const()[name = string("op_16531"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1089_cast_fp16 = reshape(shape = var_16531, x = x_1087_cast_fp16)[name = string("x_1089_cast_fp16")];
            tensor<int32, [4]> var_16535_begin_1 = const()[name = string("op_16535_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16535_end_1 = const()[name = string("op_16535_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_16535_end_mask_1 = const()[name = string("op_16535_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_16535_cast_fp16 = slice_by_index(begin = var_16535_begin_1, end = var_16535_end_1, end_mask = var_16535_end_mask_1, x = x_1089_cast_fp16)[name = string("op_16535_cast_fp16")];
            tensor<int32, [4]> var_16537 = const()[name = string("op_16537"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_16538_cast_fp16 = reshape(shape = var_16537, x = var_16535_cast_fp16)[name = string("op_16538_cast_fp16")];
            tensor<int32, [4]> var_16543_begin_1 = const()[name = string("op_16543_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16543_end_1 = const()[name = string("op_16543_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_16543_end_mask_1 = const()[name = string("op_16543_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_16543_cast_fp16 = slice_by_index(begin = var_16543_begin_1, end = var_16543_end_1, end_mask = var_16543_end_mask_1, x = var_16538_cast_fp16)[name = string("op_16543_cast_fp16")];
            fp16 var_16544_to_fp16 = const()[name = string("op_16544_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_207_cast_fp16 = mul(x = var_16543_cast_fp16, y = var_16544_to_fp16)[name = string("scores_pos_207_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_207_cast_fp16 = add(x = scores_content_207_cast_fp16, y = scores_pos_207_cast_fp16)[name = string("logits_207_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_16547_cast_fp16 = softmax(axis = var_15425, x = logits_207_cast_fp16)[name = string("op_16547_cast_fp16")];
            bool var_16549_transpose_x_1 = const()[name = string("op_16549_transpose_x_1"), val = bool(false)];
            bool var_16549_transpose_y_1 = const()[name = string("op_16549_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_16549_cast_fp16 = matmul(transpose_x = var_16549_transpose_x_1, transpose_y = var_16549_transpose_y_1, x = var_16547_cast_fp16, y = v_head_415_cast_fp16)[name = string("op_16549_cast_fp16")];
            tensor<int32, [1]> o_head_25_axes_1 = const()[name = string("o_head_25_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_25_cast_fp16 = squeeze(axes = o_head_25_axes_1, x = var_16549_cast_fp16)[name = string("o_head_25_cast_fp16")];
            bool out_25_interleave_1 = const()[name = string("out_25_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_25_cast_fp16 = concat(axis = var_15425, interleave = out_25_interleave_1, values = (var_15703_cast_fp16, var_15824_cast_fp16, var_15945_cast_fp16, var_16066_cast_fp16, var_16187_cast_fp16, var_16308_cast_fp16, var_16429_cast_fp16, o_head_25_cast_fp16))[name = string("out_25_cast_fp16")];
            tensor<int32, [3]> var_16553_perm_1 = const()[name = string("op_16553_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_589_axes_1 = const()[name = string("input_589_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_16553_cast_fp16 = transpose(perm = var_16553_perm_1, x = out_25_cast_fp16)[name = string("transpose_456")];
            tensor<fp16, [1, 1024, 51, 1]> input_589_cast_fp16 = expand_dims(axes = input_589_axes_1, x = var_16553_cast_fp16)[name = string("input_589_cast_fp16")];
            string dense_output_1041_pad_type_1 = const()[name = string("dense_output_1041_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1041_strides_1 = const()[name = string("dense_output_1041_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1041_pad_1 = const()[name = string("dense_output_1041_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1041_dilations_1 = const()[name = string("dense_output_1041_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1041_groups_1 = const()[name = string("dense_output_1041_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369885184))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370933824))))[name = string("layers_12_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1041_cast_fp16 = conv(dilations = dense_output_1041_dilations_1, groups = dense_output_1041_groups_1, pad = dense_output_1041_pad_1, pad_type = dense_output_1041_pad_type_1, strides = dense_output_1041_strides_1, weight = layers_12_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("dense_output_1041_cast_fp16")];
            string sparse_output_1041_pad_type_1 = const()[name = string("sparse_output_1041_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1041_strides_1 = const()[name = string("sparse_output_1041_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1041_pad_1 = const()[name = string("sparse_output_1041_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1041_dilations_1 = const()[name = string("sparse_output_1041_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1041_groups_1 = const()[name = string("sparse_output_1041_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370955456))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370934400))))[name = string("layers_12_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1041_cast_fp16 = conv(dilations = sparse_output_1041_dilations_1, groups = sparse_output_1041_groups_1, pad = sparse_output_1041_pad_1, pad_type = sparse_output_1041_pad_type_1, strides = sparse_output_1041_strides_1, weight = layers_12_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_589_cast_fp16)[name = string("sparse_output_1041_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_25_cast_fp16 = add(x = dense_output_1041_cast_fp16, y = sparse_output_1041_cast_fp16)[name = string("out_conv_25_cast_fp16")];
            tensor<int32, [1]> var_16570_axes_1 = const()[name = string("op_16570_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_16570_cast_fp16 = squeeze(axes = var_16570_axes_1, x = out_conv_25_cast_fp16)[name = string("op_16570_cast_fp16")];
            tensor<int32, [3]> var_16571_perm_1 = const()[name = string("op_16571_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_16571_cast_fp16 = transpose(perm = var_16571_perm_1, x = var_16570_cast_fp16)[name = string("transpose_455")];
            tensor<fp16, [1, 51, 1024]> input_591_cast_fp16 = add(x = input_579_cast_fp16, y = var_16571_cast_fp16)[name = string("input_591_cast_fp16")];
            tensor<int32, [1]> x_1093_axes_1 = const()[name = string("x_1093_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_12_norm_conv_weight_to_fp16 = const()[name = string("layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371086592)))];
            tensor<fp16, [1024]> layers_12_norm_conv_bias_to_fp16 = const()[name = string("layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371088704)))];
            tensor<fp16, [1, 51, 1024]> x_1093_cast_fp16 = layer_norm(axes = x_1093_axes_1, beta = layers_12_norm_conv_bias_to_fp16, epsilon = var_15440_to_fp16, gamma = layers_12_norm_conv_weight_to_fp16, x = input_591_cast_fp16)[name = string("x_1093_cast_fp16")];
            tensor<int32, [3]> var_16581_perm_1 = const()[name = string("op_16581_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_593_axes_1 = const()[name = string("input_593_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_16581_cast_fp16 = transpose(perm = var_16581_perm_1, x = x_1093_cast_fp16)[name = string("transpose_454")];
            tensor<fp16, [1, 1024, 51, 1]> input_593_cast_fp16 = expand_dims(axes = input_593_axes_1, x = var_16581_cast_fp16)[name = string("input_593_cast_fp16")];
            string dense_output_1043_pad_type_1 = const()[name = string("dense_output_1043_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1043_strides_1 = const()[name = string("dense_output_1043_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1043_pad_1 = const()[name = string("dense_output_1043_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1043_dilations_1 = const()[name = string("dense_output_1043_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1043_groups_1 = const()[name = string("dense_output_1043_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_12_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371090816))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373188032))))[name = string("layers_12_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1043_cast_fp16 = conv(dilations = dense_output_1043_dilations_1, groups = dense_output_1043_groups_1, pad = dense_output_1043_pad_1, pad_type = dense_output_1043_pad_type_1, strides = dense_output_1043_strides_1, weight = layers_12_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_593_cast_fp16)[name = string("dense_output_1043_cast_fp16")];
            string sparse_output_1043_pad_type_1 = const()[name = string("sparse_output_1043_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1043_strides_1 = const()[name = string("sparse_output_1043_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1043_pad_1 = const()[name = string("sparse_output_1043_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1043_dilations_1 = const()[name = string("sparse_output_1043_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1043_groups_1 = const()[name = string("sparse_output_1043_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_12_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373230656))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373188608))))[name = string("layers_12_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1043_cast_fp16 = conv(dilations = sparse_output_1043_dilations_1, groups = sparse_output_1043_groups_1, pad = sparse_output_1043_pad_1, pad_type = sparse_output_1043_pad_type_1, strides = sparse_output_1043_strides_1, weight = layers_12_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_593_cast_fp16)[name = string("sparse_output_1043_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_595_cast_fp16 = add(x = dense_output_1043_cast_fp16, y = sparse_output_1043_cast_fp16)[name = string("input_595_cast_fp16")];
            int32 input_597_split_num_splits_1 = const()[name = string("input_597_split_num_splits_1"), val = int32(2)];
            int32 input_597_split_axis_1 = const()[name = string("input_597_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_597_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_597_split_cast_fp16_1 = split(axis = input_597_split_axis_1, num_splits = input_597_split_num_splits_1, x = input_595_cast_fp16)[name = string("input_597_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_597_split_1_sigmoid_cast_fp16 = sigmoid(x = input_597_split_cast_fp16_1)[name = string("input_597_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_597_cast_fp16 = mul(x = input_597_split_cast_fp16_0, y = input_597_split_1_sigmoid_cast_fp16)[name = string("input_597_cast_fp16")];
            tensor<int32, [8]> input_599_pad_1 = const()[name = string("input_599_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_599_mode_1 = const()[name = string("input_599_mode_1"), val = string("constant")];
            fp16 const_2001_to_fp16 = const()[name = string("const_2001_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_599_cast_fp16 = pad(constant_val = const_2001_to_fp16, mode = input_599_mode_1, pad = input_599_pad_1, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")];
            string dense_output_1045_pad_type_1 = const()[name = string("dense_output_1045_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1045_strides_1 = const()[name = string("dense_output_1045_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1045_pad_1 = const()[name = string("dense_output_1045_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1045_dilations_1 = const()[name = string("dense_output_1045_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1045_groups_1 = const()[name = string("dense_output_1045_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1045_cast_fp16 = conv(dilations = dense_output_1045_dilations_1, groups = dense_output_1045_groups_1, pad = dense_output_1045_pad_1, pad_type = dense_output_1045_pad_type_1, strides = dense_output_1045_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_599_cast_fp16)[name = string("dense_output_1045_cast_fp16")];
            string sparse_output_1045_pad_type_1 = const()[name = string("sparse_output_1045_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1045_strides_1 = const()[name = string("sparse_output_1045_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1045_pad_1 = const()[name = string("sparse_output_1045_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1045_dilations_1 = const()[name = string("sparse_output_1045_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1045_groups_1 = const()[name = string("sparse_output_1045_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_12_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373511360))), nonzero_data = tensor<fp16, [9215]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373492864))))[name = string("layers_12_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1045_cast_fp16 = conv(dilations = sparse_output_1045_dilations_1, groups = sparse_output_1045_groups_1, pad = sparse_output_1045_pad_1, pad_type = sparse_output_1045_pad_type_1, strides = sparse_output_1045_strides_1, weight = layers_12_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_599_cast_fp16)[name = string("sparse_output_1045_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_601_cast_fp16 = add(x = dense_output_1045_cast_fp16, y = sparse_output_1045_cast_fp16)[name = string("input_601_cast_fp16")];
            tensor<fp16, [1024]> layers_12_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_12_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374691072)))];
            tensor<fp16, [1024]> layers_12_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_12_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374693184)))];
            tensor<fp16, [1024]> layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374695296)))];
            tensor<fp16, [1024]> layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374697408)))];
            tensor<fp16, [1, 1024, 51, 1]> input_603_cast_fp16 = batch_norm(beta = layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_15440_to_fp16, gamma = layers_12_conv_batch_norm_weight_to_fp16, mean = layers_12_conv_batch_norm_running_mean_to_fp16, variance = layers_12_conv_batch_norm_running_var_to_fp16, x = input_601_cast_fp16)[name = string("input_603_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_605_cast_fp16 = silu(x = input_603_cast_fp16)[name = string("input_605_cast_fp16")];
            string dense_output_1047_pad_type_1 = const()[name = string("dense_output_1047_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1047_strides_1 = const()[name = string("dense_output_1047_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1047_pad_1 = const()[name = string("dense_output_1047_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1047_dilations_1 = const()[name = string("dense_output_1047_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1047_groups_1 = const()[name = string("dense_output_1047_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374699520))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375748160))))[name = string("layers_12_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1047_cast_fp16 = conv(dilations = dense_output_1047_dilations_1, groups = dense_output_1047_groups_1, pad = dense_output_1047_pad_1, pad_type = dense_output_1047_pad_type_1, strides = dense_output_1047_strides_1, weight = layers_12_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_605_cast_fp16)[name = string("dense_output_1047_cast_fp16")];
            string sparse_output_1047_pad_type_1 = const()[name = string("sparse_output_1047_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1047_strides_1 = const()[name = string("sparse_output_1047_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1047_pad_1 = const()[name = string("sparse_output_1047_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1047_dilations_1 = const()[name = string("sparse_output_1047_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1047_groups_1 = const()[name = string("sparse_output_1047_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_12_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375769792))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375748736))))[name = string("layers_12_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1047_cast_fp16 = conv(dilations = sparse_output_1047_dilations_1, groups = sparse_output_1047_groups_1, pad = sparse_output_1047_pad_1, pad_type = sparse_output_1047_pad_type_1, strides = sparse_output_1047_strides_1, weight = layers_12_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_605_cast_fp16)[name = string("sparse_output_1047_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1095_cast_fp16 = add(x = dense_output_1047_cast_fp16, y = sparse_output_1047_cast_fp16)[name = string("x_1095_cast_fp16")];
            tensor<int32, [1]> var_16637_axes_1 = const()[name = string("op_16637_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_16637_cast_fp16 = squeeze(axes = var_16637_axes_1, x = x_1095_cast_fp16)[name = string("op_16637_cast_fp16")];
            tensor<int32, [3]> var_16638_perm_1 = const()[name = string("op_16638_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_16638_cast_fp16 = transpose(perm = var_16638_perm_1, x = var_16637_cast_fp16)[name = string("transpose_453")];
            tensor<fp16, [1, 51, 1024]> input_607_cast_fp16 = add(x = input_591_cast_fp16, y = var_16638_cast_fp16)[name = string("input_607_cast_fp16")];
            tensor<int32, [1]> x_1097_axes_1 = const()[name = string("x_1097_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375900928)))];
            tensor<fp16, [1024]> layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375903040)))];
            tensor<fp16, [1, 51, 1024]> x_1097_cast_fp16 = layer_norm(axes = x_1097_axes_1, beta = layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_15440_to_fp16, gamma = layers_12_norm_feed_forward2_weight_to_fp16, x = input_607_cast_fp16)[name = string("x_1097_cast_fp16")];
            tensor<int32, [4]> var_16648 = const()[name = string("op_16648"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1099_cast_fp16 = reshape(shape = var_16648, x = x_1097_cast_fp16)[name = string("x_1099_cast_fp16")];
            tensor<int32, [4]> input_609_perm_1 = const()[name = string("input_609_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1049_pad_type_1 = const()[name = string("dense_output_1049_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1049_strides_1 = const()[name = string("dense_output_1049_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1049_pad_1 = const()[name = string("dense_output_1049_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1049_dilations_1 = const()[name = string("dense_output_1049_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1049_groups_1 = const()[name = string("dense_output_1049_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_12_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375905152))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380099520))))[name = string("layers_12_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_609_cast_fp16 = transpose(perm = input_609_perm_1, x = x_1099_cast_fp16)[name = string("transpose_452")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1049_cast_fp16 = conv(dilations = dense_output_1049_dilations_1, groups = dense_output_1049_groups_1, pad = dense_output_1049_pad_1, pad_type = dense_output_1049_pad_type_1, strides = dense_output_1049_strides_1, weight = layers_12_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_609_cast_fp16)[name = string("dense_output_1049_cast_fp16")];
            string sparse_output_1049_pad_type_1 = const()[name = string("sparse_output_1049_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1049_strides_1 = const()[name = string("sparse_output_1049_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1049_pad_1 = const()[name = string("sparse_output_1049_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1049_dilations_1 = const()[name = string("sparse_output_1049_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1049_groups_1 = const()[name = string("sparse_output_1049_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_12_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380184064))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380100096))))[name = string("layers_12_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1049_cast_fp16 = conv(dilations = sparse_output_1049_dilations_1, groups = sparse_output_1049_groups_1, pad = sparse_output_1049_pad_1, pad_type = sparse_output_1049_pad_type_1, strides = sparse_output_1049_strides_1, weight = layers_12_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_609_cast_fp16)[name = string("sparse_output_1049_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_611_cast_fp16 = add(x = dense_output_1049_cast_fp16, y = sparse_output_1049_cast_fp16)[name = string("input_611_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_613_cast_fp16 = silu(x = input_611_cast_fp16)[name = string("input_613_cast_fp16")];
            string dense_output_1051_pad_type_1 = const()[name = string("dense_output_1051_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1051_strides_1 = const()[name = string("dense_output_1051_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1051_pad_1 = const()[name = string("dense_output_1051_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1051_dilations_1 = const()[name = string("dense_output_1051_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1051_groups_1 = const()[name = string("dense_output_1051_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_12_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380708416))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384902784))))[name = string("layers_12_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1051_cast_fp16 = conv(dilations = dense_output_1051_dilations_1, groups = dense_output_1051_groups_1, pad = dense_output_1051_pad_1, pad_type = dense_output_1051_pad_type_1, strides = dense_output_1051_strides_1, weight = layers_12_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_613_cast_fp16)[name = string("dense_output_1051_cast_fp16")];
            string sparse_output_1051_pad_type_1 = const()[name = string("sparse_output_1051_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1051_strides_1 = const()[name = string("sparse_output_1051_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1051_pad_1 = const()[name = string("sparse_output_1051_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1051_dilations_1 = const()[name = string("sparse_output_1051_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1051_groups_1 = const()[name = string("sparse_output_1051_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_12_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384987328))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384903360))))[name = string("layers_12_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1051_cast_fp16 = conv(dilations = sparse_output_1051_dilations_1, groups = sparse_output_1051_groups_1, pad = sparse_output_1051_pad_1, pad_type = sparse_output_1051_pad_type_1, strides = sparse_output_1051_strides_1, weight = layers_12_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_613_cast_fp16)[name = string("sparse_output_1051_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1101_cast_fp16 = add(x = dense_output_1051_cast_fp16, y = sparse_output_1051_cast_fp16)[name = string("x_1101_cast_fp16")];
            tensor<int32, [4]> x_1103_perm_1 = const()[name = string("x_1103_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_16683 = const()[name = string("op_16683"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1103_cast_fp16 = transpose(perm = x_1103_perm_1, x = x_1101_cast_fp16)[name = string("transpose_451")];
            tensor<fp16, [1, 51, 1024]> var_16684_cast_fp16 = reshape(shape = var_16683, x = x_1103_cast_fp16)[name = string("op_16684_cast_fp16")];
            fp16 var_16685_to_fp16 = const()[name = string("op_16685_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_16686_cast_fp16 = mul(x = var_16684_cast_fp16, y = var_16685_to_fp16)[name = string("op_16686_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_615_cast_fp16 = add(x = input_607_cast_fp16, y = var_16686_cast_fp16)[name = string("input_615_cast_fp16")];
            tensor<int32, [1]> input_617_axes_1 = const()[name = string("input_617_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_12_norm_out_weight_to_fp16 = const()[name = string("layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385511680)))];
            tensor<fp16, [1024]> layers_12_norm_out_bias_to_fp16 = const()[name = string("layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385513792)))];
            tensor<fp16, [1, 51, 1024]> input_617_cast_fp16 = layer_norm(axes = input_617_axes_1, beta = layers_12_norm_out_bias_to_fp16, epsilon = var_15440_to_fp16, gamma = layers_12_norm_out_weight_to_fp16, x = input_615_cast_fp16)[name = string("input_617_cast_fp16")];
            int32 var_16694 = const()[name = string("op_16694"), val = int32(-1)];
            tensor<int32, [1]> x_1105_axes_1 = const()[name = string("x_1105_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385515904)))];
            tensor<fp16, [1024]> layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385518016)))];
            fp16 var_16709_to_fp16 = const()[name = string("op_16709_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1105_cast_fp16 = layer_norm(axes = x_1105_axes_1, beta = layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_16709_to_fp16, gamma = layers_13_norm_feed_forward1_weight_to_fp16, x = input_617_cast_fp16)[name = string("x_1105_cast_fp16")];
            tensor<int32, [4]> var_16728 = const()[name = string("op_16728"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1107_cast_fp16 = reshape(shape = var_16728, x = x_1105_cast_fp16)[name = string("x_1107_cast_fp16")];
            tensor<int32, [4]> input_619_perm_1 = const()[name = string("input_619_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1053_pad_type_1 = const()[name = string("dense_output_1053_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1053_strides_1 = const()[name = string("dense_output_1053_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1053_pad_1 = const()[name = string("dense_output_1053_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1053_dilations_1 = const()[name = string("dense_output_1053_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1053_groups_1 = const()[name = string("dense_output_1053_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_13_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385520128))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389714496))))[name = string("layers_13_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_619_cast_fp16 = transpose(perm = input_619_perm_1, x = x_1107_cast_fp16)[name = string("transpose_450")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1053_cast_fp16 = conv(dilations = dense_output_1053_dilations_1, groups = dense_output_1053_groups_1, pad = dense_output_1053_pad_1, pad_type = dense_output_1053_pad_type_1, strides = dense_output_1053_strides_1, weight = layers_13_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_619_cast_fp16)[name = string("dense_output_1053_cast_fp16")];
            string sparse_output_1053_pad_type_1 = const()[name = string("sparse_output_1053_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1053_strides_1 = const()[name = string("sparse_output_1053_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1053_pad_1 = const()[name = string("sparse_output_1053_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1053_dilations_1 = const()[name = string("sparse_output_1053_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1053_groups_1 = const()[name = string("sparse_output_1053_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_13_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389799040))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(389715072))))[name = string("layers_13_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1053_cast_fp16 = conv(dilations = sparse_output_1053_dilations_1, groups = sparse_output_1053_groups_1, pad = sparse_output_1053_pad_1, pad_type = sparse_output_1053_pad_type_1, strides = sparse_output_1053_strides_1, weight = layers_13_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_619_cast_fp16)[name = string("sparse_output_1053_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_621_cast_fp16 = add(x = dense_output_1053_cast_fp16, y = sparse_output_1053_cast_fp16)[name = string("input_621_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_623_cast_fp16 = silu(x = input_621_cast_fp16)[name = string("input_623_cast_fp16")];
            string dense_output_1055_pad_type_1 = const()[name = string("dense_output_1055_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1055_strides_1 = const()[name = string("dense_output_1055_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1055_pad_1 = const()[name = string("dense_output_1055_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1055_dilations_1 = const()[name = string("dense_output_1055_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1055_groups_1 = const()[name = string("dense_output_1055_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_13_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390323392))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394517760))))[name = string("layers_13_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1055_cast_fp16 = conv(dilations = dense_output_1055_dilations_1, groups = dense_output_1055_groups_1, pad = dense_output_1055_pad_1, pad_type = dense_output_1055_pad_type_1, strides = dense_output_1055_strides_1, weight = layers_13_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_623_cast_fp16)[name = string("dense_output_1055_cast_fp16")];
            string sparse_output_1055_pad_type_1 = const()[name = string("sparse_output_1055_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1055_strides_1 = const()[name = string("sparse_output_1055_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1055_pad_1 = const()[name = string("sparse_output_1055_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1055_dilations_1 = const()[name = string("sparse_output_1055_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1055_groups_1 = const()[name = string("sparse_output_1055_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_13_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394602304))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394518336))))[name = string("layers_13_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1055_cast_fp16 = conv(dilations = sparse_output_1055_dilations_1, groups = sparse_output_1055_groups_1, pad = sparse_output_1055_pad_1, pad_type = sparse_output_1055_pad_type_1, strides = sparse_output_1055_strides_1, weight = layers_13_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_623_cast_fp16)[name = string("sparse_output_1055_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1109_cast_fp16 = add(x = dense_output_1055_cast_fp16, y = sparse_output_1055_cast_fp16)[name = string("x_1109_cast_fp16")];
            tensor<int32, [4]> x_1111_perm_1 = const()[name = string("x_1111_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_16763 = const()[name = string("op_16763"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1111_cast_fp16 = transpose(perm = x_1111_perm_1, x = x_1109_cast_fp16)[name = string("transpose_449")];
            tensor<fp16, [1, 51, 1024]> var_16764_cast_fp16 = reshape(shape = var_16763, x = x_1111_cast_fp16)[name = string("op_16764_cast_fp16")];
            fp16 var_16765_to_fp16 = const()[name = string("op_16765_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_16766_cast_fp16 = mul(x = var_16764_cast_fp16, y = var_16765_to_fp16)[name = string("op_16766_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_625_cast_fp16 = add(x = input_617_cast_fp16, y = var_16766_cast_fp16)[name = string("input_625_cast_fp16")];
            tensor<int32, [1]> q_27_axes_1 = const()[name = string("q_27_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_13_norm_self_att_weight_to_fp16 = const()[name = string("layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395126656)))];
            tensor<fp16, [1024]> layers_13_norm_self_att_bias_to_fp16 = const()[name = string("layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395128768)))];
            tensor<fp16, [1, 51, 1024]> q_27_cast_fp16 = layer_norm(axes = q_27_axes_1, beta = layers_13_norm_self_att_bias_to_fp16, epsilon = var_16709_to_fp16, gamma = layers_13_norm_self_att_weight_to_fp16, x = input_625_cast_fp16)[name = string("q_27_cast_fp16")];
            tensor<int32, [3]> var_16840 = const()[name = string("op_16840"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_627_axes_1 = const()[name = string("input_627_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_16841_cast_fp16 = transpose(perm = var_16840, x = q_27_cast_fp16)[name = string("transpose_448")];
            tensor<fp16, [1, 1024, 51, 1]> input_627_cast_fp16 = expand_dims(axes = input_627_axes_1, x = var_16841_cast_fp16)[name = string("input_627_cast_fp16")];
            string dense_output_1057_pad_type_1 = const()[name = string("dense_output_1057_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1057_strides_1 = const()[name = string("dense_output_1057_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1057_pad_1 = const()[name = string("dense_output_1057_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1057_dilations_1 = const()[name = string("dense_output_1057_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1057_groups_1 = const()[name = string("dense_output_1057_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395130880))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395262016))))[name = string("layers_13_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1057_cast_fp16 = conv(dilations = dense_output_1057_dilations_1, groups = dense_output_1057_groups_1, pad = dense_output_1057_pad_1, pad_type = dense_output_1057_pad_type_1, strides = dense_output_1057_strides_1, weight = layers_13_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1057_cast_fp16")];
            string sparse_output_1057_pad_type_1 = const()[name = string("sparse_output_1057_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1057_strides_1 = const()[name = string("sparse_output_1057_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1057_pad_1 = const()[name = string("sparse_output_1057_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1057_dilations_1 = const()[name = string("sparse_output_1057_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1057_groups_1 = const()[name = string("sparse_output_1057_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395265280))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395262592))))[name = string("layers_13_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1057_cast_fp16 = conv(dilations = sparse_output_1057_dilations_1, groups = sparse_output_1057_groups_1, pad = sparse_output_1057_pad_1, pad_type = sparse_output_1057_pad_type_1, strides = sparse_output_1057_strides_1, weight = layers_13_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1057_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16866_cast_fp16 = add(x = dense_output_1057_cast_fp16, y = sparse_output_1057_cast_fp16)[name = string("op_16866_cast_fp16")];
            tensor<int32, [4]> var_16867 = const()[name = string("op_16867"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16869 = const()[name = string("op_16869"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16868_cast_fp16 = transpose(perm = var_16867, x = var_16866_cast_fp16)[name = string("transpose_447")];
            tensor<fp16, [1, 51, 128]> q_head_209_cast_fp16 = reshape(shape = var_16869, x = var_16868_cast_fp16)[name = string("q_head_209_cast_fp16")];
            string dense_output_1059_pad_type_1 = const()[name = string("dense_output_1059_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1059_strides_1 = const()[name = string("dense_output_1059_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1059_pad_1 = const()[name = string("dense_output_1059_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1059_dilations_1 = const()[name = string("dense_output_1059_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1059_groups_1 = const()[name = string("dense_output_1059_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395281728))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395412864))))[name = string("layers_13_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1059_cast_fp16 = conv(dilations = dense_output_1059_dilations_1, groups = dense_output_1059_groups_1, pad = dense_output_1059_pad_1, pad_type = dense_output_1059_pad_type_1, strides = dense_output_1059_strides_1, weight = layers_13_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1059_cast_fp16")];
            string sparse_output_1059_pad_type_1 = const()[name = string("sparse_output_1059_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1059_strides_1 = const()[name = string("sparse_output_1059_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1059_pad_1 = const()[name = string("sparse_output_1059_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1059_dilations_1 = const()[name = string("sparse_output_1059_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1059_groups_1 = const()[name = string("sparse_output_1059_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395416128))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395413440))))[name = string("layers_13_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1059_cast_fp16 = conv(dilations = sparse_output_1059_dilations_1, groups = sparse_output_1059_groups_1, pad = sparse_output_1059_pad_1, pad_type = sparse_output_1059_pad_type_1, strides = sparse_output_1059_strides_1, weight = layers_13_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1059_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16885_cast_fp16 = add(x = dense_output_1059_cast_fp16, y = sparse_output_1059_cast_fp16)[name = string("op_16885_cast_fp16")];
            tensor<int32, [4]> var_16886 = const()[name = string("op_16886"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16888 = const()[name = string("op_16888"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16887_cast_fp16 = transpose(perm = var_16886, x = var_16885_cast_fp16)[name = string("transpose_446")];
            tensor<fp16, [1, 51, 128]> k_head_417_cast_fp16 = reshape(shape = var_16888, x = var_16887_cast_fp16)[name = string("k_head_417_cast_fp16")];
            string dense_output_1061_pad_type_1 = const()[name = string("dense_output_1061_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1061_strides_1 = const()[name = string("dense_output_1061_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1061_pad_1 = const()[name = string("dense_output_1061_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1061_dilations_1 = const()[name = string("dense_output_1061_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1061_groups_1 = const()[name = string("dense_output_1061_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395432576))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395563712))))[name = string("layers_13_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1061_cast_fp16 = conv(dilations = dense_output_1061_dilations_1, groups = dense_output_1061_groups_1, pad = dense_output_1061_pad_1, pad_type = dense_output_1061_pad_type_1, strides = dense_output_1061_strides_1, weight = layers_13_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1061_cast_fp16")];
            string sparse_output_1061_pad_type_1 = const()[name = string("sparse_output_1061_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1061_strides_1 = const()[name = string("sparse_output_1061_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1061_pad_1 = const()[name = string("sparse_output_1061_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1061_dilations_1 = const()[name = string("sparse_output_1061_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1061_groups_1 = const()[name = string("sparse_output_1061_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395566976))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395564288))))[name = string("layers_13_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1061_cast_fp16 = conv(dilations = sparse_output_1061_dilations_1, groups = sparse_output_1061_groups_1, pad = sparse_output_1061_pad_1, pad_type = sparse_output_1061_pad_type_1, strides = sparse_output_1061_strides_1, weight = layers_13_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1061_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16904_cast_fp16 = add(x = dense_output_1061_cast_fp16, y = sparse_output_1061_cast_fp16)[name = string("op_16904_cast_fp16")];
            tensor<int32, [4]> var_16905 = const()[name = string("op_16905"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16907 = const()[name = string("op_16907"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16906_cast_fp16 = transpose(perm = var_16905, x = var_16904_cast_fp16)[name = string("transpose_445")];
            tensor<fp16, [1, 51, 128]> v_head_417_cast_fp16 = reshape(shape = var_16907, x = var_16906_cast_fp16)[name = string("v_head_417_cast_fp16")];
            string dense_output_1063_pad_type_1 = const()[name = string("dense_output_1063_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1063_strides_1 = const()[name = string("dense_output_1063_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1063_pad_1 = const()[name = string("dense_output_1063_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1063_dilations_1 = const()[name = string("dense_output_1063_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1063_groups_1 = const()[name = string("dense_output_1063_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395583424))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395714560))))[name = string("layers_13_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1063_cast_fp16 = conv(dilations = dense_output_1063_dilations_1, groups = dense_output_1063_groups_1, pad = dense_output_1063_pad_1, pad_type = dense_output_1063_pad_type_1, strides = dense_output_1063_strides_1, weight = layers_13_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1063_cast_fp16")];
            string sparse_output_1063_pad_type_1 = const()[name = string("sparse_output_1063_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1063_strides_1 = const()[name = string("sparse_output_1063_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1063_pad_1 = const()[name = string("sparse_output_1063_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1063_dilations_1 = const()[name = string("sparse_output_1063_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1063_groups_1 = const()[name = string("sparse_output_1063_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395717824))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395715136))))[name = string("layers_13_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1063_cast_fp16 = conv(dilations = sparse_output_1063_dilations_1, groups = sparse_output_1063_groups_1, pad = sparse_output_1063_pad_1, pad_type = sparse_output_1063_pad_type_1, strides = sparse_output_1063_strides_1, weight = layers_13_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1063_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_16923_cast_fp16 = add(x = dense_output_1063_cast_fp16, y = sparse_output_1063_cast_fp16)[name = string("op_16923_cast_fp16")];
            tensor<int32, [4]> var_16924 = const()[name = string("op_16924"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16926 = const()[name = string("op_16926"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_16925_cast_fp16 = transpose(perm = var_16924, x = var_16923_cast_fp16)[name = string("transpose_444")];
            tensor<fp16, [1, 101, 128]> p_head_417_cast_fp16 = reshape(shape = var_16926, x = var_16925_cast_fp16)[name = string("p_head_417_cast_fp16")];
            tensor<fp16, [128]> var_16928_to_fp16 = const()[name = string("op_16928_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395734272)))];
            tensor<fp16, [1, 51, 128]> var_16929_cast_fp16 = add(x = q_head_209_cast_fp16, y = var_16928_to_fp16)[name = string("op_16929_cast_fp16")];
            tensor<int32, [1]> q_u_209_axes_1 = const()[name = string("q_u_209_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_209_cast_fp16 = expand_dims(axes = q_u_209_axes_1, x = var_16929_cast_fp16)[name = string("q_u_209_cast_fp16")];
            tensor<fp16, [128]> var_16931_to_fp16 = const()[name = string("op_16931_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395734592)))];
            tensor<fp16, [1, 51, 128]> var_16932_cast_fp16 = add(x = q_head_209_cast_fp16, y = var_16931_to_fp16)[name = string("op_16932_cast_fp16")];
            tensor<int32, [1]> q_v_209_axes_1 = const()[name = string("q_v_209_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_209_cast_fp16 = expand_dims(axes = q_v_209_axes_1, x = var_16932_cast_fp16)[name = string("q_v_209_cast_fp16")];
            tensor<int32, [1]> k_head_419_axes_1 = const()[name = string("k_head_419_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_419_cast_fp16 = expand_dims(axes = k_head_419_axes_1, x = k_head_417_cast_fp16)[name = string("k_head_419_cast_fp16")];
            tensor<int32, [1]> v_head_419_axes_1 = const()[name = string("v_head_419_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_419_cast_fp16 = expand_dims(axes = v_head_419_axes_1, x = v_head_417_cast_fp16)[name = string("v_head_419_cast_fp16")];
            tensor<int32, [1]> p_head_419_axes_1 = const()[name = string("p_head_419_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_419_cast_fp16 = expand_dims(axes = p_head_419_axes_1, x = p_head_417_cast_fp16)[name = string("p_head_419_cast_fp16")];
            bool var_16938_transpose_x_3 = const()[name = string("op_16938_transpose_x_3"), val = bool(false)];
            bool var_16938_transpose_y_3 = const()[name = string("op_16938_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_16938_cast_fp16 = matmul(transpose_x = var_16938_transpose_x_3, transpose_y = var_16938_transpose_y_3, x = q_u_209_cast_fp16, y = k_head_419_cast_fp16)[name = string("op_16938_cast_fp16")];
            fp16 var_16939_to_fp16 = const()[name = string("op_16939_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_209_cast_fp16 = mul(x = var_16938_cast_fp16, y = var_16939_to_fp16)[name = string("scores_content_209_cast_fp16")];
            bool x_1113_transpose_x_3 = const()[name = string("x_1113_transpose_x_3"), val = bool(false)];
            bool x_1113_transpose_y_3 = const()[name = string("x_1113_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1113_cast_fp16 = matmul(transpose_x = x_1113_transpose_x_3, transpose_y = x_1113_transpose_y_3, x = q_v_209_cast_fp16, y = p_head_419_cast_fp16)[name = string("x_1113_cast_fp16")];
            tensor<int32, [8]> x_1115_pad_1 = const()[name = string("x_1115_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1115_mode_1 = const()[name = string("x_1115_mode_1"), val = string("constant")];
            fp16 const_2011_to_fp16 = const()[name = string("const_2011_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1115_cast_fp16 = pad(constant_val = const_2011_to_fp16, mode = x_1115_mode_1, pad = x_1115_pad_1, x = x_1113_cast_fp16)[name = string("x_1115_cast_fp16")];
            tensor<int32, [4]> var_16953 = const()[name = string("op_16953"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1117_cast_fp16 = reshape(shape = var_16953, x = x_1115_cast_fp16)[name = string("x_1117_cast_fp16")];
            tensor<int32, [4]> var_16957_begin_1 = const()[name = string("op_16957_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_16957_end_1 = const()[name = string("op_16957_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_16957_end_mask_1 = const()[name = string("op_16957_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_16957_cast_fp16 = slice_by_index(begin = var_16957_begin_1, end = var_16957_end_1, end_mask = var_16957_end_mask_1, x = x_1117_cast_fp16)[name = string("op_16957_cast_fp16")];
            tensor<int32, [4]> var_16959 = const()[name = string("op_16959"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_16960_cast_fp16 = reshape(shape = var_16959, x = var_16957_cast_fp16)[name = string("op_16960_cast_fp16")];
            tensor<int32, [4]> var_16965_begin_1 = const()[name = string("op_16965_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_16965_end_1 = const()[name = string("op_16965_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_16965_end_mask_1 = const()[name = string("op_16965_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_16965_cast_fp16 = slice_by_index(begin = var_16965_begin_1, end = var_16965_end_1, end_mask = var_16965_end_mask_1, x = var_16960_cast_fp16)[name = string("op_16965_cast_fp16")];
            fp16 var_16966_to_fp16 = const()[name = string("op_16966_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_209_cast_fp16 = mul(x = var_16965_cast_fp16, y = var_16966_to_fp16)[name = string("scores_pos_209_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_209_cast_fp16 = add(x = scores_content_209_cast_fp16, y = scores_pos_209_cast_fp16)[name = string("logits_209_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_16969_cast_fp16 = softmax(axis = var_16694, x = logits_209_cast_fp16)[name = string("op_16969_cast_fp16")];
            bool var_16971_transpose_x_1 = const()[name = string("op_16971_transpose_x_1"), val = bool(false)];
            bool var_16971_transpose_y_1 = const()[name = string("op_16971_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_16971_cast_fp16 = matmul(transpose_x = var_16971_transpose_x_1, transpose_y = var_16971_transpose_y_1, x = var_16969_cast_fp16, y = v_head_419_cast_fp16)[name = string("op_16971_cast_fp16")];
            tensor<int32, [1]> var_16972_axes_1 = const()[name = string("op_16972_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_16972_cast_fp16 = squeeze(axes = var_16972_axes_1, x = var_16971_cast_fp16)[name = string("op_16972_cast_fp16")];
            string dense_output_1065_pad_type_1 = const()[name = string("dense_output_1065_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1065_strides_1 = const()[name = string("dense_output_1065_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1065_pad_1 = const()[name = string("dense_output_1065_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1065_dilations_1 = const()[name = string("dense_output_1065_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1065_groups_1 = const()[name = string("dense_output_1065_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395734912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395866048))))[name = string("layers_13_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1065_cast_fp16 = conv(dilations = dense_output_1065_dilations_1, groups = dense_output_1065_groups_1, pad = dense_output_1065_pad_1, pad_type = dense_output_1065_pad_type_1, strides = dense_output_1065_strides_1, weight = layers_13_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1065_cast_fp16")];
            string sparse_output_1065_pad_type_1 = const()[name = string("sparse_output_1065_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1065_strides_1 = const()[name = string("sparse_output_1065_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1065_pad_1 = const()[name = string("sparse_output_1065_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1065_dilations_1 = const()[name = string("sparse_output_1065_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1065_groups_1 = const()[name = string("sparse_output_1065_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395869312))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395866624))))[name = string("layers_13_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1065_cast_fp16 = conv(dilations = sparse_output_1065_dilations_1, groups = sparse_output_1065_groups_1, pad = sparse_output_1065_pad_1, pad_type = sparse_output_1065_pad_type_1, strides = sparse_output_1065_strides_1, weight = layers_13_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1065_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_16987_cast_fp16 = add(x = dense_output_1065_cast_fp16, y = sparse_output_1065_cast_fp16)[name = string("op_16987_cast_fp16")];
            tensor<int32, [4]> var_16988 = const()[name = string("op_16988"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_16990 = const()[name = string("op_16990"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_16989_cast_fp16 = transpose(perm = var_16988, x = var_16987_cast_fp16)[name = string("transpose_443")];
            tensor<fp16, [1, 51, 128]> q_head_211_cast_fp16 = reshape(shape = var_16990, x = var_16989_cast_fp16)[name = string("q_head_211_cast_fp16")];
            string dense_output_1067_pad_type_1 = const()[name = string("dense_output_1067_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1067_strides_1 = const()[name = string("dense_output_1067_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1067_pad_1 = const()[name = string("dense_output_1067_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1067_dilations_1 = const()[name = string("dense_output_1067_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1067_groups_1 = const()[name = string("dense_output_1067_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395885760))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396016896))))[name = string("layers_13_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1067_cast_fp16 = conv(dilations = dense_output_1067_dilations_1, groups = dense_output_1067_groups_1, pad = dense_output_1067_pad_1, pad_type = dense_output_1067_pad_type_1, strides = dense_output_1067_strides_1, weight = layers_13_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1067_cast_fp16")];
            string sparse_output_1067_pad_type_1 = const()[name = string("sparse_output_1067_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1067_strides_1 = const()[name = string("sparse_output_1067_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1067_pad_1 = const()[name = string("sparse_output_1067_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1067_dilations_1 = const()[name = string("sparse_output_1067_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1067_groups_1 = const()[name = string("sparse_output_1067_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396020160))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396017472))))[name = string("layers_13_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1067_cast_fp16 = conv(dilations = sparse_output_1067_dilations_1, groups = sparse_output_1067_groups_1, pad = sparse_output_1067_pad_1, pad_type = sparse_output_1067_pad_type_1, strides = sparse_output_1067_strides_1, weight = layers_13_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1067_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17006_cast_fp16 = add(x = dense_output_1067_cast_fp16, y = sparse_output_1067_cast_fp16)[name = string("op_17006_cast_fp16")];
            tensor<int32, [4]> var_17007 = const()[name = string("op_17007"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17009 = const()[name = string("op_17009"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17008_cast_fp16 = transpose(perm = var_17007, x = var_17006_cast_fp16)[name = string("transpose_442")];
            tensor<fp16, [1, 51, 128]> k_head_421_cast_fp16 = reshape(shape = var_17009, x = var_17008_cast_fp16)[name = string("k_head_421_cast_fp16")];
            string dense_output_1069_pad_type_1 = const()[name = string("dense_output_1069_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1069_strides_1 = const()[name = string("dense_output_1069_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1069_pad_1 = const()[name = string("dense_output_1069_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1069_dilations_1 = const()[name = string("dense_output_1069_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1069_groups_1 = const()[name = string("dense_output_1069_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396036608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396167744))))[name = string("layers_13_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1069_cast_fp16 = conv(dilations = dense_output_1069_dilations_1, groups = dense_output_1069_groups_1, pad = dense_output_1069_pad_1, pad_type = dense_output_1069_pad_type_1, strides = dense_output_1069_strides_1, weight = layers_13_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1069_cast_fp16")];
            string sparse_output_1069_pad_type_1 = const()[name = string("sparse_output_1069_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1069_strides_1 = const()[name = string("sparse_output_1069_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1069_pad_1 = const()[name = string("sparse_output_1069_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1069_dilations_1 = const()[name = string("sparse_output_1069_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1069_groups_1 = const()[name = string("sparse_output_1069_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396171008))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396168320))))[name = string("layers_13_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1069_cast_fp16 = conv(dilations = sparse_output_1069_dilations_1, groups = sparse_output_1069_groups_1, pad = sparse_output_1069_pad_1, pad_type = sparse_output_1069_pad_type_1, strides = sparse_output_1069_strides_1, weight = layers_13_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1069_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17025_cast_fp16 = add(x = dense_output_1069_cast_fp16, y = sparse_output_1069_cast_fp16)[name = string("op_17025_cast_fp16")];
            tensor<int32, [4]> var_17026 = const()[name = string("op_17026"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17028 = const()[name = string("op_17028"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17027_cast_fp16 = transpose(perm = var_17026, x = var_17025_cast_fp16)[name = string("transpose_441")];
            tensor<fp16, [1, 51, 128]> v_head_421_cast_fp16 = reshape(shape = var_17028, x = var_17027_cast_fp16)[name = string("v_head_421_cast_fp16")];
            string dense_output_1071_pad_type_1 = const()[name = string("dense_output_1071_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1071_strides_1 = const()[name = string("dense_output_1071_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1071_pad_1 = const()[name = string("dense_output_1071_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1071_dilations_1 = const()[name = string("dense_output_1071_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1071_groups_1 = const()[name = string("dense_output_1071_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396187456))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396318592))))[name = string("layers_13_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1071_cast_fp16 = conv(dilations = dense_output_1071_dilations_1, groups = dense_output_1071_groups_1, pad = dense_output_1071_pad_1, pad_type = dense_output_1071_pad_type_1, strides = dense_output_1071_strides_1, weight = layers_13_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1071_cast_fp16")];
            string sparse_output_1071_pad_type_1 = const()[name = string("sparse_output_1071_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1071_strides_1 = const()[name = string("sparse_output_1071_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1071_pad_1 = const()[name = string("sparse_output_1071_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1071_dilations_1 = const()[name = string("sparse_output_1071_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1071_groups_1 = const()[name = string("sparse_output_1071_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396321856))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396319168))))[name = string("layers_13_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1071_cast_fp16 = conv(dilations = sparse_output_1071_dilations_1, groups = sparse_output_1071_groups_1, pad = sparse_output_1071_pad_1, pad_type = sparse_output_1071_pad_type_1, strides = sparse_output_1071_strides_1, weight = layers_13_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1071_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_17044_cast_fp16 = add(x = dense_output_1071_cast_fp16, y = sparse_output_1071_cast_fp16)[name = string("op_17044_cast_fp16")];
            tensor<int32, [4]> var_17045 = const()[name = string("op_17045"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17047 = const()[name = string("op_17047"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_17046_cast_fp16 = transpose(perm = var_17045, x = var_17044_cast_fp16)[name = string("transpose_440")];
            tensor<fp16, [1, 101, 128]> p_head_421_cast_fp16 = reshape(shape = var_17047, x = var_17046_cast_fp16)[name = string("p_head_421_cast_fp16")];
            tensor<fp16, [128]> var_17049_to_fp16 = const()[name = string("op_17049_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396338304)))];
            tensor<fp16, [1, 51, 128]> var_17050_cast_fp16 = add(x = q_head_211_cast_fp16, y = var_17049_to_fp16)[name = string("op_17050_cast_fp16")];
            tensor<int32, [1]> q_u_211_axes_1 = const()[name = string("q_u_211_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_211_cast_fp16 = expand_dims(axes = q_u_211_axes_1, x = var_17050_cast_fp16)[name = string("q_u_211_cast_fp16")];
            tensor<fp16, [128]> var_17052_to_fp16 = const()[name = string("op_17052_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396338624)))];
            tensor<fp16, [1, 51, 128]> var_17053_cast_fp16 = add(x = q_head_211_cast_fp16, y = var_17052_to_fp16)[name = string("op_17053_cast_fp16")];
            tensor<int32, [1]> q_v_211_axes_1 = const()[name = string("q_v_211_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_211_cast_fp16 = expand_dims(axes = q_v_211_axes_1, x = var_17053_cast_fp16)[name = string("q_v_211_cast_fp16")];
            tensor<int32, [1]> k_head_423_axes_1 = const()[name = string("k_head_423_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_423_cast_fp16 = expand_dims(axes = k_head_423_axes_1, x = k_head_421_cast_fp16)[name = string("k_head_423_cast_fp16")];
            tensor<int32, [1]> v_head_423_axes_1 = const()[name = string("v_head_423_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_423_cast_fp16 = expand_dims(axes = v_head_423_axes_1, x = v_head_421_cast_fp16)[name = string("v_head_423_cast_fp16")];
            tensor<int32, [1]> p_head_423_axes_1 = const()[name = string("p_head_423_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_423_cast_fp16 = expand_dims(axes = p_head_423_axes_1, x = p_head_421_cast_fp16)[name = string("p_head_423_cast_fp16")];
            bool var_17059_transpose_x_3 = const()[name = string("op_17059_transpose_x_3"), val = bool(false)];
            bool var_17059_transpose_y_3 = const()[name = string("op_17059_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_17059_cast_fp16 = matmul(transpose_x = var_17059_transpose_x_3, transpose_y = var_17059_transpose_y_3, x = q_u_211_cast_fp16, y = k_head_423_cast_fp16)[name = string("op_17059_cast_fp16")];
            fp16 var_17060_to_fp16 = const()[name = string("op_17060_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_211_cast_fp16 = mul(x = var_17059_cast_fp16, y = var_17060_to_fp16)[name = string("scores_content_211_cast_fp16")];
            bool x_1121_transpose_x_3 = const()[name = string("x_1121_transpose_x_3"), val = bool(false)];
            bool x_1121_transpose_y_3 = const()[name = string("x_1121_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1121_cast_fp16 = matmul(transpose_x = x_1121_transpose_x_3, transpose_y = x_1121_transpose_y_3, x = q_v_211_cast_fp16, y = p_head_423_cast_fp16)[name = string("x_1121_cast_fp16")];
            tensor<int32, [8]> x_1123_pad_1 = const()[name = string("x_1123_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1123_mode_1 = const()[name = string("x_1123_mode_1"), val = string("constant")];
            fp16 const_2017_to_fp16 = const()[name = string("const_2017_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1123_cast_fp16 = pad(constant_val = const_2017_to_fp16, mode = x_1123_mode_1, pad = x_1123_pad_1, x = x_1121_cast_fp16)[name = string("x_1123_cast_fp16")];
            tensor<int32, [4]> var_17074 = const()[name = string("op_17074"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1125_cast_fp16 = reshape(shape = var_17074, x = x_1123_cast_fp16)[name = string("x_1125_cast_fp16")];
            tensor<int32, [4]> var_17078_begin_1 = const()[name = string("op_17078_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17078_end_1 = const()[name = string("op_17078_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_17078_end_mask_1 = const()[name = string("op_17078_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_17078_cast_fp16 = slice_by_index(begin = var_17078_begin_1, end = var_17078_end_1, end_mask = var_17078_end_mask_1, x = x_1125_cast_fp16)[name = string("op_17078_cast_fp16")];
            tensor<int32, [4]> var_17080 = const()[name = string("op_17080"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_17081_cast_fp16 = reshape(shape = var_17080, x = var_17078_cast_fp16)[name = string("op_17081_cast_fp16")];
            tensor<int32, [4]> var_17086_begin_1 = const()[name = string("op_17086_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17086_end_1 = const()[name = string("op_17086_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_17086_end_mask_1 = const()[name = string("op_17086_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_17086_cast_fp16 = slice_by_index(begin = var_17086_begin_1, end = var_17086_end_1, end_mask = var_17086_end_mask_1, x = var_17081_cast_fp16)[name = string("op_17086_cast_fp16")];
            fp16 var_17087_to_fp16 = const()[name = string("op_17087_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_211_cast_fp16 = mul(x = var_17086_cast_fp16, y = var_17087_to_fp16)[name = string("scores_pos_211_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_211_cast_fp16 = add(x = scores_content_211_cast_fp16, y = scores_pos_211_cast_fp16)[name = string("logits_211_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_17090_cast_fp16 = softmax(axis = var_16694, x = logits_211_cast_fp16)[name = string("op_17090_cast_fp16")];
            bool var_17092_transpose_x_1 = const()[name = string("op_17092_transpose_x_1"), val = bool(false)];
            bool var_17092_transpose_y_1 = const()[name = string("op_17092_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_17092_cast_fp16 = matmul(transpose_x = var_17092_transpose_x_1, transpose_y = var_17092_transpose_y_1, x = var_17090_cast_fp16, y = v_head_423_cast_fp16)[name = string("op_17092_cast_fp16")];
            tensor<int32, [1]> var_17093_axes_1 = const()[name = string("op_17093_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_17093_cast_fp16 = squeeze(axes = var_17093_axes_1, x = var_17092_cast_fp16)[name = string("op_17093_cast_fp16")];
            string dense_output_1073_pad_type_1 = const()[name = string("dense_output_1073_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1073_strides_1 = const()[name = string("dense_output_1073_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1073_pad_1 = const()[name = string("dense_output_1073_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1073_dilations_1 = const()[name = string("dense_output_1073_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1073_groups_1 = const()[name = string("dense_output_1073_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396338944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396470080))))[name = string("layers_13_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1073_cast_fp16 = conv(dilations = dense_output_1073_dilations_1, groups = dense_output_1073_groups_1, pad = dense_output_1073_pad_1, pad_type = dense_output_1073_pad_type_1, strides = dense_output_1073_strides_1, weight = layers_13_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1073_cast_fp16")];
            string sparse_output_1073_pad_type_1 = const()[name = string("sparse_output_1073_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1073_strides_1 = const()[name = string("sparse_output_1073_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1073_pad_1 = const()[name = string("sparse_output_1073_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1073_dilations_1 = const()[name = string("sparse_output_1073_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1073_groups_1 = const()[name = string("sparse_output_1073_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396473344))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396470656))))[name = string("layers_13_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1073_cast_fp16 = conv(dilations = sparse_output_1073_dilations_1, groups = sparse_output_1073_groups_1, pad = sparse_output_1073_pad_1, pad_type = sparse_output_1073_pad_type_1, strides = sparse_output_1073_strides_1, weight = layers_13_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1073_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17108_cast_fp16 = add(x = dense_output_1073_cast_fp16, y = sparse_output_1073_cast_fp16)[name = string("op_17108_cast_fp16")];
            tensor<int32, [4]> var_17109 = const()[name = string("op_17109"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17111 = const()[name = string("op_17111"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17110_cast_fp16 = transpose(perm = var_17109, x = var_17108_cast_fp16)[name = string("transpose_439")];
            tensor<fp16, [1, 51, 128]> q_head_213_cast_fp16 = reshape(shape = var_17111, x = var_17110_cast_fp16)[name = string("q_head_213_cast_fp16")];
            string dense_output_1075_pad_type_1 = const()[name = string("dense_output_1075_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1075_strides_1 = const()[name = string("dense_output_1075_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1075_pad_1 = const()[name = string("dense_output_1075_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1075_dilations_1 = const()[name = string("dense_output_1075_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1075_groups_1 = const()[name = string("dense_output_1075_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396489792))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396620928))))[name = string("layers_13_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1075_cast_fp16 = conv(dilations = dense_output_1075_dilations_1, groups = dense_output_1075_groups_1, pad = dense_output_1075_pad_1, pad_type = dense_output_1075_pad_type_1, strides = dense_output_1075_strides_1, weight = layers_13_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1075_cast_fp16")];
            string sparse_output_1075_pad_type_1 = const()[name = string("sparse_output_1075_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1075_strides_1 = const()[name = string("sparse_output_1075_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1075_pad_1 = const()[name = string("sparse_output_1075_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1075_dilations_1 = const()[name = string("sparse_output_1075_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1075_groups_1 = const()[name = string("sparse_output_1075_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396624192))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396621504))))[name = string("layers_13_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1075_cast_fp16 = conv(dilations = sparse_output_1075_dilations_1, groups = sparse_output_1075_groups_1, pad = sparse_output_1075_pad_1, pad_type = sparse_output_1075_pad_type_1, strides = sparse_output_1075_strides_1, weight = layers_13_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1075_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17127_cast_fp16 = add(x = dense_output_1075_cast_fp16, y = sparse_output_1075_cast_fp16)[name = string("op_17127_cast_fp16")];
            tensor<int32, [4]> var_17128 = const()[name = string("op_17128"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17130 = const()[name = string("op_17130"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17129_cast_fp16 = transpose(perm = var_17128, x = var_17127_cast_fp16)[name = string("transpose_438")];
            tensor<fp16, [1, 51, 128]> k_head_425_cast_fp16 = reshape(shape = var_17130, x = var_17129_cast_fp16)[name = string("k_head_425_cast_fp16")];
            string dense_output_1077_pad_type_1 = const()[name = string("dense_output_1077_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1077_strides_1 = const()[name = string("dense_output_1077_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1077_pad_1 = const()[name = string("dense_output_1077_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1077_dilations_1 = const()[name = string("dense_output_1077_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1077_groups_1 = const()[name = string("dense_output_1077_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396640640))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396771776))))[name = string("layers_13_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1077_cast_fp16 = conv(dilations = dense_output_1077_dilations_1, groups = dense_output_1077_groups_1, pad = dense_output_1077_pad_1, pad_type = dense_output_1077_pad_type_1, strides = dense_output_1077_strides_1, weight = layers_13_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1077_cast_fp16")];
            string sparse_output_1077_pad_type_1 = const()[name = string("sparse_output_1077_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1077_strides_1 = const()[name = string("sparse_output_1077_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1077_pad_1 = const()[name = string("sparse_output_1077_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1077_dilations_1 = const()[name = string("sparse_output_1077_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1077_groups_1 = const()[name = string("sparse_output_1077_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396775040))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396772352))))[name = string("layers_13_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1077_cast_fp16 = conv(dilations = sparse_output_1077_dilations_1, groups = sparse_output_1077_groups_1, pad = sparse_output_1077_pad_1, pad_type = sparse_output_1077_pad_type_1, strides = sparse_output_1077_strides_1, weight = layers_13_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1077_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17146_cast_fp16 = add(x = dense_output_1077_cast_fp16, y = sparse_output_1077_cast_fp16)[name = string("op_17146_cast_fp16")];
            tensor<int32, [4]> var_17147 = const()[name = string("op_17147"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17149 = const()[name = string("op_17149"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17148_cast_fp16 = transpose(perm = var_17147, x = var_17146_cast_fp16)[name = string("transpose_437")];
            tensor<fp16, [1, 51, 128]> v_head_425_cast_fp16 = reshape(shape = var_17149, x = var_17148_cast_fp16)[name = string("v_head_425_cast_fp16")];
            string dense_output_1079_pad_type_1 = const()[name = string("dense_output_1079_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1079_strides_1 = const()[name = string("dense_output_1079_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1079_pad_1 = const()[name = string("dense_output_1079_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1079_dilations_1 = const()[name = string("dense_output_1079_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1079_groups_1 = const()[name = string("dense_output_1079_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396791488))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396922624))))[name = string("layers_13_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1079_cast_fp16 = conv(dilations = dense_output_1079_dilations_1, groups = dense_output_1079_groups_1, pad = dense_output_1079_pad_1, pad_type = dense_output_1079_pad_type_1, strides = dense_output_1079_strides_1, weight = layers_13_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1079_cast_fp16")];
            string sparse_output_1079_pad_type_1 = const()[name = string("sparse_output_1079_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1079_strides_1 = const()[name = string("sparse_output_1079_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1079_pad_1 = const()[name = string("sparse_output_1079_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1079_dilations_1 = const()[name = string("sparse_output_1079_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1079_groups_1 = const()[name = string("sparse_output_1079_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396925888))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396923200))))[name = string("layers_13_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1079_cast_fp16 = conv(dilations = sparse_output_1079_dilations_1, groups = sparse_output_1079_groups_1, pad = sparse_output_1079_pad_1, pad_type = sparse_output_1079_pad_type_1, strides = sparse_output_1079_strides_1, weight = layers_13_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1079_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_17165_cast_fp16 = add(x = dense_output_1079_cast_fp16, y = sparse_output_1079_cast_fp16)[name = string("op_17165_cast_fp16")];
            tensor<int32, [4]> var_17166 = const()[name = string("op_17166"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17168 = const()[name = string("op_17168"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_17167_cast_fp16 = transpose(perm = var_17166, x = var_17165_cast_fp16)[name = string("transpose_436")];
            tensor<fp16, [1, 101, 128]> p_head_425_cast_fp16 = reshape(shape = var_17168, x = var_17167_cast_fp16)[name = string("p_head_425_cast_fp16")];
            tensor<fp16, [128]> var_17170_to_fp16 = const()[name = string("op_17170_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396942336)))];
            tensor<fp16, [1, 51, 128]> var_17171_cast_fp16 = add(x = q_head_213_cast_fp16, y = var_17170_to_fp16)[name = string("op_17171_cast_fp16")];
            tensor<int32, [1]> q_u_213_axes_1 = const()[name = string("q_u_213_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_213_cast_fp16 = expand_dims(axes = q_u_213_axes_1, x = var_17171_cast_fp16)[name = string("q_u_213_cast_fp16")];
            tensor<fp16, [128]> var_17173_to_fp16 = const()[name = string("op_17173_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396942656)))];
            tensor<fp16, [1, 51, 128]> var_17174_cast_fp16 = add(x = q_head_213_cast_fp16, y = var_17173_to_fp16)[name = string("op_17174_cast_fp16")];
            tensor<int32, [1]> q_v_213_axes_1 = const()[name = string("q_v_213_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_213_cast_fp16 = expand_dims(axes = q_v_213_axes_1, x = var_17174_cast_fp16)[name = string("q_v_213_cast_fp16")];
            tensor<int32, [1]> k_head_427_axes_1 = const()[name = string("k_head_427_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_427_cast_fp16 = expand_dims(axes = k_head_427_axes_1, x = k_head_425_cast_fp16)[name = string("k_head_427_cast_fp16")];
            tensor<int32, [1]> v_head_427_axes_1 = const()[name = string("v_head_427_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_427_cast_fp16 = expand_dims(axes = v_head_427_axes_1, x = v_head_425_cast_fp16)[name = string("v_head_427_cast_fp16")];
            tensor<int32, [1]> p_head_427_axes_1 = const()[name = string("p_head_427_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_427_cast_fp16 = expand_dims(axes = p_head_427_axes_1, x = p_head_425_cast_fp16)[name = string("p_head_427_cast_fp16")];
            bool var_17180_transpose_x_3 = const()[name = string("op_17180_transpose_x_3"), val = bool(false)];
            bool var_17180_transpose_y_3 = const()[name = string("op_17180_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_17180_cast_fp16 = matmul(transpose_x = var_17180_transpose_x_3, transpose_y = var_17180_transpose_y_3, x = q_u_213_cast_fp16, y = k_head_427_cast_fp16)[name = string("op_17180_cast_fp16")];
            fp16 var_17181_to_fp16 = const()[name = string("op_17181_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_213_cast_fp16 = mul(x = var_17180_cast_fp16, y = var_17181_to_fp16)[name = string("scores_content_213_cast_fp16")];
            bool x_1129_transpose_x_3 = const()[name = string("x_1129_transpose_x_3"), val = bool(false)];
            bool x_1129_transpose_y_3 = const()[name = string("x_1129_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1129_cast_fp16 = matmul(transpose_x = x_1129_transpose_x_3, transpose_y = x_1129_transpose_y_3, x = q_v_213_cast_fp16, y = p_head_427_cast_fp16)[name = string("x_1129_cast_fp16")];
            tensor<int32, [8]> x_1131_pad_1 = const()[name = string("x_1131_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1131_mode_1 = const()[name = string("x_1131_mode_1"), val = string("constant")];
            fp16 const_2023_to_fp16 = const()[name = string("const_2023_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1131_cast_fp16 = pad(constant_val = const_2023_to_fp16, mode = x_1131_mode_1, pad = x_1131_pad_1, x = x_1129_cast_fp16)[name = string("x_1131_cast_fp16")];
            tensor<int32, [4]> var_17195 = const()[name = string("op_17195"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1133_cast_fp16 = reshape(shape = var_17195, x = x_1131_cast_fp16)[name = string("x_1133_cast_fp16")];
            tensor<int32, [4]> var_17199_begin_1 = const()[name = string("op_17199_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17199_end_1 = const()[name = string("op_17199_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_17199_end_mask_1 = const()[name = string("op_17199_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_17199_cast_fp16 = slice_by_index(begin = var_17199_begin_1, end = var_17199_end_1, end_mask = var_17199_end_mask_1, x = x_1133_cast_fp16)[name = string("op_17199_cast_fp16")];
            tensor<int32, [4]> var_17201 = const()[name = string("op_17201"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_17202_cast_fp16 = reshape(shape = var_17201, x = var_17199_cast_fp16)[name = string("op_17202_cast_fp16")];
            tensor<int32, [4]> var_17207_begin_1 = const()[name = string("op_17207_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17207_end_1 = const()[name = string("op_17207_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_17207_end_mask_1 = const()[name = string("op_17207_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_17207_cast_fp16 = slice_by_index(begin = var_17207_begin_1, end = var_17207_end_1, end_mask = var_17207_end_mask_1, x = var_17202_cast_fp16)[name = string("op_17207_cast_fp16")];
            fp16 var_17208_to_fp16 = const()[name = string("op_17208_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_213_cast_fp16 = mul(x = var_17207_cast_fp16, y = var_17208_to_fp16)[name = string("scores_pos_213_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_213_cast_fp16 = add(x = scores_content_213_cast_fp16, y = scores_pos_213_cast_fp16)[name = string("logits_213_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_17211_cast_fp16 = softmax(axis = var_16694, x = logits_213_cast_fp16)[name = string("op_17211_cast_fp16")];
            bool var_17213_transpose_x_1 = const()[name = string("op_17213_transpose_x_1"), val = bool(false)];
            bool var_17213_transpose_y_1 = const()[name = string("op_17213_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_17213_cast_fp16 = matmul(transpose_x = var_17213_transpose_x_1, transpose_y = var_17213_transpose_y_1, x = var_17211_cast_fp16, y = v_head_427_cast_fp16)[name = string("op_17213_cast_fp16")];
            tensor<int32, [1]> var_17214_axes_1 = const()[name = string("op_17214_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_17214_cast_fp16 = squeeze(axes = var_17214_axes_1, x = var_17213_cast_fp16)[name = string("op_17214_cast_fp16")];
            string dense_output_1081_pad_type_1 = const()[name = string("dense_output_1081_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1081_strides_1 = const()[name = string("dense_output_1081_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1081_pad_1 = const()[name = string("dense_output_1081_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1081_dilations_1 = const()[name = string("dense_output_1081_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1081_groups_1 = const()[name = string("dense_output_1081_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396942976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397074112))))[name = string("layers_13_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1081_cast_fp16 = conv(dilations = dense_output_1081_dilations_1, groups = dense_output_1081_groups_1, pad = dense_output_1081_pad_1, pad_type = dense_output_1081_pad_type_1, strides = dense_output_1081_strides_1, weight = layers_13_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1081_cast_fp16")];
            string sparse_output_1081_pad_type_1 = const()[name = string("sparse_output_1081_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1081_strides_1 = const()[name = string("sparse_output_1081_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1081_pad_1 = const()[name = string("sparse_output_1081_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1081_dilations_1 = const()[name = string("sparse_output_1081_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1081_groups_1 = const()[name = string("sparse_output_1081_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397077376))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397074688))))[name = string("layers_13_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1081_cast_fp16 = conv(dilations = sparse_output_1081_dilations_1, groups = sparse_output_1081_groups_1, pad = sparse_output_1081_pad_1, pad_type = sparse_output_1081_pad_type_1, strides = sparse_output_1081_strides_1, weight = layers_13_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1081_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17229_cast_fp16 = add(x = dense_output_1081_cast_fp16, y = sparse_output_1081_cast_fp16)[name = string("op_17229_cast_fp16")];
            tensor<int32, [4]> var_17230 = const()[name = string("op_17230"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17232 = const()[name = string("op_17232"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17231_cast_fp16 = transpose(perm = var_17230, x = var_17229_cast_fp16)[name = string("transpose_435")];
            tensor<fp16, [1, 51, 128]> q_head_215_cast_fp16 = reshape(shape = var_17232, x = var_17231_cast_fp16)[name = string("q_head_215_cast_fp16")];
            string dense_output_1083_pad_type_1 = const()[name = string("dense_output_1083_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1083_strides_1 = const()[name = string("dense_output_1083_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1083_pad_1 = const()[name = string("dense_output_1083_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1083_dilations_1 = const()[name = string("dense_output_1083_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1083_groups_1 = const()[name = string("dense_output_1083_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397093824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397224960))))[name = string("layers_13_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1083_cast_fp16 = conv(dilations = dense_output_1083_dilations_1, groups = dense_output_1083_groups_1, pad = dense_output_1083_pad_1, pad_type = dense_output_1083_pad_type_1, strides = dense_output_1083_strides_1, weight = layers_13_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1083_cast_fp16")];
            string sparse_output_1083_pad_type_1 = const()[name = string("sparse_output_1083_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1083_strides_1 = const()[name = string("sparse_output_1083_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1083_pad_1 = const()[name = string("sparse_output_1083_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1083_dilations_1 = const()[name = string("sparse_output_1083_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1083_groups_1 = const()[name = string("sparse_output_1083_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397228224))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397225536))))[name = string("layers_13_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1083_cast_fp16 = conv(dilations = sparse_output_1083_dilations_1, groups = sparse_output_1083_groups_1, pad = sparse_output_1083_pad_1, pad_type = sparse_output_1083_pad_type_1, strides = sparse_output_1083_strides_1, weight = layers_13_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1083_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17248_cast_fp16 = add(x = dense_output_1083_cast_fp16, y = sparse_output_1083_cast_fp16)[name = string("op_17248_cast_fp16")];
            tensor<int32, [4]> var_17249 = const()[name = string("op_17249"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17251 = const()[name = string("op_17251"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17250_cast_fp16 = transpose(perm = var_17249, x = var_17248_cast_fp16)[name = string("transpose_434")];
            tensor<fp16, [1, 51, 128]> k_head_429_cast_fp16 = reshape(shape = var_17251, x = var_17250_cast_fp16)[name = string("k_head_429_cast_fp16")];
            string dense_output_1085_pad_type_1 = const()[name = string("dense_output_1085_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1085_strides_1 = const()[name = string("dense_output_1085_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1085_pad_1 = const()[name = string("dense_output_1085_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1085_dilations_1 = const()[name = string("dense_output_1085_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1085_groups_1 = const()[name = string("dense_output_1085_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397244672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397375808))))[name = string("layers_13_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1085_cast_fp16 = conv(dilations = dense_output_1085_dilations_1, groups = dense_output_1085_groups_1, pad = dense_output_1085_pad_1, pad_type = dense_output_1085_pad_type_1, strides = dense_output_1085_strides_1, weight = layers_13_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1085_cast_fp16")];
            string sparse_output_1085_pad_type_1 = const()[name = string("sparse_output_1085_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1085_strides_1 = const()[name = string("sparse_output_1085_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1085_pad_1 = const()[name = string("sparse_output_1085_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1085_dilations_1 = const()[name = string("sparse_output_1085_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1085_groups_1 = const()[name = string("sparse_output_1085_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397379072))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397376384))))[name = string("layers_13_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1085_cast_fp16 = conv(dilations = sparse_output_1085_dilations_1, groups = sparse_output_1085_groups_1, pad = sparse_output_1085_pad_1, pad_type = sparse_output_1085_pad_type_1, strides = sparse_output_1085_strides_1, weight = layers_13_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1085_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17267_cast_fp16 = add(x = dense_output_1085_cast_fp16, y = sparse_output_1085_cast_fp16)[name = string("op_17267_cast_fp16")];
            tensor<int32, [4]> var_17268 = const()[name = string("op_17268"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17270 = const()[name = string("op_17270"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17269_cast_fp16 = transpose(perm = var_17268, x = var_17267_cast_fp16)[name = string("transpose_433")];
            tensor<fp16, [1, 51, 128]> v_head_429_cast_fp16 = reshape(shape = var_17270, x = var_17269_cast_fp16)[name = string("v_head_429_cast_fp16")];
            string dense_output_1087_pad_type_1 = const()[name = string("dense_output_1087_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1087_strides_1 = const()[name = string("dense_output_1087_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1087_pad_1 = const()[name = string("dense_output_1087_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1087_dilations_1 = const()[name = string("dense_output_1087_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1087_groups_1 = const()[name = string("dense_output_1087_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397395520))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397526656))))[name = string("layers_13_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1087_cast_fp16 = conv(dilations = dense_output_1087_dilations_1, groups = dense_output_1087_groups_1, pad = dense_output_1087_pad_1, pad_type = dense_output_1087_pad_type_1, strides = dense_output_1087_strides_1, weight = layers_13_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1087_cast_fp16")];
            string sparse_output_1087_pad_type_1 = const()[name = string("sparse_output_1087_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1087_strides_1 = const()[name = string("sparse_output_1087_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1087_pad_1 = const()[name = string("sparse_output_1087_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1087_dilations_1 = const()[name = string("sparse_output_1087_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1087_groups_1 = const()[name = string("sparse_output_1087_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397529920))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397527232))))[name = string("layers_13_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1087_cast_fp16 = conv(dilations = sparse_output_1087_dilations_1, groups = sparse_output_1087_groups_1, pad = sparse_output_1087_pad_1, pad_type = sparse_output_1087_pad_type_1, strides = sparse_output_1087_strides_1, weight = layers_13_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1087_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_17286_cast_fp16 = add(x = dense_output_1087_cast_fp16, y = sparse_output_1087_cast_fp16)[name = string("op_17286_cast_fp16")];
            tensor<int32, [4]> var_17287 = const()[name = string("op_17287"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17289 = const()[name = string("op_17289"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_17288_cast_fp16 = transpose(perm = var_17287, x = var_17286_cast_fp16)[name = string("transpose_432")];
            tensor<fp16, [1, 101, 128]> p_head_429_cast_fp16 = reshape(shape = var_17289, x = var_17288_cast_fp16)[name = string("p_head_429_cast_fp16")];
            tensor<fp16, [128]> var_17291_to_fp16 = const()[name = string("op_17291_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397546368)))];
            tensor<fp16, [1, 51, 128]> var_17292_cast_fp16 = add(x = q_head_215_cast_fp16, y = var_17291_to_fp16)[name = string("op_17292_cast_fp16")];
            tensor<int32, [1]> q_u_215_axes_1 = const()[name = string("q_u_215_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_215_cast_fp16 = expand_dims(axes = q_u_215_axes_1, x = var_17292_cast_fp16)[name = string("q_u_215_cast_fp16")];
            tensor<fp16, [128]> var_17294_to_fp16 = const()[name = string("op_17294_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397546688)))];
            tensor<fp16, [1, 51, 128]> var_17295_cast_fp16 = add(x = q_head_215_cast_fp16, y = var_17294_to_fp16)[name = string("op_17295_cast_fp16")];
            tensor<int32, [1]> q_v_215_axes_1 = const()[name = string("q_v_215_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_215_cast_fp16 = expand_dims(axes = q_v_215_axes_1, x = var_17295_cast_fp16)[name = string("q_v_215_cast_fp16")];
            tensor<int32, [1]> k_head_431_axes_1 = const()[name = string("k_head_431_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_431_cast_fp16 = expand_dims(axes = k_head_431_axes_1, x = k_head_429_cast_fp16)[name = string("k_head_431_cast_fp16")];
            tensor<int32, [1]> v_head_431_axes_1 = const()[name = string("v_head_431_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_431_cast_fp16 = expand_dims(axes = v_head_431_axes_1, x = v_head_429_cast_fp16)[name = string("v_head_431_cast_fp16")];
            tensor<int32, [1]> p_head_431_axes_1 = const()[name = string("p_head_431_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_431_cast_fp16 = expand_dims(axes = p_head_431_axes_1, x = p_head_429_cast_fp16)[name = string("p_head_431_cast_fp16")];
            bool var_17301_transpose_x_3 = const()[name = string("op_17301_transpose_x_3"), val = bool(false)];
            bool var_17301_transpose_y_3 = const()[name = string("op_17301_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_17301_cast_fp16 = matmul(transpose_x = var_17301_transpose_x_3, transpose_y = var_17301_transpose_y_3, x = q_u_215_cast_fp16, y = k_head_431_cast_fp16)[name = string("op_17301_cast_fp16")];
            fp16 var_17302_to_fp16 = const()[name = string("op_17302_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_215_cast_fp16 = mul(x = var_17301_cast_fp16, y = var_17302_to_fp16)[name = string("scores_content_215_cast_fp16")];
            bool x_1137_transpose_x_3 = const()[name = string("x_1137_transpose_x_3"), val = bool(false)];
            bool x_1137_transpose_y_3 = const()[name = string("x_1137_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1137_cast_fp16 = matmul(transpose_x = x_1137_transpose_x_3, transpose_y = x_1137_transpose_y_3, x = q_v_215_cast_fp16, y = p_head_431_cast_fp16)[name = string("x_1137_cast_fp16")];
            tensor<int32, [8]> x_1139_pad_1 = const()[name = string("x_1139_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1139_mode_1 = const()[name = string("x_1139_mode_1"), val = string("constant")];
            fp16 const_2029_to_fp16 = const()[name = string("const_2029_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1139_cast_fp16 = pad(constant_val = const_2029_to_fp16, mode = x_1139_mode_1, pad = x_1139_pad_1, x = x_1137_cast_fp16)[name = string("x_1139_cast_fp16")];
            tensor<int32, [4]> var_17316 = const()[name = string("op_17316"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1141_cast_fp16 = reshape(shape = var_17316, x = x_1139_cast_fp16)[name = string("x_1141_cast_fp16")];
            tensor<int32, [4]> var_17320_begin_1 = const()[name = string("op_17320_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17320_end_1 = const()[name = string("op_17320_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_17320_end_mask_1 = const()[name = string("op_17320_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_17320_cast_fp16 = slice_by_index(begin = var_17320_begin_1, end = var_17320_end_1, end_mask = var_17320_end_mask_1, x = x_1141_cast_fp16)[name = string("op_17320_cast_fp16")];
            tensor<int32, [4]> var_17322 = const()[name = string("op_17322"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_17323_cast_fp16 = reshape(shape = var_17322, x = var_17320_cast_fp16)[name = string("op_17323_cast_fp16")];
            tensor<int32, [4]> var_17328_begin_1 = const()[name = string("op_17328_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17328_end_1 = const()[name = string("op_17328_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_17328_end_mask_1 = const()[name = string("op_17328_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_17328_cast_fp16 = slice_by_index(begin = var_17328_begin_1, end = var_17328_end_1, end_mask = var_17328_end_mask_1, x = var_17323_cast_fp16)[name = string("op_17328_cast_fp16")];
            fp16 var_17329_to_fp16 = const()[name = string("op_17329_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_215_cast_fp16 = mul(x = var_17328_cast_fp16, y = var_17329_to_fp16)[name = string("scores_pos_215_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_215_cast_fp16 = add(x = scores_content_215_cast_fp16, y = scores_pos_215_cast_fp16)[name = string("logits_215_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_17332_cast_fp16 = softmax(axis = var_16694, x = logits_215_cast_fp16)[name = string("op_17332_cast_fp16")];
            bool var_17334_transpose_x_1 = const()[name = string("op_17334_transpose_x_1"), val = bool(false)];
            bool var_17334_transpose_y_1 = const()[name = string("op_17334_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_17334_cast_fp16 = matmul(transpose_x = var_17334_transpose_x_1, transpose_y = var_17334_transpose_y_1, x = var_17332_cast_fp16, y = v_head_431_cast_fp16)[name = string("op_17334_cast_fp16")];
            tensor<int32, [1]> var_17335_axes_1 = const()[name = string("op_17335_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_17335_cast_fp16 = squeeze(axes = var_17335_axes_1, x = var_17334_cast_fp16)[name = string("op_17335_cast_fp16")];
            string dense_output_1089_pad_type_1 = const()[name = string("dense_output_1089_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1089_strides_1 = const()[name = string("dense_output_1089_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1089_pad_1 = const()[name = string("dense_output_1089_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1089_dilations_1 = const()[name = string("dense_output_1089_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1089_groups_1 = const()[name = string("dense_output_1089_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397547008))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397678144))))[name = string("layers_13_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1089_cast_fp16 = conv(dilations = dense_output_1089_dilations_1, groups = dense_output_1089_groups_1, pad = dense_output_1089_pad_1, pad_type = dense_output_1089_pad_type_1, strides = dense_output_1089_strides_1, weight = layers_13_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1089_cast_fp16")];
            string sparse_output_1089_pad_type_1 = const()[name = string("sparse_output_1089_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1089_strides_1 = const()[name = string("sparse_output_1089_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1089_pad_1 = const()[name = string("sparse_output_1089_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1089_dilations_1 = const()[name = string("sparse_output_1089_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1089_groups_1 = const()[name = string("sparse_output_1089_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397681408))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397678720))))[name = string("layers_13_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1089_cast_fp16 = conv(dilations = sparse_output_1089_dilations_1, groups = sparse_output_1089_groups_1, pad = sparse_output_1089_pad_1, pad_type = sparse_output_1089_pad_type_1, strides = sparse_output_1089_strides_1, weight = layers_13_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1089_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17350_cast_fp16 = add(x = dense_output_1089_cast_fp16, y = sparse_output_1089_cast_fp16)[name = string("op_17350_cast_fp16")];
            tensor<int32, [4]> var_17351 = const()[name = string("op_17351"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17353 = const()[name = string("op_17353"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17352_cast_fp16 = transpose(perm = var_17351, x = var_17350_cast_fp16)[name = string("transpose_431")];
            tensor<fp16, [1, 51, 128]> q_head_217_cast_fp16 = reshape(shape = var_17353, x = var_17352_cast_fp16)[name = string("q_head_217_cast_fp16")];
            string dense_output_1091_pad_type_1 = const()[name = string("dense_output_1091_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1091_strides_1 = const()[name = string("dense_output_1091_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1091_pad_1 = const()[name = string("dense_output_1091_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1091_dilations_1 = const()[name = string("dense_output_1091_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1091_groups_1 = const()[name = string("dense_output_1091_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397697856))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397828992))))[name = string("layers_13_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1091_cast_fp16 = conv(dilations = dense_output_1091_dilations_1, groups = dense_output_1091_groups_1, pad = dense_output_1091_pad_1, pad_type = dense_output_1091_pad_type_1, strides = dense_output_1091_strides_1, weight = layers_13_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1091_cast_fp16")];
            string sparse_output_1091_pad_type_1 = const()[name = string("sparse_output_1091_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1091_strides_1 = const()[name = string("sparse_output_1091_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1091_pad_1 = const()[name = string("sparse_output_1091_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1091_dilations_1 = const()[name = string("sparse_output_1091_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1091_groups_1 = const()[name = string("sparse_output_1091_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397832256))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397829568))))[name = string("layers_13_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1091_cast_fp16 = conv(dilations = sparse_output_1091_dilations_1, groups = sparse_output_1091_groups_1, pad = sparse_output_1091_pad_1, pad_type = sparse_output_1091_pad_type_1, strides = sparse_output_1091_strides_1, weight = layers_13_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1091_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17369_cast_fp16 = add(x = dense_output_1091_cast_fp16, y = sparse_output_1091_cast_fp16)[name = string("op_17369_cast_fp16")];
            tensor<int32, [4]> var_17370 = const()[name = string("op_17370"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17372 = const()[name = string("op_17372"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17371_cast_fp16 = transpose(perm = var_17370, x = var_17369_cast_fp16)[name = string("transpose_430")];
            tensor<fp16, [1, 51, 128]> k_head_433_cast_fp16 = reshape(shape = var_17372, x = var_17371_cast_fp16)[name = string("k_head_433_cast_fp16")];
            string dense_output_1093_pad_type_1 = const()[name = string("dense_output_1093_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1093_strides_1 = const()[name = string("dense_output_1093_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1093_pad_1 = const()[name = string("dense_output_1093_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1093_dilations_1 = const()[name = string("dense_output_1093_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1093_groups_1 = const()[name = string("dense_output_1093_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397848704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397979840))))[name = string("layers_13_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1093_cast_fp16 = conv(dilations = dense_output_1093_dilations_1, groups = dense_output_1093_groups_1, pad = dense_output_1093_pad_1, pad_type = dense_output_1093_pad_type_1, strides = dense_output_1093_strides_1, weight = layers_13_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1093_cast_fp16")];
            string sparse_output_1093_pad_type_1 = const()[name = string("sparse_output_1093_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1093_strides_1 = const()[name = string("sparse_output_1093_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1093_pad_1 = const()[name = string("sparse_output_1093_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1093_dilations_1 = const()[name = string("sparse_output_1093_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1093_groups_1 = const()[name = string("sparse_output_1093_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397983104))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397980416))))[name = string("layers_13_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1093_cast_fp16 = conv(dilations = sparse_output_1093_dilations_1, groups = sparse_output_1093_groups_1, pad = sparse_output_1093_pad_1, pad_type = sparse_output_1093_pad_type_1, strides = sparse_output_1093_strides_1, weight = layers_13_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1093_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17388_cast_fp16 = add(x = dense_output_1093_cast_fp16, y = sparse_output_1093_cast_fp16)[name = string("op_17388_cast_fp16")];
            tensor<int32, [4]> var_17389 = const()[name = string("op_17389"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17391 = const()[name = string("op_17391"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17390_cast_fp16 = transpose(perm = var_17389, x = var_17388_cast_fp16)[name = string("transpose_429")];
            tensor<fp16, [1, 51, 128]> v_head_433_cast_fp16 = reshape(shape = var_17391, x = var_17390_cast_fp16)[name = string("v_head_433_cast_fp16")];
            string dense_output_1095_pad_type_1 = const()[name = string("dense_output_1095_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1095_strides_1 = const()[name = string("dense_output_1095_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1095_pad_1 = const()[name = string("dense_output_1095_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1095_dilations_1 = const()[name = string("dense_output_1095_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1095_groups_1 = const()[name = string("dense_output_1095_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397999552))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398130688))))[name = string("layers_13_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1095_cast_fp16 = conv(dilations = dense_output_1095_dilations_1, groups = dense_output_1095_groups_1, pad = dense_output_1095_pad_1, pad_type = dense_output_1095_pad_type_1, strides = dense_output_1095_strides_1, weight = layers_13_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1095_cast_fp16")];
            string sparse_output_1095_pad_type_1 = const()[name = string("sparse_output_1095_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1095_strides_1 = const()[name = string("sparse_output_1095_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1095_pad_1 = const()[name = string("sparse_output_1095_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1095_dilations_1 = const()[name = string("sparse_output_1095_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1095_groups_1 = const()[name = string("sparse_output_1095_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398133952))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398131264))))[name = string("layers_13_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1095_cast_fp16 = conv(dilations = sparse_output_1095_dilations_1, groups = sparse_output_1095_groups_1, pad = sparse_output_1095_pad_1, pad_type = sparse_output_1095_pad_type_1, strides = sparse_output_1095_strides_1, weight = layers_13_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1095_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_17407_cast_fp16 = add(x = dense_output_1095_cast_fp16, y = sparse_output_1095_cast_fp16)[name = string("op_17407_cast_fp16")];
            tensor<int32, [4]> var_17408 = const()[name = string("op_17408"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17410 = const()[name = string("op_17410"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_17409_cast_fp16 = transpose(perm = var_17408, x = var_17407_cast_fp16)[name = string("transpose_428")];
            tensor<fp16, [1, 101, 128]> p_head_433_cast_fp16 = reshape(shape = var_17410, x = var_17409_cast_fp16)[name = string("p_head_433_cast_fp16")];
            tensor<fp16, [128]> var_17412_to_fp16 = const()[name = string("op_17412_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398150400)))];
            tensor<fp16, [1, 51, 128]> var_17413_cast_fp16 = add(x = q_head_217_cast_fp16, y = var_17412_to_fp16)[name = string("op_17413_cast_fp16")];
            tensor<int32, [1]> q_u_217_axes_1 = const()[name = string("q_u_217_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_217_cast_fp16 = expand_dims(axes = q_u_217_axes_1, x = var_17413_cast_fp16)[name = string("q_u_217_cast_fp16")];
            tensor<fp16, [128]> var_17415_to_fp16 = const()[name = string("op_17415_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398150720)))];
            tensor<fp16, [1, 51, 128]> var_17416_cast_fp16 = add(x = q_head_217_cast_fp16, y = var_17415_to_fp16)[name = string("op_17416_cast_fp16")];
            tensor<int32, [1]> q_v_217_axes_1 = const()[name = string("q_v_217_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_217_cast_fp16 = expand_dims(axes = q_v_217_axes_1, x = var_17416_cast_fp16)[name = string("q_v_217_cast_fp16")];
            tensor<int32, [1]> k_head_435_axes_1 = const()[name = string("k_head_435_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_435_cast_fp16 = expand_dims(axes = k_head_435_axes_1, x = k_head_433_cast_fp16)[name = string("k_head_435_cast_fp16")];
            tensor<int32, [1]> v_head_435_axes_1 = const()[name = string("v_head_435_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_435_cast_fp16 = expand_dims(axes = v_head_435_axes_1, x = v_head_433_cast_fp16)[name = string("v_head_435_cast_fp16")];
            tensor<int32, [1]> p_head_435_axes_1 = const()[name = string("p_head_435_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_435_cast_fp16 = expand_dims(axes = p_head_435_axes_1, x = p_head_433_cast_fp16)[name = string("p_head_435_cast_fp16")];
            bool var_17422_transpose_x_3 = const()[name = string("op_17422_transpose_x_3"), val = bool(false)];
            bool var_17422_transpose_y_3 = const()[name = string("op_17422_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_17422_cast_fp16 = matmul(transpose_x = var_17422_transpose_x_3, transpose_y = var_17422_transpose_y_3, x = q_u_217_cast_fp16, y = k_head_435_cast_fp16)[name = string("op_17422_cast_fp16")];
            fp16 var_17423_to_fp16 = const()[name = string("op_17423_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_217_cast_fp16 = mul(x = var_17422_cast_fp16, y = var_17423_to_fp16)[name = string("scores_content_217_cast_fp16")];
            bool x_1145_transpose_x_3 = const()[name = string("x_1145_transpose_x_3"), val = bool(false)];
            bool x_1145_transpose_y_3 = const()[name = string("x_1145_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1145_cast_fp16 = matmul(transpose_x = x_1145_transpose_x_3, transpose_y = x_1145_transpose_y_3, x = q_v_217_cast_fp16, y = p_head_435_cast_fp16)[name = string("x_1145_cast_fp16")];
            tensor<int32, [8]> x_1147_pad_1 = const()[name = string("x_1147_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1147_mode_1 = const()[name = string("x_1147_mode_1"), val = string("constant")];
            fp16 const_2035_to_fp16 = const()[name = string("const_2035_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1147_cast_fp16 = pad(constant_val = const_2035_to_fp16, mode = x_1147_mode_1, pad = x_1147_pad_1, x = x_1145_cast_fp16)[name = string("x_1147_cast_fp16")];
            tensor<int32, [4]> var_17437 = const()[name = string("op_17437"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1149_cast_fp16 = reshape(shape = var_17437, x = x_1147_cast_fp16)[name = string("x_1149_cast_fp16")];
            tensor<int32, [4]> var_17441_begin_1 = const()[name = string("op_17441_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17441_end_1 = const()[name = string("op_17441_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_17441_end_mask_1 = const()[name = string("op_17441_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_17441_cast_fp16 = slice_by_index(begin = var_17441_begin_1, end = var_17441_end_1, end_mask = var_17441_end_mask_1, x = x_1149_cast_fp16)[name = string("op_17441_cast_fp16")];
            tensor<int32, [4]> var_17443 = const()[name = string("op_17443"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_17444_cast_fp16 = reshape(shape = var_17443, x = var_17441_cast_fp16)[name = string("op_17444_cast_fp16")];
            tensor<int32, [4]> var_17449_begin_1 = const()[name = string("op_17449_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17449_end_1 = const()[name = string("op_17449_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_17449_end_mask_1 = const()[name = string("op_17449_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_17449_cast_fp16 = slice_by_index(begin = var_17449_begin_1, end = var_17449_end_1, end_mask = var_17449_end_mask_1, x = var_17444_cast_fp16)[name = string("op_17449_cast_fp16")];
            fp16 var_17450_to_fp16 = const()[name = string("op_17450_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_217_cast_fp16 = mul(x = var_17449_cast_fp16, y = var_17450_to_fp16)[name = string("scores_pos_217_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_217_cast_fp16 = add(x = scores_content_217_cast_fp16, y = scores_pos_217_cast_fp16)[name = string("logits_217_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_17453_cast_fp16 = softmax(axis = var_16694, x = logits_217_cast_fp16)[name = string("op_17453_cast_fp16")];
            bool var_17455_transpose_x_1 = const()[name = string("op_17455_transpose_x_1"), val = bool(false)];
            bool var_17455_transpose_y_1 = const()[name = string("op_17455_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_17455_cast_fp16 = matmul(transpose_x = var_17455_transpose_x_1, transpose_y = var_17455_transpose_y_1, x = var_17453_cast_fp16, y = v_head_435_cast_fp16)[name = string("op_17455_cast_fp16")];
            tensor<int32, [1]> var_17456_axes_1 = const()[name = string("op_17456_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_17456_cast_fp16 = squeeze(axes = var_17456_axes_1, x = var_17455_cast_fp16)[name = string("op_17456_cast_fp16")];
            string dense_output_1097_pad_type_1 = const()[name = string("dense_output_1097_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1097_strides_1 = const()[name = string("dense_output_1097_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1097_pad_1 = const()[name = string("dense_output_1097_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1097_dilations_1 = const()[name = string("dense_output_1097_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1097_groups_1 = const()[name = string("dense_output_1097_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398151040))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398282176))))[name = string("layers_13_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1097_cast_fp16 = conv(dilations = dense_output_1097_dilations_1, groups = dense_output_1097_groups_1, pad = dense_output_1097_pad_1, pad_type = dense_output_1097_pad_type_1, strides = dense_output_1097_strides_1, weight = layers_13_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1097_cast_fp16")];
            string sparse_output_1097_pad_type_1 = const()[name = string("sparse_output_1097_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1097_strides_1 = const()[name = string("sparse_output_1097_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1097_pad_1 = const()[name = string("sparse_output_1097_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1097_dilations_1 = const()[name = string("sparse_output_1097_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1097_groups_1 = const()[name = string("sparse_output_1097_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398285440))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398282752))))[name = string("layers_13_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1097_cast_fp16 = conv(dilations = sparse_output_1097_dilations_1, groups = sparse_output_1097_groups_1, pad = sparse_output_1097_pad_1, pad_type = sparse_output_1097_pad_type_1, strides = sparse_output_1097_strides_1, weight = layers_13_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1097_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17471_cast_fp16 = add(x = dense_output_1097_cast_fp16, y = sparse_output_1097_cast_fp16)[name = string("op_17471_cast_fp16")];
            tensor<int32, [4]> var_17472 = const()[name = string("op_17472"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17474 = const()[name = string("op_17474"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17473_cast_fp16 = transpose(perm = var_17472, x = var_17471_cast_fp16)[name = string("transpose_427")];
            tensor<fp16, [1, 51, 128]> q_head_219_cast_fp16 = reshape(shape = var_17474, x = var_17473_cast_fp16)[name = string("q_head_219_cast_fp16")];
            string dense_output_1099_pad_type_1 = const()[name = string("dense_output_1099_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1099_strides_1 = const()[name = string("dense_output_1099_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1099_pad_1 = const()[name = string("dense_output_1099_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1099_dilations_1 = const()[name = string("dense_output_1099_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1099_groups_1 = const()[name = string("dense_output_1099_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398301888))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398433024))))[name = string("layers_13_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1099_cast_fp16 = conv(dilations = dense_output_1099_dilations_1, groups = dense_output_1099_groups_1, pad = dense_output_1099_pad_1, pad_type = dense_output_1099_pad_type_1, strides = dense_output_1099_strides_1, weight = layers_13_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1099_cast_fp16")];
            string sparse_output_1099_pad_type_1 = const()[name = string("sparse_output_1099_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1099_strides_1 = const()[name = string("sparse_output_1099_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1099_pad_1 = const()[name = string("sparse_output_1099_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1099_dilations_1 = const()[name = string("sparse_output_1099_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1099_groups_1 = const()[name = string("sparse_output_1099_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398436288))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398433600))))[name = string("layers_13_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1099_cast_fp16 = conv(dilations = sparse_output_1099_dilations_1, groups = sparse_output_1099_groups_1, pad = sparse_output_1099_pad_1, pad_type = sparse_output_1099_pad_type_1, strides = sparse_output_1099_strides_1, weight = layers_13_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1099_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17490_cast_fp16 = add(x = dense_output_1099_cast_fp16, y = sparse_output_1099_cast_fp16)[name = string("op_17490_cast_fp16")];
            tensor<int32, [4]> var_17491 = const()[name = string("op_17491"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17493 = const()[name = string("op_17493"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17492_cast_fp16 = transpose(perm = var_17491, x = var_17490_cast_fp16)[name = string("transpose_426")];
            tensor<fp16, [1, 51, 128]> k_head_437_cast_fp16 = reshape(shape = var_17493, x = var_17492_cast_fp16)[name = string("k_head_437_cast_fp16")];
            string dense_output_1101_pad_type_1 = const()[name = string("dense_output_1101_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1101_strides_1 = const()[name = string("dense_output_1101_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1101_pad_1 = const()[name = string("dense_output_1101_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1101_dilations_1 = const()[name = string("dense_output_1101_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1101_groups_1 = const()[name = string("dense_output_1101_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398452736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398583872))))[name = string("layers_13_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1101_cast_fp16 = conv(dilations = dense_output_1101_dilations_1, groups = dense_output_1101_groups_1, pad = dense_output_1101_pad_1, pad_type = dense_output_1101_pad_type_1, strides = dense_output_1101_strides_1, weight = layers_13_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1101_cast_fp16")];
            string sparse_output_1101_pad_type_1 = const()[name = string("sparse_output_1101_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1101_strides_1 = const()[name = string("sparse_output_1101_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1101_pad_1 = const()[name = string("sparse_output_1101_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1101_dilations_1 = const()[name = string("sparse_output_1101_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1101_groups_1 = const()[name = string("sparse_output_1101_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398587136))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398584448))))[name = string("layers_13_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1101_cast_fp16 = conv(dilations = sparse_output_1101_dilations_1, groups = sparse_output_1101_groups_1, pad = sparse_output_1101_pad_1, pad_type = sparse_output_1101_pad_type_1, strides = sparse_output_1101_strides_1, weight = layers_13_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1101_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17509_cast_fp16 = add(x = dense_output_1101_cast_fp16, y = sparse_output_1101_cast_fp16)[name = string("op_17509_cast_fp16")];
            tensor<int32, [4]> var_17510 = const()[name = string("op_17510"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17512 = const()[name = string("op_17512"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17511_cast_fp16 = transpose(perm = var_17510, x = var_17509_cast_fp16)[name = string("transpose_425")];
            tensor<fp16, [1, 51, 128]> v_head_437_cast_fp16 = reshape(shape = var_17512, x = var_17511_cast_fp16)[name = string("v_head_437_cast_fp16")];
            string dense_output_1103_pad_type_1 = const()[name = string("dense_output_1103_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1103_strides_1 = const()[name = string("dense_output_1103_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1103_pad_1 = const()[name = string("dense_output_1103_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1103_dilations_1 = const()[name = string("dense_output_1103_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1103_groups_1 = const()[name = string("dense_output_1103_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398603584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398734720))))[name = string("layers_13_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1103_cast_fp16 = conv(dilations = dense_output_1103_dilations_1, groups = dense_output_1103_groups_1, pad = dense_output_1103_pad_1, pad_type = dense_output_1103_pad_type_1, strides = dense_output_1103_strides_1, weight = layers_13_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1103_cast_fp16")];
            string sparse_output_1103_pad_type_1 = const()[name = string("sparse_output_1103_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1103_strides_1 = const()[name = string("sparse_output_1103_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1103_pad_1 = const()[name = string("sparse_output_1103_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1103_dilations_1 = const()[name = string("sparse_output_1103_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1103_groups_1 = const()[name = string("sparse_output_1103_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398737984))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398735296))))[name = string("layers_13_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1103_cast_fp16 = conv(dilations = sparse_output_1103_dilations_1, groups = sparse_output_1103_groups_1, pad = sparse_output_1103_pad_1, pad_type = sparse_output_1103_pad_type_1, strides = sparse_output_1103_strides_1, weight = layers_13_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1103_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_17528_cast_fp16 = add(x = dense_output_1103_cast_fp16, y = sparse_output_1103_cast_fp16)[name = string("op_17528_cast_fp16")];
            tensor<int32, [4]> var_17529 = const()[name = string("op_17529"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17531 = const()[name = string("op_17531"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_17530_cast_fp16 = transpose(perm = var_17529, x = var_17528_cast_fp16)[name = string("transpose_424")];
            tensor<fp16, [1, 101, 128]> p_head_437_cast_fp16 = reshape(shape = var_17531, x = var_17530_cast_fp16)[name = string("p_head_437_cast_fp16")];
            tensor<fp16, [128]> var_17533_to_fp16 = const()[name = string("op_17533_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398754432)))];
            tensor<fp16, [1, 51, 128]> var_17534_cast_fp16 = add(x = q_head_219_cast_fp16, y = var_17533_to_fp16)[name = string("op_17534_cast_fp16")];
            tensor<int32, [1]> q_u_219_axes_1 = const()[name = string("q_u_219_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_219_cast_fp16 = expand_dims(axes = q_u_219_axes_1, x = var_17534_cast_fp16)[name = string("q_u_219_cast_fp16")];
            tensor<fp16, [128]> var_17536_to_fp16 = const()[name = string("op_17536_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398754752)))];
            tensor<fp16, [1, 51, 128]> var_17537_cast_fp16 = add(x = q_head_219_cast_fp16, y = var_17536_to_fp16)[name = string("op_17537_cast_fp16")];
            tensor<int32, [1]> q_v_219_axes_1 = const()[name = string("q_v_219_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_219_cast_fp16 = expand_dims(axes = q_v_219_axes_1, x = var_17537_cast_fp16)[name = string("q_v_219_cast_fp16")];
            tensor<int32, [1]> k_head_439_axes_1 = const()[name = string("k_head_439_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_439_cast_fp16 = expand_dims(axes = k_head_439_axes_1, x = k_head_437_cast_fp16)[name = string("k_head_439_cast_fp16")];
            tensor<int32, [1]> v_head_439_axes_1 = const()[name = string("v_head_439_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_439_cast_fp16 = expand_dims(axes = v_head_439_axes_1, x = v_head_437_cast_fp16)[name = string("v_head_439_cast_fp16")];
            tensor<int32, [1]> p_head_439_axes_1 = const()[name = string("p_head_439_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_439_cast_fp16 = expand_dims(axes = p_head_439_axes_1, x = p_head_437_cast_fp16)[name = string("p_head_439_cast_fp16")];
            bool var_17543_transpose_x_3 = const()[name = string("op_17543_transpose_x_3"), val = bool(false)];
            bool var_17543_transpose_y_3 = const()[name = string("op_17543_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_17543_cast_fp16 = matmul(transpose_x = var_17543_transpose_x_3, transpose_y = var_17543_transpose_y_3, x = q_u_219_cast_fp16, y = k_head_439_cast_fp16)[name = string("op_17543_cast_fp16")];
            fp16 var_17544_to_fp16 = const()[name = string("op_17544_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_219_cast_fp16 = mul(x = var_17543_cast_fp16, y = var_17544_to_fp16)[name = string("scores_content_219_cast_fp16")];
            bool x_1153_transpose_x_3 = const()[name = string("x_1153_transpose_x_3"), val = bool(false)];
            bool x_1153_transpose_y_3 = const()[name = string("x_1153_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1153_cast_fp16 = matmul(transpose_x = x_1153_transpose_x_3, transpose_y = x_1153_transpose_y_3, x = q_v_219_cast_fp16, y = p_head_439_cast_fp16)[name = string("x_1153_cast_fp16")];
            tensor<int32, [8]> x_1155_pad_1 = const()[name = string("x_1155_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1155_mode_1 = const()[name = string("x_1155_mode_1"), val = string("constant")];
            fp16 const_2041_to_fp16 = const()[name = string("const_2041_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1155_cast_fp16 = pad(constant_val = const_2041_to_fp16, mode = x_1155_mode_1, pad = x_1155_pad_1, x = x_1153_cast_fp16)[name = string("x_1155_cast_fp16")];
            tensor<int32, [4]> var_17558 = const()[name = string("op_17558"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1157_cast_fp16 = reshape(shape = var_17558, x = x_1155_cast_fp16)[name = string("x_1157_cast_fp16")];
            tensor<int32, [4]> var_17562_begin_1 = const()[name = string("op_17562_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17562_end_1 = const()[name = string("op_17562_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_17562_end_mask_1 = const()[name = string("op_17562_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_17562_cast_fp16 = slice_by_index(begin = var_17562_begin_1, end = var_17562_end_1, end_mask = var_17562_end_mask_1, x = x_1157_cast_fp16)[name = string("op_17562_cast_fp16")];
            tensor<int32, [4]> var_17564 = const()[name = string("op_17564"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_17565_cast_fp16 = reshape(shape = var_17564, x = var_17562_cast_fp16)[name = string("op_17565_cast_fp16")];
            tensor<int32, [4]> var_17570_begin_1 = const()[name = string("op_17570_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17570_end_1 = const()[name = string("op_17570_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_17570_end_mask_1 = const()[name = string("op_17570_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_17570_cast_fp16 = slice_by_index(begin = var_17570_begin_1, end = var_17570_end_1, end_mask = var_17570_end_mask_1, x = var_17565_cast_fp16)[name = string("op_17570_cast_fp16")];
            fp16 var_17571_to_fp16 = const()[name = string("op_17571_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_219_cast_fp16 = mul(x = var_17570_cast_fp16, y = var_17571_to_fp16)[name = string("scores_pos_219_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_219_cast_fp16 = add(x = scores_content_219_cast_fp16, y = scores_pos_219_cast_fp16)[name = string("logits_219_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_17574_cast_fp16 = softmax(axis = var_16694, x = logits_219_cast_fp16)[name = string("op_17574_cast_fp16")];
            bool var_17576_transpose_x_1 = const()[name = string("op_17576_transpose_x_1"), val = bool(false)];
            bool var_17576_transpose_y_1 = const()[name = string("op_17576_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_17576_cast_fp16 = matmul(transpose_x = var_17576_transpose_x_1, transpose_y = var_17576_transpose_y_1, x = var_17574_cast_fp16, y = v_head_439_cast_fp16)[name = string("op_17576_cast_fp16")];
            tensor<int32, [1]> var_17577_axes_1 = const()[name = string("op_17577_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_17577_cast_fp16 = squeeze(axes = var_17577_axes_1, x = var_17576_cast_fp16)[name = string("op_17577_cast_fp16")];
            string dense_output_1105_pad_type_1 = const()[name = string("dense_output_1105_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1105_strides_1 = const()[name = string("dense_output_1105_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1105_pad_1 = const()[name = string("dense_output_1105_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1105_dilations_1 = const()[name = string("dense_output_1105_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1105_groups_1 = const()[name = string("dense_output_1105_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398755072))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398886208))))[name = string("layers_13_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1105_cast_fp16 = conv(dilations = dense_output_1105_dilations_1, groups = dense_output_1105_groups_1, pad = dense_output_1105_pad_1, pad_type = dense_output_1105_pad_type_1, strides = dense_output_1105_strides_1, weight = layers_13_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1105_cast_fp16")];
            string sparse_output_1105_pad_type_1 = const()[name = string("sparse_output_1105_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1105_strides_1 = const()[name = string("sparse_output_1105_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1105_pad_1 = const()[name = string("sparse_output_1105_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1105_dilations_1 = const()[name = string("sparse_output_1105_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1105_groups_1 = const()[name = string("sparse_output_1105_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398889472))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398886784))))[name = string("layers_13_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1105_cast_fp16 = conv(dilations = sparse_output_1105_dilations_1, groups = sparse_output_1105_groups_1, pad = sparse_output_1105_pad_1, pad_type = sparse_output_1105_pad_type_1, strides = sparse_output_1105_strides_1, weight = layers_13_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1105_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17592_cast_fp16 = add(x = dense_output_1105_cast_fp16, y = sparse_output_1105_cast_fp16)[name = string("op_17592_cast_fp16")];
            tensor<int32, [4]> var_17593 = const()[name = string("op_17593"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17595 = const()[name = string("op_17595"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17594_cast_fp16 = transpose(perm = var_17593, x = var_17592_cast_fp16)[name = string("transpose_423")];
            tensor<fp16, [1, 51, 128]> q_head_221_cast_fp16 = reshape(shape = var_17595, x = var_17594_cast_fp16)[name = string("q_head_221_cast_fp16")];
            string dense_output_1107_pad_type_1 = const()[name = string("dense_output_1107_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1107_strides_1 = const()[name = string("dense_output_1107_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1107_pad_1 = const()[name = string("dense_output_1107_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1107_dilations_1 = const()[name = string("dense_output_1107_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1107_groups_1 = const()[name = string("dense_output_1107_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398905920))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399037056))))[name = string("layers_13_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1107_cast_fp16 = conv(dilations = dense_output_1107_dilations_1, groups = dense_output_1107_groups_1, pad = dense_output_1107_pad_1, pad_type = dense_output_1107_pad_type_1, strides = dense_output_1107_strides_1, weight = layers_13_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1107_cast_fp16")];
            string sparse_output_1107_pad_type_1 = const()[name = string("sparse_output_1107_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1107_strides_1 = const()[name = string("sparse_output_1107_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1107_pad_1 = const()[name = string("sparse_output_1107_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1107_dilations_1 = const()[name = string("sparse_output_1107_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1107_groups_1 = const()[name = string("sparse_output_1107_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399040320))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399037632))))[name = string("layers_13_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1107_cast_fp16 = conv(dilations = sparse_output_1107_dilations_1, groups = sparse_output_1107_groups_1, pad = sparse_output_1107_pad_1, pad_type = sparse_output_1107_pad_type_1, strides = sparse_output_1107_strides_1, weight = layers_13_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1107_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17611_cast_fp16 = add(x = dense_output_1107_cast_fp16, y = sparse_output_1107_cast_fp16)[name = string("op_17611_cast_fp16")];
            tensor<int32, [4]> var_17612 = const()[name = string("op_17612"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17614 = const()[name = string("op_17614"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17613_cast_fp16 = transpose(perm = var_17612, x = var_17611_cast_fp16)[name = string("transpose_422")];
            tensor<fp16, [1, 51, 128]> k_head_441_cast_fp16 = reshape(shape = var_17614, x = var_17613_cast_fp16)[name = string("k_head_441_cast_fp16")];
            string dense_output_1109_pad_type_1 = const()[name = string("dense_output_1109_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1109_strides_1 = const()[name = string("dense_output_1109_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1109_pad_1 = const()[name = string("dense_output_1109_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1109_dilations_1 = const()[name = string("dense_output_1109_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1109_groups_1 = const()[name = string("dense_output_1109_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399056768))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399187904))))[name = string("layers_13_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1109_cast_fp16 = conv(dilations = dense_output_1109_dilations_1, groups = dense_output_1109_groups_1, pad = dense_output_1109_pad_1, pad_type = dense_output_1109_pad_type_1, strides = dense_output_1109_strides_1, weight = layers_13_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1109_cast_fp16")];
            string sparse_output_1109_pad_type_1 = const()[name = string("sparse_output_1109_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1109_strides_1 = const()[name = string("sparse_output_1109_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1109_pad_1 = const()[name = string("sparse_output_1109_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1109_dilations_1 = const()[name = string("sparse_output_1109_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1109_groups_1 = const()[name = string("sparse_output_1109_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399191168))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399188480))))[name = string("layers_13_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1109_cast_fp16 = conv(dilations = sparse_output_1109_dilations_1, groups = sparse_output_1109_groups_1, pad = sparse_output_1109_pad_1, pad_type = sparse_output_1109_pad_type_1, strides = sparse_output_1109_strides_1, weight = layers_13_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1109_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17630_cast_fp16 = add(x = dense_output_1109_cast_fp16, y = sparse_output_1109_cast_fp16)[name = string("op_17630_cast_fp16")];
            tensor<int32, [4]> var_17631 = const()[name = string("op_17631"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17633 = const()[name = string("op_17633"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17632_cast_fp16 = transpose(perm = var_17631, x = var_17630_cast_fp16)[name = string("transpose_421")];
            tensor<fp16, [1, 51, 128]> v_head_441_cast_fp16 = reshape(shape = var_17633, x = var_17632_cast_fp16)[name = string("v_head_441_cast_fp16")];
            string dense_output_1111_pad_type_1 = const()[name = string("dense_output_1111_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1111_strides_1 = const()[name = string("dense_output_1111_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1111_pad_1 = const()[name = string("dense_output_1111_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1111_dilations_1 = const()[name = string("dense_output_1111_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1111_groups_1 = const()[name = string("dense_output_1111_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399207616))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399338752))))[name = string("layers_13_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1111_cast_fp16 = conv(dilations = dense_output_1111_dilations_1, groups = dense_output_1111_groups_1, pad = dense_output_1111_pad_1, pad_type = dense_output_1111_pad_type_1, strides = dense_output_1111_strides_1, weight = layers_13_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1111_cast_fp16")];
            string sparse_output_1111_pad_type_1 = const()[name = string("sparse_output_1111_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1111_strides_1 = const()[name = string("sparse_output_1111_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1111_pad_1 = const()[name = string("sparse_output_1111_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1111_dilations_1 = const()[name = string("sparse_output_1111_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1111_groups_1 = const()[name = string("sparse_output_1111_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399342016))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399339328))))[name = string("layers_13_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1111_cast_fp16 = conv(dilations = sparse_output_1111_dilations_1, groups = sparse_output_1111_groups_1, pad = sparse_output_1111_pad_1, pad_type = sparse_output_1111_pad_type_1, strides = sparse_output_1111_strides_1, weight = layers_13_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1111_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_17649_cast_fp16 = add(x = dense_output_1111_cast_fp16, y = sparse_output_1111_cast_fp16)[name = string("op_17649_cast_fp16")];
            tensor<int32, [4]> var_17650 = const()[name = string("op_17650"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17652 = const()[name = string("op_17652"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_17651_cast_fp16 = transpose(perm = var_17650, x = var_17649_cast_fp16)[name = string("transpose_420")];
            tensor<fp16, [1, 101, 128]> p_head_441_cast_fp16 = reshape(shape = var_17652, x = var_17651_cast_fp16)[name = string("p_head_441_cast_fp16")];
            tensor<fp16, [128]> var_17654_to_fp16 = const()[name = string("op_17654_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399358464)))];
            tensor<fp16, [1, 51, 128]> var_17655_cast_fp16 = add(x = q_head_221_cast_fp16, y = var_17654_to_fp16)[name = string("op_17655_cast_fp16")];
            tensor<int32, [1]> q_u_221_axes_1 = const()[name = string("q_u_221_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_221_cast_fp16 = expand_dims(axes = q_u_221_axes_1, x = var_17655_cast_fp16)[name = string("q_u_221_cast_fp16")];
            tensor<fp16, [128]> var_17657_to_fp16 = const()[name = string("op_17657_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399358784)))];
            tensor<fp16, [1, 51, 128]> var_17658_cast_fp16 = add(x = q_head_221_cast_fp16, y = var_17657_to_fp16)[name = string("op_17658_cast_fp16")];
            tensor<int32, [1]> q_v_221_axes_1 = const()[name = string("q_v_221_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_221_cast_fp16 = expand_dims(axes = q_v_221_axes_1, x = var_17658_cast_fp16)[name = string("q_v_221_cast_fp16")];
            tensor<int32, [1]> k_head_443_axes_1 = const()[name = string("k_head_443_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_443_cast_fp16 = expand_dims(axes = k_head_443_axes_1, x = k_head_441_cast_fp16)[name = string("k_head_443_cast_fp16")];
            tensor<int32, [1]> v_head_443_axes_1 = const()[name = string("v_head_443_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_443_cast_fp16 = expand_dims(axes = v_head_443_axes_1, x = v_head_441_cast_fp16)[name = string("v_head_443_cast_fp16")];
            tensor<int32, [1]> p_head_443_axes_1 = const()[name = string("p_head_443_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_443_cast_fp16 = expand_dims(axes = p_head_443_axes_1, x = p_head_441_cast_fp16)[name = string("p_head_443_cast_fp16")];
            bool var_17664_transpose_x_3 = const()[name = string("op_17664_transpose_x_3"), val = bool(false)];
            bool var_17664_transpose_y_3 = const()[name = string("op_17664_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_17664_cast_fp16 = matmul(transpose_x = var_17664_transpose_x_3, transpose_y = var_17664_transpose_y_3, x = q_u_221_cast_fp16, y = k_head_443_cast_fp16)[name = string("op_17664_cast_fp16")];
            fp16 var_17665_to_fp16 = const()[name = string("op_17665_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_221_cast_fp16 = mul(x = var_17664_cast_fp16, y = var_17665_to_fp16)[name = string("scores_content_221_cast_fp16")];
            bool x_1161_transpose_x_3 = const()[name = string("x_1161_transpose_x_3"), val = bool(false)];
            bool x_1161_transpose_y_3 = const()[name = string("x_1161_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1161_cast_fp16 = matmul(transpose_x = x_1161_transpose_x_3, transpose_y = x_1161_transpose_y_3, x = q_v_221_cast_fp16, y = p_head_443_cast_fp16)[name = string("x_1161_cast_fp16")];
            tensor<int32, [8]> x_1163_pad_1 = const()[name = string("x_1163_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1163_mode_1 = const()[name = string("x_1163_mode_1"), val = string("constant")];
            fp16 const_2047_to_fp16 = const()[name = string("const_2047_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1163_cast_fp16 = pad(constant_val = const_2047_to_fp16, mode = x_1163_mode_1, pad = x_1163_pad_1, x = x_1161_cast_fp16)[name = string("x_1163_cast_fp16")];
            tensor<int32, [4]> var_17679 = const()[name = string("op_17679"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1165_cast_fp16 = reshape(shape = var_17679, x = x_1163_cast_fp16)[name = string("x_1165_cast_fp16")];
            tensor<int32, [4]> var_17683_begin_1 = const()[name = string("op_17683_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17683_end_1 = const()[name = string("op_17683_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_17683_end_mask_1 = const()[name = string("op_17683_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_17683_cast_fp16 = slice_by_index(begin = var_17683_begin_1, end = var_17683_end_1, end_mask = var_17683_end_mask_1, x = x_1165_cast_fp16)[name = string("op_17683_cast_fp16")];
            tensor<int32, [4]> var_17685 = const()[name = string("op_17685"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_17686_cast_fp16 = reshape(shape = var_17685, x = var_17683_cast_fp16)[name = string("op_17686_cast_fp16")];
            tensor<int32, [4]> var_17691_begin_1 = const()[name = string("op_17691_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17691_end_1 = const()[name = string("op_17691_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_17691_end_mask_1 = const()[name = string("op_17691_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_17691_cast_fp16 = slice_by_index(begin = var_17691_begin_1, end = var_17691_end_1, end_mask = var_17691_end_mask_1, x = var_17686_cast_fp16)[name = string("op_17691_cast_fp16")];
            fp16 var_17692_to_fp16 = const()[name = string("op_17692_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_221_cast_fp16 = mul(x = var_17691_cast_fp16, y = var_17692_to_fp16)[name = string("scores_pos_221_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_221_cast_fp16 = add(x = scores_content_221_cast_fp16, y = scores_pos_221_cast_fp16)[name = string("logits_221_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_17695_cast_fp16 = softmax(axis = var_16694, x = logits_221_cast_fp16)[name = string("op_17695_cast_fp16")];
            bool var_17697_transpose_x_1 = const()[name = string("op_17697_transpose_x_1"), val = bool(false)];
            bool var_17697_transpose_y_1 = const()[name = string("op_17697_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_17697_cast_fp16 = matmul(transpose_x = var_17697_transpose_x_1, transpose_y = var_17697_transpose_y_1, x = var_17695_cast_fp16, y = v_head_443_cast_fp16)[name = string("op_17697_cast_fp16")];
            tensor<int32, [1]> var_17698_axes_1 = const()[name = string("op_17698_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_17698_cast_fp16 = squeeze(axes = var_17698_axes_1, x = var_17697_cast_fp16)[name = string("op_17698_cast_fp16")];
            string dense_output_1113_pad_type_1 = const()[name = string("dense_output_1113_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1113_strides_1 = const()[name = string("dense_output_1113_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1113_pad_1 = const()[name = string("dense_output_1113_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1113_dilations_1 = const()[name = string("dense_output_1113_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1113_groups_1 = const()[name = string("dense_output_1113_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399359104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399490240))))[name = string("layers_13_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1113_cast_fp16 = conv(dilations = dense_output_1113_dilations_1, groups = dense_output_1113_groups_1, pad = dense_output_1113_pad_1, pad_type = dense_output_1113_pad_type_1, strides = dense_output_1113_strides_1, weight = layers_13_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1113_cast_fp16")];
            string sparse_output_1113_pad_type_1 = const()[name = string("sparse_output_1113_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1113_strides_1 = const()[name = string("sparse_output_1113_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1113_pad_1 = const()[name = string("sparse_output_1113_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1113_dilations_1 = const()[name = string("sparse_output_1113_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1113_groups_1 = const()[name = string("sparse_output_1113_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399493504))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399490816))))[name = string("layers_13_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1113_cast_fp16 = conv(dilations = sparse_output_1113_dilations_1, groups = sparse_output_1113_groups_1, pad = sparse_output_1113_pad_1, pad_type = sparse_output_1113_pad_type_1, strides = sparse_output_1113_strides_1, weight = layers_13_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1113_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17713_cast_fp16 = add(x = dense_output_1113_cast_fp16, y = sparse_output_1113_cast_fp16)[name = string("op_17713_cast_fp16")];
            tensor<int32, [4]> var_17714 = const()[name = string("op_17714"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17716 = const()[name = string("op_17716"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17715_cast_fp16 = transpose(perm = var_17714, x = var_17713_cast_fp16)[name = string("transpose_419")];
            tensor<fp16, [1, 51, 128]> q_head_223_cast_fp16 = reshape(shape = var_17716, x = var_17715_cast_fp16)[name = string("q_head_223_cast_fp16")];
            string dense_output_1115_pad_type_1 = const()[name = string("dense_output_1115_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1115_strides_1 = const()[name = string("dense_output_1115_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1115_pad_1 = const()[name = string("dense_output_1115_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1115_dilations_1 = const()[name = string("dense_output_1115_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1115_groups_1 = const()[name = string("dense_output_1115_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399509952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399641088))))[name = string("layers_13_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1115_cast_fp16 = conv(dilations = dense_output_1115_dilations_1, groups = dense_output_1115_groups_1, pad = dense_output_1115_pad_1, pad_type = dense_output_1115_pad_type_1, strides = dense_output_1115_strides_1, weight = layers_13_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1115_cast_fp16")];
            string sparse_output_1115_pad_type_1 = const()[name = string("sparse_output_1115_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1115_strides_1 = const()[name = string("sparse_output_1115_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1115_pad_1 = const()[name = string("sparse_output_1115_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1115_dilations_1 = const()[name = string("sparse_output_1115_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1115_groups_1 = const()[name = string("sparse_output_1115_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399644352))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399641664))))[name = string("layers_13_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1115_cast_fp16 = conv(dilations = sparse_output_1115_dilations_1, groups = sparse_output_1115_groups_1, pad = sparse_output_1115_pad_1, pad_type = sparse_output_1115_pad_type_1, strides = sparse_output_1115_strides_1, weight = layers_13_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1115_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17732_cast_fp16 = add(x = dense_output_1115_cast_fp16, y = sparse_output_1115_cast_fp16)[name = string("op_17732_cast_fp16")];
            tensor<int32, [4]> var_17733 = const()[name = string("op_17733"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17735 = const()[name = string("op_17735"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17734_cast_fp16 = transpose(perm = var_17733, x = var_17732_cast_fp16)[name = string("transpose_418")];
            tensor<fp16, [1, 51, 128]> k_head_445_cast_fp16 = reshape(shape = var_17735, x = var_17734_cast_fp16)[name = string("k_head_445_cast_fp16")];
            string dense_output_1117_pad_type_1 = const()[name = string("dense_output_1117_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1117_strides_1 = const()[name = string("dense_output_1117_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1117_pad_1 = const()[name = string("dense_output_1117_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1117_dilations_1 = const()[name = string("dense_output_1117_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1117_groups_1 = const()[name = string("dense_output_1117_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399660800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399791936))))[name = string("layers_13_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1117_cast_fp16 = conv(dilations = dense_output_1117_dilations_1, groups = dense_output_1117_groups_1, pad = dense_output_1117_pad_1, pad_type = dense_output_1117_pad_type_1, strides = dense_output_1117_strides_1, weight = layers_13_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = string("dense_output_1117_cast_fp16")];
            string sparse_output_1117_pad_type_1 = const()[name = string("sparse_output_1117_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1117_strides_1 = const()[name = string("sparse_output_1117_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1117_pad_1 = const()[name = string("sparse_output_1117_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1117_dilations_1 = const()[name = string("sparse_output_1117_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1117_groups_1 = const()[name = string("sparse_output_1117_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399795200))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399792512))))[name = string("layers_13_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1117_cast_fp16 = conv(dilations = sparse_output_1117_dilations_1, groups = sparse_output_1117_groups_1, pad = sparse_output_1117_pad_1, pad_type = sparse_output_1117_pad_type_1, strides = sparse_output_1117_strides_1, weight = layers_13_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_627_cast_fp16)[name = string("sparse_output_1117_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_17751_cast_fp16 = add(x = dense_output_1117_cast_fp16, y = sparse_output_1117_cast_fp16)[name = string("op_17751_cast_fp16")];
            tensor<int32, [4]> var_17752 = const()[name = string("op_17752"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17754 = const()[name = string("op_17754"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_17753_cast_fp16 = transpose(perm = var_17752, x = var_17751_cast_fp16)[name = string("transpose_417")];
            tensor<fp16, [1, 51, 128]> v_head_445_cast_fp16 = reshape(shape = var_17754, x = var_17753_cast_fp16)[name = string("v_head_445_cast_fp16")];
            string dense_output_1119_pad_type_1 = const()[name = string("dense_output_1119_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1119_strides_1 = const()[name = string("dense_output_1119_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1119_pad_1 = const()[name = string("dense_output_1119_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1119_dilations_1 = const()[name = string("dense_output_1119_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1119_groups_1 = const()[name = string("dense_output_1119_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399811648))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399942784))))[name = string("layers_13_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1119_cast_fp16 = conv(dilations = dense_output_1119_dilations_1, groups = dense_output_1119_groups_1, pad = dense_output_1119_pad_1, pad_type = dense_output_1119_pad_type_1, strides = dense_output_1119_strides_1, weight = layers_13_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1119_cast_fp16")];
            string sparse_output_1119_pad_type_1 = const()[name = string("sparse_output_1119_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1119_strides_1 = const()[name = string("sparse_output_1119_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1119_pad_1 = const()[name = string("sparse_output_1119_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1119_dilations_1 = const()[name = string("sparse_output_1119_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1119_groups_1 = const()[name = string("sparse_output_1119_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_13_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399946048))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399943360))))[name = string("layers_13_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1119_cast_fp16 = conv(dilations = sparse_output_1119_dilations_1, groups = sparse_output_1119_groups_1, pad = sparse_output_1119_pad_1, pad_type = sparse_output_1119_pad_type_1, strides = sparse_output_1119_strides_1, weight = layers_13_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1119_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_17770_cast_fp16 = add(x = dense_output_1119_cast_fp16, y = sparse_output_1119_cast_fp16)[name = string("op_17770_cast_fp16")];
            tensor<int32, [4]> var_17771 = const()[name = string("op_17771"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_17773 = const()[name = string("op_17773"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_17772_cast_fp16 = transpose(perm = var_17771, x = var_17770_cast_fp16)[name = string("transpose_416")];
            tensor<fp16, [1, 101, 128]> p_head_445_cast_fp16 = reshape(shape = var_17773, x = var_17772_cast_fp16)[name = string("p_head_445_cast_fp16")];
            tensor<fp16, [128]> var_17775_to_fp16 = const()[name = string("op_17775_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399962496)))];
            tensor<fp16, [1, 51, 128]> var_17776_cast_fp16 = add(x = q_head_223_cast_fp16, y = var_17775_to_fp16)[name = string("op_17776_cast_fp16")];
            tensor<int32, [1]> q_u_223_axes_1 = const()[name = string("q_u_223_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_223_cast_fp16 = expand_dims(axes = q_u_223_axes_1, x = var_17776_cast_fp16)[name = string("q_u_223_cast_fp16")];
            tensor<fp16, [128]> var_17778_to_fp16 = const()[name = string("op_17778_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399962816)))];
            tensor<fp16, [1, 51, 128]> var_17779_cast_fp16 = add(x = q_head_223_cast_fp16, y = var_17778_to_fp16)[name = string("op_17779_cast_fp16")];
            tensor<int32, [1]> q_v_223_axes_1 = const()[name = string("q_v_223_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_223_cast_fp16 = expand_dims(axes = q_v_223_axes_1, x = var_17779_cast_fp16)[name = string("q_v_223_cast_fp16")];
            tensor<int32, [1]> k_head_447_axes_1 = const()[name = string("k_head_447_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_447_cast_fp16 = expand_dims(axes = k_head_447_axes_1, x = k_head_445_cast_fp16)[name = string("k_head_447_cast_fp16")];
            tensor<int32, [1]> v_head_447_axes_1 = const()[name = string("v_head_447_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_447_cast_fp16 = expand_dims(axes = v_head_447_axes_1, x = v_head_445_cast_fp16)[name = string("v_head_447_cast_fp16")];
            tensor<int32, [1]> p_head_447_axes_1 = const()[name = string("p_head_447_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_447_cast_fp16 = expand_dims(axes = p_head_447_axes_1, x = p_head_445_cast_fp16)[name = string("p_head_447_cast_fp16")];
            bool var_17785_transpose_x_3 = const()[name = string("op_17785_transpose_x_3"), val = bool(false)];
            bool var_17785_transpose_y_3 = const()[name = string("op_17785_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_17785_cast_fp16 = matmul(transpose_x = var_17785_transpose_x_3, transpose_y = var_17785_transpose_y_3, x = q_u_223_cast_fp16, y = k_head_447_cast_fp16)[name = string("op_17785_cast_fp16")];
            fp16 var_17786_to_fp16 = const()[name = string("op_17786_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_223_cast_fp16 = mul(x = var_17785_cast_fp16, y = var_17786_to_fp16)[name = string("scores_content_223_cast_fp16")];
            bool x_1169_transpose_x_3 = const()[name = string("x_1169_transpose_x_3"), val = bool(false)];
            bool x_1169_transpose_y_3 = const()[name = string("x_1169_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1169_cast_fp16 = matmul(transpose_x = x_1169_transpose_x_3, transpose_y = x_1169_transpose_y_3, x = q_v_223_cast_fp16, y = p_head_447_cast_fp16)[name = string("x_1169_cast_fp16")];
            tensor<int32, [8]> x_1171_pad_1 = const()[name = string("x_1171_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1171_mode_1 = const()[name = string("x_1171_mode_1"), val = string("constant")];
            fp16 const_2053_to_fp16 = const()[name = string("const_2053_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1171_cast_fp16 = pad(constant_val = const_2053_to_fp16, mode = x_1171_mode_1, pad = x_1171_pad_1, x = x_1169_cast_fp16)[name = string("x_1171_cast_fp16")];
            tensor<int32, [4]> var_17800 = const()[name = string("op_17800"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1173_cast_fp16 = reshape(shape = var_17800, x = x_1171_cast_fp16)[name = string("x_1173_cast_fp16")];
            tensor<int32, [4]> var_17804_begin_1 = const()[name = string("op_17804_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_17804_end_1 = const()[name = string("op_17804_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_17804_end_mask_1 = const()[name = string("op_17804_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_17804_cast_fp16 = slice_by_index(begin = var_17804_begin_1, end = var_17804_end_1, end_mask = var_17804_end_mask_1, x = x_1173_cast_fp16)[name = string("op_17804_cast_fp16")];
            tensor<int32, [4]> var_17806 = const()[name = string("op_17806"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_17807_cast_fp16 = reshape(shape = var_17806, x = var_17804_cast_fp16)[name = string("op_17807_cast_fp16")];
            tensor<int32, [4]> var_17812_begin_1 = const()[name = string("op_17812_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_17812_end_1 = const()[name = string("op_17812_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_17812_end_mask_1 = const()[name = string("op_17812_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_17812_cast_fp16 = slice_by_index(begin = var_17812_begin_1, end = var_17812_end_1, end_mask = var_17812_end_mask_1, x = var_17807_cast_fp16)[name = string("op_17812_cast_fp16")];
            fp16 var_17813_to_fp16 = const()[name = string("op_17813_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_223_cast_fp16 = mul(x = var_17812_cast_fp16, y = var_17813_to_fp16)[name = string("scores_pos_223_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_223_cast_fp16 = add(x = scores_content_223_cast_fp16, y = scores_pos_223_cast_fp16)[name = string("logits_223_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_17816_cast_fp16 = softmax(axis = var_16694, x = logits_223_cast_fp16)[name = string("op_17816_cast_fp16")];
            bool var_17818_transpose_x_1 = const()[name = string("op_17818_transpose_x_1"), val = bool(false)];
            bool var_17818_transpose_y_1 = const()[name = string("op_17818_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_17818_cast_fp16 = matmul(transpose_x = var_17818_transpose_x_1, transpose_y = var_17818_transpose_y_1, x = var_17816_cast_fp16, y = v_head_447_cast_fp16)[name = string("op_17818_cast_fp16")];
            tensor<int32, [1]> o_head_27_axes_1 = const()[name = string("o_head_27_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_27_cast_fp16 = squeeze(axes = o_head_27_axes_1, x = var_17818_cast_fp16)[name = string("o_head_27_cast_fp16")];
            bool out_27_interleave_1 = const()[name = string("out_27_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_27_cast_fp16 = concat(axis = var_16694, interleave = out_27_interleave_1, values = (var_16972_cast_fp16, var_17093_cast_fp16, var_17214_cast_fp16, var_17335_cast_fp16, var_17456_cast_fp16, var_17577_cast_fp16, var_17698_cast_fp16, o_head_27_cast_fp16))[name = string("out_27_cast_fp16")];
            tensor<int32, [3]> var_17822_perm_1 = const()[name = string("op_17822_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_635_axes_1 = const()[name = string("input_635_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_17822_cast_fp16 = transpose(perm = var_17822_perm_1, x = out_27_cast_fp16)[name = string("transpose_415")];
            tensor<fp16, [1, 1024, 51, 1]> input_635_cast_fp16 = expand_dims(axes = input_635_axes_1, x = var_17822_cast_fp16)[name = string("input_635_cast_fp16")];
            string dense_output_1121_pad_type_1 = const()[name = string("dense_output_1121_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1121_strides_1 = const()[name = string("dense_output_1121_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1121_pad_1 = const()[name = string("dense_output_1121_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1121_dilations_1 = const()[name = string("dense_output_1121_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1121_groups_1 = const()[name = string("dense_output_1121_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399963136))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401011776))))[name = string("layers_13_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1121_cast_fp16 = conv(dilations = dense_output_1121_dilations_1, groups = dense_output_1121_groups_1, pad = dense_output_1121_pad_1, pad_type = dense_output_1121_pad_type_1, strides = dense_output_1121_strides_1, weight = layers_13_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_635_cast_fp16)[name = string("dense_output_1121_cast_fp16")];
            string sparse_output_1121_pad_type_1 = const()[name = string("sparse_output_1121_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1121_strides_1 = const()[name = string("sparse_output_1121_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1121_pad_1 = const()[name = string("sparse_output_1121_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1121_dilations_1 = const()[name = string("sparse_output_1121_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1121_groups_1 = const()[name = string("sparse_output_1121_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401033408))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401012352))))[name = string("layers_13_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1121_cast_fp16 = conv(dilations = sparse_output_1121_dilations_1, groups = sparse_output_1121_groups_1, pad = sparse_output_1121_pad_1, pad_type = sparse_output_1121_pad_type_1, strides = sparse_output_1121_strides_1, weight = layers_13_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_635_cast_fp16)[name = string("sparse_output_1121_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_27_cast_fp16 = add(x = dense_output_1121_cast_fp16, y = sparse_output_1121_cast_fp16)[name = string("out_conv_27_cast_fp16")];
            tensor<int32, [1]> var_17839_axes_1 = const()[name = string("op_17839_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_17839_cast_fp16 = squeeze(axes = var_17839_axes_1, x = out_conv_27_cast_fp16)[name = string("op_17839_cast_fp16")];
            tensor<int32, [3]> var_17840_perm_1 = const()[name = string("op_17840_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_17840_cast_fp16 = transpose(perm = var_17840_perm_1, x = var_17839_cast_fp16)[name = string("transpose_414")];
            tensor<fp16, [1, 51, 1024]> input_637_cast_fp16 = add(x = input_625_cast_fp16, y = var_17840_cast_fp16)[name = string("input_637_cast_fp16")];
            tensor<int32, [1]> x_1177_axes_1 = const()[name = string("x_1177_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_13_norm_conv_weight_to_fp16 = const()[name = string("layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401164544)))];
            tensor<fp16, [1024]> layers_13_norm_conv_bias_to_fp16 = const()[name = string("layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401166656)))];
            tensor<fp16, [1, 51, 1024]> x_1177_cast_fp16 = layer_norm(axes = x_1177_axes_1, beta = layers_13_norm_conv_bias_to_fp16, epsilon = var_16709_to_fp16, gamma = layers_13_norm_conv_weight_to_fp16, x = input_637_cast_fp16)[name = string("x_1177_cast_fp16")];
            tensor<int32, [3]> var_17850_perm_1 = const()[name = string("op_17850_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_639_axes_1 = const()[name = string("input_639_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_17850_cast_fp16 = transpose(perm = var_17850_perm_1, x = x_1177_cast_fp16)[name = string("transpose_413")];
            tensor<fp16, [1, 1024, 51, 1]> input_639_cast_fp16 = expand_dims(axes = input_639_axes_1, x = var_17850_cast_fp16)[name = string("input_639_cast_fp16")];
            string dense_output_1123_pad_type_1 = const()[name = string("dense_output_1123_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1123_strides_1 = const()[name = string("dense_output_1123_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1123_pad_1 = const()[name = string("dense_output_1123_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1123_dilations_1 = const()[name = string("dense_output_1123_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1123_groups_1 = const()[name = string("dense_output_1123_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_13_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401168768))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403265984))))[name = string("layers_13_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1123_cast_fp16 = conv(dilations = dense_output_1123_dilations_1, groups = dense_output_1123_groups_1, pad = dense_output_1123_pad_1, pad_type = dense_output_1123_pad_type_1, strides = dense_output_1123_strides_1, weight = layers_13_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_639_cast_fp16)[name = string("dense_output_1123_cast_fp16")];
            string sparse_output_1123_pad_type_1 = const()[name = string("sparse_output_1123_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1123_strides_1 = const()[name = string("sparse_output_1123_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1123_pad_1 = const()[name = string("sparse_output_1123_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1123_dilations_1 = const()[name = string("sparse_output_1123_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1123_groups_1 = const()[name = string("sparse_output_1123_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_13_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403308608))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403266560))))[name = string("layers_13_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1123_cast_fp16 = conv(dilations = sparse_output_1123_dilations_1, groups = sparse_output_1123_groups_1, pad = sparse_output_1123_pad_1, pad_type = sparse_output_1123_pad_type_1, strides = sparse_output_1123_strides_1, weight = layers_13_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_639_cast_fp16)[name = string("sparse_output_1123_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_641_cast_fp16 = add(x = dense_output_1123_cast_fp16, y = sparse_output_1123_cast_fp16)[name = string("input_641_cast_fp16")];
            int32 input_643_split_num_splits_1 = const()[name = string("input_643_split_num_splits_1"), val = int32(2)];
            int32 input_643_split_axis_1 = const()[name = string("input_643_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_643_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_643_split_cast_fp16_1 = split(axis = input_643_split_axis_1, num_splits = input_643_split_num_splits_1, x = input_641_cast_fp16)[name = string("input_643_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_643_split_1_sigmoid_cast_fp16 = sigmoid(x = input_643_split_cast_fp16_1)[name = string("input_643_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_643_cast_fp16 = mul(x = input_643_split_cast_fp16_0, y = input_643_split_1_sigmoid_cast_fp16)[name = string("input_643_cast_fp16")];
            tensor<int32, [8]> input_645_pad_1 = const()[name = string("input_645_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_645_mode_1 = const()[name = string("input_645_mode_1"), val = string("constant")];
            fp16 const_2055_to_fp16 = const()[name = string("const_2055_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_645_cast_fp16 = pad(constant_val = const_2055_to_fp16, mode = input_645_mode_1, pad = input_645_pad_1, x = input_643_cast_fp16)[name = string("input_645_cast_fp16")];
            string dense_output_1125_pad_type_1 = const()[name = string("dense_output_1125_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1125_strides_1 = const()[name = string("dense_output_1125_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1125_pad_1 = const()[name = string("dense_output_1125_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1125_dilations_1 = const()[name = string("dense_output_1125_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1125_groups_1 = const()[name = string("dense_output_1125_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1125_cast_fp16 = conv(dilations = dense_output_1125_dilations_1, groups = dense_output_1125_groups_1, pad = dense_output_1125_pad_1, pad_type = dense_output_1125_pad_type_1, strides = dense_output_1125_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_645_cast_fp16)[name = string("dense_output_1125_cast_fp16")];
            string sparse_output_1125_pad_type_1 = const()[name = string("sparse_output_1125_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1125_strides_1 = const()[name = string("sparse_output_1125_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1125_pad_1 = const()[name = string("sparse_output_1125_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1125_dilations_1 = const()[name = string("sparse_output_1125_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1125_groups_1 = const()[name = string("sparse_output_1125_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_13_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403589312))), nonzero_data = tensor<fp16, [9215]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403570816))))[name = string("layers_13_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1125_cast_fp16 = conv(dilations = sparse_output_1125_dilations_1, groups = sparse_output_1125_groups_1, pad = sparse_output_1125_pad_1, pad_type = sparse_output_1125_pad_type_1, strides = sparse_output_1125_strides_1, weight = layers_13_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_645_cast_fp16)[name = string("sparse_output_1125_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_647_cast_fp16 = add(x = dense_output_1125_cast_fp16, y = sparse_output_1125_cast_fp16)[name = string("input_647_cast_fp16")];
            tensor<fp16, [1024]> layers_13_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_13_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404769024)))];
            tensor<fp16, [1024]> layers_13_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_13_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404771136)))];
            tensor<fp16, [1024]> layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404773248)))];
            tensor<fp16, [1024]> layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404775360)))];
            tensor<fp16, [1, 1024, 51, 1]> input_649_cast_fp16 = batch_norm(beta = layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_16709_to_fp16, gamma = layers_13_conv_batch_norm_weight_to_fp16, mean = layers_13_conv_batch_norm_running_mean_to_fp16, variance = layers_13_conv_batch_norm_running_var_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_651_cast_fp16 = silu(x = input_649_cast_fp16)[name = string("input_651_cast_fp16")];
            string dense_output_1127_pad_type_1 = const()[name = string("dense_output_1127_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1127_strides_1 = const()[name = string("dense_output_1127_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1127_pad_1 = const()[name = string("dense_output_1127_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1127_dilations_1 = const()[name = string("dense_output_1127_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1127_groups_1 = const()[name = string("dense_output_1127_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404777472))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405826112))))[name = string("layers_13_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1127_cast_fp16 = conv(dilations = dense_output_1127_dilations_1, groups = dense_output_1127_groups_1, pad = dense_output_1127_pad_1, pad_type = dense_output_1127_pad_type_1, strides = dense_output_1127_strides_1, weight = layers_13_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("dense_output_1127_cast_fp16")];
            string sparse_output_1127_pad_type_1 = const()[name = string("sparse_output_1127_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1127_strides_1 = const()[name = string("sparse_output_1127_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1127_pad_1 = const()[name = string("sparse_output_1127_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1127_dilations_1 = const()[name = string("sparse_output_1127_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1127_groups_1 = const()[name = string("sparse_output_1127_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_13_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405847744))), nonzero_data = tensor<fp16, [10487]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405826688))))[name = string("layers_13_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1127_cast_fp16 = conv(dilations = sparse_output_1127_dilations_1, groups = sparse_output_1127_groups_1, pad = sparse_output_1127_pad_1, pad_type = sparse_output_1127_pad_type_1, strides = sparse_output_1127_strides_1, weight = layers_13_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_651_cast_fp16)[name = string("sparse_output_1127_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1179_cast_fp16 = add(x = dense_output_1127_cast_fp16, y = sparse_output_1127_cast_fp16)[name = string("x_1179_cast_fp16")];
            tensor<int32, [1]> var_17906_axes_1 = const()[name = string("op_17906_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_17906_cast_fp16 = squeeze(axes = var_17906_axes_1, x = x_1179_cast_fp16)[name = string("op_17906_cast_fp16")];
            tensor<int32, [3]> var_17907_perm_1 = const()[name = string("op_17907_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_17907_cast_fp16 = transpose(perm = var_17907_perm_1, x = var_17906_cast_fp16)[name = string("transpose_412")];
            tensor<fp16, [1, 51, 1024]> input_653_cast_fp16 = add(x = input_637_cast_fp16, y = var_17907_cast_fp16)[name = string("input_653_cast_fp16")];
            tensor<int32, [1]> x_1181_axes_1 = const()[name = string("x_1181_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405978880)))];
            tensor<fp16, [1024]> layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405980992)))];
            tensor<fp16, [1, 51, 1024]> x_1181_cast_fp16 = layer_norm(axes = x_1181_axes_1, beta = layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_16709_to_fp16, gamma = layers_13_norm_feed_forward2_weight_to_fp16, x = input_653_cast_fp16)[name = string("x_1181_cast_fp16")];
            tensor<int32, [4]> var_17917 = const()[name = string("op_17917"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1183_cast_fp16 = reshape(shape = var_17917, x = x_1181_cast_fp16)[name = string("x_1183_cast_fp16")];
            tensor<int32, [4]> input_655_perm_1 = const()[name = string("input_655_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1129_pad_type_1 = const()[name = string("dense_output_1129_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1129_strides_1 = const()[name = string("dense_output_1129_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1129_pad_1 = const()[name = string("dense_output_1129_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1129_dilations_1 = const()[name = string("dense_output_1129_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1129_groups_1 = const()[name = string("dense_output_1129_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_13_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405983104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410177472))))[name = string("layers_13_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_655_cast_fp16 = transpose(perm = input_655_perm_1, x = x_1183_cast_fp16)[name = string("transpose_411")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1129_cast_fp16 = conv(dilations = dense_output_1129_dilations_1, groups = dense_output_1129_groups_1, pad = dense_output_1129_pad_1, pad_type = dense_output_1129_pad_type_1, strides = dense_output_1129_strides_1, weight = layers_13_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("dense_output_1129_cast_fp16")];
            string sparse_output_1129_pad_type_1 = const()[name = string("sparse_output_1129_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1129_strides_1 = const()[name = string("sparse_output_1129_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1129_pad_1 = const()[name = string("sparse_output_1129_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1129_dilations_1 = const()[name = string("sparse_output_1129_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1129_groups_1 = const()[name = string("sparse_output_1129_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_13_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410262016))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410178048))))[name = string("layers_13_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1129_cast_fp16 = conv(dilations = sparse_output_1129_dilations_1, groups = sparse_output_1129_groups_1, pad = sparse_output_1129_pad_1, pad_type = sparse_output_1129_pad_type_1, strides = sparse_output_1129_strides_1, weight = layers_13_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_655_cast_fp16)[name = string("sparse_output_1129_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_657_cast_fp16 = add(x = dense_output_1129_cast_fp16, y = sparse_output_1129_cast_fp16)[name = string("input_657_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_659_cast_fp16 = silu(x = input_657_cast_fp16)[name = string("input_659_cast_fp16")];
            string dense_output_1131_pad_type_1 = const()[name = string("dense_output_1131_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1131_strides_1 = const()[name = string("dense_output_1131_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1131_pad_1 = const()[name = string("dense_output_1131_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1131_dilations_1 = const()[name = string("dense_output_1131_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1131_groups_1 = const()[name = string("dense_output_1131_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_13_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410786368))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(414980736))))[name = string("layers_13_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1131_cast_fp16 = conv(dilations = dense_output_1131_dilations_1, groups = dense_output_1131_groups_1, pad = dense_output_1131_pad_1, pad_type = dense_output_1131_pad_type_1, strides = dense_output_1131_strides_1, weight = layers_13_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_659_cast_fp16)[name = string("dense_output_1131_cast_fp16")];
            string sparse_output_1131_pad_type_1 = const()[name = string("sparse_output_1131_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1131_strides_1 = const()[name = string("sparse_output_1131_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1131_pad_1 = const()[name = string("sparse_output_1131_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1131_dilations_1 = const()[name = string("sparse_output_1131_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1131_groups_1 = const()[name = string("sparse_output_1131_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_13_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415065280))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(414981312))))[name = string("layers_13_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1131_cast_fp16 = conv(dilations = sparse_output_1131_dilations_1, groups = sparse_output_1131_groups_1, pad = sparse_output_1131_pad_1, pad_type = sparse_output_1131_pad_type_1, strides = sparse_output_1131_strides_1, weight = layers_13_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_659_cast_fp16)[name = string("sparse_output_1131_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1185_cast_fp16 = add(x = dense_output_1131_cast_fp16, y = sparse_output_1131_cast_fp16)[name = string("x_1185_cast_fp16")];
            tensor<int32, [4]> x_1187_perm_1 = const()[name = string("x_1187_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_17952 = const()[name = string("op_17952"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1187_cast_fp16 = transpose(perm = x_1187_perm_1, x = x_1185_cast_fp16)[name = string("transpose_410")];
            tensor<fp16, [1, 51, 1024]> var_17953_cast_fp16 = reshape(shape = var_17952, x = x_1187_cast_fp16)[name = string("op_17953_cast_fp16")];
            fp16 var_17954_to_fp16 = const()[name = string("op_17954_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_17955_cast_fp16 = mul(x = var_17953_cast_fp16, y = var_17954_to_fp16)[name = string("op_17955_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_661_cast_fp16 = add(x = input_653_cast_fp16, y = var_17955_cast_fp16)[name = string("input_661_cast_fp16")];
            tensor<int32, [1]> input_663_axes_1 = const()[name = string("input_663_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_13_norm_out_weight_to_fp16 = const()[name = string("layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415589632)))];
            tensor<fp16, [1024]> layers_13_norm_out_bias_to_fp16 = const()[name = string("layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415591744)))];
            tensor<fp16, [1, 51, 1024]> input_663_cast_fp16 = layer_norm(axes = input_663_axes_1, beta = layers_13_norm_out_bias_to_fp16, epsilon = var_16709_to_fp16, gamma = layers_13_norm_out_weight_to_fp16, x = input_661_cast_fp16)[name = string("input_663_cast_fp16")];
            int32 var_17963 = const()[name = string("op_17963"), val = int32(-1)];
            tensor<int32, [1]> x_1189_axes_1 = const()[name = string("x_1189_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415593856)))];
            tensor<fp16, [1024]> layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415595968)))];
            fp16 var_17978_to_fp16 = const()[name = string("op_17978_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1189_cast_fp16 = layer_norm(axes = x_1189_axes_1, beta = layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_17978_to_fp16, gamma = layers_14_norm_feed_forward1_weight_to_fp16, x = input_663_cast_fp16)[name = string("x_1189_cast_fp16")];
            tensor<int32, [4]> var_17997 = const()[name = string("op_17997"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1191_cast_fp16 = reshape(shape = var_17997, x = x_1189_cast_fp16)[name = string("x_1191_cast_fp16")];
            tensor<int32, [4]> input_665_perm_1 = const()[name = string("input_665_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1133_pad_type_1 = const()[name = string("dense_output_1133_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1133_strides_1 = const()[name = string("dense_output_1133_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1133_pad_1 = const()[name = string("dense_output_1133_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1133_dilations_1 = const()[name = string("dense_output_1133_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1133_groups_1 = const()[name = string("dense_output_1133_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_14_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415598080))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419792448))))[name = string("layers_14_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_665_cast_fp16 = transpose(perm = input_665_perm_1, x = x_1191_cast_fp16)[name = string("transpose_409")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1133_cast_fp16 = conv(dilations = dense_output_1133_dilations_1, groups = dense_output_1133_groups_1, pad = dense_output_1133_pad_1, pad_type = dense_output_1133_pad_type_1, strides = dense_output_1133_strides_1, weight = layers_14_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_665_cast_fp16)[name = string("dense_output_1133_cast_fp16")];
            string sparse_output_1133_pad_type_1 = const()[name = string("sparse_output_1133_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1133_strides_1 = const()[name = string("sparse_output_1133_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1133_pad_1 = const()[name = string("sparse_output_1133_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1133_dilations_1 = const()[name = string("sparse_output_1133_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1133_groups_1 = const()[name = string("sparse_output_1133_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_14_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419876992))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419793024))))[name = string("layers_14_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1133_cast_fp16 = conv(dilations = sparse_output_1133_dilations_1, groups = sparse_output_1133_groups_1, pad = sparse_output_1133_pad_1, pad_type = sparse_output_1133_pad_type_1, strides = sparse_output_1133_strides_1, weight = layers_14_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_665_cast_fp16)[name = string("sparse_output_1133_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_667_cast_fp16 = add(x = dense_output_1133_cast_fp16, y = sparse_output_1133_cast_fp16)[name = string("input_667_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_669_cast_fp16 = silu(x = input_667_cast_fp16)[name = string("input_669_cast_fp16")];
            string dense_output_1135_pad_type_1 = const()[name = string("dense_output_1135_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1135_strides_1 = const()[name = string("dense_output_1135_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1135_pad_1 = const()[name = string("dense_output_1135_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1135_dilations_1 = const()[name = string("dense_output_1135_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1135_groups_1 = const()[name = string("dense_output_1135_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_14_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420401344))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424595712))))[name = string("layers_14_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1135_cast_fp16 = conv(dilations = dense_output_1135_dilations_1, groups = dense_output_1135_groups_1, pad = dense_output_1135_pad_1, pad_type = dense_output_1135_pad_type_1, strides = dense_output_1135_strides_1, weight = layers_14_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_669_cast_fp16)[name = string("dense_output_1135_cast_fp16")];
            string sparse_output_1135_pad_type_1 = const()[name = string("sparse_output_1135_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1135_strides_1 = const()[name = string("sparse_output_1135_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1135_pad_1 = const()[name = string("sparse_output_1135_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1135_dilations_1 = const()[name = string("sparse_output_1135_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1135_groups_1 = const()[name = string("sparse_output_1135_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_14_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424680256))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424596288))))[name = string("layers_14_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1135_cast_fp16 = conv(dilations = sparse_output_1135_dilations_1, groups = sparse_output_1135_groups_1, pad = sparse_output_1135_pad_1, pad_type = sparse_output_1135_pad_type_1, strides = sparse_output_1135_strides_1, weight = layers_14_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_669_cast_fp16)[name = string("sparse_output_1135_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1193_cast_fp16 = add(x = dense_output_1135_cast_fp16, y = sparse_output_1135_cast_fp16)[name = string("x_1193_cast_fp16")];
            tensor<int32, [4]> x_1195_perm_1 = const()[name = string("x_1195_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_18032 = const()[name = string("op_18032"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1195_cast_fp16 = transpose(perm = x_1195_perm_1, x = x_1193_cast_fp16)[name = string("transpose_408")];
            tensor<fp16, [1, 51, 1024]> var_18033_cast_fp16 = reshape(shape = var_18032, x = x_1195_cast_fp16)[name = string("op_18033_cast_fp16")];
            fp16 var_18034_to_fp16 = const()[name = string("op_18034_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_18035_cast_fp16 = mul(x = var_18033_cast_fp16, y = var_18034_to_fp16)[name = string("op_18035_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_671_cast_fp16 = add(x = input_663_cast_fp16, y = var_18035_cast_fp16)[name = string("input_671_cast_fp16")];
            tensor<int32, [1]> q_29_axes_1 = const()[name = string("q_29_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_14_norm_self_att_weight_to_fp16 = const()[name = string("layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425204608)))];
            tensor<fp16, [1024]> layers_14_norm_self_att_bias_to_fp16 = const()[name = string("layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425206720)))];
            tensor<fp16, [1, 51, 1024]> q_29_cast_fp16 = layer_norm(axes = q_29_axes_1, beta = layers_14_norm_self_att_bias_to_fp16, epsilon = var_17978_to_fp16, gamma = layers_14_norm_self_att_weight_to_fp16, x = input_671_cast_fp16)[name = string("q_29_cast_fp16")];
            tensor<int32, [3]> var_18109 = const()[name = string("op_18109"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_673_axes_1 = const()[name = string("input_673_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_18110_cast_fp16 = transpose(perm = var_18109, x = q_29_cast_fp16)[name = string("transpose_407")];
            tensor<fp16, [1, 1024, 51, 1]> input_673_cast_fp16 = expand_dims(axes = input_673_axes_1, x = var_18110_cast_fp16)[name = string("input_673_cast_fp16")];
            string dense_output_1137_pad_type_1 = const()[name = string("dense_output_1137_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1137_strides_1 = const()[name = string("dense_output_1137_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1137_pad_1 = const()[name = string("dense_output_1137_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1137_dilations_1 = const()[name = string("dense_output_1137_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1137_groups_1 = const()[name = string("dense_output_1137_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425208832))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425339968))))[name = string("layers_14_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1137_cast_fp16 = conv(dilations = dense_output_1137_dilations_1, groups = dense_output_1137_groups_1, pad = dense_output_1137_pad_1, pad_type = dense_output_1137_pad_type_1, strides = dense_output_1137_strides_1, weight = layers_14_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1137_cast_fp16")];
            string sparse_output_1137_pad_type_1 = const()[name = string("sparse_output_1137_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1137_strides_1 = const()[name = string("sparse_output_1137_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1137_pad_1 = const()[name = string("sparse_output_1137_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1137_dilations_1 = const()[name = string("sparse_output_1137_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1137_groups_1 = const()[name = string("sparse_output_1137_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425343232))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425340544))))[name = string("layers_14_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1137_cast_fp16 = conv(dilations = sparse_output_1137_dilations_1, groups = sparse_output_1137_groups_1, pad = sparse_output_1137_pad_1, pad_type = sparse_output_1137_pad_type_1, strides = sparse_output_1137_strides_1, weight = layers_14_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1137_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18135_cast_fp16 = add(x = dense_output_1137_cast_fp16, y = sparse_output_1137_cast_fp16)[name = string("op_18135_cast_fp16")];
            tensor<int32, [4]> var_18136 = const()[name = string("op_18136"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18138 = const()[name = string("op_18138"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18137_cast_fp16 = transpose(perm = var_18136, x = var_18135_cast_fp16)[name = string("transpose_406")];
            tensor<fp16, [1, 51, 128]> q_head_225_cast_fp16 = reshape(shape = var_18138, x = var_18137_cast_fp16)[name = string("q_head_225_cast_fp16")];
            string dense_output_1139_pad_type_1 = const()[name = string("dense_output_1139_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1139_strides_1 = const()[name = string("dense_output_1139_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1139_pad_1 = const()[name = string("dense_output_1139_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1139_dilations_1 = const()[name = string("dense_output_1139_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1139_groups_1 = const()[name = string("dense_output_1139_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425359680))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425490816))))[name = string("layers_14_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1139_cast_fp16 = conv(dilations = dense_output_1139_dilations_1, groups = dense_output_1139_groups_1, pad = dense_output_1139_pad_1, pad_type = dense_output_1139_pad_type_1, strides = dense_output_1139_strides_1, weight = layers_14_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1139_cast_fp16")];
            string sparse_output_1139_pad_type_1 = const()[name = string("sparse_output_1139_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1139_strides_1 = const()[name = string("sparse_output_1139_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1139_pad_1 = const()[name = string("sparse_output_1139_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1139_dilations_1 = const()[name = string("sparse_output_1139_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1139_groups_1 = const()[name = string("sparse_output_1139_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425494080))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425491392))))[name = string("layers_14_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1139_cast_fp16 = conv(dilations = sparse_output_1139_dilations_1, groups = sparse_output_1139_groups_1, pad = sparse_output_1139_pad_1, pad_type = sparse_output_1139_pad_type_1, strides = sparse_output_1139_strides_1, weight = layers_14_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1139_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18154_cast_fp16 = add(x = dense_output_1139_cast_fp16, y = sparse_output_1139_cast_fp16)[name = string("op_18154_cast_fp16")];
            tensor<int32, [4]> var_18155 = const()[name = string("op_18155"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18157 = const()[name = string("op_18157"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18156_cast_fp16 = transpose(perm = var_18155, x = var_18154_cast_fp16)[name = string("transpose_405")];
            tensor<fp16, [1, 51, 128]> k_head_449_cast_fp16 = reshape(shape = var_18157, x = var_18156_cast_fp16)[name = string("k_head_449_cast_fp16")];
            string dense_output_1141_pad_type_1 = const()[name = string("dense_output_1141_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1141_strides_1 = const()[name = string("dense_output_1141_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1141_pad_1 = const()[name = string("dense_output_1141_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1141_dilations_1 = const()[name = string("dense_output_1141_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1141_groups_1 = const()[name = string("dense_output_1141_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425510528))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425641664))))[name = string("layers_14_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1141_cast_fp16 = conv(dilations = dense_output_1141_dilations_1, groups = dense_output_1141_groups_1, pad = dense_output_1141_pad_1, pad_type = dense_output_1141_pad_type_1, strides = dense_output_1141_strides_1, weight = layers_14_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1141_cast_fp16")];
            string sparse_output_1141_pad_type_1 = const()[name = string("sparse_output_1141_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1141_strides_1 = const()[name = string("sparse_output_1141_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1141_pad_1 = const()[name = string("sparse_output_1141_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1141_dilations_1 = const()[name = string("sparse_output_1141_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1141_groups_1 = const()[name = string("sparse_output_1141_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425644928))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425642240))))[name = string("layers_14_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1141_cast_fp16 = conv(dilations = sparse_output_1141_dilations_1, groups = sparse_output_1141_groups_1, pad = sparse_output_1141_pad_1, pad_type = sparse_output_1141_pad_type_1, strides = sparse_output_1141_strides_1, weight = layers_14_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1141_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18173_cast_fp16 = add(x = dense_output_1141_cast_fp16, y = sparse_output_1141_cast_fp16)[name = string("op_18173_cast_fp16")];
            tensor<int32, [4]> var_18174 = const()[name = string("op_18174"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18176 = const()[name = string("op_18176"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18175_cast_fp16 = transpose(perm = var_18174, x = var_18173_cast_fp16)[name = string("transpose_404")];
            tensor<fp16, [1, 51, 128]> v_head_449_cast_fp16 = reshape(shape = var_18176, x = var_18175_cast_fp16)[name = string("v_head_449_cast_fp16")];
            string dense_output_1143_pad_type_1 = const()[name = string("dense_output_1143_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1143_strides_1 = const()[name = string("dense_output_1143_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1143_pad_1 = const()[name = string("dense_output_1143_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1143_dilations_1 = const()[name = string("dense_output_1143_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1143_groups_1 = const()[name = string("dense_output_1143_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425661376))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425792512))))[name = string("layers_14_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1143_cast_fp16 = conv(dilations = dense_output_1143_dilations_1, groups = dense_output_1143_groups_1, pad = dense_output_1143_pad_1, pad_type = dense_output_1143_pad_type_1, strides = dense_output_1143_strides_1, weight = layers_14_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1143_cast_fp16")];
            string sparse_output_1143_pad_type_1 = const()[name = string("sparse_output_1143_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1143_strides_1 = const()[name = string("sparse_output_1143_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1143_pad_1 = const()[name = string("sparse_output_1143_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1143_dilations_1 = const()[name = string("sparse_output_1143_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1143_groups_1 = const()[name = string("sparse_output_1143_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425795776))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425793088))))[name = string("layers_14_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1143_cast_fp16 = conv(dilations = sparse_output_1143_dilations_1, groups = sparse_output_1143_groups_1, pad = sparse_output_1143_pad_1, pad_type = sparse_output_1143_pad_type_1, strides = sparse_output_1143_strides_1, weight = layers_14_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1143_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_18192_cast_fp16 = add(x = dense_output_1143_cast_fp16, y = sparse_output_1143_cast_fp16)[name = string("op_18192_cast_fp16")];
            tensor<int32, [4]> var_18193 = const()[name = string("op_18193"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18195 = const()[name = string("op_18195"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_18194_cast_fp16 = transpose(perm = var_18193, x = var_18192_cast_fp16)[name = string("transpose_403")];
            tensor<fp16, [1, 101, 128]> p_head_449_cast_fp16 = reshape(shape = var_18195, x = var_18194_cast_fp16)[name = string("p_head_449_cast_fp16")];
            tensor<fp16, [128]> var_18197_to_fp16 = const()[name = string("op_18197_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425812224)))];
            tensor<fp16, [1, 51, 128]> var_18198_cast_fp16 = add(x = q_head_225_cast_fp16, y = var_18197_to_fp16)[name = string("op_18198_cast_fp16")];
            tensor<int32, [1]> q_u_225_axes_1 = const()[name = string("q_u_225_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_225_cast_fp16 = expand_dims(axes = q_u_225_axes_1, x = var_18198_cast_fp16)[name = string("q_u_225_cast_fp16")];
            tensor<fp16, [128]> var_18200_to_fp16 = const()[name = string("op_18200_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425812544)))];
            tensor<fp16, [1, 51, 128]> var_18201_cast_fp16 = add(x = q_head_225_cast_fp16, y = var_18200_to_fp16)[name = string("op_18201_cast_fp16")];
            tensor<int32, [1]> q_v_225_axes_1 = const()[name = string("q_v_225_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_225_cast_fp16 = expand_dims(axes = q_v_225_axes_1, x = var_18201_cast_fp16)[name = string("q_v_225_cast_fp16")];
            tensor<int32, [1]> k_head_451_axes_1 = const()[name = string("k_head_451_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_451_cast_fp16 = expand_dims(axes = k_head_451_axes_1, x = k_head_449_cast_fp16)[name = string("k_head_451_cast_fp16")];
            tensor<int32, [1]> v_head_451_axes_1 = const()[name = string("v_head_451_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_451_cast_fp16 = expand_dims(axes = v_head_451_axes_1, x = v_head_449_cast_fp16)[name = string("v_head_451_cast_fp16")];
            tensor<int32, [1]> p_head_451_axes_1 = const()[name = string("p_head_451_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_451_cast_fp16 = expand_dims(axes = p_head_451_axes_1, x = p_head_449_cast_fp16)[name = string("p_head_451_cast_fp16")];
            bool var_18207_transpose_x_3 = const()[name = string("op_18207_transpose_x_3"), val = bool(false)];
            bool var_18207_transpose_y_3 = const()[name = string("op_18207_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_18207_cast_fp16 = matmul(transpose_x = var_18207_transpose_x_3, transpose_y = var_18207_transpose_y_3, x = q_u_225_cast_fp16, y = k_head_451_cast_fp16)[name = string("op_18207_cast_fp16")];
            fp16 var_18208_to_fp16 = const()[name = string("op_18208_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_225_cast_fp16 = mul(x = var_18207_cast_fp16, y = var_18208_to_fp16)[name = string("scores_content_225_cast_fp16")];
            bool x_1197_transpose_x_3 = const()[name = string("x_1197_transpose_x_3"), val = bool(false)];
            bool x_1197_transpose_y_3 = const()[name = string("x_1197_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1197_cast_fp16 = matmul(transpose_x = x_1197_transpose_x_3, transpose_y = x_1197_transpose_y_3, x = q_v_225_cast_fp16, y = p_head_451_cast_fp16)[name = string("x_1197_cast_fp16")];
            tensor<int32, [8]> x_1199_pad_1 = const()[name = string("x_1199_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1199_mode_1 = const()[name = string("x_1199_mode_1"), val = string("constant")];
            fp16 const_2065_to_fp16 = const()[name = string("const_2065_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1199_cast_fp16 = pad(constant_val = const_2065_to_fp16, mode = x_1199_mode_1, pad = x_1199_pad_1, x = x_1197_cast_fp16)[name = string("x_1199_cast_fp16")];
            tensor<int32, [4]> var_18222 = const()[name = string("op_18222"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1201_cast_fp16 = reshape(shape = var_18222, x = x_1199_cast_fp16)[name = string("x_1201_cast_fp16")];
            tensor<int32, [4]> var_18226_begin_1 = const()[name = string("op_18226_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_18226_end_1 = const()[name = string("op_18226_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_18226_end_mask_1 = const()[name = string("op_18226_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_18226_cast_fp16 = slice_by_index(begin = var_18226_begin_1, end = var_18226_end_1, end_mask = var_18226_end_mask_1, x = x_1201_cast_fp16)[name = string("op_18226_cast_fp16")];
            tensor<int32, [4]> var_18228 = const()[name = string("op_18228"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_18229_cast_fp16 = reshape(shape = var_18228, x = var_18226_cast_fp16)[name = string("op_18229_cast_fp16")];
            tensor<int32, [4]> var_18234_begin_1 = const()[name = string("op_18234_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_18234_end_1 = const()[name = string("op_18234_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_18234_end_mask_1 = const()[name = string("op_18234_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_18234_cast_fp16 = slice_by_index(begin = var_18234_begin_1, end = var_18234_end_1, end_mask = var_18234_end_mask_1, x = var_18229_cast_fp16)[name = string("op_18234_cast_fp16")];
            fp16 var_18235_to_fp16 = const()[name = string("op_18235_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_225_cast_fp16 = mul(x = var_18234_cast_fp16, y = var_18235_to_fp16)[name = string("scores_pos_225_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_225_cast_fp16 = add(x = scores_content_225_cast_fp16, y = scores_pos_225_cast_fp16)[name = string("logits_225_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_18238_cast_fp16 = softmax(axis = var_17963, x = logits_225_cast_fp16)[name = string("op_18238_cast_fp16")];
            bool var_18240_transpose_x_1 = const()[name = string("op_18240_transpose_x_1"), val = bool(false)];
            bool var_18240_transpose_y_1 = const()[name = string("op_18240_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_18240_cast_fp16 = matmul(transpose_x = var_18240_transpose_x_1, transpose_y = var_18240_transpose_y_1, x = var_18238_cast_fp16, y = v_head_451_cast_fp16)[name = string("op_18240_cast_fp16")];
            tensor<int32, [1]> var_18241_axes_1 = const()[name = string("op_18241_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_18241_cast_fp16 = squeeze(axes = var_18241_axes_1, x = var_18240_cast_fp16)[name = string("op_18241_cast_fp16")];
            string dense_output_1145_pad_type_1 = const()[name = string("dense_output_1145_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1145_strides_1 = const()[name = string("dense_output_1145_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1145_pad_1 = const()[name = string("dense_output_1145_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1145_dilations_1 = const()[name = string("dense_output_1145_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1145_groups_1 = const()[name = string("dense_output_1145_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425812864))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425944000))))[name = string("layers_14_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1145_cast_fp16 = conv(dilations = dense_output_1145_dilations_1, groups = dense_output_1145_groups_1, pad = dense_output_1145_pad_1, pad_type = dense_output_1145_pad_type_1, strides = dense_output_1145_strides_1, weight = layers_14_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1145_cast_fp16")];
            string sparse_output_1145_pad_type_1 = const()[name = string("sparse_output_1145_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1145_strides_1 = const()[name = string("sparse_output_1145_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1145_pad_1 = const()[name = string("sparse_output_1145_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1145_dilations_1 = const()[name = string("sparse_output_1145_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1145_groups_1 = const()[name = string("sparse_output_1145_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425947264))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425944576))))[name = string("layers_14_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1145_cast_fp16 = conv(dilations = sparse_output_1145_dilations_1, groups = sparse_output_1145_groups_1, pad = sparse_output_1145_pad_1, pad_type = sparse_output_1145_pad_type_1, strides = sparse_output_1145_strides_1, weight = layers_14_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1145_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18256_cast_fp16 = add(x = dense_output_1145_cast_fp16, y = sparse_output_1145_cast_fp16)[name = string("op_18256_cast_fp16")];
            tensor<int32, [4]> var_18257 = const()[name = string("op_18257"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18259 = const()[name = string("op_18259"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18258_cast_fp16 = transpose(perm = var_18257, x = var_18256_cast_fp16)[name = string("transpose_402")];
            tensor<fp16, [1, 51, 128]> q_head_227_cast_fp16 = reshape(shape = var_18259, x = var_18258_cast_fp16)[name = string("q_head_227_cast_fp16")];
            string dense_output_1147_pad_type_1 = const()[name = string("dense_output_1147_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1147_strides_1 = const()[name = string("dense_output_1147_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1147_pad_1 = const()[name = string("dense_output_1147_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1147_dilations_1 = const()[name = string("dense_output_1147_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1147_groups_1 = const()[name = string("dense_output_1147_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425963712))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426094848))))[name = string("layers_14_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1147_cast_fp16 = conv(dilations = dense_output_1147_dilations_1, groups = dense_output_1147_groups_1, pad = dense_output_1147_pad_1, pad_type = dense_output_1147_pad_type_1, strides = dense_output_1147_strides_1, weight = layers_14_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1147_cast_fp16")];
            string sparse_output_1147_pad_type_1 = const()[name = string("sparse_output_1147_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1147_strides_1 = const()[name = string("sparse_output_1147_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1147_pad_1 = const()[name = string("sparse_output_1147_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1147_dilations_1 = const()[name = string("sparse_output_1147_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1147_groups_1 = const()[name = string("sparse_output_1147_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426098112))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426095424))))[name = string("layers_14_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1147_cast_fp16 = conv(dilations = sparse_output_1147_dilations_1, groups = sparse_output_1147_groups_1, pad = sparse_output_1147_pad_1, pad_type = sparse_output_1147_pad_type_1, strides = sparse_output_1147_strides_1, weight = layers_14_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1147_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18275_cast_fp16 = add(x = dense_output_1147_cast_fp16, y = sparse_output_1147_cast_fp16)[name = string("op_18275_cast_fp16")];
            tensor<int32, [4]> var_18276 = const()[name = string("op_18276"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18278 = const()[name = string("op_18278"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18277_cast_fp16 = transpose(perm = var_18276, x = var_18275_cast_fp16)[name = string("transpose_401")];
            tensor<fp16, [1, 51, 128]> k_head_453_cast_fp16 = reshape(shape = var_18278, x = var_18277_cast_fp16)[name = string("k_head_453_cast_fp16")];
            string dense_output_1149_pad_type_1 = const()[name = string("dense_output_1149_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1149_strides_1 = const()[name = string("dense_output_1149_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1149_pad_1 = const()[name = string("dense_output_1149_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1149_dilations_1 = const()[name = string("dense_output_1149_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1149_groups_1 = const()[name = string("dense_output_1149_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426114560))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426245696))))[name = string("layers_14_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1149_cast_fp16 = conv(dilations = dense_output_1149_dilations_1, groups = dense_output_1149_groups_1, pad = dense_output_1149_pad_1, pad_type = dense_output_1149_pad_type_1, strides = dense_output_1149_strides_1, weight = layers_14_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1149_cast_fp16")];
            string sparse_output_1149_pad_type_1 = const()[name = string("sparse_output_1149_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1149_strides_1 = const()[name = string("sparse_output_1149_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1149_pad_1 = const()[name = string("sparse_output_1149_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1149_dilations_1 = const()[name = string("sparse_output_1149_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1149_groups_1 = const()[name = string("sparse_output_1149_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426248960))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426246272))))[name = string("layers_14_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1149_cast_fp16 = conv(dilations = sparse_output_1149_dilations_1, groups = sparse_output_1149_groups_1, pad = sparse_output_1149_pad_1, pad_type = sparse_output_1149_pad_type_1, strides = sparse_output_1149_strides_1, weight = layers_14_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1149_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18294_cast_fp16 = add(x = dense_output_1149_cast_fp16, y = sparse_output_1149_cast_fp16)[name = string("op_18294_cast_fp16")];
            tensor<int32, [4]> var_18295 = const()[name = string("op_18295"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18297 = const()[name = string("op_18297"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18296_cast_fp16 = transpose(perm = var_18295, x = var_18294_cast_fp16)[name = string("transpose_400")];
            tensor<fp16, [1, 51, 128]> v_head_453_cast_fp16 = reshape(shape = var_18297, x = var_18296_cast_fp16)[name = string("v_head_453_cast_fp16")];
            string dense_output_1151_pad_type_1 = const()[name = string("dense_output_1151_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1151_strides_1 = const()[name = string("dense_output_1151_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1151_pad_1 = const()[name = string("dense_output_1151_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1151_dilations_1 = const()[name = string("dense_output_1151_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1151_groups_1 = const()[name = string("dense_output_1151_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426265408))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426396544))))[name = string("layers_14_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1151_cast_fp16 = conv(dilations = dense_output_1151_dilations_1, groups = dense_output_1151_groups_1, pad = dense_output_1151_pad_1, pad_type = dense_output_1151_pad_type_1, strides = dense_output_1151_strides_1, weight = layers_14_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1151_cast_fp16")];
            string sparse_output_1151_pad_type_1 = const()[name = string("sparse_output_1151_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1151_strides_1 = const()[name = string("sparse_output_1151_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1151_pad_1 = const()[name = string("sparse_output_1151_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1151_dilations_1 = const()[name = string("sparse_output_1151_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1151_groups_1 = const()[name = string("sparse_output_1151_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426399808))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426397120))))[name = string("layers_14_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1151_cast_fp16 = conv(dilations = sparse_output_1151_dilations_1, groups = sparse_output_1151_groups_1, pad = sparse_output_1151_pad_1, pad_type = sparse_output_1151_pad_type_1, strides = sparse_output_1151_strides_1, weight = layers_14_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1151_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_18313_cast_fp16 = add(x = dense_output_1151_cast_fp16, y = sparse_output_1151_cast_fp16)[name = string("op_18313_cast_fp16")];
            tensor<int32, [4]> var_18314 = const()[name = string("op_18314"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18316 = const()[name = string("op_18316"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_18315_cast_fp16 = transpose(perm = var_18314, x = var_18313_cast_fp16)[name = string("transpose_399")];
            tensor<fp16, [1, 101, 128]> p_head_453_cast_fp16 = reshape(shape = var_18316, x = var_18315_cast_fp16)[name = string("p_head_453_cast_fp16")];
            tensor<fp16, [128]> var_18318_to_fp16 = const()[name = string("op_18318_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426416256)))];
            tensor<fp16, [1, 51, 128]> var_18319_cast_fp16 = add(x = q_head_227_cast_fp16, y = var_18318_to_fp16)[name = string("op_18319_cast_fp16")];
            tensor<int32, [1]> q_u_227_axes_1 = const()[name = string("q_u_227_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_227_cast_fp16 = expand_dims(axes = q_u_227_axes_1, x = var_18319_cast_fp16)[name = string("q_u_227_cast_fp16")];
            tensor<fp16, [128]> var_18321_to_fp16 = const()[name = string("op_18321_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426416576)))];
            tensor<fp16, [1, 51, 128]> var_18322_cast_fp16 = add(x = q_head_227_cast_fp16, y = var_18321_to_fp16)[name = string("op_18322_cast_fp16")];
            tensor<int32, [1]> q_v_227_axes_1 = const()[name = string("q_v_227_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_227_cast_fp16 = expand_dims(axes = q_v_227_axes_1, x = var_18322_cast_fp16)[name = string("q_v_227_cast_fp16")];
            tensor<int32, [1]> k_head_455_axes_1 = const()[name = string("k_head_455_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_455_cast_fp16 = expand_dims(axes = k_head_455_axes_1, x = k_head_453_cast_fp16)[name = string("k_head_455_cast_fp16")];
            tensor<int32, [1]> v_head_455_axes_1 = const()[name = string("v_head_455_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_455_cast_fp16 = expand_dims(axes = v_head_455_axes_1, x = v_head_453_cast_fp16)[name = string("v_head_455_cast_fp16")];
            tensor<int32, [1]> p_head_455_axes_1 = const()[name = string("p_head_455_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_455_cast_fp16 = expand_dims(axes = p_head_455_axes_1, x = p_head_453_cast_fp16)[name = string("p_head_455_cast_fp16")];
            bool var_18328_transpose_x_3 = const()[name = string("op_18328_transpose_x_3"), val = bool(false)];
            bool var_18328_transpose_y_3 = const()[name = string("op_18328_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_18328_cast_fp16 = matmul(transpose_x = var_18328_transpose_x_3, transpose_y = var_18328_transpose_y_3, x = q_u_227_cast_fp16, y = k_head_455_cast_fp16)[name = string("op_18328_cast_fp16")];
            fp16 var_18329_to_fp16 = const()[name = string("op_18329_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_227_cast_fp16 = mul(x = var_18328_cast_fp16, y = var_18329_to_fp16)[name = string("scores_content_227_cast_fp16")];
            bool x_1205_transpose_x_3 = const()[name = string("x_1205_transpose_x_3"), val = bool(false)];
            bool x_1205_transpose_y_3 = const()[name = string("x_1205_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1205_cast_fp16 = matmul(transpose_x = x_1205_transpose_x_3, transpose_y = x_1205_transpose_y_3, x = q_v_227_cast_fp16, y = p_head_455_cast_fp16)[name = string("x_1205_cast_fp16")];
            tensor<int32, [8]> x_1207_pad_1 = const()[name = string("x_1207_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1207_mode_1 = const()[name = string("x_1207_mode_1"), val = string("constant")];
            fp16 const_2071_to_fp16 = const()[name = string("const_2071_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1207_cast_fp16 = pad(constant_val = const_2071_to_fp16, mode = x_1207_mode_1, pad = x_1207_pad_1, x = x_1205_cast_fp16)[name = string("x_1207_cast_fp16")];
            tensor<int32, [4]> var_18343 = const()[name = string("op_18343"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1209_cast_fp16 = reshape(shape = var_18343, x = x_1207_cast_fp16)[name = string("x_1209_cast_fp16")];
            tensor<int32, [4]> var_18347_begin_1 = const()[name = string("op_18347_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_18347_end_1 = const()[name = string("op_18347_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_18347_end_mask_1 = const()[name = string("op_18347_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_18347_cast_fp16 = slice_by_index(begin = var_18347_begin_1, end = var_18347_end_1, end_mask = var_18347_end_mask_1, x = x_1209_cast_fp16)[name = string("op_18347_cast_fp16")];
            tensor<int32, [4]> var_18349 = const()[name = string("op_18349"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_18350_cast_fp16 = reshape(shape = var_18349, x = var_18347_cast_fp16)[name = string("op_18350_cast_fp16")];
            tensor<int32, [4]> var_18355_begin_1 = const()[name = string("op_18355_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_18355_end_1 = const()[name = string("op_18355_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_18355_end_mask_1 = const()[name = string("op_18355_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_18355_cast_fp16 = slice_by_index(begin = var_18355_begin_1, end = var_18355_end_1, end_mask = var_18355_end_mask_1, x = var_18350_cast_fp16)[name = string("op_18355_cast_fp16")];
            fp16 var_18356_to_fp16 = const()[name = string("op_18356_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_227_cast_fp16 = mul(x = var_18355_cast_fp16, y = var_18356_to_fp16)[name = string("scores_pos_227_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_227_cast_fp16 = add(x = scores_content_227_cast_fp16, y = scores_pos_227_cast_fp16)[name = string("logits_227_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_18359_cast_fp16 = softmax(axis = var_17963, x = logits_227_cast_fp16)[name = string("op_18359_cast_fp16")];
            bool var_18361_transpose_x_1 = const()[name = string("op_18361_transpose_x_1"), val = bool(false)];
            bool var_18361_transpose_y_1 = const()[name = string("op_18361_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_18361_cast_fp16 = matmul(transpose_x = var_18361_transpose_x_1, transpose_y = var_18361_transpose_y_1, x = var_18359_cast_fp16, y = v_head_455_cast_fp16)[name = string("op_18361_cast_fp16")];
            tensor<int32, [1]> var_18362_axes_1 = const()[name = string("op_18362_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_18362_cast_fp16 = squeeze(axes = var_18362_axes_1, x = var_18361_cast_fp16)[name = string("op_18362_cast_fp16")];
            string dense_output_1153_pad_type_1 = const()[name = string("dense_output_1153_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1153_strides_1 = const()[name = string("dense_output_1153_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1153_pad_1 = const()[name = string("dense_output_1153_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1153_dilations_1 = const()[name = string("dense_output_1153_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1153_groups_1 = const()[name = string("dense_output_1153_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426416896))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426548032))))[name = string("layers_14_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1153_cast_fp16 = conv(dilations = dense_output_1153_dilations_1, groups = dense_output_1153_groups_1, pad = dense_output_1153_pad_1, pad_type = dense_output_1153_pad_type_1, strides = dense_output_1153_strides_1, weight = layers_14_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1153_cast_fp16")];
            string sparse_output_1153_pad_type_1 = const()[name = string("sparse_output_1153_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1153_strides_1 = const()[name = string("sparse_output_1153_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1153_pad_1 = const()[name = string("sparse_output_1153_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1153_dilations_1 = const()[name = string("sparse_output_1153_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1153_groups_1 = const()[name = string("sparse_output_1153_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426551296))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426548608))))[name = string("layers_14_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1153_cast_fp16 = conv(dilations = sparse_output_1153_dilations_1, groups = sparse_output_1153_groups_1, pad = sparse_output_1153_pad_1, pad_type = sparse_output_1153_pad_type_1, strides = sparse_output_1153_strides_1, weight = layers_14_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1153_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18377_cast_fp16 = add(x = dense_output_1153_cast_fp16, y = sparse_output_1153_cast_fp16)[name = string("op_18377_cast_fp16")];
            tensor<int32, [4]> var_18378 = const()[name = string("op_18378"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18380 = const()[name = string("op_18380"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18379_cast_fp16 = transpose(perm = var_18378, x = var_18377_cast_fp16)[name = string("transpose_398")];
            tensor<fp16, [1, 51, 128]> q_head_229_cast_fp16 = reshape(shape = var_18380, x = var_18379_cast_fp16)[name = string("q_head_229_cast_fp16")];
            string dense_output_1155_pad_type_1 = const()[name = string("dense_output_1155_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1155_strides_1 = const()[name = string("dense_output_1155_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1155_pad_1 = const()[name = string("dense_output_1155_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1155_dilations_1 = const()[name = string("dense_output_1155_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1155_groups_1 = const()[name = string("dense_output_1155_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426567744))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426698880))))[name = string("layers_14_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1155_cast_fp16 = conv(dilations = dense_output_1155_dilations_1, groups = dense_output_1155_groups_1, pad = dense_output_1155_pad_1, pad_type = dense_output_1155_pad_type_1, strides = dense_output_1155_strides_1, weight = layers_14_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1155_cast_fp16")];
            string sparse_output_1155_pad_type_1 = const()[name = string("sparse_output_1155_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1155_strides_1 = const()[name = string("sparse_output_1155_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1155_pad_1 = const()[name = string("sparse_output_1155_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1155_dilations_1 = const()[name = string("sparse_output_1155_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1155_groups_1 = const()[name = string("sparse_output_1155_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426702144))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426699456))))[name = string("layers_14_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1155_cast_fp16 = conv(dilations = sparse_output_1155_dilations_1, groups = sparse_output_1155_groups_1, pad = sparse_output_1155_pad_1, pad_type = sparse_output_1155_pad_type_1, strides = sparse_output_1155_strides_1, weight = layers_14_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1155_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18396_cast_fp16 = add(x = dense_output_1155_cast_fp16, y = sparse_output_1155_cast_fp16)[name = string("op_18396_cast_fp16")];
            tensor<int32, [4]> var_18397 = const()[name = string("op_18397"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18399 = const()[name = string("op_18399"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18398_cast_fp16 = transpose(perm = var_18397, x = var_18396_cast_fp16)[name = string("transpose_397")];
            tensor<fp16, [1, 51, 128]> k_head_457_cast_fp16 = reshape(shape = var_18399, x = var_18398_cast_fp16)[name = string("k_head_457_cast_fp16")];
            string dense_output_1157_pad_type_1 = const()[name = string("dense_output_1157_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1157_strides_1 = const()[name = string("dense_output_1157_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1157_pad_1 = const()[name = string("dense_output_1157_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1157_dilations_1 = const()[name = string("dense_output_1157_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1157_groups_1 = const()[name = string("dense_output_1157_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426718592))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426849728))))[name = string("layers_14_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1157_cast_fp16 = conv(dilations = dense_output_1157_dilations_1, groups = dense_output_1157_groups_1, pad = dense_output_1157_pad_1, pad_type = dense_output_1157_pad_type_1, strides = dense_output_1157_strides_1, weight = layers_14_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1157_cast_fp16")];
            string sparse_output_1157_pad_type_1 = const()[name = string("sparse_output_1157_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1157_strides_1 = const()[name = string("sparse_output_1157_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1157_pad_1 = const()[name = string("sparse_output_1157_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1157_dilations_1 = const()[name = string("sparse_output_1157_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1157_groups_1 = const()[name = string("sparse_output_1157_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426852992))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426850304))))[name = string("layers_14_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1157_cast_fp16 = conv(dilations = sparse_output_1157_dilations_1, groups = sparse_output_1157_groups_1, pad = sparse_output_1157_pad_1, pad_type = sparse_output_1157_pad_type_1, strides = sparse_output_1157_strides_1, weight = layers_14_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1157_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18415_cast_fp16 = add(x = dense_output_1157_cast_fp16, y = sparse_output_1157_cast_fp16)[name = string("op_18415_cast_fp16")];
            tensor<int32, [4]> var_18416 = const()[name = string("op_18416"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18418 = const()[name = string("op_18418"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18417_cast_fp16 = transpose(perm = var_18416, x = var_18415_cast_fp16)[name = string("transpose_396")];
            tensor<fp16, [1, 51, 128]> v_head_457_cast_fp16 = reshape(shape = var_18418, x = var_18417_cast_fp16)[name = string("v_head_457_cast_fp16")];
            string dense_output_1159_pad_type_1 = const()[name = string("dense_output_1159_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1159_strides_1 = const()[name = string("dense_output_1159_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1159_pad_1 = const()[name = string("dense_output_1159_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1159_dilations_1 = const()[name = string("dense_output_1159_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1159_groups_1 = const()[name = string("dense_output_1159_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426869440))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427000576))))[name = string("layers_14_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1159_cast_fp16 = conv(dilations = dense_output_1159_dilations_1, groups = dense_output_1159_groups_1, pad = dense_output_1159_pad_1, pad_type = dense_output_1159_pad_type_1, strides = dense_output_1159_strides_1, weight = layers_14_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1159_cast_fp16")];
            string sparse_output_1159_pad_type_1 = const()[name = string("sparse_output_1159_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1159_strides_1 = const()[name = string("sparse_output_1159_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1159_pad_1 = const()[name = string("sparse_output_1159_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1159_dilations_1 = const()[name = string("sparse_output_1159_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1159_groups_1 = const()[name = string("sparse_output_1159_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427003840))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427001152))))[name = string("layers_14_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1159_cast_fp16 = conv(dilations = sparse_output_1159_dilations_1, groups = sparse_output_1159_groups_1, pad = sparse_output_1159_pad_1, pad_type = sparse_output_1159_pad_type_1, strides = sparse_output_1159_strides_1, weight = layers_14_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1159_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_18434_cast_fp16 = add(x = dense_output_1159_cast_fp16, y = sparse_output_1159_cast_fp16)[name = string("op_18434_cast_fp16")];
            tensor<int32, [4]> var_18435 = const()[name = string("op_18435"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18437 = const()[name = string("op_18437"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_18436_cast_fp16 = transpose(perm = var_18435, x = var_18434_cast_fp16)[name = string("transpose_395")];
            tensor<fp16, [1, 101, 128]> p_head_457_cast_fp16 = reshape(shape = var_18437, x = var_18436_cast_fp16)[name = string("p_head_457_cast_fp16")];
            tensor<fp16, [128]> var_18439_to_fp16 = const()[name = string("op_18439_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427020288)))];
            tensor<fp16, [1, 51, 128]> var_18440_cast_fp16 = add(x = q_head_229_cast_fp16, y = var_18439_to_fp16)[name = string("op_18440_cast_fp16")];
            tensor<int32, [1]> q_u_229_axes_1 = const()[name = string("q_u_229_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_229_cast_fp16 = expand_dims(axes = q_u_229_axes_1, x = var_18440_cast_fp16)[name = string("q_u_229_cast_fp16")];
            tensor<fp16, [128]> var_18442_to_fp16 = const()[name = string("op_18442_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427020608)))];
            tensor<fp16, [1, 51, 128]> var_18443_cast_fp16 = add(x = q_head_229_cast_fp16, y = var_18442_to_fp16)[name = string("op_18443_cast_fp16")];
            tensor<int32, [1]> q_v_229_axes_1 = const()[name = string("q_v_229_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_229_cast_fp16 = expand_dims(axes = q_v_229_axes_1, x = var_18443_cast_fp16)[name = string("q_v_229_cast_fp16")];
            tensor<int32, [1]> k_head_459_axes_1 = const()[name = string("k_head_459_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_459_cast_fp16 = expand_dims(axes = k_head_459_axes_1, x = k_head_457_cast_fp16)[name = string("k_head_459_cast_fp16")];
            tensor<int32, [1]> v_head_459_axes_1 = const()[name = string("v_head_459_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_459_cast_fp16 = expand_dims(axes = v_head_459_axes_1, x = v_head_457_cast_fp16)[name = string("v_head_459_cast_fp16")];
            tensor<int32, [1]> p_head_459_axes_1 = const()[name = string("p_head_459_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_459_cast_fp16 = expand_dims(axes = p_head_459_axes_1, x = p_head_457_cast_fp16)[name = string("p_head_459_cast_fp16")];
            bool var_18449_transpose_x_3 = const()[name = string("op_18449_transpose_x_3"), val = bool(false)];
            bool var_18449_transpose_y_3 = const()[name = string("op_18449_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_18449_cast_fp16 = matmul(transpose_x = var_18449_transpose_x_3, transpose_y = var_18449_transpose_y_3, x = q_u_229_cast_fp16, y = k_head_459_cast_fp16)[name = string("op_18449_cast_fp16")];
            fp16 var_18450_to_fp16 = const()[name = string("op_18450_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_229_cast_fp16 = mul(x = var_18449_cast_fp16, y = var_18450_to_fp16)[name = string("scores_content_229_cast_fp16")];
            bool x_1213_transpose_x_3 = const()[name = string("x_1213_transpose_x_3"), val = bool(false)];
            bool x_1213_transpose_y_3 = const()[name = string("x_1213_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1213_cast_fp16 = matmul(transpose_x = x_1213_transpose_x_3, transpose_y = x_1213_transpose_y_3, x = q_v_229_cast_fp16, y = p_head_459_cast_fp16)[name = string("x_1213_cast_fp16")];
            tensor<int32, [8]> x_1215_pad_1 = const()[name = string("x_1215_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1215_mode_1 = const()[name = string("x_1215_mode_1"), val = string("constant")];
            fp16 const_2077_to_fp16 = const()[name = string("const_2077_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1215_cast_fp16 = pad(constant_val = const_2077_to_fp16, mode = x_1215_mode_1, pad = x_1215_pad_1, x = x_1213_cast_fp16)[name = string("x_1215_cast_fp16")];
            tensor<int32, [4]> var_18464 = const()[name = string("op_18464"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1217_cast_fp16 = reshape(shape = var_18464, x = x_1215_cast_fp16)[name = string("x_1217_cast_fp16")];
            tensor<int32, [4]> var_18468_begin_1 = const()[name = string("op_18468_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_18468_end_1 = const()[name = string("op_18468_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_18468_end_mask_1 = const()[name = string("op_18468_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_18468_cast_fp16 = slice_by_index(begin = var_18468_begin_1, end = var_18468_end_1, end_mask = var_18468_end_mask_1, x = x_1217_cast_fp16)[name = string("op_18468_cast_fp16")];
            tensor<int32, [4]> var_18470 = const()[name = string("op_18470"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_18471_cast_fp16 = reshape(shape = var_18470, x = var_18468_cast_fp16)[name = string("op_18471_cast_fp16")];
            tensor<int32, [4]> var_18476_begin_1 = const()[name = string("op_18476_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_18476_end_1 = const()[name = string("op_18476_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_18476_end_mask_1 = const()[name = string("op_18476_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_18476_cast_fp16 = slice_by_index(begin = var_18476_begin_1, end = var_18476_end_1, end_mask = var_18476_end_mask_1, x = var_18471_cast_fp16)[name = string("op_18476_cast_fp16")];
            fp16 var_18477_to_fp16 = const()[name = string("op_18477_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_229_cast_fp16 = mul(x = var_18476_cast_fp16, y = var_18477_to_fp16)[name = string("scores_pos_229_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_229_cast_fp16 = add(x = scores_content_229_cast_fp16, y = scores_pos_229_cast_fp16)[name = string("logits_229_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_18480_cast_fp16 = softmax(axis = var_17963, x = logits_229_cast_fp16)[name = string("op_18480_cast_fp16")];
            bool var_18482_transpose_x_1 = const()[name = string("op_18482_transpose_x_1"), val = bool(false)];
            bool var_18482_transpose_y_1 = const()[name = string("op_18482_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_18482_cast_fp16 = matmul(transpose_x = var_18482_transpose_x_1, transpose_y = var_18482_transpose_y_1, x = var_18480_cast_fp16, y = v_head_459_cast_fp16)[name = string("op_18482_cast_fp16")];
            tensor<int32, [1]> var_18483_axes_1 = const()[name = string("op_18483_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_18483_cast_fp16 = squeeze(axes = var_18483_axes_1, x = var_18482_cast_fp16)[name = string("op_18483_cast_fp16")];
            string dense_output_1161_pad_type_1 = const()[name = string("dense_output_1161_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1161_strides_1 = const()[name = string("dense_output_1161_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1161_pad_1 = const()[name = string("dense_output_1161_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1161_dilations_1 = const()[name = string("dense_output_1161_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1161_groups_1 = const()[name = string("dense_output_1161_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427020928))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427152064))))[name = string("layers_14_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1161_cast_fp16 = conv(dilations = dense_output_1161_dilations_1, groups = dense_output_1161_groups_1, pad = dense_output_1161_pad_1, pad_type = dense_output_1161_pad_type_1, strides = dense_output_1161_strides_1, weight = layers_14_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1161_cast_fp16")];
            string sparse_output_1161_pad_type_1 = const()[name = string("sparse_output_1161_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1161_strides_1 = const()[name = string("sparse_output_1161_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1161_pad_1 = const()[name = string("sparse_output_1161_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1161_dilations_1 = const()[name = string("sparse_output_1161_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1161_groups_1 = const()[name = string("sparse_output_1161_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427155328))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427152640))))[name = string("layers_14_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1161_cast_fp16 = conv(dilations = sparse_output_1161_dilations_1, groups = sparse_output_1161_groups_1, pad = sparse_output_1161_pad_1, pad_type = sparse_output_1161_pad_type_1, strides = sparse_output_1161_strides_1, weight = layers_14_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1161_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18498_cast_fp16 = add(x = dense_output_1161_cast_fp16, y = sparse_output_1161_cast_fp16)[name = string("op_18498_cast_fp16")];
            tensor<int32, [4]> var_18499 = const()[name = string("op_18499"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18501 = const()[name = string("op_18501"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18500_cast_fp16 = transpose(perm = var_18499, x = var_18498_cast_fp16)[name = string("transpose_394")];
            tensor<fp16, [1, 51, 128]> q_head_231_cast_fp16 = reshape(shape = var_18501, x = var_18500_cast_fp16)[name = string("q_head_231_cast_fp16")];
            string dense_output_1163_pad_type_1 = const()[name = string("dense_output_1163_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1163_strides_1 = const()[name = string("dense_output_1163_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1163_pad_1 = const()[name = string("dense_output_1163_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1163_dilations_1 = const()[name = string("dense_output_1163_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1163_groups_1 = const()[name = string("dense_output_1163_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427171776))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427302912))))[name = string("layers_14_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1163_cast_fp16 = conv(dilations = dense_output_1163_dilations_1, groups = dense_output_1163_groups_1, pad = dense_output_1163_pad_1, pad_type = dense_output_1163_pad_type_1, strides = dense_output_1163_strides_1, weight = layers_14_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1163_cast_fp16")];
            string sparse_output_1163_pad_type_1 = const()[name = string("sparse_output_1163_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1163_strides_1 = const()[name = string("sparse_output_1163_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1163_pad_1 = const()[name = string("sparse_output_1163_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1163_dilations_1 = const()[name = string("sparse_output_1163_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1163_groups_1 = const()[name = string("sparse_output_1163_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427306176))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427303488))))[name = string("layers_14_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1163_cast_fp16 = conv(dilations = sparse_output_1163_dilations_1, groups = sparse_output_1163_groups_1, pad = sparse_output_1163_pad_1, pad_type = sparse_output_1163_pad_type_1, strides = sparse_output_1163_strides_1, weight = layers_14_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1163_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18517_cast_fp16 = add(x = dense_output_1163_cast_fp16, y = sparse_output_1163_cast_fp16)[name = string("op_18517_cast_fp16")];
            tensor<int32, [4]> var_18518 = const()[name = string("op_18518"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18520 = const()[name = string("op_18520"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18519_cast_fp16 = transpose(perm = var_18518, x = var_18517_cast_fp16)[name = string("transpose_393")];
            tensor<fp16, [1, 51, 128]> k_head_461_cast_fp16 = reshape(shape = var_18520, x = var_18519_cast_fp16)[name = string("k_head_461_cast_fp16")];
            string dense_output_1165_pad_type_1 = const()[name = string("dense_output_1165_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1165_strides_1 = const()[name = string("dense_output_1165_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1165_pad_1 = const()[name = string("dense_output_1165_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1165_dilations_1 = const()[name = string("dense_output_1165_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1165_groups_1 = const()[name = string("dense_output_1165_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427322624))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427453760))))[name = string("layers_14_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1165_cast_fp16 = conv(dilations = dense_output_1165_dilations_1, groups = dense_output_1165_groups_1, pad = dense_output_1165_pad_1, pad_type = dense_output_1165_pad_type_1, strides = dense_output_1165_strides_1, weight = layers_14_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1165_cast_fp16")];
            string sparse_output_1165_pad_type_1 = const()[name = string("sparse_output_1165_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1165_strides_1 = const()[name = string("sparse_output_1165_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1165_pad_1 = const()[name = string("sparse_output_1165_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1165_dilations_1 = const()[name = string("sparse_output_1165_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1165_groups_1 = const()[name = string("sparse_output_1165_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427457024))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427454336))))[name = string("layers_14_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1165_cast_fp16 = conv(dilations = sparse_output_1165_dilations_1, groups = sparse_output_1165_groups_1, pad = sparse_output_1165_pad_1, pad_type = sparse_output_1165_pad_type_1, strides = sparse_output_1165_strides_1, weight = layers_14_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1165_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18536_cast_fp16 = add(x = dense_output_1165_cast_fp16, y = sparse_output_1165_cast_fp16)[name = string("op_18536_cast_fp16")];
            tensor<int32, [4]> var_18537 = const()[name = string("op_18537"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18539 = const()[name = string("op_18539"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18538_cast_fp16 = transpose(perm = var_18537, x = var_18536_cast_fp16)[name = string("transpose_392")];
            tensor<fp16, [1, 51, 128]> v_head_461_cast_fp16 = reshape(shape = var_18539, x = var_18538_cast_fp16)[name = string("v_head_461_cast_fp16")];
            string dense_output_1167_pad_type_1 = const()[name = string("dense_output_1167_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1167_strides_1 = const()[name = string("dense_output_1167_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1167_pad_1 = const()[name = string("dense_output_1167_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1167_dilations_1 = const()[name = string("dense_output_1167_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1167_groups_1 = const()[name = string("dense_output_1167_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427473472))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427604608))))[name = string("layers_14_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1167_cast_fp16 = conv(dilations = dense_output_1167_dilations_1, groups = dense_output_1167_groups_1, pad = dense_output_1167_pad_1, pad_type = dense_output_1167_pad_type_1, strides = dense_output_1167_strides_1, weight = layers_14_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1167_cast_fp16")];
            string sparse_output_1167_pad_type_1 = const()[name = string("sparse_output_1167_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1167_strides_1 = const()[name = string("sparse_output_1167_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1167_pad_1 = const()[name = string("sparse_output_1167_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1167_dilations_1 = const()[name = string("sparse_output_1167_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1167_groups_1 = const()[name = string("sparse_output_1167_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427607872))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427605184))))[name = string("layers_14_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1167_cast_fp16 = conv(dilations = sparse_output_1167_dilations_1, groups = sparse_output_1167_groups_1, pad = sparse_output_1167_pad_1, pad_type = sparse_output_1167_pad_type_1, strides = sparse_output_1167_strides_1, weight = layers_14_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1167_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_18555_cast_fp16 = add(x = dense_output_1167_cast_fp16, y = sparse_output_1167_cast_fp16)[name = string("op_18555_cast_fp16")];
            tensor<int32, [4]> var_18556 = const()[name = string("op_18556"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18558 = const()[name = string("op_18558"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_18557_cast_fp16 = transpose(perm = var_18556, x = var_18555_cast_fp16)[name = string("transpose_391")];
            tensor<fp16, [1, 101, 128]> p_head_461_cast_fp16 = reshape(shape = var_18558, x = var_18557_cast_fp16)[name = string("p_head_461_cast_fp16")];
            tensor<fp16, [128]> var_18560_to_fp16 = const()[name = string("op_18560_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427624320)))];
            tensor<fp16, [1, 51, 128]> var_18561_cast_fp16 = add(x = q_head_231_cast_fp16, y = var_18560_to_fp16)[name = string("op_18561_cast_fp16")];
            tensor<int32, [1]> q_u_231_axes_1 = const()[name = string("q_u_231_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_231_cast_fp16 = expand_dims(axes = q_u_231_axes_1, x = var_18561_cast_fp16)[name = string("q_u_231_cast_fp16")];
            tensor<fp16, [128]> var_18563_to_fp16 = const()[name = string("op_18563_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427624640)))];
            tensor<fp16, [1, 51, 128]> var_18564_cast_fp16 = add(x = q_head_231_cast_fp16, y = var_18563_to_fp16)[name = string("op_18564_cast_fp16")];
            tensor<int32, [1]> q_v_231_axes_1 = const()[name = string("q_v_231_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_231_cast_fp16 = expand_dims(axes = q_v_231_axes_1, x = var_18564_cast_fp16)[name = string("q_v_231_cast_fp16")];
            tensor<int32, [1]> k_head_463_axes_1 = const()[name = string("k_head_463_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_463_cast_fp16 = expand_dims(axes = k_head_463_axes_1, x = k_head_461_cast_fp16)[name = string("k_head_463_cast_fp16")];
            tensor<int32, [1]> v_head_463_axes_1 = const()[name = string("v_head_463_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_463_cast_fp16 = expand_dims(axes = v_head_463_axes_1, x = v_head_461_cast_fp16)[name = string("v_head_463_cast_fp16")];
            tensor<int32, [1]> p_head_463_axes_1 = const()[name = string("p_head_463_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_463_cast_fp16 = expand_dims(axes = p_head_463_axes_1, x = p_head_461_cast_fp16)[name = string("p_head_463_cast_fp16")];
            bool var_18570_transpose_x_3 = const()[name = string("op_18570_transpose_x_3"), val = bool(false)];
            bool var_18570_transpose_y_3 = const()[name = string("op_18570_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_18570_cast_fp16 = matmul(transpose_x = var_18570_transpose_x_3, transpose_y = var_18570_transpose_y_3, x = q_u_231_cast_fp16, y = k_head_463_cast_fp16)[name = string("op_18570_cast_fp16")];
            fp16 var_18571_to_fp16 = const()[name = string("op_18571_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_231_cast_fp16 = mul(x = var_18570_cast_fp16, y = var_18571_to_fp16)[name = string("scores_content_231_cast_fp16")];
            bool x_1221_transpose_x_3 = const()[name = string("x_1221_transpose_x_3"), val = bool(false)];
            bool x_1221_transpose_y_3 = const()[name = string("x_1221_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1221_cast_fp16 = matmul(transpose_x = x_1221_transpose_x_3, transpose_y = x_1221_transpose_y_3, x = q_v_231_cast_fp16, y = p_head_463_cast_fp16)[name = string("x_1221_cast_fp16")];
            tensor<int32, [8]> x_1223_pad_1 = const()[name = string("x_1223_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1223_mode_1 = const()[name = string("x_1223_mode_1"), val = string("constant")];
            fp16 const_2083_to_fp16 = const()[name = string("const_2083_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1223_cast_fp16 = pad(constant_val = const_2083_to_fp16, mode = x_1223_mode_1, pad = x_1223_pad_1, x = x_1221_cast_fp16)[name = string("x_1223_cast_fp16")];
            tensor<int32, [4]> var_18585 = const()[name = string("op_18585"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1225_cast_fp16 = reshape(shape = var_18585, x = x_1223_cast_fp16)[name = string("x_1225_cast_fp16")];
            tensor<int32, [4]> var_18589_begin_1 = const()[name = string("op_18589_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_18589_end_1 = const()[name = string("op_18589_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_18589_end_mask_1 = const()[name = string("op_18589_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_18589_cast_fp16 = slice_by_index(begin = var_18589_begin_1, end = var_18589_end_1, end_mask = var_18589_end_mask_1, x = x_1225_cast_fp16)[name = string("op_18589_cast_fp16")];
            tensor<int32, [4]> var_18591 = const()[name = string("op_18591"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_18592_cast_fp16 = reshape(shape = var_18591, x = var_18589_cast_fp16)[name = string("op_18592_cast_fp16")];
            tensor<int32, [4]> var_18597_begin_1 = const()[name = string("op_18597_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_18597_end_1 = const()[name = string("op_18597_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_18597_end_mask_1 = const()[name = string("op_18597_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_18597_cast_fp16 = slice_by_index(begin = var_18597_begin_1, end = var_18597_end_1, end_mask = var_18597_end_mask_1, x = var_18592_cast_fp16)[name = string("op_18597_cast_fp16")];
            fp16 var_18598_to_fp16 = const()[name = string("op_18598_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_231_cast_fp16 = mul(x = var_18597_cast_fp16, y = var_18598_to_fp16)[name = string("scores_pos_231_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_231_cast_fp16 = add(x = scores_content_231_cast_fp16, y = scores_pos_231_cast_fp16)[name = string("logits_231_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_18601_cast_fp16 = softmax(axis = var_17963, x = logits_231_cast_fp16)[name = string("op_18601_cast_fp16")];
            bool var_18603_transpose_x_1 = const()[name = string("op_18603_transpose_x_1"), val = bool(false)];
            bool var_18603_transpose_y_1 = const()[name = string("op_18603_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_18603_cast_fp16 = matmul(transpose_x = var_18603_transpose_x_1, transpose_y = var_18603_transpose_y_1, x = var_18601_cast_fp16, y = v_head_463_cast_fp16)[name = string("op_18603_cast_fp16")];
            tensor<int32, [1]> var_18604_axes_1 = const()[name = string("op_18604_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_18604_cast_fp16 = squeeze(axes = var_18604_axes_1, x = var_18603_cast_fp16)[name = string("op_18604_cast_fp16")];
            string dense_output_1169_pad_type_1 = const()[name = string("dense_output_1169_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1169_strides_1 = const()[name = string("dense_output_1169_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1169_pad_1 = const()[name = string("dense_output_1169_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1169_dilations_1 = const()[name = string("dense_output_1169_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1169_groups_1 = const()[name = string("dense_output_1169_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427624960))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427756096))))[name = string("layers_14_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1169_cast_fp16 = conv(dilations = dense_output_1169_dilations_1, groups = dense_output_1169_groups_1, pad = dense_output_1169_pad_1, pad_type = dense_output_1169_pad_type_1, strides = dense_output_1169_strides_1, weight = layers_14_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1169_cast_fp16")];
            string sparse_output_1169_pad_type_1 = const()[name = string("sparse_output_1169_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1169_strides_1 = const()[name = string("sparse_output_1169_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1169_pad_1 = const()[name = string("sparse_output_1169_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1169_dilations_1 = const()[name = string("sparse_output_1169_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1169_groups_1 = const()[name = string("sparse_output_1169_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427759360))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427756672))))[name = string("layers_14_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1169_cast_fp16 = conv(dilations = sparse_output_1169_dilations_1, groups = sparse_output_1169_groups_1, pad = sparse_output_1169_pad_1, pad_type = sparse_output_1169_pad_type_1, strides = sparse_output_1169_strides_1, weight = layers_14_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1169_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18619_cast_fp16 = add(x = dense_output_1169_cast_fp16, y = sparse_output_1169_cast_fp16)[name = string("op_18619_cast_fp16")];
            tensor<int32, [4]> var_18620 = const()[name = string("op_18620"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18622 = const()[name = string("op_18622"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18621_cast_fp16 = transpose(perm = var_18620, x = var_18619_cast_fp16)[name = string("transpose_390")];
            tensor<fp16, [1, 51, 128]> q_head_233_cast_fp16 = reshape(shape = var_18622, x = var_18621_cast_fp16)[name = string("q_head_233_cast_fp16")];
            string dense_output_1171_pad_type_1 = const()[name = string("dense_output_1171_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1171_strides_1 = const()[name = string("dense_output_1171_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1171_pad_1 = const()[name = string("dense_output_1171_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1171_dilations_1 = const()[name = string("dense_output_1171_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1171_groups_1 = const()[name = string("dense_output_1171_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427775808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427906944))))[name = string("layers_14_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1171_cast_fp16 = conv(dilations = dense_output_1171_dilations_1, groups = dense_output_1171_groups_1, pad = dense_output_1171_pad_1, pad_type = dense_output_1171_pad_type_1, strides = dense_output_1171_strides_1, weight = layers_14_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1171_cast_fp16")];
            string sparse_output_1171_pad_type_1 = const()[name = string("sparse_output_1171_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1171_strides_1 = const()[name = string("sparse_output_1171_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1171_pad_1 = const()[name = string("sparse_output_1171_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1171_dilations_1 = const()[name = string("sparse_output_1171_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1171_groups_1 = const()[name = string("sparse_output_1171_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427910208))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427907520))))[name = string("layers_14_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1171_cast_fp16 = conv(dilations = sparse_output_1171_dilations_1, groups = sparse_output_1171_groups_1, pad = sparse_output_1171_pad_1, pad_type = sparse_output_1171_pad_type_1, strides = sparse_output_1171_strides_1, weight = layers_14_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1171_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18638_cast_fp16 = add(x = dense_output_1171_cast_fp16, y = sparse_output_1171_cast_fp16)[name = string("op_18638_cast_fp16")];
            tensor<int32, [4]> var_18639 = const()[name = string("op_18639"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18641 = const()[name = string("op_18641"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18640_cast_fp16 = transpose(perm = var_18639, x = var_18638_cast_fp16)[name = string("transpose_389")];
            tensor<fp16, [1, 51, 128]> k_head_465_cast_fp16 = reshape(shape = var_18641, x = var_18640_cast_fp16)[name = string("k_head_465_cast_fp16")];
            string dense_output_1173_pad_type_1 = const()[name = string("dense_output_1173_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1173_strides_1 = const()[name = string("dense_output_1173_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1173_pad_1 = const()[name = string("dense_output_1173_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1173_dilations_1 = const()[name = string("dense_output_1173_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1173_groups_1 = const()[name = string("dense_output_1173_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427926656))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428057792))))[name = string("layers_14_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1173_cast_fp16 = conv(dilations = dense_output_1173_dilations_1, groups = dense_output_1173_groups_1, pad = dense_output_1173_pad_1, pad_type = dense_output_1173_pad_type_1, strides = dense_output_1173_strides_1, weight = layers_14_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1173_cast_fp16")];
            string sparse_output_1173_pad_type_1 = const()[name = string("sparse_output_1173_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1173_strides_1 = const()[name = string("sparse_output_1173_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1173_pad_1 = const()[name = string("sparse_output_1173_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1173_dilations_1 = const()[name = string("sparse_output_1173_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1173_groups_1 = const()[name = string("sparse_output_1173_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428061056))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428058368))))[name = string("layers_14_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1173_cast_fp16 = conv(dilations = sparse_output_1173_dilations_1, groups = sparse_output_1173_groups_1, pad = sparse_output_1173_pad_1, pad_type = sparse_output_1173_pad_type_1, strides = sparse_output_1173_strides_1, weight = layers_14_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1173_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18657_cast_fp16 = add(x = dense_output_1173_cast_fp16, y = sparse_output_1173_cast_fp16)[name = string("op_18657_cast_fp16")];
            tensor<int32, [4]> var_18658 = const()[name = string("op_18658"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18660 = const()[name = string("op_18660"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18659_cast_fp16 = transpose(perm = var_18658, x = var_18657_cast_fp16)[name = string("transpose_388")];
            tensor<fp16, [1, 51, 128]> v_head_465_cast_fp16 = reshape(shape = var_18660, x = var_18659_cast_fp16)[name = string("v_head_465_cast_fp16")];
            string dense_output_1175_pad_type_1 = const()[name = string("dense_output_1175_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1175_strides_1 = const()[name = string("dense_output_1175_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1175_pad_1 = const()[name = string("dense_output_1175_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1175_dilations_1 = const()[name = string("dense_output_1175_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1175_groups_1 = const()[name = string("dense_output_1175_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428077504))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428208640))))[name = string("layers_14_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1175_cast_fp16 = conv(dilations = dense_output_1175_dilations_1, groups = dense_output_1175_groups_1, pad = dense_output_1175_pad_1, pad_type = dense_output_1175_pad_type_1, strides = dense_output_1175_strides_1, weight = layers_14_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1175_cast_fp16")];
            string sparse_output_1175_pad_type_1 = const()[name = string("sparse_output_1175_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1175_strides_1 = const()[name = string("sparse_output_1175_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1175_pad_1 = const()[name = string("sparse_output_1175_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1175_dilations_1 = const()[name = string("sparse_output_1175_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1175_groups_1 = const()[name = string("sparse_output_1175_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428211904))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428209216))))[name = string("layers_14_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1175_cast_fp16 = conv(dilations = sparse_output_1175_dilations_1, groups = sparse_output_1175_groups_1, pad = sparse_output_1175_pad_1, pad_type = sparse_output_1175_pad_type_1, strides = sparse_output_1175_strides_1, weight = layers_14_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1175_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_18676_cast_fp16 = add(x = dense_output_1175_cast_fp16, y = sparse_output_1175_cast_fp16)[name = string("op_18676_cast_fp16")];
            tensor<int32, [4]> var_18677 = const()[name = string("op_18677"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18679 = const()[name = string("op_18679"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_18678_cast_fp16 = transpose(perm = var_18677, x = var_18676_cast_fp16)[name = string("transpose_387")];
            tensor<fp16, [1, 101, 128]> p_head_465_cast_fp16 = reshape(shape = var_18679, x = var_18678_cast_fp16)[name = string("p_head_465_cast_fp16")];
            tensor<fp16, [128]> var_18681_to_fp16 = const()[name = string("op_18681_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428228352)))];
            tensor<fp16, [1, 51, 128]> var_18682_cast_fp16 = add(x = q_head_233_cast_fp16, y = var_18681_to_fp16)[name = string("op_18682_cast_fp16")];
            tensor<int32, [1]> q_u_233_axes_1 = const()[name = string("q_u_233_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_233_cast_fp16 = expand_dims(axes = q_u_233_axes_1, x = var_18682_cast_fp16)[name = string("q_u_233_cast_fp16")];
            tensor<fp16, [128]> var_18684_to_fp16 = const()[name = string("op_18684_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428228672)))];
            tensor<fp16, [1, 51, 128]> var_18685_cast_fp16 = add(x = q_head_233_cast_fp16, y = var_18684_to_fp16)[name = string("op_18685_cast_fp16")];
            tensor<int32, [1]> q_v_233_axes_1 = const()[name = string("q_v_233_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_233_cast_fp16 = expand_dims(axes = q_v_233_axes_1, x = var_18685_cast_fp16)[name = string("q_v_233_cast_fp16")];
            tensor<int32, [1]> k_head_467_axes_1 = const()[name = string("k_head_467_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_467_cast_fp16 = expand_dims(axes = k_head_467_axes_1, x = k_head_465_cast_fp16)[name = string("k_head_467_cast_fp16")];
            tensor<int32, [1]> v_head_467_axes_1 = const()[name = string("v_head_467_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_467_cast_fp16 = expand_dims(axes = v_head_467_axes_1, x = v_head_465_cast_fp16)[name = string("v_head_467_cast_fp16")];
            tensor<int32, [1]> p_head_467_axes_1 = const()[name = string("p_head_467_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_467_cast_fp16 = expand_dims(axes = p_head_467_axes_1, x = p_head_465_cast_fp16)[name = string("p_head_467_cast_fp16")];
            bool var_18691_transpose_x_3 = const()[name = string("op_18691_transpose_x_3"), val = bool(false)];
            bool var_18691_transpose_y_3 = const()[name = string("op_18691_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_18691_cast_fp16 = matmul(transpose_x = var_18691_transpose_x_3, transpose_y = var_18691_transpose_y_3, x = q_u_233_cast_fp16, y = k_head_467_cast_fp16)[name = string("op_18691_cast_fp16")];
            fp16 var_18692_to_fp16 = const()[name = string("op_18692_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_233_cast_fp16 = mul(x = var_18691_cast_fp16, y = var_18692_to_fp16)[name = string("scores_content_233_cast_fp16")];
            bool x_1229_transpose_x_3 = const()[name = string("x_1229_transpose_x_3"), val = bool(false)];
            bool x_1229_transpose_y_3 = const()[name = string("x_1229_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1229_cast_fp16 = matmul(transpose_x = x_1229_transpose_x_3, transpose_y = x_1229_transpose_y_3, x = q_v_233_cast_fp16, y = p_head_467_cast_fp16)[name = string("x_1229_cast_fp16")];
            tensor<int32, [8]> x_1231_pad_1 = const()[name = string("x_1231_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1231_mode_1 = const()[name = string("x_1231_mode_1"), val = string("constant")];
            fp16 const_2089_to_fp16 = const()[name = string("const_2089_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1231_cast_fp16 = pad(constant_val = const_2089_to_fp16, mode = x_1231_mode_1, pad = x_1231_pad_1, x = x_1229_cast_fp16)[name = string("x_1231_cast_fp16")];
            tensor<int32, [4]> var_18706 = const()[name = string("op_18706"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1233_cast_fp16 = reshape(shape = var_18706, x = x_1231_cast_fp16)[name = string("x_1233_cast_fp16")];
            tensor<int32, [4]> var_18710_begin_1 = const()[name = string("op_18710_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_18710_end_1 = const()[name = string("op_18710_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_18710_end_mask_1 = const()[name = string("op_18710_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_18710_cast_fp16 = slice_by_index(begin = var_18710_begin_1, end = var_18710_end_1, end_mask = var_18710_end_mask_1, x = x_1233_cast_fp16)[name = string("op_18710_cast_fp16")];
            tensor<int32, [4]> var_18712 = const()[name = string("op_18712"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_18713_cast_fp16 = reshape(shape = var_18712, x = var_18710_cast_fp16)[name = string("op_18713_cast_fp16")];
            tensor<int32, [4]> var_18718_begin_1 = const()[name = string("op_18718_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_18718_end_1 = const()[name = string("op_18718_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_18718_end_mask_1 = const()[name = string("op_18718_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_18718_cast_fp16 = slice_by_index(begin = var_18718_begin_1, end = var_18718_end_1, end_mask = var_18718_end_mask_1, x = var_18713_cast_fp16)[name = string("op_18718_cast_fp16")];
            fp16 var_18719_to_fp16 = const()[name = string("op_18719_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_233_cast_fp16 = mul(x = var_18718_cast_fp16, y = var_18719_to_fp16)[name = string("scores_pos_233_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_233_cast_fp16 = add(x = scores_content_233_cast_fp16, y = scores_pos_233_cast_fp16)[name = string("logits_233_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_18722_cast_fp16 = softmax(axis = var_17963, x = logits_233_cast_fp16)[name = string("op_18722_cast_fp16")];
            bool var_18724_transpose_x_1 = const()[name = string("op_18724_transpose_x_1"), val = bool(false)];
            bool var_18724_transpose_y_1 = const()[name = string("op_18724_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_18724_cast_fp16 = matmul(transpose_x = var_18724_transpose_x_1, transpose_y = var_18724_transpose_y_1, x = var_18722_cast_fp16, y = v_head_467_cast_fp16)[name = string("op_18724_cast_fp16")];
            tensor<int32, [1]> var_18725_axes_1 = const()[name = string("op_18725_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_18725_cast_fp16 = squeeze(axes = var_18725_axes_1, x = var_18724_cast_fp16)[name = string("op_18725_cast_fp16")];
            string dense_output_1177_pad_type_1 = const()[name = string("dense_output_1177_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1177_strides_1 = const()[name = string("dense_output_1177_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1177_pad_1 = const()[name = string("dense_output_1177_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1177_dilations_1 = const()[name = string("dense_output_1177_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1177_groups_1 = const()[name = string("dense_output_1177_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428228992))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428360128))))[name = string("layers_14_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1177_cast_fp16 = conv(dilations = dense_output_1177_dilations_1, groups = dense_output_1177_groups_1, pad = dense_output_1177_pad_1, pad_type = dense_output_1177_pad_type_1, strides = dense_output_1177_strides_1, weight = layers_14_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1177_cast_fp16")];
            string sparse_output_1177_pad_type_1 = const()[name = string("sparse_output_1177_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1177_strides_1 = const()[name = string("sparse_output_1177_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1177_pad_1 = const()[name = string("sparse_output_1177_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1177_dilations_1 = const()[name = string("sparse_output_1177_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1177_groups_1 = const()[name = string("sparse_output_1177_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428363392))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428360704))))[name = string("layers_14_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1177_cast_fp16 = conv(dilations = sparse_output_1177_dilations_1, groups = sparse_output_1177_groups_1, pad = sparse_output_1177_pad_1, pad_type = sparse_output_1177_pad_type_1, strides = sparse_output_1177_strides_1, weight = layers_14_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1177_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18740_cast_fp16 = add(x = dense_output_1177_cast_fp16, y = sparse_output_1177_cast_fp16)[name = string("op_18740_cast_fp16")];
            tensor<int32, [4]> var_18741 = const()[name = string("op_18741"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18743 = const()[name = string("op_18743"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18742_cast_fp16 = transpose(perm = var_18741, x = var_18740_cast_fp16)[name = string("transpose_386")];
            tensor<fp16, [1, 51, 128]> q_head_235_cast_fp16 = reshape(shape = var_18743, x = var_18742_cast_fp16)[name = string("q_head_235_cast_fp16")];
            string dense_output_1179_pad_type_1 = const()[name = string("dense_output_1179_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1179_strides_1 = const()[name = string("dense_output_1179_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1179_pad_1 = const()[name = string("dense_output_1179_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1179_dilations_1 = const()[name = string("dense_output_1179_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1179_groups_1 = const()[name = string("dense_output_1179_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428379840))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428510976))))[name = string("layers_14_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1179_cast_fp16 = conv(dilations = dense_output_1179_dilations_1, groups = dense_output_1179_groups_1, pad = dense_output_1179_pad_1, pad_type = dense_output_1179_pad_type_1, strides = dense_output_1179_strides_1, weight = layers_14_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1179_cast_fp16")];
            string sparse_output_1179_pad_type_1 = const()[name = string("sparse_output_1179_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1179_strides_1 = const()[name = string("sparse_output_1179_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1179_pad_1 = const()[name = string("sparse_output_1179_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1179_dilations_1 = const()[name = string("sparse_output_1179_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1179_groups_1 = const()[name = string("sparse_output_1179_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428514240))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428511552))))[name = string("layers_14_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1179_cast_fp16 = conv(dilations = sparse_output_1179_dilations_1, groups = sparse_output_1179_groups_1, pad = sparse_output_1179_pad_1, pad_type = sparse_output_1179_pad_type_1, strides = sparse_output_1179_strides_1, weight = layers_14_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1179_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18759_cast_fp16 = add(x = dense_output_1179_cast_fp16, y = sparse_output_1179_cast_fp16)[name = string("op_18759_cast_fp16")];
            tensor<int32, [4]> var_18760 = const()[name = string("op_18760"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18762 = const()[name = string("op_18762"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18761_cast_fp16 = transpose(perm = var_18760, x = var_18759_cast_fp16)[name = string("transpose_385")];
            tensor<fp16, [1, 51, 128]> k_head_469_cast_fp16 = reshape(shape = var_18762, x = var_18761_cast_fp16)[name = string("k_head_469_cast_fp16")];
            string dense_output_1181_pad_type_1 = const()[name = string("dense_output_1181_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1181_strides_1 = const()[name = string("dense_output_1181_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1181_pad_1 = const()[name = string("dense_output_1181_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1181_dilations_1 = const()[name = string("dense_output_1181_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1181_groups_1 = const()[name = string("dense_output_1181_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428530688))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428661824))))[name = string("layers_14_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1181_cast_fp16 = conv(dilations = dense_output_1181_dilations_1, groups = dense_output_1181_groups_1, pad = dense_output_1181_pad_1, pad_type = dense_output_1181_pad_type_1, strides = dense_output_1181_strides_1, weight = layers_14_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1181_cast_fp16")];
            string sparse_output_1181_pad_type_1 = const()[name = string("sparse_output_1181_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1181_strides_1 = const()[name = string("sparse_output_1181_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1181_pad_1 = const()[name = string("sparse_output_1181_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1181_dilations_1 = const()[name = string("sparse_output_1181_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1181_groups_1 = const()[name = string("sparse_output_1181_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428665088))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428662400))))[name = string("layers_14_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1181_cast_fp16 = conv(dilations = sparse_output_1181_dilations_1, groups = sparse_output_1181_groups_1, pad = sparse_output_1181_pad_1, pad_type = sparse_output_1181_pad_type_1, strides = sparse_output_1181_strides_1, weight = layers_14_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1181_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18778_cast_fp16 = add(x = dense_output_1181_cast_fp16, y = sparse_output_1181_cast_fp16)[name = string("op_18778_cast_fp16")];
            tensor<int32, [4]> var_18779 = const()[name = string("op_18779"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18781 = const()[name = string("op_18781"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18780_cast_fp16 = transpose(perm = var_18779, x = var_18778_cast_fp16)[name = string("transpose_384")];
            tensor<fp16, [1, 51, 128]> v_head_469_cast_fp16 = reshape(shape = var_18781, x = var_18780_cast_fp16)[name = string("v_head_469_cast_fp16")];
            string dense_output_1183_pad_type_1 = const()[name = string("dense_output_1183_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1183_strides_1 = const()[name = string("dense_output_1183_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1183_pad_1 = const()[name = string("dense_output_1183_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1183_dilations_1 = const()[name = string("dense_output_1183_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1183_groups_1 = const()[name = string("dense_output_1183_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428681536))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428812672))))[name = string("layers_14_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1183_cast_fp16 = conv(dilations = dense_output_1183_dilations_1, groups = dense_output_1183_groups_1, pad = dense_output_1183_pad_1, pad_type = dense_output_1183_pad_type_1, strides = dense_output_1183_strides_1, weight = layers_14_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1183_cast_fp16")];
            string sparse_output_1183_pad_type_1 = const()[name = string("sparse_output_1183_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1183_strides_1 = const()[name = string("sparse_output_1183_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1183_pad_1 = const()[name = string("sparse_output_1183_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1183_dilations_1 = const()[name = string("sparse_output_1183_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1183_groups_1 = const()[name = string("sparse_output_1183_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428815936))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428813248))))[name = string("layers_14_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1183_cast_fp16 = conv(dilations = sparse_output_1183_dilations_1, groups = sparse_output_1183_groups_1, pad = sparse_output_1183_pad_1, pad_type = sparse_output_1183_pad_type_1, strides = sparse_output_1183_strides_1, weight = layers_14_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1183_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_18797_cast_fp16 = add(x = dense_output_1183_cast_fp16, y = sparse_output_1183_cast_fp16)[name = string("op_18797_cast_fp16")];
            tensor<int32, [4]> var_18798 = const()[name = string("op_18798"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18800 = const()[name = string("op_18800"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_18799_cast_fp16 = transpose(perm = var_18798, x = var_18797_cast_fp16)[name = string("transpose_383")];
            tensor<fp16, [1, 101, 128]> p_head_469_cast_fp16 = reshape(shape = var_18800, x = var_18799_cast_fp16)[name = string("p_head_469_cast_fp16")];
            tensor<fp16, [128]> var_18802_to_fp16 = const()[name = string("op_18802_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428832384)))];
            tensor<fp16, [1, 51, 128]> var_18803_cast_fp16 = add(x = q_head_235_cast_fp16, y = var_18802_to_fp16)[name = string("op_18803_cast_fp16")];
            tensor<int32, [1]> q_u_235_axes_1 = const()[name = string("q_u_235_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_235_cast_fp16 = expand_dims(axes = q_u_235_axes_1, x = var_18803_cast_fp16)[name = string("q_u_235_cast_fp16")];
            tensor<fp16, [128]> var_18805_to_fp16 = const()[name = string("op_18805_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428832704)))];
            tensor<fp16, [1, 51, 128]> var_18806_cast_fp16 = add(x = q_head_235_cast_fp16, y = var_18805_to_fp16)[name = string("op_18806_cast_fp16")];
            tensor<int32, [1]> q_v_235_axes_1 = const()[name = string("q_v_235_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_235_cast_fp16 = expand_dims(axes = q_v_235_axes_1, x = var_18806_cast_fp16)[name = string("q_v_235_cast_fp16")];
            tensor<int32, [1]> k_head_471_axes_1 = const()[name = string("k_head_471_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_471_cast_fp16 = expand_dims(axes = k_head_471_axes_1, x = k_head_469_cast_fp16)[name = string("k_head_471_cast_fp16")];
            tensor<int32, [1]> v_head_471_axes_1 = const()[name = string("v_head_471_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_471_cast_fp16 = expand_dims(axes = v_head_471_axes_1, x = v_head_469_cast_fp16)[name = string("v_head_471_cast_fp16")];
            tensor<int32, [1]> p_head_471_axes_1 = const()[name = string("p_head_471_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_471_cast_fp16 = expand_dims(axes = p_head_471_axes_1, x = p_head_469_cast_fp16)[name = string("p_head_471_cast_fp16")];
            bool var_18812_transpose_x_3 = const()[name = string("op_18812_transpose_x_3"), val = bool(false)];
            bool var_18812_transpose_y_3 = const()[name = string("op_18812_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_18812_cast_fp16 = matmul(transpose_x = var_18812_transpose_x_3, transpose_y = var_18812_transpose_y_3, x = q_u_235_cast_fp16, y = k_head_471_cast_fp16)[name = string("op_18812_cast_fp16")];
            fp16 var_18813_to_fp16 = const()[name = string("op_18813_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_235_cast_fp16 = mul(x = var_18812_cast_fp16, y = var_18813_to_fp16)[name = string("scores_content_235_cast_fp16")];
            bool x_1237_transpose_x_3 = const()[name = string("x_1237_transpose_x_3"), val = bool(false)];
            bool x_1237_transpose_y_3 = const()[name = string("x_1237_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1237_cast_fp16 = matmul(transpose_x = x_1237_transpose_x_3, transpose_y = x_1237_transpose_y_3, x = q_v_235_cast_fp16, y = p_head_471_cast_fp16)[name = string("x_1237_cast_fp16")];
            tensor<int32, [8]> x_1239_pad_1 = const()[name = string("x_1239_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1239_mode_1 = const()[name = string("x_1239_mode_1"), val = string("constant")];
            fp16 const_2095_to_fp16 = const()[name = string("const_2095_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1239_cast_fp16 = pad(constant_val = const_2095_to_fp16, mode = x_1239_mode_1, pad = x_1239_pad_1, x = x_1237_cast_fp16)[name = string("x_1239_cast_fp16")];
            tensor<int32, [4]> var_18827 = const()[name = string("op_18827"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1241_cast_fp16 = reshape(shape = var_18827, x = x_1239_cast_fp16)[name = string("x_1241_cast_fp16")];
            tensor<int32, [4]> var_18831_begin_1 = const()[name = string("op_18831_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_18831_end_1 = const()[name = string("op_18831_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_18831_end_mask_1 = const()[name = string("op_18831_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_18831_cast_fp16 = slice_by_index(begin = var_18831_begin_1, end = var_18831_end_1, end_mask = var_18831_end_mask_1, x = x_1241_cast_fp16)[name = string("op_18831_cast_fp16")];
            tensor<int32, [4]> var_18833 = const()[name = string("op_18833"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_18834_cast_fp16 = reshape(shape = var_18833, x = var_18831_cast_fp16)[name = string("op_18834_cast_fp16")];
            tensor<int32, [4]> var_18839_begin_1 = const()[name = string("op_18839_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_18839_end_1 = const()[name = string("op_18839_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_18839_end_mask_1 = const()[name = string("op_18839_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_18839_cast_fp16 = slice_by_index(begin = var_18839_begin_1, end = var_18839_end_1, end_mask = var_18839_end_mask_1, x = var_18834_cast_fp16)[name = string("op_18839_cast_fp16")];
            fp16 var_18840_to_fp16 = const()[name = string("op_18840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_235_cast_fp16 = mul(x = var_18839_cast_fp16, y = var_18840_to_fp16)[name = string("scores_pos_235_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_235_cast_fp16 = add(x = scores_content_235_cast_fp16, y = scores_pos_235_cast_fp16)[name = string("logits_235_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_18843_cast_fp16 = softmax(axis = var_17963, x = logits_235_cast_fp16)[name = string("op_18843_cast_fp16")];
            bool var_18845_transpose_x_1 = const()[name = string("op_18845_transpose_x_1"), val = bool(false)];
            bool var_18845_transpose_y_1 = const()[name = string("op_18845_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_18845_cast_fp16 = matmul(transpose_x = var_18845_transpose_x_1, transpose_y = var_18845_transpose_y_1, x = var_18843_cast_fp16, y = v_head_471_cast_fp16)[name = string("op_18845_cast_fp16")];
            tensor<int32, [1]> var_18846_axes_1 = const()[name = string("op_18846_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_18846_cast_fp16 = squeeze(axes = var_18846_axes_1, x = var_18845_cast_fp16)[name = string("op_18846_cast_fp16")];
            string dense_output_1185_pad_type_1 = const()[name = string("dense_output_1185_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1185_strides_1 = const()[name = string("dense_output_1185_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1185_pad_1 = const()[name = string("dense_output_1185_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1185_dilations_1 = const()[name = string("dense_output_1185_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1185_groups_1 = const()[name = string("dense_output_1185_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428833024))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428964160))))[name = string("layers_14_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1185_cast_fp16 = conv(dilations = dense_output_1185_dilations_1, groups = dense_output_1185_groups_1, pad = dense_output_1185_pad_1, pad_type = dense_output_1185_pad_type_1, strides = dense_output_1185_strides_1, weight = layers_14_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1185_cast_fp16")];
            string sparse_output_1185_pad_type_1 = const()[name = string("sparse_output_1185_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1185_strides_1 = const()[name = string("sparse_output_1185_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1185_pad_1 = const()[name = string("sparse_output_1185_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1185_dilations_1 = const()[name = string("sparse_output_1185_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1185_groups_1 = const()[name = string("sparse_output_1185_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428967424))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428964736))))[name = string("layers_14_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1185_cast_fp16 = conv(dilations = sparse_output_1185_dilations_1, groups = sparse_output_1185_groups_1, pad = sparse_output_1185_pad_1, pad_type = sparse_output_1185_pad_type_1, strides = sparse_output_1185_strides_1, weight = layers_14_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1185_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18861_cast_fp16 = add(x = dense_output_1185_cast_fp16, y = sparse_output_1185_cast_fp16)[name = string("op_18861_cast_fp16")];
            tensor<int32, [4]> var_18862 = const()[name = string("op_18862"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18864 = const()[name = string("op_18864"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18863_cast_fp16 = transpose(perm = var_18862, x = var_18861_cast_fp16)[name = string("transpose_382")];
            tensor<fp16, [1, 51, 128]> q_head_237_cast_fp16 = reshape(shape = var_18864, x = var_18863_cast_fp16)[name = string("q_head_237_cast_fp16")];
            string dense_output_1187_pad_type_1 = const()[name = string("dense_output_1187_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1187_strides_1 = const()[name = string("dense_output_1187_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1187_pad_1 = const()[name = string("dense_output_1187_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1187_dilations_1 = const()[name = string("dense_output_1187_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1187_groups_1 = const()[name = string("dense_output_1187_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428983872))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429115008))))[name = string("layers_14_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1187_cast_fp16 = conv(dilations = dense_output_1187_dilations_1, groups = dense_output_1187_groups_1, pad = dense_output_1187_pad_1, pad_type = dense_output_1187_pad_type_1, strides = dense_output_1187_strides_1, weight = layers_14_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1187_cast_fp16")];
            string sparse_output_1187_pad_type_1 = const()[name = string("sparse_output_1187_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1187_strides_1 = const()[name = string("sparse_output_1187_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1187_pad_1 = const()[name = string("sparse_output_1187_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1187_dilations_1 = const()[name = string("sparse_output_1187_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1187_groups_1 = const()[name = string("sparse_output_1187_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429118272))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429115584))))[name = string("layers_14_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1187_cast_fp16 = conv(dilations = sparse_output_1187_dilations_1, groups = sparse_output_1187_groups_1, pad = sparse_output_1187_pad_1, pad_type = sparse_output_1187_pad_type_1, strides = sparse_output_1187_strides_1, weight = layers_14_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1187_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18880_cast_fp16 = add(x = dense_output_1187_cast_fp16, y = sparse_output_1187_cast_fp16)[name = string("op_18880_cast_fp16")];
            tensor<int32, [4]> var_18881 = const()[name = string("op_18881"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18883 = const()[name = string("op_18883"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18882_cast_fp16 = transpose(perm = var_18881, x = var_18880_cast_fp16)[name = string("transpose_381")];
            tensor<fp16, [1, 51, 128]> k_head_473_cast_fp16 = reshape(shape = var_18883, x = var_18882_cast_fp16)[name = string("k_head_473_cast_fp16")];
            string dense_output_1189_pad_type_1 = const()[name = string("dense_output_1189_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1189_strides_1 = const()[name = string("dense_output_1189_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1189_pad_1 = const()[name = string("dense_output_1189_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1189_dilations_1 = const()[name = string("dense_output_1189_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1189_groups_1 = const()[name = string("dense_output_1189_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429134720))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429265856))))[name = string("layers_14_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1189_cast_fp16 = conv(dilations = dense_output_1189_dilations_1, groups = dense_output_1189_groups_1, pad = dense_output_1189_pad_1, pad_type = dense_output_1189_pad_type_1, strides = dense_output_1189_strides_1, weight = layers_14_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1189_cast_fp16")];
            string sparse_output_1189_pad_type_1 = const()[name = string("sparse_output_1189_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1189_strides_1 = const()[name = string("sparse_output_1189_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1189_pad_1 = const()[name = string("sparse_output_1189_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1189_dilations_1 = const()[name = string("sparse_output_1189_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1189_groups_1 = const()[name = string("sparse_output_1189_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429269120))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429266432))))[name = string("layers_14_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1189_cast_fp16 = conv(dilations = sparse_output_1189_dilations_1, groups = sparse_output_1189_groups_1, pad = sparse_output_1189_pad_1, pad_type = sparse_output_1189_pad_type_1, strides = sparse_output_1189_strides_1, weight = layers_14_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1189_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18899_cast_fp16 = add(x = dense_output_1189_cast_fp16, y = sparse_output_1189_cast_fp16)[name = string("op_18899_cast_fp16")];
            tensor<int32, [4]> var_18900 = const()[name = string("op_18900"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18902 = const()[name = string("op_18902"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18901_cast_fp16 = transpose(perm = var_18900, x = var_18899_cast_fp16)[name = string("transpose_380")];
            tensor<fp16, [1, 51, 128]> v_head_473_cast_fp16 = reshape(shape = var_18902, x = var_18901_cast_fp16)[name = string("v_head_473_cast_fp16")];
            string dense_output_1191_pad_type_1 = const()[name = string("dense_output_1191_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1191_strides_1 = const()[name = string("dense_output_1191_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1191_pad_1 = const()[name = string("dense_output_1191_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1191_dilations_1 = const()[name = string("dense_output_1191_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1191_groups_1 = const()[name = string("dense_output_1191_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429285568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429416704))))[name = string("layers_14_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1191_cast_fp16 = conv(dilations = dense_output_1191_dilations_1, groups = dense_output_1191_groups_1, pad = dense_output_1191_pad_1, pad_type = dense_output_1191_pad_type_1, strides = dense_output_1191_strides_1, weight = layers_14_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1191_cast_fp16")];
            string sparse_output_1191_pad_type_1 = const()[name = string("sparse_output_1191_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1191_strides_1 = const()[name = string("sparse_output_1191_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1191_pad_1 = const()[name = string("sparse_output_1191_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1191_dilations_1 = const()[name = string("sparse_output_1191_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1191_groups_1 = const()[name = string("sparse_output_1191_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429419968))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429417280))))[name = string("layers_14_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1191_cast_fp16 = conv(dilations = sparse_output_1191_dilations_1, groups = sparse_output_1191_groups_1, pad = sparse_output_1191_pad_1, pad_type = sparse_output_1191_pad_type_1, strides = sparse_output_1191_strides_1, weight = layers_14_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1191_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_18918_cast_fp16 = add(x = dense_output_1191_cast_fp16, y = sparse_output_1191_cast_fp16)[name = string("op_18918_cast_fp16")];
            tensor<int32, [4]> var_18919 = const()[name = string("op_18919"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18921 = const()[name = string("op_18921"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_18920_cast_fp16 = transpose(perm = var_18919, x = var_18918_cast_fp16)[name = string("transpose_379")];
            tensor<fp16, [1, 101, 128]> p_head_473_cast_fp16 = reshape(shape = var_18921, x = var_18920_cast_fp16)[name = string("p_head_473_cast_fp16")];
            tensor<fp16, [128]> var_18923_to_fp16 = const()[name = string("op_18923_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429436416)))];
            tensor<fp16, [1, 51, 128]> var_18924_cast_fp16 = add(x = q_head_237_cast_fp16, y = var_18923_to_fp16)[name = string("op_18924_cast_fp16")];
            tensor<int32, [1]> q_u_237_axes_1 = const()[name = string("q_u_237_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_237_cast_fp16 = expand_dims(axes = q_u_237_axes_1, x = var_18924_cast_fp16)[name = string("q_u_237_cast_fp16")];
            tensor<fp16, [128]> var_18926_to_fp16 = const()[name = string("op_18926_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429436736)))];
            tensor<fp16, [1, 51, 128]> var_18927_cast_fp16 = add(x = q_head_237_cast_fp16, y = var_18926_to_fp16)[name = string("op_18927_cast_fp16")];
            tensor<int32, [1]> q_v_237_axes_1 = const()[name = string("q_v_237_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_237_cast_fp16 = expand_dims(axes = q_v_237_axes_1, x = var_18927_cast_fp16)[name = string("q_v_237_cast_fp16")];
            tensor<int32, [1]> k_head_475_axes_1 = const()[name = string("k_head_475_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_475_cast_fp16 = expand_dims(axes = k_head_475_axes_1, x = k_head_473_cast_fp16)[name = string("k_head_475_cast_fp16")];
            tensor<int32, [1]> v_head_475_axes_1 = const()[name = string("v_head_475_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_475_cast_fp16 = expand_dims(axes = v_head_475_axes_1, x = v_head_473_cast_fp16)[name = string("v_head_475_cast_fp16")];
            tensor<int32, [1]> p_head_475_axes_1 = const()[name = string("p_head_475_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_475_cast_fp16 = expand_dims(axes = p_head_475_axes_1, x = p_head_473_cast_fp16)[name = string("p_head_475_cast_fp16")];
            bool var_18933_transpose_x_3 = const()[name = string("op_18933_transpose_x_3"), val = bool(false)];
            bool var_18933_transpose_y_3 = const()[name = string("op_18933_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_18933_cast_fp16 = matmul(transpose_x = var_18933_transpose_x_3, transpose_y = var_18933_transpose_y_3, x = q_u_237_cast_fp16, y = k_head_475_cast_fp16)[name = string("op_18933_cast_fp16")];
            fp16 var_18934_to_fp16 = const()[name = string("op_18934_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_237_cast_fp16 = mul(x = var_18933_cast_fp16, y = var_18934_to_fp16)[name = string("scores_content_237_cast_fp16")];
            bool x_1245_transpose_x_3 = const()[name = string("x_1245_transpose_x_3"), val = bool(false)];
            bool x_1245_transpose_y_3 = const()[name = string("x_1245_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1245_cast_fp16 = matmul(transpose_x = x_1245_transpose_x_3, transpose_y = x_1245_transpose_y_3, x = q_v_237_cast_fp16, y = p_head_475_cast_fp16)[name = string("x_1245_cast_fp16")];
            tensor<int32, [8]> x_1247_pad_1 = const()[name = string("x_1247_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1247_mode_1 = const()[name = string("x_1247_mode_1"), val = string("constant")];
            fp16 const_2101_to_fp16 = const()[name = string("const_2101_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1247_cast_fp16 = pad(constant_val = const_2101_to_fp16, mode = x_1247_mode_1, pad = x_1247_pad_1, x = x_1245_cast_fp16)[name = string("x_1247_cast_fp16")];
            tensor<int32, [4]> var_18948 = const()[name = string("op_18948"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1249_cast_fp16 = reshape(shape = var_18948, x = x_1247_cast_fp16)[name = string("x_1249_cast_fp16")];
            tensor<int32, [4]> var_18952_begin_1 = const()[name = string("op_18952_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_18952_end_1 = const()[name = string("op_18952_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_18952_end_mask_1 = const()[name = string("op_18952_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_18952_cast_fp16 = slice_by_index(begin = var_18952_begin_1, end = var_18952_end_1, end_mask = var_18952_end_mask_1, x = x_1249_cast_fp16)[name = string("op_18952_cast_fp16")];
            tensor<int32, [4]> var_18954 = const()[name = string("op_18954"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_18955_cast_fp16 = reshape(shape = var_18954, x = var_18952_cast_fp16)[name = string("op_18955_cast_fp16")];
            tensor<int32, [4]> var_18960_begin_1 = const()[name = string("op_18960_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_18960_end_1 = const()[name = string("op_18960_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_18960_end_mask_1 = const()[name = string("op_18960_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_18960_cast_fp16 = slice_by_index(begin = var_18960_begin_1, end = var_18960_end_1, end_mask = var_18960_end_mask_1, x = var_18955_cast_fp16)[name = string("op_18960_cast_fp16")];
            fp16 var_18961_to_fp16 = const()[name = string("op_18961_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_237_cast_fp16 = mul(x = var_18960_cast_fp16, y = var_18961_to_fp16)[name = string("scores_pos_237_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_237_cast_fp16 = add(x = scores_content_237_cast_fp16, y = scores_pos_237_cast_fp16)[name = string("logits_237_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_18964_cast_fp16 = softmax(axis = var_17963, x = logits_237_cast_fp16)[name = string("op_18964_cast_fp16")];
            bool var_18966_transpose_x_1 = const()[name = string("op_18966_transpose_x_1"), val = bool(false)];
            bool var_18966_transpose_y_1 = const()[name = string("op_18966_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_18966_cast_fp16 = matmul(transpose_x = var_18966_transpose_x_1, transpose_y = var_18966_transpose_y_1, x = var_18964_cast_fp16, y = v_head_475_cast_fp16)[name = string("op_18966_cast_fp16")];
            tensor<int32, [1]> var_18967_axes_1 = const()[name = string("op_18967_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_18967_cast_fp16 = squeeze(axes = var_18967_axes_1, x = var_18966_cast_fp16)[name = string("op_18967_cast_fp16")];
            string dense_output_1193_pad_type_1 = const()[name = string("dense_output_1193_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1193_strides_1 = const()[name = string("dense_output_1193_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1193_pad_1 = const()[name = string("dense_output_1193_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1193_dilations_1 = const()[name = string("dense_output_1193_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1193_groups_1 = const()[name = string("dense_output_1193_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429437056))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429568192))))[name = string("layers_14_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1193_cast_fp16 = conv(dilations = dense_output_1193_dilations_1, groups = dense_output_1193_groups_1, pad = dense_output_1193_pad_1, pad_type = dense_output_1193_pad_type_1, strides = dense_output_1193_strides_1, weight = layers_14_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1193_cast_fp16")];
            string sparse_output_1193_pad_type_1 = const()[name = string("sparse_output_1193_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1193_strides_1 = const()[name = string("sparse_output_1193_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1193_pad_1 = const()[name = string("sparse_output_1193_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1193_dilations_1 = const()[name = string("sparse_output_1193_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1193_groups_1 = const()[name = string("sparse_output_1193_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429571456))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429568768))))[name = string("layers_14_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1193_cast_fp16 = conv(dilations = sparse_output_1193_dilations_1, groups = sparse_output_1193_groups_1, pad = sparse_output_1193_pad_1, pad_type = sparse_output_1193_pad_type_1, strides = sparse_output_1193_strides_1, weight = layers_14_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1193_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_18982_cast_fp16 = add(x = dense_output_1193_cast_fp16, y = sparse_output_1193_cast_fp16)[name = string("op_18982_cast_fp16")];
            tensor<int32, [4]> var_18983 = const()[name = string("op_18983"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_18985 = const()[name = string("op_18985"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_18984_cast_fp16 = transpose(perm = var_18983, x = var_18982_cast_fp16)[name = string("transpose_378")];
            tensor<fp16, [1, 51, 128]> q_head_239_cast_fp16 = reshape(shape = var_18985, x = var_18984_cast_fp16)[name = string("q_head_239_cast_fp16")];
            string dense_output_1195_pad_type_1 = const()[name = string("dense_output_1195_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1195_strides_1 = const()[name = string("dense_output_1195_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1195_pad_1 = const()[name = string("dense_output_1195_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1195_dilations_1 = const()[name = string("dense_output_1195_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1195_groups_1 = const()[name = string("dense_output_1195_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429587904))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429719040))))[name = string("layers_14_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1195_cast_fp16 = conv(dilations = dense_output_1195_dilations_1, groups = dense_output_1195_groups_1, pad = dense_output_1195_pad_1, pad_type = dense_output_1195_pad_type_1, strides = dense_output_1195_strides_1, weight = layers_14_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1195_cast_fp16")];
            string sparse_output_1195_pad_type_1 = const()[name = string("sparse_output_1195_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1195_strides_1 = const()[name = string("sparse_output_1195_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1195_pad_1 = const()[name = string("sparse_output_1195_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1195_dilations_1 = const()[name = string("sparse_output_1195_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1195_groups_1 = const()[name = string("sparse_output_1195_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429722304))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429719616))))[name = string("layers_14_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1195_cast_fp16 = conv(dilations = sparse_output_1195_dilations_1, groups = sparse_output_1195_groups_1, pad = sparse_output_1195_pad_1, pad_type = sparse_output_1195_pad_type_1, strides = sparse_output_1195_strides_1, weight = layers_14_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1195_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19001_cast_fp16 = add(x = dense_output_1195_cast_fp16, y = sparse_output_1195_cast_fp16)[name = string("op_19001_cast_fp16")];
            tensor<int32, [4]> var_19002 = const()[name = string("op_19002"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19004 = const()[name = string("op_19004"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19003_cast_fp16 = transpose(perm = var_19002, x = var_19001_cast_fp16)[name = string("transpose_377")];
            tensor<fp16, [1, 51, 128]> k_head_477_cast_fp16 = reshape(shape = var_19004, x = var_19003_cast_fp16)[name = string("k_head_477_cast_fp16")];
            string dense_output_1197_pad_type_1 = const()[name = string("dense_output_1197_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1197_strides_1 = const()[name = string("dense_output_1197_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1197_pad_1 = const()[name = string("dense_output_1197_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1197_dilations_1 = const()[name = string("dense_output_1197_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1197_groups_1 = const()[name = string("dense_output_1197_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429738752))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429869888))))[name = string("layers_14_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1197_cast_fp16 = conv(dilations = dense_output_1197_dilations_1, groups = dense_output_1197_groups_1, pad = dense_output_1197_pad_1, pad_type = dense_output_1197_pad_type_1, strides = dense_output_1197_strides_1, weight = layers_14_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_673_cast_fp16)[name = string("dense_output_1197_cast_fp16")];
            string sparse_output_1197_pad_type_1 = const()[name = string("sparse_output_1197_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1197_strides_1 = const()[name = string("sparse_output_1197_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1197_pad_1 = const()[name = string("sparse_output_1197_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1197_dilations_1 = const()[name = string("sparse_output_1197_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1197_groups_1 = const()[name = string("sparse_output_1197_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429873152))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429870464))))[name = string("layers_14_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1197_cast_fp16 = conv(dilations = sparse_output_1197_dilations_1, groups = sparse_output_1197_groups_1, pad = sparse_output_1197_pad_1, pad_type = sparse_output_1197_pad_type_1, strides = sparse_output_1197_strides_1, weight = layers_14_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_673_cast_fp16)[name = string("sparse_output_1197_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19020_cast_fp16 = add(x = dense_output_1197_cast_fp16, y = sparse_output_1197_cast_fp16)[name = string("op_19020_cast_fp16")];
            tensor<int32, [4]> var_19021 = const()[name = string("op_19021"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19023 = const()[name = string("op_19023"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19022_cast_fp16 = transpose(perm = var_19021, x = var_19020_cast_fp16)[name = string("transpose_376")];
            tensor<fp16, [1, 51, 128]> v_head_477_cast_fp16 = reshape(shape = var_19023, x = var_19022_cast_fp16)[name = string("v_head_477_cast_fp16")];
            string dense_output_1199_pad_type_1 = const()[name = string("dense_output_1199_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1199_strides_1 = const()[name = string("dense_output_1199_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1199_pad_1 = const()[name = string("dense_output_1199_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1199_dilations_1 = const()[name = string("dense_output_1199_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1199_groups_1 = const()[name = string("dense_output_1199_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429889600))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430020736))))[name = string("layers_14_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1199_cast_fp16 = conv(dilations = dense_output_1199_dilations_1, groups = dense_output_1199_groups_1, pad = dense_output_1199_pad_1, pad_type = dense_output_1199_pad_type_1, strides = dense_output_1199_strides_1, weight = layers_14_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1199_cast_fp16")];
            string sparse_output_1199_pad_type_1 = const()[name = string("sparse_output_1199_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1199_strides_1 = const()[name = string("sparse_output_1199_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1199_pad_1 = const()[name = string("sparse_output_1199_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1199_dilations_1 = const()[name = string("sparse_output_1199_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1199_groups_1 = const()[name = string("sparse_output_1199_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_14_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430024000))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430021312))))[name = string("layers_14_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1199_cast_fp16 = conv(dilations = sparse_output_1199_dilations_1, groups = sparse_output_1199_groups_1, pad = sparse_output_1199_pad_1, pad_type = sparse_output_1199_pad_type_1, strides = sparse_output_1199_strides_1, weight = layers_14_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1199_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_19039_cast_fp16 = add(x = dense_output_1199_cast_fp16, y = sparse_output_1199_cast_fp16)[name = string("op_19039_cast_fp16")];
            tensor<int32, [4]> var_19040 = const()[name = string("op_19040"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19042 = const()[name = string("op_19042"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_19041_cast_fp16 = transpose(perm = var_19040, x = var_19039_cast_fp16)[name = string("transpose_375")];
            tensor<fp16, [1, 101, 128]> p_head_477_cast_fp16 = reshape(shape = var_19042, x = var_19041_cast_fp16)[name = string("p_head_477_cast_fp16")];
            tensor<fp16, [128]> var_19044_to_fp16 = const()[name = string("op_19044_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430040448)))];
            tensor<fp16, [1, 51, 128]> var_19045_cast_fp16 = add(x = q_head_239_cast_fp16, y = var_19044_to_fp16)[name = string("op_19045_cast_fp16")];
            tensor<int32, [1]> q_u_239_axes_1 = const()[name = string("q_u_239_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_239_cast_fp16 = expand_dims(axes = q_u_239_axes_1, x = var_19045_cast_fp16)[name = string("q_u_239_cast_fp16")];
            tensor<fp16, [128]> var_19047_to_fp16 = const()[name = string("op_19047_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430040768)))];
            tensor<fp16, [1, 51, 128]> var_19048_cast_fp16 = add(x = q_head_239_cast_fp16, y = var_19047_to_fp16)[name = string("op_19048_cast_fp16")];
            tensor<int32, [1]> q_v_239_axes_1 = const()[name = string("q_v_239_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_239_cast_fp16 = expand_dims(axes = q_v_239_axes_1, x = var_19048_cast_fp16)[name = string("q_v_239_cast_fp16")];
            tensor<int32, [1]> k_head_479_axes_1 = const()[name = string("k_head_479_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_479_cast_fp16 = expand_dims(axes = k_head_479_axes_1, x = k_head_477_cast_fp16)[name = string("k_head_479_cast_fp16")];
            tensor<int32, [1]> v_head_479_axes_1 = const()[name = string("v_head_479_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_479_cast_fp16 = expand_dims(axes = v_head_479_axes_1, x = v_head_477_cast_fp16)[name = string("v_head_479_cast_fp16")];
            tensor<int32, [1]> p_head_479_axes_1 = const()[name = string("p_head_479_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_479_cast_fp16 = expand_dims(axes = p_head_479_axes_1, x = p_head_477_cast_fp16)[name = string("p_head_479_cast_fp16")];
            bool var_19054_transpose_x_3 = const()[name = string("op_19054_transpose_x_3"), val = bool(false)];
            bool var_19054_transpose_y_3 = const()[name = string("op_19054_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_19054_cast_fp16 = matmul(transpose_x = var_19054_transpose_x_3, transpose_y = var_19054_transpose_y_3, x = q_u_239_cast_fp16, y = k_head_479_cast_fp16)[name = string("op_19054_cast_fp16")];
            fp16 var_19055_to_fp16 = const()[name = string("op_19055_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_239_cast_fp16 = mul(x = var_19054_cast_fp16, y = var_19055_to_fp16)[name = string("scores_content_239_cast_fp16")];
            bool x_1253_transpose_x_3 = const()[name = string("x_1253_transpose_x_3"), val = bool(false)];
            bool x_1253_transpose_y_3 = const()[name = string("x_1253_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1253_cast_fp16 = matmul(transpose_x = x_1253_transpose_x_3, transpose_y = x_1253_transpose_y_3, x = q_v_239_cast_fp16, y = p_head_479_cast_fp16)[name = string("x_1253_cast_fp16")];
            tensor<int32, [8]> x_1255_pad_1 = const()[name = string("x_1255_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1255_mode_1 = const()[name = string("x_1255_mode_1"), val = string("constant")];
            fp16 const_2107_to_fp16 = const()[name = string("const_2107_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1255_cast_fp16 = pad(constant_val = const_2107_to_fp16, mode = x_1255_mode_1, pad = x_1255_pad_1, x = x_1253_cast_fp16)[name = string("x_1255_cast_fp16")];
            tensor<int32, [4]> var_19069 = const()[name = string("op_19069"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1257_cast_fp16 = reshape(shape = var_19069, x = x_1255_cast_fp16)[name = string("x_1257_cast_fp16")];
            tensor<int32, [4]> var_19073_begin_1 = const()[name = string("op_19073_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_19073_end_1 = const()[name = string("op_19073_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_19073_end_mask_1 = const()[name = string("op_19073_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_19073_cast_fp16 = slice_by_index(begin = var_19073_begin_1, end = var_19073_end_1, end_mask = var_19073_end_mask_1, x = x_1257_cast_fp16)[name = string("op_19073_cast_fp16")];
            tensor<int32, [4]> var_19075 = const()[name = string("op_19075"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_19076_cast_fp16 = reshape(shape = var_19075, x = var_19073_cast_fp16)[name = string("op_19076_cast_fp16")];
            tensor<int32, [4]> var_19081_begin_1 = const()[name = string("op_19081_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_19081_end_1 = const()[name = string("op_19081_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_19081_end_mask_1 = const()[name = string("op_19081_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_19081_cast_fp16 = slice_by_index(begin = var_19081_begin_1, end = var_19081_end_1, end_mask = var_19081_end_mask_1, x = var_19076_cast_fp16)[name = string("op_19081_cast_fp16")];
            fp16 var_19082_to_fp16 = const()[name = string("op_19082_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_239_cast_fp16 = mul(x = var_19081_cast_fp16, y = var_19082_to_fp16)[name = string("scores_pos_239_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_239_cast_fp16 = add(x = scores_content_239_cast_fp16, y = scores_pos_239_cast_fp16)[name = string("logits_239_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_19085_cast_fp16 = softmax(axis = var_17963, x = logits_239_cast_fp16)[name = string("op_19085_cast_fp16")];
            bool var_19087_transpose_x_1 = const()[name = string("op_19087_transpose_x_1"), val = bool(false)];
            bool var_19087_transpose_y_1 = const()[name = string("op_19087_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_19087_cast_fp16 = matmul(transpose_x = var_19087_transpose_x_1, transpose_y = var_19087_transpose_y_1, x = var_19085_cast_fp16, y = v_head_479_cast_fp16)[name = string("op_19087_cast_fp16")];
            tensor<int32, [1]> o_head_29_axes_1 = const()[name = string("o_head_29_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_29_cast_fp16 = squeeze(axes = o_head_29_axes_1, x = var_19087_cast_fp16)[name = string("o_head_29_cast_fp16")];
            bool out_29_interleave_1 = const()[name = string("out_29_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_29_cast_fp16 = concat(axis = var_17963, interleave = out_29_interleave_1, values = (var_18241_cast_fp16, var_18362_cast_fp16, var_18483_cast_fp16, var_18604_cast_fp16, var_18725_cast_fp16, var_18846_cast_fp16, var_18967_cast_fp16, o_head_29_cast_fp16))[name = string("out_29_cast_fp16")];
            tensor<int32, [3]> var_19091_perm_1 = const()[name = string("op_19091_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_681_axes_1 = const()[name = string("input_681_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_19091_cast_fp16 = transpose(perm = var_19091_perm_1, x = out_29_cast_fp16)[name = string("transpose_374")];
            tensor<fp16, [1, 1024, 51, 1]> input_681_cast_fp16 = expand_dims(axes = input_681_axes_1, x = var_19091_cast_fp16)[name = string("input_681_cast_fp16")];
            string dense_output_1201_pad_type_1 = const()[name = string("dense_output_1201_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1201_strides_1 = const()[name = string("dense_output_1201_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1201_pad_1 = const()[name = string("dense_output_1201_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1201_dilations_1 = const()[name = string("dense_output_1201_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1201_groups_1 = const()[name = string("dense_output_1201_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430041088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431089728))))[name = string("layers_14_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1201_cast_fp16 = conv(dilations = dense_output_1201_dilations_1, groups = dense_output_1201_groups_1, pad = dense_output_1201_pad_1, pad_type = dense_output_1201_pad_type_1, strides = dense_output_1201_strides_1, weight = layers_14_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_681_cast_fp16)[name = string("dense_output_1201_cast_fp16")];
            string sparse_output_1201_pad_type_1 = const()[name = string("sparse_output_1201_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1201_strides_1 = const()[name = string("sparse_output_1201_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1201_pad_1 = const()[name = string("sparse_output_1201_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1201_dilations_1 = const()[name = string("sparse_output_1201_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1201_groups_1 = const()[name = string("sparse_output_1201_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431111360))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431090304))))[name = string("layers_14_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1201_cast_fp16 = conv(dilations = sparse_output_1201_dilations_1, groups = sparse_output_1201_groups_1, pad = sparse_output_1201_pad_1, pad_type = sparse_output_1201_pad_type_1, strides = sparse_output_1201_strides_1, weight = layers_14_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_681_cast_fp16)[name = string("sparse_output_1201_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_29_cast_fp16 = add(x = dense_output_1201_cast_fp16, y = sparse_output_1201_cast_fp16)[name = string("out_conv_29_cast_fp16")];
            tensor<int32, [1]> var_19108_axes_1 = const()[name = string("op_19108_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_19108_cast_fp16 = squeeze(axes = var_19108_axes_1, x = out_conv_29_cast_fp16)[name = string("op_19108_cast_fp16")];
            tensor<int32, [3]> var_19109_perm_1 = const()[name = string("op_19109_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_19109_cast_fp16 = transpose(perm = var_19109_perm_1, x = var_19108_cast_fp16)[name = string("transpose_373")];
            tensor<fp16, [1, 51, 1024]> input_683_cast_fp16 = add(x = input_671_cast_fp16, y = var_19109_cast_fp16)[name = string("input_683_cast_fp16")];
            tensor<int32, [1]> x_1261_axes_1 = const()[name = string("x_1261_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_14_norm_conv_weight_to_fp16 = const()[name = string("layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431242496)))];
            tensor<fp16, [1024]> layers_14_norm_conv_bias_to_fp16 = const()[name = string("layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431244608)))];
            tensor<fp16, [1, 51, 1024]> x_1261_cast_fp16 = layer_norm(axes = x_1261_axes_1, beta = layers_14_norm_conv_bias_to_fp16, epsilon = var_17978_to_fp16, gamma = layers_14_norm_conv_weight_to_fp16, x = input_683_cast_fp16)[name = string("x_1261_cast_fp16")];
            tensor<int32, [3]> var_19119_perm_1 = const()[name = string("op_19119_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_685_axes_1 = const()[name = string("input_685_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_19119_cast_fp16 = transpose(perm = var_19119_perm_1, x = x_1261_cast_fp16)[name = string("transpose_372")];
            tensor<fp16, [1, 1024, 51, 1]> input_685_cast_fp16 = expand_dims(axes = input_685_axes_1, x = var_19119_cast_fp16)[name = string("input_685_cast_fp16")];
            string dense_output_1203_pad_type_1 = const()[name = string("dense_output_1203_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1203_strides_1 = const()[name = string("dense_output_1203_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1203_pad_1 = const()[name = string("dense_output_1203_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1203_dilations_1 = const()[name = string("dense_output_1203_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1203_groups_1 = const()[name = string("dense_output_1203_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_14_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431246720))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433343936))))[name = string("layers_14_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1203_cast_fp16 = conv(dilations = dense_output_1203_dilations_1, groups = dense_output_1203_groups_1, pad = dense_output_1203_pad_1, pad_type = dense_output_1203_pad_type_1, strides = dense_output_1203_strides_1, weight = layers_14_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_685_cast_fp16)[name = string("dense_output_1203_cast_fp16")];
            string sparse_output_1203_pad_type_1 = const()[name = string("sparse_output_1203_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1203_strides_1 = const()[name = string("sparse_output_1203_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1203_pad_1 = const()[name = string("sparse_output_1203_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1203_dilations_1 = const()[name = string("sparse_output_1203_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1203_groups_1 = const()[name = string("sparse_output_1203_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_14_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433386560))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433344512))))[name = string("layers_14_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1203_cast_fp16 = conv(dilations = sparse_output_1203_dilations_1, groups = sparse_output_1203_groups_1, pad = sparse_output_1203_pad_1, pad_type = sparse_output_1203_pad_type_1, strides = sparse_output_1203_strides_1, weight = layers_14_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_685_cast_fp16)[name = string("sparse_output_1203_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_687_cast_fp16 = add(x = dense_output_1203_cast_fp16, y = sparse_output_1203_cast_fp16)[name = string("input_687_cast_fp16")];
            int32 input_689_split_num_splits_1 = const()[name = string("input_689_split_num_splits_1"), val = int32(2)];
            int32 input_689_split_axis_1 = const()[name = string("input_689_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_689_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_689_split_cast_fp16_1 = split(axis = input_689_split_axis_1, num_splits = input_689_split_num_splits_1, x = input_687_cast_fp16)[name = string("input_689_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_689_split_1_sigmoid_cast_fp16 = sigmoid(x = input_689_split_cast_fp16_1)[name = string("input_689_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_689_cast_fp16 = mul(x = input_689_split_cast_fp16_0, y = input_689_split_1_sigmoid_cast_fp16)[name = string("input_689_cast_fp16")];
            tensor<int32, [8]> input_691_pad_1 = const()[name = string("input_691_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_691_mode_1 = const()[name = string("input_691_mode_1"), val = string("constant")];
            fp16 const_2109_to_fp16 = const()[name = string("const_2109_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_691_cast_fp16 = pad(constant_val = const_2109_to_fp16, mode = input_691_mode_1, pad = input_691_pad_1, x = input_689_cast_fp16)[name = string("input_691_cast_fp16")];
            string dense_output_1205_pad_type_1 = const()[name = string("dense_output_1205_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1205_strides_1 = const()[name = string("dense_output_1205_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1205_pad_1 = const()[name = string("dense_output_1205_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1205_dilations_1 = const()[name = string("dense_output_1205_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1205_groups_1 = const()[name = string("dense_output_1205_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1205_cast_fp16 = conv(dilations = dense_output_1205_dilations_1, groups = dense_output_1205_groups_1, pad = dense_output_1205_pad_1, pad_type = dense_output_1205_pad_type_1, strides = dense_output_1205_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_691_cast_fp16)[name = string("dense_output_1205_cast_fp16")];
            string sparse_output_1205_pad_type_1 = const()[name = string("sparse_output_1205_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1205_strides_1 = const()[name = string("sparse_output_1205_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1205_pad_1 = const()[name = string("sparse_output_1205_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1205_dilations_1 = const()[name = string("sparse_output_1205_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1205_groups_1 = const()[name = string("sparse_output_1205_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_14_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433648768))))[name = string("layers_14_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1205_cast_fp16 = conv(dilations = sparse_output_1205_dilations_1, groups = sparse_output_1205_groups_1, pad = sparse_output_1205_pad_1, pad_type = sparse_output_1205_pad_type_1, strides = sparse_output_1205_strides_1, weight = layers_14_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_691_cast_fp16)[name = string("sparse_output_1205_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_693_cast_fp16 = add(x = dense_output_1205_cast_fp16, y = sparse_output_1205_cast_fp16)[name = string("input_693_cast_fp16")];
            tensor<fp16, [1024]> layers_14_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_14_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433667264)))];
            tensor<fp16, [1024]> layers_14_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_14_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433669376)))];
            tensor<fp16, [1024]> layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433671488)))];
            tensor<fp16, [1024]> layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433673600)))];
            tensor<fp16, [1, 1024, 51, 1]> input_695_cast_fp16 = batch_norm(beta = layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_17978_to_fp16, gamma = layers_14_conv_batch_norm_weight_to_fp16, mean = layers_14_conv_batch_norm_running_mean_to_fp16, variance = layers_14_conv_batch_norm_running_var_to_fp16, x = input_693_cast_fp16)[name = string("input_695_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_697_cast_fp16 = silu(x = input_695_cast_fp16)[name = string("input_697_cast_fp16")];
            string dense_output_1207_pad_type_1 = const()[name = string("dense_output_1207_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1207_strides_1 = const()[name = string("dense_output_1207_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1207_pad_1 = const()[name = string("dense_output_1207_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1207_dilations_1 = const()[name = string("dense_output_1207_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1207_groups_1 = const()[name = string("dense_output_1207_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433675712))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434724352))))[name = string("layers_14_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1207_cast_fp16 = conv(dilations = dense_output_1207_dilations_1, groups = dense_output_1207_groups_1, pad = dense_output_1207_pad_1, pad_type = dense_output_1207_pad_type_1, strides = dense_output_1207_strides_1, weight = layers_14_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_697_cast_fp16)[name = string("dense_output_1207_cast_fp16")];
            string sparse_output_1207_pad_type_1 = const()[name = string("sparse_output_1207_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1207_strides_1 = const()[name = string("sparse_output_1207_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1207_pad_1 = const()[name = string("sparse_output_1207_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1207_dilations_1 = const()[name = string("sparse_output_1207_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1207_groups_1 = const()[name = string("sparse_output_1207_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_14_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434745984))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434724928))))[name = string("layers_14_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1207_cast_fp16 = conv(dilations = sparse_output_1207_dilations_1, groups = sparse_output_1207_groups_1, pad = sparse_output_1207_pad_1, pad_type = sparse_output_1207_pad_type_1, strides = sparse_output_1207_strides_1, weight = layers_14_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_697_cast_fp16)[name = string("sparse_output_1207_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1263_cast_fp16 = add(x = dense_output_1207_cast_fp16, y = sparse_output_1207_cast_fp16)[name = string("x_1263_cast_fp16")];
            tensor<int32, [1]> var_19175_axes_1 = const()[name = string("op_19175_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_19175_cast_fp16 = squeeze(axes = var_19175_axes_1, x = x_1263_cast_fp16)[name = string("op_19175_cast_fp16")];
            tensor<int32, [3]> var_19176_perm_1 = const()[name = string("op_19176_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_19176_cast_fp16 = transpose(perm = var_19176_perm_1, x = var_19175_cast_fp16)[name = string("transpose_371")];
            tensor<fp16, [1, 51, 1024]> input_699_cast_fp16 = add(x = input_683_cast_fp16, y = var_19176_cast_fp16)[name = string("input_699_cast_fp16")];
            tensor<int32, [1]> x_1265_axes_1 = const()[name = string("x_1265_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434877120)))];
            tensor<fp16, [1024]> layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434879232)))];
            tensor<fp16, [1, 51, 1024]> x_1265_cast_fp16 = layer_norm(axes = x_1265_axes_1, beta = layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_17978_to_fp16, gamma = layers_14_norm_feed_forward2_weight_to_fp16, x = input_699_cast_fp16)[name = string("x_1265_cast_fp16")];
            tensor<int32, [4]> var_19186 = const()[name = string("op_19186"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1267_cast_fp16 = reshape(shape = var_19186, x = x_1265_cast_fp16)[name = string("x_1267_cast_fp16")];
            tensor<int32, [4]> input_701_perm_1 = const()[name = string("input_701_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1209_pad_type_1 = const()[name = string("dense_output_1209_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1209_strides_1 = const()[name = string("dense_output_1209_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1209_pad_1 = const()[name = string("dense_output_1209_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1209_dilations_1 = const()[name = string("dense_output_1209_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1209_groups_1 = const()[name = string("dense_output_1209_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_14_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434881344))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439075712))))[name = string("layers_14_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_701_cast_fp16 = transpose(perm = input_701_perm_1, x = x_1267_cast_fp16)[name = string("transpose_370")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1209_cast_fp16 = conv(dilations = dense_output_1209_dilations_1, groups = dense_output_1209_groups_1, pad = dense_output_1209_pad_1, pad_type = dense_output_1209_pad_type_1, strides = dense_output_1209_strides_1, weight = layers_14_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_701_cast_fp16)[name = string("dense_output_1209_cast_fp16")];
            string sparse_output_1209_pad_type_1 = const()[name = string("sparse_output_1209_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1209_strides_1 = const()[name = string("sparse_output_1209_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1209_pad_1 = const()[name = string("sparse_output_1209_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1209_dilations_1 = const()[name = string("sparse_output_1209_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1209_groups_1 = const()[name = string("sparse_output_1209_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_14_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439160256))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439076288))))[name = string("layers_14_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1209_cast_fp16 = conv(dilations = sparse_output_1209_dilations_1, groups = sparse_output_1209_groups_1, pad = sparse_output_1209_pad_1, pad_type = sparse_output_1209_pad_type_1, strides = sparse_output_1209_strides_1, weight = layers_14_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_701_cast_fp16)[name = string("sparse_output_1209_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_703_cast_fp16 = add(x = dense_output_1209_cast_fp16, y = sparse_output_1209_cast_fp16)[name = string("input_703_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_705_cast_fp16 = silu(x = input_703_cast_fp16)[name = string("input_705_cast_fp16")];
            string dense_output_1211_pad_type_1 = const()[name = string("dense_output_1211_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1211_strides_1 = const()[name = string("dense_output_1211_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1211_pad_1 = const()[name = string("dense_output_1211_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1211_dilations_1 = const()[name = string("dense_output_1211_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1211_groups_1 = const()[name = string("dense_output_1211_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_14_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439684608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443878976))))[name = string("layers_14_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1211_cast_fp16 = conv(dilations = dense_output_1211_dilations_1, groups = dense_output_1211_groups_1, pad = dense_output_1211_pad_1, pad_type = dense_output_1211_pad_type_1, strides = dense_output_1211_strides_1, weight = layers_14_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_705_cast_fp16)[name = string("dense_output_1211_cast_fp16")];
            string sparse_output_1211_pad_type_1 = const()[name = string("sparse_output_1211_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1211_strides_1 = const()[name = string("sparse_output_1211_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1211_pad_1 = const()[name = string("sparse_output_1211_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1211_dilations_1 = const()[name = string("sparse_output_1211_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1211_groups_1 = const()[name = string("sparse_output_1211_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_14_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443963520))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443879552))))[name = string("layers_14_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1211_cast_fp16 = conv(dilations = sparse_output_1211_dilations_1, groups = sparse_output_1211_groups_1, pad = sparse_output_1211_pad_1, pad_type = sparse_output_1211_pad_type_1, strides = sparse_output_1211_strides_1, weight = layers_14_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_705_cast_fp16)[name = string("sparse_output_1211_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1269_cast_fp16 = add(x = dense_output_1211_cast_fp16, y = sparse_output_1211_cast_fp16)[name = string("x_1269_cast_fp16")];
            tensor<int32, [4]> x_1271_perm_1 = const()[name = string("x_1271_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_19221 = const()[name = string("op_19221"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1271_cast_fp16 = transpose(perm = x_1271_perm_1, x = x_1269_cast_fp16)[name = string("transpose_369")];
            tensor<fp16, [1, 51, 1024]> var_19222_cast_fp16 = reshape(shape = var_19221, x = x_1271_cast_fp16)[name = string("op_19222_cast_fp16")];
            fp16 var_19223_to_fp16 = const()[name = string("op_19223_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_19224_cast_fp16 = mul(x = var_19222_cast_fp16, y = var_19223_to_fp16)[name = string("op_19224_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_707_cast_fp16 = add(x = input_699_cast_fp16, y = var_19224_cast_fp16)[name = string("input_707_cast_fp16")];
            tensor<int32, [1]> input_709_axes_1 = const()[name = string("input_709_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_14_norm_out_weight_to_fp16 = const()[name = string("layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444487872)))];
            tensor<fp16, [1024]> layers_14_norm_out_bias_to_fp16 = const()[name = string("layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444489984)))];
            tensor<fp16, [1, 51, 1024]> input_709_cast_fp16 = layer_norm(axes = input_709_axes_1, beta = layers_14_norm_out_bias_to_fp16, epsilon = var_17978_to_fp16, gamma = layers_14_norm_out_weight_to_fp16, x = input_707_cast_fp16)[name = string("input_709_cast_fp16")];
            int32 var_19232 = const()[name = string("op_19232"), val = int32(-1)];
            tensor<int32, [1]> x_1273_axes_1 = const()[name = string("x_1273_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444492096)))];
            tensor<fp16, [1024]> layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444494208)))];
            fp16 var_19247_to_fp16 = const()[name = string("op_19247_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1273_cast_fp16 = layer_norm(axes = x_1273_axes_1, beta = layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_19247_to_fp16, gamma = layers_15_norm_feed_forward1_weight_to_fp16, x = input_709_cast_fp16)[name = string("x_1273_cast_fp16")];
            tensor<int32, [4]> var_19266 = const()[name = string("op_19266"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1275_cast_fp16 = reshape(shape = var_19266, x = x_1273_cast_fp16)[name = string("x_1275_cast_fp16")];
            tensor<int32, [4]> input_711_perm_1 = const()[name = string("input_711_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1213_pad_type_1 = const()[name = string("dense_output_1213_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1213_strides_1 = const()[name = string("dense_output_1213_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1213_pad_1 = const()[name = string("dense_output_1213_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1213_dilations_1 = const()[name = string("dense_output_1213_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1213_groups_1 = const()[name = string("dense_output_1213_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_15_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444496320))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448690688))))[name = string("layers_15_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_711_cast_fp16 = transpose(perm = input_711_perm_1, x = x_1275_cast_fp16)[name = string("transpose_368")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1213_cast_fp16 = conv(dilations = dense_output_1213_dilations_1, groups = dense_output_1213_groups_1, pad = dense_output_1213_pad_1, pad_type = dense_output_1213_pad_type_1, strides = dense_output_1213_strides_1, weight = layers_15_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_711_cast_fp16)[name = string("dense_output_1213_cast_fp16")];
            string sparse_output_1213_pad_type_1 = const()[name = string("sparse_output_1213_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1213_strides_1 = const()[name = string("sparse_output_1213_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1213_pad_1 = const()[name = string("sparse_output_1213_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1213_dilations_1 = const()[name = string("sparse_output_1213_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1213_groups_1 = const()[name = string("sparse_output_1213_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_15_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448775232))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448691264))))[name = string("layers_15_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1213_cast_fp16 = conv(dilations = sparse_output_1213_dilations_1, groups = sparse_output_1213_groups_1, pad = sparse_output_1213_pad_1, pad_type = sparse_output_1213_pad_type_1, strides = sparse_output_1213_strides_1, weight = layers_15_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_711_cast_fp16)[name = string("sparse_output_1213_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_713_cast_fp16 = add(x = dense_output_1213_cast_fp16, y = sparse_output_1213_cast_fp16)[name = string("input_713_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_715_cast_fp16 = silu(x = input_713_cast_fp16)[name = string("input_715_cast_fp16")];
            string dense_output_1215_pad_type_1 = const()[name = string("dense_output_1215_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1215_strides_1 = const()[name = string("dense_output_1215_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1215_pad_1 = const()[name = string("dense_output_1215_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1215_dilations_1 = const()[name = string("dense_output_1215_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1215_groups_1 = const()[name = string("dense_output_1215_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_15_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449299584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453493952))))[name = string("layers_15_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1215_cast_fp16 = conv(dilations = dense_output_1215_dilations_1, groups = dense_output_1215_groups_1, pad = dense_output_1215_pad_1, pad_type = dense_output_1215_pad_type_1, strides = dense_output_1215_strides_1, weight = layers_15_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("dense_output_1215_cast_fp16")];
            string sparse_output_1215_pad_type_1 = const()[name = string("sparse_output_1215_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1215_strides_1 = const()[name = string("sparse_output_1215_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1215_pad_1 = const()[name = string("sparse_output_1215_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1215_dilations_1 = const()[name = string("sparse_output_1215_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1215_groups_1 = const()[name = string("sparse_output_1215_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_15_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453578496))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453494528))))[name = string("layers_15_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1215_cast_fp16 = conv(dilations = sparse_output_1215_dilations_1, groups = sparse_output_1215_groups_1, pad = sparse_output_1215_pad_1, pad_type = sparse_output_1215_pad_type_1, strides = sparse_output_1215_strides_1, weight = layers_15_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_715_cast_fp16)[name = string("sparse_output_1215_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1277_cast_fp16 = add(x = dense_output_1215_cast_fp16, y = sparse_output_1215_cast_fp16)[name = string("x_1277_cast_fp16")];
            tensor<int32, [4]> x_1279_perm_1 = const()[name = string("x_1279_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_19301 = const()[name = string("op_19301"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1279_cast_fp16 = transpose(perm = x_1279_perm_1, x = x_1277_cast_fp16)[name = string("transpose_367")];
            tensor<fp16, [1, 51, 1024]> var_19302_cast_fp16 = reshape(shape = var_19301, x = x_1279_cast_fp16)[name = string("op_19302_cast_fp16")];
            fp16 var_19303_to_fp16 = const()[name = string("op_19303_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_19304_cast_fp16 = mul(x = var_19302_cast_fp16, y = var_19303_to_fp16)[name = string("op_19304_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_717_cast_fp16 = add(x = input_709_cast_fp16, y = var_19304_cast_fp16)[name = string("input_717_cast_fp16")];
            tensor<int32, [1]> q_31_axes_1 = const()[name = string("q_31_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_15_norm_self_att_weight_to_fp16 = const()[name = string("layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454102848)))];
            tensor<fp16, [1024]> layers_15_norm_self_att_bias_to_fp16 = const()[name = string("layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454104960)))];
            tensor<fp16, [1, 51, 1024]> q_31_cast_fp16 = layer_norm(axes = q_31_axes_1, beta = layers_15_norm_self_att_bias_to_fp16, epsilon = var_19247_to_fp16, gamma = layers_15_norm_self_att_weight_to_fp16, x = input_717_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<int32, [3]> var_19378 = const()[name = string("op_19378"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_719_axes_1 = const()[name = string("input_719_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_19379_cast_fp16 = transpose(perm = var_19378, x = q_31_cast_fp16)[name = string("transpose_366")];
            tensor<fp16, [1, 1024, 51, 1]> input_719_cast_fp16 = expand_dims(axes = input_719_axes_1, x = var_19379_cast_fp16)[name = string("input_719_cast_fp16")];
            string dense_output_1217_pad_type_1 = const()[name = string("dense_output_1217_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1217_strides_1 = const()[name = string("dense_output_1217_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1217_pad_1 = const()[name = string("dense_output_1217_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1217_dilations_1 = const()[name = string("dense_output_1217_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1217_groups_1 = const()[name = string("dense_output_1217_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454107072))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454238208))))[name = string("layers_15_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1217_cast_fp16 = conv(dilations = dense_output_1217_dilations_1, groups = dense_output_1217_groups_1, pad = dense_output_1217_pad_1, pad_type = dense_output_1217_pad_type_1, strides = dense_output_1217_strides_1, weight = layers_15_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1217_cast_fp16")];
            string sparse_output_1217_pad_type_1 = const()[name = string("sparse_output_1217_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1217_strides_1 = const()[name = string("sparse_output_1217_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1217_pad_1 = const()[name = string("sparse_output_1217_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1217_dilations_1 = const()[name = string("sparse_output_1217_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1217_groups_1 = const()[name = string("sparse_output_1217_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454241472))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454238784))))[name = string("layers_15_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1217_cast_fp16 = conv(dilations = sparse_output_1217_dilations_1, groups = sparse_output_1217_groups_1, pad = sparse_output_1217_pad_1, pad_type = sparse_output_1217_pad_type_1, strides = sparse_output_1217_strides_1, weight = layers_15_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1217_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19404_cast_fp16 = add(x = dense_output_1217_cast_fp16, y = sparse_output_1217_cast_fp16)[name = string("op_19404_cast_fp16")];
            tensor<int32, [4]> var_19405 = const()[name = string("op_19405"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19407 = const()[name = string("op_19407"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19406_cast_fp16 = transpose(perm = var_19405, x = var_19404_cast_fp16)[name = string("transpose_365")];
            tensor<fp16, [1, 51, 128]> q_head_241_cast_fp16 = reshape(shape = var_19407, x = var_19406_cast_fp16)[name = string("q_head_241_cast_fp16")];
            string dense_output_1219_pad_type_1 = const()[name = string("dense_output_1219_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1219_strides_1 = const()[name = string("dense_output_1219_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1219_pad_1 = const()[name = string("dense_output_1219_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1219_dilations_1 = const()[name = string("dense_output_1219_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1219_groups_1 = const()[name = string("dense_output_1219_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454257920))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454389056))))[name = string("layers_15_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1219_cast_fp16 = conv(dilations = dense_output_1219_dilations_1, groups = dense_output_1219_groups_1, pad = dense_output_1219_pad_1, pad_type = dense_output_1219_pad_type_1, strides = dense_output_1219_strides_1, weight = layers_15_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1219_cast_fp16")];
            string sparse_output_1219_pad_type_1 = const()[name = string("sparse_output_1219_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1219_strides_1 = const()[name = string("sparse_output_1219_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1219_pad_1 = const()[name = string("sparse_output_1219_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1219_dilations_1 = const()[name = string("sparse_output_1219_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1219_groups_1 = const()[name = string("sparse_output_1219_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454392320))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454389632))))[name = string("layers_15_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1219_cast_fp16 = conv(dilations = sparse_output_1219_dilations_1, groups = sparse_output_1219_groups_1, pad = sparse_output_1219_pad_1, pad_type = sparse_output_1219_pad_type_1, strides = sparse_output_1219_strides_1, weight = layers_15_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1219_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19423_cast_fp16 = add(x = dense_output_1219_cast_fp16, y = sparse_output_1219_cast_fp16)[name = string("op_19423_cast_fp16")];
            tensor<int32, [4]> var_19424 = const()[name = string("op_19424"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19426 = const()[name = string("op_19426"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19425_cast_fp16 = transpose(perm = var_19424, x = var_19423_cast_fp16)[name = string("transpose_364")];
            tensor<fp16, [1, 51, 128]> k_head_481_cast_fp16 = reshape(shape = var_19426, x = var_19425_cast_fp16)[name = string("k_head_481_cast_fp16")];
            string dense_output_1221_pad_type_1 = const()[name = string("dense_output_1221_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1221_strides_1 = const()[name = string("dense_output_1221_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1221_pad_1 = const()[name = string("dense_output_1221_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1221_dilations_1 = const()[name = string("dense_output_1221_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1221_groups_1 = const()[name = string("dense_output_1221_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454408768))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454539904))))[name = string("layers_15_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1221_cast_fp16 = conv(dilations = dense_output_1221_dilations_1, groups = dense_output_1221_groups_1, pad = dense_output_1221_pad_1, pad_type = dense_output_1221_pad_type_1, strides = dense_output_1221_strides_1, weight = layers_15_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1221_cast_fp16")];
            string sparse_output_1221_pad_type_1 = const()[name = string("sparse_output_1221_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1221_strides_1 = const()[name = string("sparse_output_1221_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1221_pad_1 = const()[name = string("sparse_output_1221_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1221_dilations_1 = const()[name = string("sparse_output_1221_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1221_groups_1 = const()[name = string("sparse_output_1221_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454543168))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454540480))))[name = string("layers_15_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1221_cast_fp16 = conv(dilations = sparse_output_1221_dilations_1, groups = sparse_output_1221_groups_1, pad = sparse_output_1221_pad_1, pad_type = sparse_output_1221_pad_type_1, strides = sparse_output_1221_strides_1, weight = layers_15_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1221_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19442_cast_fp16 = add(x = dense_output_1221_cast_fp16, y = sparse_output_1221_cast_fp16)[name = string("op_19442_cast_fp16")];
            tensor<int32, [4]> var_19443 = const()[name = string("op_19443"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19445 = const()[name = string("op_19445"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19444_cast_fp16 = transpose(perm = var_19443, x = var_19442_cast_fp16)[name = string("transpose_363")];
            tensor<fp16, [1, 51, 128]> v_head_481_cast_fp16 = reshape(shape = var_19445, x = var_19444_cast_fp16)[name = string("v_head_481_cast_fp16")];
            string dense_output_1223_pad_type_1 = const()[name = string("dense_output_1223_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1223_strides_1 = const()[name = string("dense_output_1223_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1223_pad_1 = const()[name = string("dense_output_1223_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1223_dilations_1 = const()[name = string("dense_output_1223_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1223_groups_1 = const()[name = string("dense_output_1223_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454559616))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454690752))))[name = string("layers_15_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1223_cast_fp16 = conv(dilations = dense_output_1223_dilations_1, groups = dense_output_1223_groups_1, pad = dense_output_1223_pad_1, pad_type = dense_output_1223_pad_type_1, strides = dense_output_1223_strides_1, weight = layers_15_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1223_cast_fp16")];
            string sparse_output_1223_pad_type_1 = const()[name = string("sparse_output_1223_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1223_strides_1 = const()[name = string("sparse_output_1223_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1223_pad_1 = const()[name = string("sparse_output_1223_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1223_dilations_1 = const()[name = string("sparse_output_1223_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1223_groups_1 = const()[name = string("sparse_output_1223_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454694016))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454691328))))[name = string("layers_15_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1223_cast_fp16 = conv(dilations = sparse_output_1223_dilations_1, groups = sparse_output_1223_groups_1, pad = sparse_output_1223_pad_1, pad_type = sparse_output_1223_pad_type_1, strides = sparse_output_1223_strides_1, weight = layers_15_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1223_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_19461_cast_fp16 = add(x = dense_output_1223_cast_fp16, y = sparse_output_1223_cast_fp16)[name = string("op_19461_cast_fp16")];
            tensor<int32, [4]> var_19462 = const()[name = string("op_19462"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19464 = const()[name = string("op_19464"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_19463_cast_fp16 = transpose(perm = var_19462, x = var_19461_cast_fp16)[name = string("transpose_362")];
            tensor<fp16, [1, 101, 128]> p_head_481_cast_fp16 = reshape(shape = var_19464, x = var_19463_cast_fp16)[name = string("p_head_481_cast_fp16")];
            tensor<fp16, [128]> var_19466_to_fp16 = const()[name = string("op_19466_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454710464)))];
            tensor<fp16, [1, 51, 128]> var_19467_cast_fp16 = add(x = q_head_241_cast_fp16, y = var_19466_to_fp16)[name = string("op_19467_cast_fp16")];
            tensor<int32, [1]> q_u_241_axes_1 = const()[name = string("q_u_241_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_241_cast_fp16 = expand_dims(axes = q_u_241_axes_1, x = var_19467_cast_fp16)[name = string("q_u_241_cast_fp16")];
            tensor<fp16, [128]> var_19469_to_fp16 = const()[name = string("op_19469_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454710784)))];
            tensor<fp16, [1, 51, 128]> var_19470_cast_fp16 = add(x = q_head_241_cast_fp16, y = var_19469_to_fp16)[name = string("op_19470_cast_fp16")];
            tensor<int32, [1]> q_v_241_axes_1 = const()[name = string("q_v_241_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_241_cast_fp16 = expand_dims(axes = q_v_241_axes_1, x = var_19470_cast_fp16)[name = string("q_v_241_cast_fp16")];
            tensor<int32, [1]> k_head_483_axes_1 = const()[name = string("k_head_483_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_483_cast_fp16 = expand_dims(axes = k_head_483_axes_1, x = k_head_481_cast_fp16)[name = string("k_head_483_cast_fp16")];
            tensor<int32, [1]> v_head_483_axes_1 = const()[name = string("v_head_483_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_483_cast_fp16 = expand_dims(axes = v_head_483_axes_1, x = v_head_481_cast_fp16)[name = string("v_head_483_cast_fp16")];
            tensor<int32, [1]> p_head_483_axes_1 = const()[name = string("p_head_483_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_483_cast_fp16 = expand_dims(axes = p_head_483_axes_1, x = p_head_481_cast_fp16)[name = string("p_head_483_cast_fp16")];
            bool var_19476_transpose_x_3 = const()[name = string("op_19476_transpose_x_3"), val = bool(false)];
            bool var_19476_transpose_y_3 = const()[name = string("op_19476_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_19476_cast_fp16 = matmul(transpose_x = var_19476_transpose_x_3, transpose_y = var_19476_transpose_y_3, x = q_u_241_cast_fp16, y = k_head_483_cast_fp16)[name = string("op_19476_cast_fp16")];
            fp16 var_19477_to_fp16 = const()[name = string("op_19477_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_241_cast_fp16 = mul(x = var_19476_cast_fp16, y = var_19477_to_fp16)[name = string("scores_content_241_cast_fp16")];
            bool x_1281_transpose_x_3 = const()[name = string("x_1281_transpose_x_3"), val = bool(false)];
            bool x_1281_transpose_y_3 = const()[name = string("x_1281_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1281_cast_fp16 = matmul(transpose_x = x_1281_transpose_x_3, transpose_y = x_1281_transpose_y_3, x = q_v_241_cast_fp16, y = p_head_483_cast_fp16)[name = string("x_1281_cast_fp16")];
            tensor<int32, [8]> x_1283_pad_1 = const()[name = string("x_1283_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1283_mode_1 = const()[name = string("x_1283_mode_1"), val = string("constant")];
            fp16 const_2119_to_fp16 = const()[name = string("const_2119_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1283_cast_fp16 = pad(constant_val = const_2119_to_fp16, mode = x_1283_mode_1, pad = x_1283_pad_1, x = x_1281_cast_fp16)[name = string("x_1283_cast_fp16")];
            tensor<int32, [4]> var_19491 = const()[name = string("op_19491"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1285_cast_fp16 = reshape(shape = var_19491, x = x_1283_cast_fp16)[name = string("x_1285_cast_fp16")];
            tensor<int32, [4]> var_19495_begin_1 = const()[name = string("op_19495_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_19495_end_1 = const()[name = string("op_19495_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_19495_end_mask_1 = const()[name = string("op_19495_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_19495_cast_fp16 = slice_by_index(begin = var_19495_begin_1, end = var_19495_end_1, end_mask = var_19495_end_mask_1, x = x_1285_cast_fp16)[name = string("op_19495_cast_fp16")];
            tensor<int32, [4]> var_19497 = const()[name = string("op_19497"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_19498_cast_fp16 = reshape(shape = var_19497, x = var_19495_cast_fp16)[name = string("op_19498_cast_fp16")];
            tensor<int32, [4]> var_19503_begin_1 = const()[name = string("op_19503_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_19503_end_1 = const()[name = string("op_19503_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_19503_end_mask_1 = const()[name = string("op_19503_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_19503_cast_fp16 = slice_by_index(begin = var_19503_begin_1, end = var_19503_end_1, end_mask = var_19503_end_mask_1, x = var_19498_cast_fp16)[name = string("op_19503_cast_fp16")];
            fp16 var_19504_to_fp16 = const()[name = string("op_19504_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_241_cast_fp16 = mul(x = var_19503_cast_fp16, y = var_19504_to_fp16)[name = string("scores_pos_241_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_241_cast_fp16 = add(x = scores_content_241_cast_fp16, y = scores_pos_241_cast_fp16)[name = string("logits_241_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_19507_cast_fp16 = softmax(axis = var_19232, x = logits_241_cast_fp16)[name = string("op_19507_cast_fp16")];
            bool var_19509_transpose_x_1 = const()[name = string("op_19509_transpose_x_1"), val = bool(false)];
            bool var_19509_transpose_y_1 = const()[name = string("op_19509_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_19509_cast_fp16 = matmul(transpose_x = var_19509_transpose_x_1, transpose_y = var_19509_transpose_y_1, x = var_19507_cast_fp16, y = v_head_483_cast_fp16)[name = string("op_19509_cast_fp16")];
            tensor<int32, [1]> var_19510_axes_1 = const()[name = string("op_19510_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_19510_cast_fp16 = squeeze(axes = var_19510_axes_1, x = var_19509_cast_fp16)[name = string("op_19510_cast_fp16")];
            string dense_output_1225_pad_type_1 = const()[name = string("dense_output_1225_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1225_strides_1 = const()[name = string("dense_output_1225_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1225_pad_1 = const()[name = string("dense_output_1225_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1225_dilations_1 = const()[name = string("dense_output_1225_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1225_groups_1 = const()[name = string("dense_output_1225_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454711104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454842240))))[name = string("layers_15_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1225_cast_fp16 = conv(dilations = dense_output_1225_dilations_1, groups = dense_output_1225_groups_1, pad = dense_output_1225_pad_1, pad_type = dense_output_1225_pad_type_1, strides = dense_output_1225_strides_1, weight = layers_15_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1225_cast_fp16")];
            string sparse_output_1225_pad_type_1 = const()[name = string("sparse_output_1225_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1225_strides_1 = const()[name = string("sparse_output_1225_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1225_pad_1 = const()[name = string("sparse_output_1225_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1225_dilations_1 = const()[name = string("sparse_output_1225_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1225_groups_1 = const()[name = string("sparse_output_1225_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454845504))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454842816))))[name = string("layers_15_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1225_cast_fp16 = conv(dilations = sparse_output_1225_dilations_1, groups = sparse_output_1225_groups_1, pad = sparse_output_1225_pad_1, pad_type = sparse_output_1225_pad_type_1, strides = sparse_output_1225_strides_1, weight = layers_15_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1225_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19525_cast_fp16 = add(x = dense_output_1225_cast_fp16, y = sparse_output_1225_cast_fp16)[name = string("op_19525_cast_fp16")];
            tensor<int32, [4]> var_19526 = const()[name = string("op_19526"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19528 = const()[name = string("op_19528"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19527_cast_fp16 = transpose(perm = var_19526, x = var_19525_cast_fp16)[name = string("transpose_361")];
            tensor<fp16, [1, 51, 128]> q_head_243_cast_fp16 = reshape(shape = var_19528, x = var_19527_cast_fp16)[name = string("q_head_243_cast_fp16")];
            string dense_output_1227_pad_type_1 = const()[name = string("dense_output_1227_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1227_strides_1 = const()[name = string("dense_output_1227_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1227_pad_1 = const()[name = string("dense_output_1227_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1227_dilations_1 = const()[name = string("dense_output_1227_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1227_groups_1 = const()[name = string("dense_output_1227_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454861952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454993088))))[name = string("layers_15_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1227_cast_fp16 = conv(dilations = dense_output_1227_dilations_1, groups = dense_output_1227_groups_1, pad = dense_output_1227_pad_1, pad_type = dense_output_1227_pad_type_1, strides = dense_output_1227_strides_1, weight = layers_15_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1227_cast_fp16")];
            string sparse_output_1227_pad_type_1 = const()[name = string("sparse_output_1227_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1227_strides_1 = const()[name = string("sparse_output_1227_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1227_pad_1 = const()[name = string("sparse_output_1227_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1227_dilations_1 = const()[name = string("sparse_output_1227_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1227_groups_1 = const()[name = string("sparse_output_1227_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454996352))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454993664))))[name = string("layers_15_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1227_cast_fp16 = conv(dilations = sparse_output_1227_dilations_1, groups = sparse_output_1227_groups_1, pad = sparse_output_1227_pad_1, pad_type = sparse_output_1227_pad_type_1, strides = sparse_output_1227_strides_1, weight = layers_15_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1227_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19544_cast_fp16 = add(x = dense_output_1227_cast_fp16, y = sparse_output_1227_cast_fp16)[name = string("op_19544_cast_fp16")];
            tensor<int32, [4]> var_19545 = const()[name = string("op_19545"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19547 = const()[name = string("op_19547"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19546_cast_fp16 = transpose(perm = var_19545, x = var_19544_cast_fp16)[name = string("transpose_360")];
            tensor<fp16, [1, 51, 128]> k_head_485_cast_fp16 = reshape(shape = var_19547, x = var_19546_cast_fp16)[name = string("k_head_485_cast_fp16")];
            string dense_output_1229_pad_type_1 = const()[name = string("dense_output_1229_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1229_strides_1 = const()[name = string("dense_output_1229_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1229_pad_1 = const()[name = string("dense_output_1229_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1229_dilations_1 = const()[name = string("dense_output_1229_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1229_groups_1 = const()[name = string("dense_output_1229_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455012800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455143936))))[name = string("layers_15_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1229_cast_fp16 = conv(dilations = dense_output_1229_dilations_1, groups = dense_output_1229_groups_1, pad = dense_output_1229_pad_1, pad_type = dense_output_1229_pad_type_1, strides = dense_output_1229_strides_1, weight = layers_15_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1229_cast_fp16")];
            string sparse_output_1229_pad_type_1 = const()[name = string("sparse_output_1229_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1229_strides_1 = const()[name = string("sparse_output_1229_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1229_pad_1 = const()[name = string("sparse_output_1229_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1229_dilations_1 = const()[name = string("sparse_output_1229_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1229_groups_1 = const()[name = string("sparse_output_1229_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455147200))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455144512))))[name = string("layers_15_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1229_cast_fp16 = conv(dilations = sparse_output_1229_dilations_1, groups = sparse_output_1229_groups_1, pad = sparse_output_1229_pad_1, pad_type = sparse_output_1229_pad_type_1, strides = sparse_output_1229_strides_1, weight = layers_15_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1229_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19563_cast_fp16 = add(x = dense_output_1229_cast_fp16, y = sparse_output_1229_cast_fp16)[name = string("op_19563_cast_fp16")];
            tensor<int32, [4]> var_19564 = const()[name = string("op_19564"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19566 = const()[name = string("op_19566"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19565_cast_fp16 = transpose(perm = var_19564, x = var_19563_cast_fp16)[name = string("transpose_359")];
            tensor<fp16, [1, 51, 128]> v_head_485_cast_fp16 = reshape(shape = var_19566, x = var_19565_cast_fp16)[name = string("v_head_485_cast_fp16")];
            string dense_output_1231_pad_type_1 = const()[name = string("dense_output_1231_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1231_strides_1 = const()[name = string("dense_output_1231_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1231_pad_1 = const()[name = string("dense_output_1231_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1231_dilations_1 = const()[name = string("dense_output_1231_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1231_groups_1 = const()[name = string("dense_output_1231_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455163648))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455294784))))[name = string("layers_15_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1231_cast_fp16 = conv(dilations = dense_output_1231_dilations_1, groups = dense_output_1231_groups_1, pad = dense_output_1231_pad_1, pad_type = dense_output_1231_pad_type_1, strides = dense_output_1231_strides_1, weight = layers_15_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1231_cast_fp16")];
            string sparse_output_1231_pad_type_1 = const()[name = string("sparse_output_1231_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1231_strides_1 = const()[name = string("sparse_output_1231_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1231_pad_1 = const()[name = string("sparse_output_1231_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1231_dilations_1 = const()[name = string("sparse_output_1231_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1231_groups_1 = const()[name = string("sparse_output_1231_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455298048))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455295360))))[name = string("layers_15_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1231_cast_fp16 = conv(dilations = sparse_output_1231_dilations_1, groups = sparse_output_1231_groups_1, pad = sparse_output_1231_pad_1, pad_type = sparse_output_1231_pad_type_1, strides = sparse_output_1231_strides_1, weight = layers_15_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1231_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_19582_cast_fp16 = add(x = dense_output_1231_cast_fp16, y = sparse_output_1231_cast_fp16)[name = string("op_19582_cast_fp16")];
            tensor<int32, [4]> var_19583 = const()[name = string("op_19583"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19585 = const()[name = string("op_19585"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_19584_cast_fp16 = transpose(perm = var_19583, x = var_19582_cast_fp16)[name = string("transpose_358")];
            tensor<fp16, [1, 101, 128]> p_head_485_cast_fp16 = reshape(shape = var_19585, x = var_19584_cast_fp16)[name = string("p_head_485_cast_fp16")];
            tensor<fp16, [128]> var_19587_to_fp16 = const()[name = string("op_19587_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455314496)))];
            tensor<fp16, [1, 51, 128]> var_19588_cast_fp16 = add(x = q_head_243_cast_fp16, y = var_19587_to_fp16)[name = string("op_19588_cast_fp16")];
            tensor<int32, [1]> q_u_243_axes_1 = const()[name = string("q_u_243_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_243_cast_fp16 = expand_dims(axes = q_u_243_axes_1, x = var_19588_cast_fp16)[name = string("q_u_243_cast_fp16")];
            tensor<fp16, [128]> var_19590_to_fp16 = const()[name = string("op_19590_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455314816)))];
            tensor<fp16, [1, 51, 128]> var_19591_cast_fp16 = add(x = q_head_243_cast_fp16, y = var_19590_to_fp16)[name = string("op_19591_cast_fp16")];
            tensor<int32, [1]> q_v_243_axes_1 = const()[name = string("q_v_243_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_243_cast_fp16 = expand_dims(axes = q_v_243_axes_1, x = var_19591_cast_fp16)[name = string("q_v_243_cast_fp16")];
            tensor<int32, [1]> k_head_487_axes_1 = const()[name = string("k_head_487_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_487_cast_fp16 = expand_dims(axes = k_head_487_axes_1, x = k_head_485_cast_fp16)[name = string("k_head_487_cast_fp16")];
            tensor<int32, [1]> v_head_487_axes_1 = const()[name = string("v_head_487_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_487_cast_fp16 = expand_dims(axes = v_head_487_axes_1, x = v_head_485_cast_fp16)[name = string("v_head_487_cast_fp16")];
            tensor<int32, [1]> p_head_487_axes_1 = const()[name = string("p_head_487_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_487_cast_fp16 = expand_dims(axes = p_head_487_axes_1, x = p_head_485_cast_fp16)[name = string("p_head_487_cast_fp16")];
            bool var_19597_transpose_x_3 = const()[name = string("op_19597_transpose_x_3"), val = bool(false)];
            bool var_19597_transpose_y_3 = const()[name = string("op_19597_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_19597_cast_fp16 = matmul(transpose_x = var_19597_transpose_x_3, transpose_y = var_19597_transpose_y_3, x = q_u_243_cast_fp16, y = k_head_487_cast_fp16)[name = string("op_19597_cast_fp16")];
            fp16 var_19598_to_fp16 = const()[name = string("op_19598_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_243_cast_fp16 = mul(x = var_19597_cast_fp16, y = var_19598_to_fp16)[name = string("scores_content_243_cast_fp16")];
            bool x_1289_transpose_x_3 = const()[name = string("x_1289_transpose_x_3"), val = bool(false)];
            bool x_1289_transpose_y_3 = const()[name = string("x_1289_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1289_cast_fp16 = matmul(transpose_x = x_1289_transpose_x_3, transpose_y = x_1289_transpose_y_3, x = q_v_243_cast_fp16, y = p_head_487_cast_fp16)[name = string("x_1289_cast_fp16")];
            tensor<int32, [8]> x_1291_pad_1 = const()[name = string("x_1291_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1291_mode_1 = const()[name = string("x_1291_mode_1"), val = string("constant")];
            fp16 const_2125_to_fp16 = const()[name = string("const_2125_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1291_cast_fp16 = pad(constant_val = const_2125_to_fp16, mode = x_1291_mode_1, pad = x_1291_pad_1, x = x_1289_cast_fp16)[name = string("x_1291_cast_fp16")];
            tensor<int32, [4]> var_19612 = const()[name = string("op_19612"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1293_cast_fp16 = reshape(shape = var_19612, x = x_1291_cast_fp16)[name = string("x_1293_cast_fp16")];
            tensor<int32, [4]> var_19616_begin_1 = const()[name = string("op_19616_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_19616_end_1 = const()[name = string("op_19616_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_19616_end_mask_1 = const()[name = string("op_19616_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_19616_cast_fp16 = slice_by_index(begin = var_19616_begin_1, end = var_19616_end_1, end_mask = var_19616_end_mask_1, x = x_1293_cast_fp16)[name = string("op_19616_cast_fp16")];
            tensor<int32, [4]> var_19618 = const()[name = string("op_19618"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_19619_cast_fp16 = reshape(shape = var_19618, x = var_19616_cast_fp16)[name = string("op_19619_cast_fp16")];
            tensor<int32, [4]> var_19624_begin_1 = const()[name = string("op_19624_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_19624_end_1 = const()[name = string("op_19624_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_19624_end_mask_1 = const()[name = string("op_19624_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_19624_cast_fp16 = slice_by_index(begin = var_19624_begin_1, end = var_19624_end_1, end_mask = var_19624_end_mask_1, x = var_19619_cast_fp16)[name = string("op_19624_cast_fp16")];
            fp16 var_19625_to_fp16 = const()[name = string("op_19625_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_243_cast_fp16 = mul(x = var_19624_cast_fp16, y = var_19625_to_fp16)[name = string("scores_pos_243_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_243_cast_fp16 = add(x = scores_content_243_cast_fp16, y = scores_pos_243_cast_fp16)[name = string("logits_243_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_19628_cast_fp16 = softmax(axis = var_19232, x = logits_243_cast_fp16)[name = string("op_19628_cast_fp16")];
            bool var_19630_transpose_x_1 = const()[name = string("op_19630_transpose_x_1"), val = bool(false)];
            bool var_19630_transpose_y_1 = const()[name = string("op_19630_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_19630_cast_fp16 = matmul(transpose_x = var_19630_transpose_x_1, transpose_y = var_19630_transpose_y_1, x = var_19628_cast_fp16, y = v_head_487_cast_fp16)[name = string("op_19630_cast_fp16")];
            tensor<int32, [1]> var_19631_axes_1 = const()[name = string("op_19631_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_19631_cast_fp16 = squeeze(axes = var_19631_axes_1, x = var_19630_cast_fp16)[name = string("op_19631_cast_fp16")];
            string dense_output_1233_pad_type_1 = const()[name = string("dense_output_1233_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1233_strides_1 = const()[name = string("dense_output_1233_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1233_pad_1 = const()[name = string("dense_output_1233_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1233_dilations_1 = const()[name = string("dense_output_1233_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1233_groups_1 = const()[name = string("dense_output_1233_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455315136))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455446272))))[name = string("layers_15_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1233_cast_fp16 = conv(dilations = dense_output_1233_dilations_1, groups = dense_output_1233_groups_1, pad = dense_output_1233_pad_1, pad_type = dense_output_1233_pad_type_1, strides = dense_output_1233_strides_1, weight = layers_15_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1233_cast_fp16")];
            string sparse_output_1233_pad_type_1 = const()[name = string("sparse_output_1233_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1233_strides_1 = const()[name = string("sparse_output_1233_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1233_pad_1 = const()[name = string("sparse_output_1233_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1233_dilations_1 = const()[name = string("sparse_output_1233_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1233_groups_1 = const()[name = string("sparse_output_1233_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455449536))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455446848))))[name = string("layers_15_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1233_cast_fp16 = conv(dilations = sparse_output_1233_dilations_1, groups = sparse_output_1233_groups_1, pad = sparse_output_1233_pad_1, pad_type = sparse_output_1233_pad_type_1, strides = sparse_output_1233_strides_1, weight = layers_15_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1233_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19646_cast_fp16 = add(x = dense_output_1233_cast_fp16, y = sparse_output_1233_cast_fp16)[name = string("op_19646_cast_fp16")];
            tensor<int32, [4]> var_19647 = const()[name = string("op_19647"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19649 = const()[name = string("op_19649"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19648_cast_fp16 = transpose(perm = var_19647, x = var_19646_cast_fp16)[name = string("transpose_357")];
            tensor<fp16, [1, 51, 128]> q_head_245_cast_fp16 = reshape(shape = var_19649, x = var_19648_cast_fp16)[name = string("q_head_245_cast_fp16")];
            string dense_output_1235_pad_type_1 = const()[name = string("dense_output_1235_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1235_strides_1 = const()[name = string("dense_output_1235_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1235_pad_1 = const()[name = string("dense_output_1235_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1235_dilations_1 = const()[name = string("dense_output_1235_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1235_groups_1 = const()[name = string("dense_output_1235_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455465984))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455597120))))[name = string("layers_15_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1235_cast_fp16 = conv(dilations = dense_output_1235_dilations_1, groups = dense_output_1235_groups_1, pad = dense_output_1235_pad_1, pad_type = dense_output_1235_pad_type_1, strides = dense_output_1235_strides_1, weight = layers_15_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1235_cast_fp16")];
            string sparse_output_1235_pad_type_1 = const()[name = string("sparse_output_1235_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1235_strides_1 = const()[name = string("sparse_output_1235_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1235_pad_1 = const()[name = string("sparse_output_1235_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1235_dilations_1 = const()[name = string("sparse_output_1235_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1235_groups_1 = const()[name = string("sparse_output_1235_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455600384))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455597696))))[name = string("layers_15_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1235_cast_fp16 = conv(dilations = sparse_output_1235_dilations_1, groups = sparse_output_1235_groups_1, pad = sparse_output_1235_pad_1, pad_type = sparse_output_1235_pad_type_1, strides = sparse_output_1235_strides_1, weight = layers_15_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1235_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19665_cast_fp16 = add(x = dense_output_1235_cast_fp16, y = sparse_output_1235_cast_fp16)[name = string("op_19665_cast_fp16")];
            tensor<int32, [4]> var_19666 = const()[name = string("op_19666"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19668 = const()[name = string("op_19668"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19667_cast_fp16 = transpose(perm = var_19666, x = var_19665_cast_fp16)[name = string("transpose_356")];
            tensor<fp16, [1, 51, 128]> k_head_489_cast_fp16 = reshape(shape = var_19668, x = var_19667_cast_fp16)[name = string("k_head_489_cast_fp16")];
            string dense_output_1237_pad_type_1 = const()[name = string("dense_output_1237_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1237_strides_1 = const()[name = string("dense_output_1237_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1237_pad_1 = const()[name = string("dense_output_1237_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1237_dilations_1 = const()[name = string("dense_output_1237_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1237_groups_1 = const()[name = string("dense_output_1237_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455616832))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455747968))))[name = string("layers_15_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1237_cast_fp16 = conv(dilations = dense_output_1237_dilations_1, groups = dense_output_1237_groups_1, pad = dense_output_1237_pad_1, pad_type = dense_output_1237_pad_type_1, strides = dense_output_1237_strides_1, weight = layers_15_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1237_cast_fp16")];
            string sparse_output_1237_pad_type_1 = const()[name = string("sparse_output_1237_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1237_strides_1 = const()[name = string("sparse_output_1237_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1237_pad_1 = const()[name = string("sparse_output_1237_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1237_dilations_1 = const()[name = string("sparse_output_1237_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1237_groups_1 = const()[name = string("sparse_output_1237_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455751232))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455748544))))[name = string("layers_15_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1237_cast_fp16 = conv(dilations = sparse_output_1237_dilations_1, groups = sparse_output_1237_groups_1, pad = sparse_output_1237_pad_1, pad_type = sparse_output_1237_pad_type_1, strides = sparse_output_1237_strides_1, weight = layers_15_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1237_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19684_cast_fp16 = add(x = dense_output_1237_cast_fp16, y = sparse_output_1237_cast_fp16)[name = string("op_19684_cast_fp16")];
            tensor<int32, [4]> var_19685 = const()[name = string("op_19685"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19687 = const()[name = string("op_19687"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19686_cast_fp16 = transpose(perm = var_19685, x = var_19684_cast_fp16)[name = string("transpose_355")];
            tensor<fp16, [1, 51, 128]> v_head_489_cast_fp16 = reshape(shape = var_19687, x = var_19686_cast_fp16)[name = string("v_head_489_cast_fp16")];
            string dense_output_1239_pad_type_1 = const()[name = string("dense_output_1239_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1239_strides_1 = const()[name = string("dense_output_1239_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1239_pad_1 = const()[name = string("dense_output_1239_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1239_dilations_1 = const()[name = string("dense_output_1239_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1239_groups_1 = const()[name = string("dense_output_1239_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455767680))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455898816))))[name = string("layers_15_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1239_cast_fp16 = conv(dilations = dense_output_1239_dilations_1, groups = dense_output_1239_groups_1, pad = dense_output_1239_pad_1, pad_type = dense_output_1239_pad_type_1, strides = dense_output_1239_strides_1, weight = layers_15_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1239_cast_fp16")];
            string sparse_output_1239_pad_type_1 = const()[name = string("sparse_output_1239_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1239_strides_1 = const()[name = string("sparse_output_1239_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1239_pad_1 = const()[name = string("sparse_output_1239_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1239_dilations_1 = const()[name = string("sparse_output_1239_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1239_groups_1 = const()[name = string("sparse_output_1239_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455902080))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455899392))))[name = string("layers_15_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1239_cast_fp16 = conv(dilations = sparse_output_1239_dilations_1, groups = sparse_output_1239_groups_1, pad = sparse_output_1239_pad_1, pad_type = sparse_output_1239_pad_type_1, strides = sparse_output_1239_strides_1, weight = layers_15_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1239_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_19703_cast_fp16 = add(x = dense_output_1239_cast_fp16, y = sparse_output_1239_cast_fp16)[name = string("op_19703_cast_fp16")];
            tensor<int32, [4]> var_19704 = const()[name = string("op_19704"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19706 = const()[name = string("op_19706"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_19705_cast_fp16 = transpose(perm = var_19704, x = var_19703_cast_fp16)[name = string("transpose_354")];
            tensor<fp16, [1, 101, 128]> p_head_489_cast_fp16 = reshape(shape = var_19706, x = var_19705_cast_fp16)[name = string("p_head_489_cast_fp16")];
            tensor<fp16, [128]> var_19708_to_fp16 = const()[name = string("op_19708_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455918528)))];
            tensor<fp16, [1, 51, 128]> var_19709_cast_fp16 = add(x = q_head_245_cast_fp16, y = var_19708_to_fp16)[name = string("op_19709_cast_fp16")];
            tensor<int32, [1]> q_u_245_axes_1 = const()[name = string("q_u_245_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_245_cast_fp16 = expand_dims(axes = q_u_245_axes_1, x = var_19709_cast_fp16)[name = string("q_u_245_cast_fp16")];
            tensor<fp16, [128]> var_19711_to_fp16 = const()[name = string("op_19711_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455918848)))];
            tensor<fp16, [1, 51, 128]> var_19712_cast_fp16 = add(x = q_head_245_cast_fp16, y = var_19711_to_fp16)[name = string("op_19712_cast_fp16")];
            tensor<int32, [1]> q_v_245_axes_1 = const()[name = string("q_v_245_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_245_cast_fp16 = expand_dims(axes = q_v_245_axes_1, x = var_19712_cast_fp16)[name = string("q_v_245_cast_fp16")];
            tensor<int32, [1]> k_head_491_axes_1 = const()[name = string("k_head_491_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_491_cast_fp16 = expand_dims(axes = k_head_491_axes_1, x = k_head_489_cast_fp16)[name = string("k_head_491_cast_fp16")];
            tensor<int32, [1]> v_head_491_axes_1 = const()[name = string("v_head_491_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_491_cast_fp16 = expand_dims(axes = v_head_491_axes_1, x = v_head_489_cast_fp16)[name = string("v_head_491_cast_fp16")];
            tensor<int32, [1]> p_head_491_axes_1 = const()[name = string("p_head_491_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_491_cast_fp16 = expand_dims(axes = p_head_491_axes_1, x = p_head_489_cast_fp16)[name = string("p_head_491_cast_fp16")];
            bool var_19718_transpose_x_3 = const()[name = string("op_19718_transpose_x_3"), val = bool(false)];
            bool var_19718_transpose_y_3 = const()[name = string("op_19718_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_19718_cast_fp16 = matmul(transpose_x = var_19718_transpose_x_3, transpose_y = var_19718_transpose_y_3, x = q_u_245_cast_fp16, y = k_head_491_cast_fp16)[name = string("op_19718_cast_fp16")];
            fp16 var_19719_to_fp16 = const()[name = string("op_19719_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_245_cast_fp16 = mul(x = var_19718_cast_fp16, y = var_19719_to_fp16)[name = string("scores_content_245_cast_fp16")];
            bool x_1297_transpose_x_3 = const()[name = string("x_1297_transpose_x_3"), val = bool(false)];
            bool x_1297_transpose_y_3 = const()[name = string("x_1297_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1297_cast_fp16 = matmul(transpose_x = x_1297_transpose_x_3, transpose_y = x_1297_transpose_y_3, x = q_v_245_cast_fp16, y = p_head_491_cast_fp16)[name = string("x_1297_cast_fp16")];
            tensor<int32, [8]> x_1299_pad_1 = const()[name = string("x_1299_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1299_mode_1 = const()[name = string("x_1299_mode_1"), val = string("constant")];
            fp16 const_2131_to_fp16 = const()[name = string("const_2131_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1299_cast_fp16 = pad(constant_val = const_2131_to_fp16, mode = x_1299_mode_1, pad = x_1299_pad_1, x = x_1297_cast_fp16)[name = string("x_1299_cast_fp16")];
            tensor<int32, [4]> var_19733 = const()[name = string("op_19733"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1301_cast_fp16 = reshape(shape = var_19733, x = x_1299_cast_fp16)[name = string("x_1301_cast_fp16")];
            tensor<int32, [4]> var_19737_begin_1 = const()[name = string("op_19737_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_19737_end_1 = const()[name = string("op_19737_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_19737_end_mask_1 = const()[name = string("op_19737_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_19737_cast_fp16 = slice_by_index(begin = var_19737_begin_1, end = var_19737_end_1, end_mask = var_19737_end_mask_1, x = x_1301_cast_fp16)[name = string("op_19737_cast_fp16")];
            tensor<int32, [4]> var_19739 = const()[name = string("op_19739"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_19740_cast_fp16 = reshape(shape = var_19739, x = var_19737_cast_fp16)[name = string("op_19740_cast_fp16")];
            tensor<int32, [4]> var_19745_begin_1 = const()[name = string("op_19745_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_19745_end_1 = const()[name = string("op_19745_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_19745_end_mask_1 = const()[name = string("op_19745_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_19745_cast_fp16 = slice_by_index(begin = var_19745_begin_1, end = var_19745_end_1, end_mask = var_19745_end_mask_1, x = var_19740_cast_fp16)[name = string("op_19745_cast_fp16")];
            fp16 var_19746_to_fp16 = const()[name = string("op_19746_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_245_cast_fp16 = mul(x = var_19745_cast_fp16, y = var_19746_to_fp16)[name = string("scores_pos_245_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_245_cast_fp16 = add(x = scores_content_245_cast_fp16, y = scores_pos_245_cast_fp16)[name = string("logits_245_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_19749_cast_fp16 = softmax(axis = var_19232, x = logits_245_cast_fp16)[name = string("op_19749_cast_fp16")];
            bool var_19751_transpose_x_1 = const()[name = string("op_19751_transpose_x_1"), val = bool(false)];
            bool var_19751_transpose_y_1 = const()[name = string("op_19751_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_19751_cast_fp16 = matmul(transpose_x = var_19751_transpose_x_1, transpose_y = var_19751_transpose_y_1, x = var_19749_cast_fp16, y = v_head_491_cast_fp16)[name = string("op_19751_cast_fp16")];
            tensor<int32, [1]> var_19752_axes_1 = const()[name = string("op_19752_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_19752_cast_fp16 = squeeze(axes = var_19752_axes_1, x = var_19751_cast_fp16)[name = string("op_19752_cast_fp16")];
            string dense_output_1241_pad_type_1 = const()[name = string("dense_output_1241_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1241_strides_1 = const()[name = string("dense_output_1241_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1241_pad_1 = const()[name = string("dense_output_1241_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1241_dilations_1 = const()[name = string("dense_output_1241_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1241_groups_1 = const()[name = string("dense_output_1241_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455919168))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456050304))))[name = string("layers_15_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1241_cast_fp16 = conv(dilations = dense_output_1241_dilations_1, groups = dense_output_1241_groups_1, pad = dense_output_1241_pad_1, pad_type = dense_output_1241_pad_type_1, strides = dense_output_1241_strides_1, weight = layers_15_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1241_cast_fp16")];
            string sparse_output_1241_pad_type_1 = const()[name = string("sparse_output_1241_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1241_strides_1 = const()[name = string("sparse_output_1241_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1241_pad_1 = const()[name = string("sparse_output_1241_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1241_dilations_1 = const()[name = string("sparse_output_1241_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1241_groups_1 = const()[name = string("sparse_output_1241_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456053568))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456050880))))[name = string("layers_15_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1241_cast_fp16 = conv(dilations = sparse_output_1241_dilations_1, groups = sparse_output_1241_groups_1, pad = sparse_output_1241_pad_1, pad_type = sparse_output_1241_pad_type_1, strides = sparse_output_1241_strides_1, weight = layers_15_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1241_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19767_cast_fp16 = add(x = dense_output_1241_cast_fp16, y = sparse_output_1241_cast_fp16)[name = string("op_19767_cast_fp16")];
            tensor<int32, [4]> var_19768 = const()[name = string("op_19768"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19770 = const()[name = string("op_19770"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19769_cast_fp16 = transpose(perm = var_19768, x = var_19767_cast_fp16)[name = string("transpose_353")];
            tensor<fp16, [1, 51, 128]> q_head_247_cast_fp16 = reshape(shape = var_19770, x = var_19769_cast_fp16)[name = string("q_head_247_cast_fp16")];
            string dense_output_1243_pad_type_1 = const()[name = string("dense_output_1243_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1243_strides_1 = const()[name = string("dense_output_1243_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1243_pad_1 = const()[name = string("dense_output_1243_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1243_dilations_1 = const()[name = string("dense_output_1243_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1243_groups_1 = const()[name = string("dense_output_1243_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456070016))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456201152))))[name = string("layers_15_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1243_cast_fp16 = conv(dilations = dense_output_1243_dilations_1, groups = dense_output_1243_groups_1, pad = dense_output_1243_pad_1, pad_type = dense_output_1243_pad_type_1, strides = dense_output_1243_strides_1, weight = layers_15_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1243_cast_fp16")];
            string sparse_output_1243_pad_type_1 = const()[name = string("sparse_output_1243_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1243_strides_1 = const()[name = string("sparse_output_1243_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1243_pad_1 = const()[name = string("sparse_output_1243_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1243_dilations_1 = const()[name = string("sparse_output_1243_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1243_groups_1 = const()[name = string("sparse_output_1243_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456204416))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456201728))))[name = string("layers_15_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1243_cast_fp16 = conv(dilations = sparse_output_1243_dilations_1, groups = sparse_output_1243_groups_1, pad = sparse_output_1243_pad_1, pad_type = sparse_output_1243_pad_type_1, strides = sparse_output_1243_strides_1, weight = layers_15_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1243_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19786_cast_fp16 = add(x = dense_output_1243_cast_fp16, y = sparse_output_1243_cast_fp16)[name = string("op_19786_cast_fp16")];
            tensor<int32, [4]> var_19787 = const()[name = string("op_19787"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19789 = const()[name = string("op_19789"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19788_cast_fp16 = transpose(perm = var_19787, x = var_19786_cast_fp16)[name = string("transpose_352")];
            tensor<fp16, [1, 51, 128]> k_head_493_cast_fp16 = reshape(shape = var_19789, x = var_19788_cast_fp16)[name = string("k_head_493_cast_fp16")];
            string dense_output_1245_pad_type_1 = const()[name = string("dense_output_1245_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1245_strides_1 = const()[name = string("dense_output_1245_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1245_pad_1 = const()[name = string("dense_output_1245_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1245_dilations_1 = const()[name = string("dense_output_1245_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1245_groups_1 = const()[name = string("dense_output_1245_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456220864))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456352000))))[name = string("layers_15_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1245_cast_fp16 = conv(dilations = dense_output_1245_dilations_1, groups = dense_output_1245_groups_1, pad = dense_output_1245_pad_1, pad_type = dense_output_1245_pad_type_1, strides = dense_output_1245_strides_1, weight = layers_15_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1245_cast_fp16")];
            string sparse_output_1245_pad_type_1 = const()[name = string("sparse_output_1245_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1245_strides_1 = const()[name = string("sparse_output_1245_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1245_pad_1 = const()[name = string("sparse_output_1245_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1245_dilations_1 = const()[name = string("sparse_output_1245_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1245_groups_1 = const()[name = string("sparse_output_1245_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456355264))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456352576))))[name = string("layers_15_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1245_cast_fp16 = conv(dilations = sparse_output_1245_dilations_1, groups = sparse_output_1245_groups_1, pad = sparse_output_1245_pad_1, pad_type = sparse_output_1245_pad_type_1, strides = sparse_output_1245_strides_1, weight = layers_15_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1245_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19805_cast_fp16 = add(x = dense_output_1245_cast_fp16, y = sparse_output_1245_cast_fp16)[name = string("op_19805_cast_fp16")];
            tensor<int32, [4]> var_19806 = const()[name = string("op_19806"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19808 = const()[name = string("op_19808"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19807_cast_fp16 = transpose(perm = var_19806, x = var_19805_cast_fp16)[name = string("transpose_351")];
            tensor<fp16, [1, 51, 128]> v_head_493_cast_fp16 = reshape(shape = var_19808, x = var_19807_cast_fp16)[name = string("v_head_493_cast_fp16")];
            string dense_output_1247_pad_type_1 = const()[name = string("dense_output_1247_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1247_strides_1 = const()[name = string("dense_output_1247_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1247_pad_1 = const()[name = string("dense_output_1247_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1247_dilations_1 = const()[name = string("dense_output_1247_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1247_groups_1 = const()[name = string("dense_output_1247_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456371712))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456502848))))[name = string("layers_15_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1247_cast_fp16 = conv(dilations = dense_output_1247_dilations_1, groups = dense_output_1247_groups_1, pad = dense_output_1247_pad_1, pad_type = dense_output_1247_pad_type_1, strides = dense_output_1247_strides_1, weight = layers_15_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1247_cast_fp16")];
            string sparse_output_1247_pad_type_1 = const()[name = string("sparse_output_1247_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1247_strides_1 = const()[name = string("sparse_output_1247_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1247_pad_1 = const()[name = string("sparse_output_1247_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1247_dilations_1 = const()[name = string("sparse_output_1247_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1247_groups_1 = const()[name = string("sparse_output_1247_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456506112))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456503424))))[name = string("layers_15_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1247_cast_fp16 = conv(dilations = sparse_output_1247_dilations_1, groups = sparse_output_1247_groups_1, pad = sparse_output_1247_pad_1, pad_type = sparse_output_1247_pad_type_1, strides = sparse_output_1247_strides_1, weight = layers_15_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1247_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_19824_cast_fp16 = add(x = dense_output_1247_cast_fp16, y = sparse_output_1247_cast_fp16)[name = string("op_19824_cast_fp16")];
            tensor<int32, [4]> var_19825 = const()[name = string("op_19825"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19827 = const()[name = string("op_19827"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_19826_cast_fp16 = transpose(perm = var_19825, x = var_19824_cast_fp16)[name = string("transpose_350")];
            tensor<fp16, [1, 101, 128]> p_head_493_cast_fp16 = reshape(shape = var_19827, x = var_19826_cast_fp16)[name = string("p_head_493_cast_fp16")];
            tensor<fp16, [128]> var_19829_to_fp16 = const()[name = string("op_19829_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456522560)))];
            tensor<fp16, [1, 51, 128]> var_19830_cast_fp16 = add(x = q_head_247_cast_fp16, y = var_19829_to_fp16)[name = string("op_19830_cast_fp16")];
            tensor<int32, [1]> q_u_247_axes_1 = const()[name = string("q_u_247_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_247_cast_fp16 = expand_dims(axes = q_u_247_axes_1, x = var_19830_cast_fp16)[name = string("q_u_247_cast_fp16")];
            tensor<fp16, [128]> var_19832_to_fp16 = const()[name = string("op_19832_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456522880)))];
            tensor<fp16, [1, 51, 128]> var_19833_cast_fp16 = add(x = q_head_247_cast_fp16, y = var_19832_to_fp16)[name = string("op_19833_cast_fp16")];
            tensor<int32, [1]> q_v_247_axes_1 = const()[name = string("q_v_247_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_247_cast_fp16 = expand_dims(axes = q_v_247_axes_1, x = var_19833_cast_fp16)[name = string("q_v_247_cast_fp16")];
            tensor<int32, [1]> k_head_495_axes_1 = const()[name = string("k_head_495_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_495_cast_fp16 = expand_dims(axes = k_head_495_axes_1, x = k_head_493_cast_fp16)[name = string("k_head_495_cast_fp16")];
            tensor<int32, [1]> v_head_495_axes_1 = const()[name = string("v_head_495_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_495_cast_fp16 = expand_dims(axes = v_head_495_axes_1, x = v_head_493_cast_fp16)[name = string("v_head_495_cast_fp16")];
            tensor<int32, [1]> p_head_495_axes_1 = const()[name = string("p_head_495_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_495_cast_fp16 = expand_dims(axes = p_head_495_axes_1, x = p_head_493_cast_fp16)[name = string("p_head_495_cast_fp16")];
            bool var_19839_transpose_x_3 = const()[name = string("op_19839_transpose_x_3"), val = bool(false)];
            bool var_19839_transpose_y_3 = const()[name = string("op_19839_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_19839_cast_fp16 = matmul(transpose_x = var_19839_transpose_x_3, transpose_y = var_19839_transpose_y_3, x = q_u_247_cast_fp16, y = k_head_495_cast_fp16)[name = string("op_19839_cast_fp16")];
            fp16 var_19840_to_fp16 = const()[name = string("op_19840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_247_cast_fp16 = mul(x = var_19839_cast_fp16, y = var_19840_to_fp16)[name = string("scores_content_247_cast_fp16")];
            bool x_1305_transpose_x_3 = const()[name = string("x_1305_transpose_x_3"), val = bool(false)];
            bool x_1305_transpose_y_3 = const()[name = string("x_1305_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1305_cast_fp16 = matmul(transpose_x = x_1305_transpose_x_3, transpose_y = x_1305_transpose_y_3, x = q_v_247_cast_fp16, y = p_head_495_cast_fp16)[name = string("x_1305_cast_fp16")];
            tensor<int32, [8]> x_1307_pad_1 = const()[name = string("x_1307_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1307_mode_1 = const()[name = string("x_1307_mode_1"), val = string("constant")];
            fp16 const_2137_to_fp16 = const()[name = string("const_2137_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1307_cast_fp16 = pad(constant_val = const_2137_to_fp16, mode = x_1307_mode_1, pad = x_1307_pad_1, x = x_1305_cast_fp16)[name = string("x_1307_cast_fp16")];
            tensor<int32, [4]> var_19854 = const()[name = string("op_19854"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1309_cast_fp16 = reshape(shape = var_19854, x = x_1307_cast_fp16)[name = string("x_1309_cast_fp16")];
            tensor<int32, [4]> var_19858_begin_1 = const()[name = string("op_19858_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_19858_end_1 = const()[name = string("op_19858_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_19858_end_mask_1 = const()[name = string("op_19858_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_19858_cast_fp16 = slice_by_index(begin = var_19858_begin_1, end = var_19858_end_1, end_mask = var_19858_end_mask_1, x = x_1309_cast_fp16)[name = string("op_19858_cast_fp16")];
            tensor<int32, [4]> var_19860 = const()[name = string("op_19860"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_19861_cast_fp16 = reshape(shape = var_19860, x = var_19858_cast_fp16)[name = string("op_19861_cast_fp16")];
            tensor<int32, [4]> var_19866_begin_1 = const()[name = string("op_19866_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_19866_end_1 = const()[name = string("op_19866_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_19866_end_mask_1 = const()[name = string("op_19866_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_19866_cast_fp16 = slice_by_index(begin = var_19866_begin_1, end = var_19866_end_1, end_mask = var_19866_end_mask_1, x = var_19861_cast_fp16)[name = string("op_19866_cast_fp16")];
            fp16 var_19867_to_fp16 = const()[name = string("op_19867_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_247_cast_fp16 = mul(x = var_19866_cast_fp16, y = var_19867_to_fp16)[name = string("scores_pos_247_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_247_cast_fp16 = add(x = scores_content_247_cast_fp16, y = scores_pos_247_cast_fp16)[name = string("logits_247_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_19870_cast_fp16 = softmax(axis = var_19232, x = logits_247_cast_fp16)[name = string("op_19870_cast_fp16")];
            bool var_19872_transpose_x_1 = const()[name = string("op_19872_transpose_x_1"), val = bool(false)];
            bool var_19872_transpose_y_1 = const()[name = string("op_19872_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_19872_cast_fp16 = matmul(transpose_x = var_19872_transpose_x_1, transpose_y = var_19872_transpose_y_1, x = var_19870_cast_fp16, y = v_head_495_cast_fp16)[name = string("op_19872_cast_fp16")];
            tensor<int32, [1]> var_19873_axes_1 = const()[name = string("op_19873_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_19873_cast_fp16 = squeeze(axes = var_19873_axes_1, x = var_19872_cast_fp16)[name = string("op_19873_cast_fp16")];
            string dense_output_1249_pad_type_1 = const()[name = string("dense_output_1249_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1249_strides_1 = const()[name = string("dense_output_1249_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1249_pad_1 = const()[name = string("dense_output_1249_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1249_dilations_1 = const()[name = string("dense_output_1249_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1249_groups_1 = const()[name = string("dense_output_1249_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456523200))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456654336))))[name = string("layers_15_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1249_cast_fp16 = conv(dilations = dense_output_1249_dilations_1, groups = dense_output_1249_groups_1, pad = dense_output_1249_pad_1, pad_type = dense_output_1249_pad_type_1, strides = dense_output_1249_strides_1, weight = layers_15_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1249_cast_fp16")];
            string sparse_output_1249_pad_type_1 = const()[name = string("sparse_output_1249_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1249_strides_1 = const()[name = string("sparse_output_1249_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1249_pad_1 = const()[name = string("sparse_output_1249_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1249_dilations_1 = const()[name = string("sparse_output_1249_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1249_groups_1 = const()[name = string("sparse_output_1249_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456657600))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456654912))))[name = string("layers_15_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1249_cast_fp16 = conv(dilations = sparse_output_1249_dilations_1, groups = sparse_output_1249_groups_1, pad = sparse_output_1249_pad_1, pad_type = sparse_output_1249_pad_type_1, strides = sparse_output_1249_strides_1, weight = layers_15_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1249_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19888_cast_fp16 = add(x = dense_output_1249_cast_fp16, y = sparse_output_1249_cast_fp16)[name = string("op_19888_cast_fp16")];
            tensor<int32, [4]> var_19889 = const()[name = string("op_19889"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19891 = const()[name = string("op_19891"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19890_cast_fp16 = transpose(perm = var_19889, x = var_19888_cast_fp16)[name = string("transpose_349")];
            tensor<fp16, [1, 51, 128]> q_head_249_cast_fp16 = reshape(shape = var_19891, x = var_19890_cast_fp16)[name = string("q_head_249_cast_fp16")];
            string dense_output_1251_pad_type_1 = const()[name = string("dense_output_1251_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1251_strides_1 = const()[name = string("dense_output_1251_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1251_pad_1 = const()[name = string("dense_output_1251_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1251_dilations_1 = const()[name = string("dense_output_1251_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1251_groups_1 = const()[name = string("dense_output_1251_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456674048))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456805184))))[name = string("layers_15_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1251_cast_fp16 = conv(dilations = dense_output_1251_dilations_1, groups = dense_output_1251_groups_1, pad = dense_output_1251_pad_1, pad_type = dense_output_1251_pad_type_1, strides = dense_output_1251_strides_1, weight = layers_15_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1251_cast_fp16")];
            string sparse_output_1251_pad_type_1 = const()[name = string("sparse_output_1251_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1251_strides_1 = const()[name = string("sparse_output_1251_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1251_pad_1 = const()[name = string("sparse_output_1251_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1251_dilations_1 = const()[name = string("sparse_output_1251_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1251_groups_1 = const()[name = string("sparse_output_1251_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456808448))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456805760))))[name = string("layers_15_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1251_cast_fp16 = conv(dilations = sparse_output_1251_dilations_1, groups = sparse_output_1251_groups_1, pad = sparse_output_1251_pad_1, pad_type = sparse_output_1251_pad_type_1, strides = sparse_output_1251_strides_1, weight = layers_15_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1251_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19907_cast_fp16 = add(x = dense_output_1251_cast_fp16, y = sparse_output_1251_cast_fp16)[name = string("op_19907_cast_fp16")];
            tensor<int32, [4]> var_19908 = const()[name = string("op_19908"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19910 = const()[name = string("op_19910"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19909_cast_fp16 = transpose(perm = var_19908, x = var_19907_cast_fp16)[name = string("transpose_348")];
            tensor<fp16, [1, 51, 128]> k_head_497_cast_fp16 = reshape(shape = var_19910, x = var_19909_cast_fp16)[name = string("k_head_497_cast_fp16")];
            string dense_output_1253_pad_type_1 = const()[name = string("dense_output_1253_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1253_strides_1 = const()[name = string("dense_output_1253_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1253_pad_1 = const()[name = string("dense_output_1253_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1253_dilations_1 = const()[name = string("dense_output_1253_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1253_groups_1 = const()[name = string("dense_output_1253_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456824896))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456956032))))[name = string("layers_15_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1253_cast_fp16 = conv(dilations = dense_output_1253_dilations_1, groups = dense_output_1253_groups_1, pad = dense_output_1253_pad_1, pad_type = dense_output_1253_pad_type_1, strides = dense_output_1253_strides_1, weight = layers_15_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1253_cast_fp16")];
            string sparse_output_1253_pad_type_1 = const()[name = string("sparse_output_1253_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1253_strides_1 = const()[name = string("sparse_output_1253_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1253_pad_1 = const()[name = string("sparse_output_1253_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1253_dilations_1 = const()[name = string("sparse_output_1253_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1253_groups_1 = const()[name = string("sparse_output_1253_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456959296))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456956608))))[name = string("layers_15_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1253_cast_fp16 = conv(dilations = sparse_output_1253_dilations_1, groups = sparse_output_1253_groups_1, pad = sparse_output_1253_pad_1, pad_type = sparse_output_1253_pad_type_1, strides = sparse_output_1253_strides_1, weight = layers_15_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1253_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_19926_cast_fp16 = add(x = dense_output_1253_cast_fp16, y = sparse_output_1253_cast_fp16)[name = string("op_19926_cast_fp16")];
            tensor<int32, [4]> var_19927 = const()[name = string("op_19927"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19929 = const()[name = string("op_19929"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_19928_cast_fp16 = transpose(perm = var_19927, x = var_19926_cast_fp16)[name = string("transpose_347")];
            tensor<fp16, [1, 51, 128]> v_head_497_cast_fp16 = reshape(shape = var_19929, x = var_19928_cast_fp16)[name = string("v_head_497_cast_fp16")];
            string dense_output_1255_pad_type_1 = const()[name = string("dense_output_1255_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1255_strides_1 = const()[name = string("dense_output_1255_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1255_pad_1 = const()[name = string("dense_output_1255_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1255_dilations_1 = const()[name = string("dense_output_1255_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1255_groups_1 = const()[name = string("dense_output_1255_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456975744))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457106880))))[name = string("layers_15_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1255_cast_fp16 = conv(dilations = dense_output_1255_dilations_1, groups = dense_output_1255_groups_1, pad = dense_output_1255_pad_1, pad_type = dense_output_1255_pad_type_1, strides = dense_output_1255_strides_1, weight = layers_15_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1255_cast_fp16")];
            string sparse_output_1255_pad_type_1 = const()[name = string("sparse_output_1255_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1255_strides_1 = const()[name = string("sparse_output_1255_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1255_pad_1 = const()[name = string("sparse_output_1255_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1255_dilations_1 = const()[name = string("sparse_output_1255_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1255_groups_1 = const()[name = string("sparse_output_1255_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457110144))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457107456))))[name = string("layers_15_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1255_cast_fp16 = conv(dilations = sparse_output_1255_dilations_1, groups = sparse_output_1255_groups_1, pad = sparse_output_1255_pad_1, pad_type = sparse_output_1255_pad_type_1, strides = sparse_output_1255_strides_1, weight = layers_15_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1255_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_19945_cast_fp16 = add(x = dense_output_1255_cast_fp16, y = sparse_output_1255_cast_fp16)[name = string("op_19945_cast_fp16")];
            tensor<int32, [4]> var_19946 = const()[name = string("op_19946"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_19948 = const()[name = string("op_19948"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_19947_cast_fp16 = transpose(perm = var_19946, x = var_19945_cast_fp16)[name = string("transpose_346")];
            tensor<fp16, [1, 101, 128]> p_head_497_cast_fp16 = reshape(shape = var_19948, x = var_19947_cast_fp16)[name = string("p_head_497_cast_fp16")];
            tensor<fp16, [128]> var_19950_to_fp16 = const()[name = string("op_19950_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457126592)))];
            tensor<fp16, [1, 51, 128]> var_19951_cast_fp16 = add(x = q_head_249_cast_fp16, y = var_19950_to_fp16)[name = string("op_19951_cast_fp16")];
            tensor<int32, [1]> q_u_249_axes_1 = const()[name = string("q_u_249_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_249_cast_fp16 = expand_dims(axes = q_u_249_axes_1, x = var_19951_cast_fp16)[name = string("q_u_249_cast_fp16")];
            tensor<fp16, [128]> var_19953_to_fp16 = const()[name = string("op_19953_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457126912)))];
            tensor<fp16, [1, 51, 128]> var_19954_cast_fp16 = add(x = q_head_249_cast_fp16, y = var_19953_to_fp16)[name = string("op_19954_cast_fp16")];
            tensor<int32, [1]> q_v_249_axes_1 = const()[name = string("q_v_249_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_249_cast_fp16 = expand_dims(axes = q_v_249_axes_1, x = var_19954_cast_fp16)[name = string("q_v_249_cast_fp16")];
            tensor<int32, [1]> k_head_499_axes_1 = const()[name = string("k_head_499_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_499_cast_fp16 = expand_dims(axes = k_head_499_axes_1, x = k_head_497_cast_fp16)[name = string("k_head_499_cast_fp16")];
            tensor<int32, [1]> v_head_499_axes_1 = const()[name = string("v_head_499_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_499_cast_fp16 = expand_dims(axes = v_head_499_axes_1, x = v_head_497_cast_fp16)[name = string("v_head_499_cast_fp16")];
            tensor<int32, [1]> p_head_499_axes_1 = const()[name = string("p_head_499_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_499_cast_fp16 = expand_dims(axes = p_head_499_axes_1, x = p_head_497_cast_fp16)[name = string("p_head_499_cast_fp16")];
            bool var_19960_transpose_x_3 = const()[name = string("op_19960_transpose_x_3"), val = bool(false)];
            bool var_19960_transpose_y_3 = const()[name = string("op_19960_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_19960_cast_fp16 = matmul(transpose_x = var_19960_transpose_x_3, transpose_y = var_19960_transpose_y_3, x = q_u_249_cast_fp16, y = k_head_499_cast_fp16)[name = string("op_19960_cast_fp16")];
            fp16 var_19961_to_fp16 = const()[name = string("op_19961_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_249_cast_fp16 = mul(x = var_19960_cast_fp16, y = var_19961_to_fp16)[name = string("scores_content_249_cast_fp16")];
            bool x_1313_transpose_x_3 = const()[name = string("x_1313_transpose_x_3"), val = bool(false)];
            bool x_1313_transpose_y_3 = const()[name = string("x_1313_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1313_cast_fp16 = matmul(transpose_x = x_1313_transpose_x_3, transpose_y = x_1313_transpose_y_3, x = q_v_249_cast_fp16, y = p_head_499_cast_fp16)[name = string("x_1313_cast_fp16")];
            tensor<int32, [8]> x_1315_pad_1 = const()[name = string("x_1315_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1315_mode_1 = const()[name = string("x_1315_mode_1"), val = string("constant")];
            fp16 const_2143_to_fp16 = const()[name = string("const_2143_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1315_cast_fp16 = pad(constant_val = const_2143_to_fp16, mode = x_1315_mode_1, pad = x_1315_pad_1, x = x_1313_cast_fp16)[name = string("x_1315_cast_fp16")];
            tensor<int32, [4]> var_19975 = const()[name = string("op_19975"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1317_cast_fp16 = reshape(shape = var_19975, x = x_1315_cast_fp16)[name = string("x_1317_cast_fp16")];
            tensor<int32, [4]> var_19979_begin_1 = const()[name = string("op_19979_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_19979_end_1 = const()[name = string("op_19979_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_19979_end_mask_1 = const()[name = string("op_19979_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_19979_cast_fp16 = slice_by_index(begin = var_19979_begin_1, end = var_19979_end_1, end_mask = var_19979_end_mask_1, x = x_1317_cast_fp16)[name = string("op_19979_cast_fp16")];
            tensor<int32, [4]> var_19981 = const()[name = string("op_19981"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_19982_cast_fp16 = reshape(shape = var_19981, x = var_19979_cast_fp16)[name = string("op_19982_cast_fp16")];
            tensor<int32, [4]> var_19987_begin_1 = const()[name = string("op_19987_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_19987_end_1 = const()[name = string("op_19987_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_19987_end_mask_1 = const()[name = string("op_19987_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_19987_cast_fp16 = slice_by_index(begin = var_19987_begin_1, end = var_19987_end_1, end_mask = var_19987_end_mask_1, x = var_19982_cast_fp16)[name = string("op_19987_cast_fp16")];
            fp16 var_19988_to_fp16 = const()[name = string("op_19988_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_249_cast_fp16 = mul(x = var_19987_cast_fp16, y = var_19988_to_fp16)[name = string("scores_pos_249_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_249_cast_fp16 = add(x = scores_content_249_cast_fp16, y = scores_pos_249_cast_fp16)[name = string("logits_249_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_19991_cast_fp16 = softmax(axis = var_19232, x = logits_249_cast_fp16)[name = string("op_19991_cast_fp16")];
            bool var_19993_transpose_x_1 = const()[name = string("op_19993_transpose_x_1"), val = bool(false)];
            bool var_19993_transpose_y_1 = const()[name = string("op_19993_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_19993_cast_fp16 = matmul(transpose_x = var_19993_transpose_x_1, transpose_y = var_19993_transpose_y_1, x = var_19991_cast_fp16, y = v_head_499_cast_fp16)[name = string("op_19993_cast_fp16")];
            tensor<int32, [1]> var_19994_axes_1 = const()[name = string("op_19994_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_19994_cast_fp16 = squeeze(axes = var_19994_axes_1, x = var_19993_cast_fp16)[name = string("op_19994_cast_fp16")];
            string dense_output_1257_pad_type_1 = const()[name = string("dense_output_1257_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1257_strides_1 = const()[name = string("dense_output_1257_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1257_pad_1 = const()[name = string("dense_output_1257_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1257_dilations_1 = const()[name = string("dense_output_1257_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1257_groups_1 = const()[name = string("dense_output_1257_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457127232))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457258368))))[name = string("layers_15_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1257_cast_fp16 = conv(dilations = dense_output_1257_dilations_1, groups = dense_output_1257_groups_1, pad = dense_output_1257_pad_1, pad_type = dense_output_1257_pad_type_1, strides = dense_output_1257_strides_1, weight = layers_15_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1257_cast_fp16")];
            string sparse_output_1257_pad_type_1 = const()[name = string("sparse_output_1257_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1257_strides_1 = const()[name = string("sparse_output_1257_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1257_pad_1 = const()[name = string("sparse_output_1257_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1257_dilations_1 = const()[name = string("sparse_output_1257_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1257_groups_1 = const()[name = string("sparse_output_1257_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457261632))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457258944))))[name = string("layers_15_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1257_cast_fp16 = conv(dilations = sparse_output_1257_dilations_1, groups = sparse_output_1257_groups_1, pad = sparse_output_1257_pad_1, pad_type = sparse_output_1257_pad_type_1, strides = sparse_output_1257_strides_1, weight = layers_15_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1257_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20009_cast_fp16 = add(x = dense_output_1257_cast_fp16, y = sparse_output_1257_cast_fp16)[name = string("op_20009_cast_fp16")];
            tensor<int32, [4]> var_20010 = const()[name = string("op_20010"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20012 = const()[name = string("op_20012"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20011_cast_fp16 = transpose(perm = var_20010, x = var_20009_cast_fp16)[name = string("transpose_345")];
            tensor<fp16, [1, 51, 128]> q_head_251_cast_fp16 = reshape(shape = var_20012, x = var_20011_cast_fp16)[name = string("q_head_251_cast_fp16")];
            string dense_output_1259_pad_type_1 = const()[name = string("dense_output_1259_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1259_strides_1 = const()[name = string("dense_output_1259_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1259_pad_1 = const()[name = string("dense_output_1259_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1259_dilations_1 = const()[name = string("dense_output_1259_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1259_groups_1 = const()[name = string("dense_output_1259_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457278080))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457409216))))[name = string("layers_15_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1259_cast_fp16 = conv(dilations = dense_output_1259_dilations_1, groups = dense_output_1259_groups_1, pad = dense_output_1259_pad_1, pad_type = dense_output_1259_pad_type_1, strides = dense_output_1259_strides_1, weight = layers_15_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1259_cast_fp16")];
            string sparse_output_1259_pad_type_1 = const()[name = string("sparse_output_1259_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1259_strides_1 = const()[name = string("sparse_output_1259_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1259_pad_1 = const()[name = string("sparse_output_1259_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1259_dilations_1 = const()[name = string("sparse_output_1259_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1259_groups_1 = const()[name = string("sparse_output_1259_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457412480))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457409792))))[name = string("layers_15_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1259_cast_fp16 = conv(dilations = sparse_output_1259_dilations_1, groups = sparse_output_1259_groups_1, pad = sparse_output_1259_pad_1, pad_type = sparse_output_1259_pad_type_1, strides = sparse_output_1259_strides_1, weight = layers_15_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1259_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20028_cast_fp16 = add(x = dense_output_1259_cast_fp16, y = sparse_output_1259_cast_fp16)[name = string("op_20028_cast_fp16")];
            tensor<int32, [4]> var_20029 = const()[name = string("op_20029"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20031 = const()[name = string("op_20031"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20030_cast_fp16 = transpose(perm = var_20029, x = var_20028_cast_fp16)[name = string("transpose_344")];
            tensor<fp16, [1, 51, 128]> k_head_501_cast_fp16 = reshape(shape = var_20031, x = var_20030_cast_fp16)[name = string("k_head_501_cast_fp16")];
            string dense_output_1261_pad_type_1 = const()[name = string("dense_output_1261_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1261_strides_1 = const()[name = string("dense_output_1261_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1261_pad_1 = const()[name = string("dense_output_1261_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1261_dilations_1 = const()[name = string("dense_output_1261_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1261_groups_1 = const()[name = string("dense_output_1261_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457428928))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457560064))))[name = string("layers_15_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1261_cast_fp16 = conv(dilations = dense_output_1261_dilations_1, groups = dense_output_1261_groups_1, pad = dense_output_1261_pad_1, pad_type = dense_output_1261_pad_type_1, strides = dense_output_1261_strides_1, weight = layers_15_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1261_cast_fp16")];
            string sparse_output_1261_pad_type_1 = const()[name = string("sparse_output_1261_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1261_strides_1 = const()[name = string("sparse_output_1261_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1261_pad_1 = const()[name = string("sparse_output_1261_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1261_dilations_1 = const()[name = string("sparse_output_1261_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1261_groups_1 = const()[name = string("sparse_output_1261_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457563328))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457560640))))[name = string("layers_15_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1261_cast_fp16 = conv(dilations = sparse_output_1261_dilations_1, groups = sparse_output_1261_groups_1, pad = sparse_output_1261_pad_1, pad_type = sparse_output_1261_pad_type_1, strides = sparse_output_1261_strides_1, weight = layers_15_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1261_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20047_cast_fp16 = add(x = dense_output_1261_cast_fp16, y = sparse_output_1261_cast_fp16)[name = string("op_20047_cast_fp16")];
            tensor<int32, [4]> var_20048 = const()[name = string("op_20048"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20050 = const()[name = string("op_20050"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20049_cast_fp16 = transpose(perm = var_20048, x = var_20047_cast_fp16)[name = string("transpose_343")];
            tensor<fp16, [1, 51, 128]> v_head_501_cast_fp16 = reshape(shape = var_20050, x = var_20049_cast_fp16)[name = string("v_head_501_cast_fp16")];
            string dense_output_1263_pad_type_1 = const()[name = string("dense_output_1263_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1263_strides_1 = const()[name = string("dense_output_1263_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1263_pad_1 = const()[name = string("dense_output_1263_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1263_dilations_1 = const()[name = string("dense_output_1263_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1263_groups_1 = const()[name = string("dense_output_1263_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457579776))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457710912))))[name = string("layers_15_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1263_cast_fp16 = conv(dilations = dense_output_1263_dilations_1, groups = dense_output_1263_groups_1, pad = dense_output_1263_pad_1, pad_type = dense_output_1263_pad_type_1, strides = dense_output_1263_strides_1, weight = layers_15_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1263_cast_fp16")];
            string sparse_output_1263_pad_type_1 = const()[name = string("sparse_output_1263_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1263_strides_1 = const()[name = string("sparse_output_1263_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1263_pad_1 = const()[name = string("sparse_output_1263_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1263_dilations_1 = const()[name = string("sparse_output_1263_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1263_groups_1 = const()[name = string("sparse_output_1263_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457714176))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457711488))))[name = string("layers_15_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1263_cast_fp16 = conv(dilations = sparse_output_1263_dilations_1, groups = sparse_output_1263_groups_1, pad = sparse_output_1263_pad_1, pad_type = sparse_output_1263_pad_type_1, strides = sparse_output_1263_strides_1, weight = layers_15_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1263_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_20066_cast_fp16 = add(x = dense_output_1263_cast_fp16, y = sparse_output_1263_cast_fp16)[name = string("op_20066_cast_fp16")];
            tensor<int32, [4]> var_20067 = const()[name = string("op_20067"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20069 = const()[name = string("op_20069"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_20068_cast_fp16 = transpose(perm = var_20067, x = var_20066_cast_fp16)[name = string("transpose_342")];
            tensor<fp16, [1, 101, 128]> p_head_501_cast_fp16 = reshape(shape = var_20069, x = var_20068_cast_fp16)[name = string("p_head_501_cast_fp16")];
            tensor<fp16, [128]> var_20071_to_fp16 = const()[name = string("op_20071_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457730624)))];
            tensor<fp16, [1, 51, 128]> var_20072_cast_fp16 = add(x = q_head_251_cast_fp16, y = var_20071_to_fp16)[name = string("op_20072_cast_fp16")];
            tensor<int32, [1]> q_u_251_axes_1 = const()[name = string("q_u_251_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_251_cast_fp16 = expand_dims(axes = q_u_251_axes_1, x = var_20072_cast_fp16)[name = string("q_u_251_cast_fp16")];
            tensor<fp16, [128]> var_20074_to_fp16 = const()[name = string("op_20074_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457730944)))];
            tensor<fp16, [1, 51, 128]> var_20075_cast_fp16 = add(x = q_head_251_cast_fp16, y = var_20074_to_fp16)[name = string("op_20075_cast_fp16")];
            tensor<int32, [1]> q_v_251_axes_1 = const()[name = string("q_v_251_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_251_cast_fp16 = expand_dims(axes = q_v_251_axes_1, x = var_20075_cast_fp16)[name = string("q_v_251_cast_fp16")];
            tensor<int32, [1]> k_head_503_axes_1 = const()[name = string("k_head_503_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_503_cast_fp16 = expand_dims(axes = k_head_503_axes_1, x = k_head_501_cast_fp16)[name = string("k_head_503_cast_fp16")];
            tensor<int32, [1]> v_head_503_axes_1 = const()[name = string("v_head_503_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_503_cast_fp16 = expand_dims(axes = v_head_503_axes_1, x = v_head_501_cast_fp16)[name = string("v_head_503_cast_fp16")];
            tensor<int32, [1]> p_head_503_axes_1 = const()[name = string("p_head_503_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_503_cast_fp16 = expand_dims(axes = p_head_503_axes_1, x = p_head_501_cast_fp16)[name = string("p_head_503_cast_fp16")];
            bool var_20081_transpose_x_3 = const()[name = string("op_20081_transpose_x_3"), val = bool(false)];
            bool var_20081_transpose_y_3 = const()[name = string("op_20081_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_20081_cast_fp16 = matmul(transpose_x = var_20081_transpose_x_3, transpose_y = var_20081_transpose_y_3, x = q_u_251_cast_fp16, y = k_head_503_cast_fp16)[name = string("op_20081_cast_fp16")];
            fp16 var_20082_to_fp16 = const()[name = string("op_20082_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_251_cast_fp16 = mul(x = var_20081_cast_fp16, y = var_20082_to_fp16)[name = string("scores_content_251_cast_fp16")];
            bool x_1321_transpose_x_3 = const()[name = string("x_1321_transpose_x_3"), val = bool(false)];
            bool x_1321_transpose_y_3 = const()[name = string("x_1321_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1321_cast_fp16 = matmul(transpose_x = x_1321_transpose_x_3, transpose_y = x_1321_transpose_y_3, x = q_v_251_cast_fp16, y = p_head_503_cast_fp16)[name = string("x_1321_cast_fp16")];
            tensor<int32, [8]> x_1323_pad_1 = const()[name = string("x_1323_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1323_mode_1 = const()[name = string("x_1323_mode_1"), val = string("constant")];
            fp16 const_2149_to_fp16 = const()[name = string("const_2149_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1323_cast_fp16 = pad(constant_val = const_2149_to_fp16, mode = x_1323_mode_1, pad = x_1323_pad_1, x = x_1321_cast_fp16)[name = string("x_1323_cast_fp16")];
            tensor<int32, [4]> var_20096 = const()[name = string("op_20096"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1325_cast_fp16 = reshape(shape = var_20096, x = x_1323_cast_fp16)[name = string("x_1325_cast_fp16")];
            tensor<int32, [4]> var_20100_begin_1 = const()[name = string("op_20100_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_20100_end_1 = const()[name = string("op_20100_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_20100_end_mask_1 = const()[name = string("op_20100_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_20100_cast_fp16 = slice_by_index(begin = var_20100_begin_1, end = var_20100_end_1, end_mask = var_20100_end_mask_1, x = x_1325_cast_fp16)[name = string("op_20100_cast_fp16")];
            tensor<int32, [4]> var_20102 = const()[name = string("op_20102"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_20103_cast_fp16 = reshape(shape = var_20102, x = var_20100_cast_fp16)[name = string("op_20103_cast_fp16")];
            tensor<int32, [4]> var_20108_begin_1 = const()[name = string("op_20108_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_20108_end_1 = const()[name = string("op_20108_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_20108_end_mask_1 = const()[name = string("op_20108_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_20108_cast_fp16 = slice_by_index(begin = var_20108_begin_1, end = var_20108_end_1, end_mask = var_20108_end_mask_1, x = var_20103_cast_fp16)[name = string("op_20108_cast_fp16")];
            fp16 var_20109_to_fp16 = const()[name = string("op_20109_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_251_cast_fp16 = mul(x = var_20108_cast_fp16, y = var_20109_to_fp16)[name = string("scores_pos_251_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_251_cast_fp16 = add(x = scores_content_251_cast_fp16, y = scores_pos_251_cast_fp16)[name = string("logits_251_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_20112_cast_fp16 = softmax(axis = var_19232, x = logits_251_cast_fp16)[name = string("op_20112_cast_fp16")];
            bool var_20114_transpose_x_1 = const()[name = string("op_20114_transpose_x_1"), val = bool(false)];
            bool var_20114_transpose_y_1 = const()[name = string("op_20114_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_20114_cast_fp16 = matmul(transpose_x = var_20114_transpose_x_1, transpose_y = var_20114_transpose_y_1, x = var_20112_cast_fp16, y = v_head_503_cast_fp16)[name = string("op_20114_cast_fp16")];
            tensor<int32, [1]> var_20115_axes_1 = const()[name = string("op_20115_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_20115_cast_fp16 = squeeze(axes = var_20115_axes_1, x = var_20114_cast_fp16)[name = string("op_20115_cast_fp16")];
            string dense_output_1265_pad_type_1 = const()[name = string("dense_output_1265_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1265_strides_1 = const()[name = string("dense_output_1265_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1265_pad_1 = const()[name = string("dense_output_1265_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1265_dilations_1 = const()[name = string("dense_output_1265_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1265_groups_1 = const()[name = string("dense_output_1265_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457731264))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457862400))))[name = string("layers_15_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1265_cast_fp16 = conv(dilations = dense_output_1265_dilations_1, groups = dense_output_1265_groups_1, pad = dense_output_1265_pad_1, pad_type = dense_output_1265_pad_type_1, strides = dense_output_1265_strides_1, weight = layers_15_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1265_cast_fp16")];
            string sparse_output_1265_pad_type_1 = const()[name = string("sparse_output_1265_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1265_strides_1 = const()[name = string("sparse_output_1265_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1265_pad_1 = const()[name = string("sparse_output_1265_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1265_dilations_1 = const()[name = string("sparse_output_1265_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1265_groups_1 = const()[name = string("sparse_output_1265_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457865664))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457862976))))[name = string("layers_15_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1265_cast_fp16 = conv(dilations = sparse_output_1265_dilations_1, groups = sparse_output_1265_groups_1, pad = sparse_output_1265_pad_1, pad_type = sparse_output_1265_pad_type_1, strides = sparse_output_1265_strides_1, weight = layers_15_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1265_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20130_cast_fp16 = add(x = dense_output_1265_cast_fp16, y = sparse_output_1265_cast_fp16)[name = string("op_20130_cast_fp16")];
            tensor<int32, [4]> var_20131 = const()[name = string("op_20131"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20133 = const()[name = string("op_20133"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20132_cast_fp16 = transpose(perm = var_20131, x = var_20130_cast_fp16)[name = string("transpose_341")];
            tensor<fp16, [1, 51, 128]> q_head_253_cast_fp16 = reshape(shape = var_20133, x = var_20132_cast_fp16)[name = string("q_head_253_cast_fp16")];
            string dense_output_1267_pad_type_1 = const()[name = string("dense_output_1267_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1267_strides_1 = const()[name = string("dense_output_1267_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1267_pad_1 = const()[name = string("dense_output_1267_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1267_dilations_1 = const()[name = string("dense_output_1267_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1267_groups_1 = const()[name = string("dense_output_1267_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457882112))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458013248))))[name = string("layers_15_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1267_cast_fp16 = conv(dilations = dense_output_1267_dilations_1, groups = dense_output_1267_groups_1, pad = dense_output_1267_pad_1, pad_type = dense_output_1267_pad_type_1, strides = dense_output_1267_strides_1, weight = layers_15_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1267_cast_fp16")];
            string sparse_output_1267_pad_type_1 = const()[name = string("sparse_output_1267_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1267_strides_1 = const()[name = string("sparse_output_1267_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1267_pad_1 = const()[name = string("sparse_output_1267_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1267_dilations_1 = const()[name = string("sparse_output_1267_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1267_groups_1 = const()[name = string("sparse_output_1267_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458016512))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458013824))))[name = string("layers_15_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1267_cast_fp16 = conv(dilations = sparse_output_1267_dilations_1, groups = sparse_output_1267_groups_1, pad = sparse_output_1267_pad_1, pad_type = sparse_output_1267_pad_type_1, strides = sparse_output_1267_strides_1, weight = layers_15_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1267_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20149_cast_fp16 = add(x = dense_output_1267_cast_fp16, y = sparse_output_1267_cast_fp16)[name = string("op_20149_cast_fp16")];
            tensor<int32, [4]> var_20150 = const()[name = string("op_20150"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20152 = const()[name = string("op_20152"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20151_cast_fp16 = transpose(perm = var_20150, x = var_20149_cast_fp16)[name = string("transpose_340")];
            tensor<fp16, [1, 51, 128]> k_head_505_cast_fp16 = reshape(shape = var_20152, x = var_20151_cast_fp16)[name = string("k_head_505_cast_fp16")];
            string dense_output_1269_pad_type_1 = const()[name = string("dense_output_1269_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1269_strides_1 = const()[name = string("dense_output_1269_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1269_pad_1 = const()[name = string("dense_output_1269_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1269_dilations_1 = const()[name = string("dense_output_1269_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1269_groups_1 = const()[name = string("dense_output_1269_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458032960))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458164096))))[name = string("layers_15_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1269_cast_fp16 = conv(dilations = dense_output_1269_dilations_1, groups = dense_output_1269_groups_1, pad = dense_output_1269_pad_1, pad_type = dense_output_1269_pad_type_1, strides = dense_output_1269_strides_1, weight = layers_15_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1269_cast_fp16")];
            string sparse_output_1269_pad_type_1 = const()[name = string("sparse_output_1269_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1269_strides_1 = const()[name = string("sparse_output_1269_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1269_pad_1 = const()[name = string("sparse_output_1269_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1269_dilations_1 = const()[name = string("sparse_output_1269_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1269_groups_1 = const()[name = string("sparse_output_1269_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458167360))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458164672))))[name = string("layers_15_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1269_cast_fp16 = conv(dilations = sparse_output_1269_dilations_1, groups = sparse_output_1269_groups_1, pad = sparse_output_1269_pad_1, pad_type = sparse_output_1269_pad_type_1, strides = sparse_output_1269_strides_1, weight = layers_15_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1269_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20168_cast_fp16 = add(x = dense_output_1269_cast_fp16, y = sparse_output_1269_cast_fp16)[name = string("op_20168_cast_fp16")];
            tensor<int32, [4]> var_20169 = const()[name = string("op_20169"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20171 = const()[name = string("op_20171"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20170_cast_fp16 = transpose(perm = var_20169, x = var_20168_cast_fp16)[name = string("transpose_339")];
            tensor<fp16, [1, 51, 128]> v_head_505_cast_fp16 = reshape(shape = var_20171, x = var_20170_cast_fp16)[name = string("v_head_505_cast_fp16")];
            string dense_output_1271_pad_type_1 = const()[name = string("dense_output_1271_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1271_strides_1 = const()[name = string("dense_output_1271_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1271_pad_1 = const()[name = string("dense_output_1271_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1271_dilations_1 = const()[name = string("dense_output_1271_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1271_groups_1 = const()[name = string("dense_output_1271_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458183808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458314944))))[name = string("layers_15_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1271_cast_fp16 = conv(dilations = dense_output_1271_dilations_1, groups = dense_output_1271_groups_1, pad = dense_output_1271_pad_1, pad_type = dense_output_1271_pad_type_1, strides = dense_output_1271_strides_1, weight = layers_15_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1271_cast_fp16")];
            string sparse_output_1271_pad_type_1 = const()[name = string("sparse_output_1271_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1271_strides_1 = const()[name = string("sparse_output_1271_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1271_pad_1 = const()[name = string("sparse_output_1271_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1271_dilations_1 = const()[name = string("sparse_output_1271_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1271_groups_1 = const()[name = string("sparse_output_1271_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458318208))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458315520))))[name = string("layers_15_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1271_cast_fp16 = conv(dilations = sparse_output_1271_dilations_1, groups = sparse_output_1271_groups_1, pad = sparse_output_1271_pad_1, pad_type = sparse_output_1271_pad_type_1, strides = sparse_output_1271_strides_1, weight = layers_15_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1271_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_20187_cast_fp16 = add(x = dense_output_1271_cast_fp16, y = sparse_output_1271_cast_fp16)[name = string("op_20187_cast_fp16")];
            tensor<int32, [4]> var_20188 = const()[name = string("op_20188"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20190 = const()[name = string("op_20190"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_20189_cast_fp16 = transpose(perm = var_20188, x = var_20187_cast_fp16)[name = string("transpose_338")];
            tensor<fp16, [1, 101, 128]> p_head_505_cast_fp16 = reshape(shape = var_20190, x = var_20189_cast_fp16)[name = string("p_head_505_cast_fp16")];
            tensor<fp16, [128]> var_20192_to_fp16 = const()[name = string("op_20192_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458334656)))];
            tensor<fp16, [1, 51, 128]> var_20193_cast_fp16 = add(x = q_head_253_cast_fp16, y = var_20192_to_fp16)[name = string("op_20193_cast_fp16")];
            tensor<int32, [1]> q_u_253_axes_1 = const()[name = string("q_u_253_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_253_cast_fp16 = expand_dims(axes = q_u_253_axes_1, x = var_20193_cast_fp16)[name = string("q_u_253_cast_fp16")];
            tensor<fp16, [128]> var_20195_to_fp16 = const()[name = string("op_20195_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458334976)))];
            tensor<fp16, [1, 51, 128]> var_20196_cast_fp16 = add(x = q_head_253_cast_fp16, y = var_20195_to_fp16)[name = string("op_20196_cast_fp16")];
            tensor<int32, [1]> q_v_253_axes_1 = const()[name = string("q_v_253_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_253_cast_fp16 = expand_dims(axes = q_v_253_axes_1, x = var_20196_cast_fp16)[name = string("q_v_253_cast_fp16")];
            tensor<int32, [1]> k_head_507_axes_1 = const()[name = string("k_head_507_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_507_cast_fp16 = expand_dims(axes = k_head_507_axes_1, x = k_head_505_cast_fp16)[name = string("k_head_507_cast_fp16")];
            tensor<int32, [1]> v_head_507_axes_1 = const()[name = string("v_head_507_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_507_cast_fp16 = expand_dims(axes = v_head_507_axes_1, x = v_head_505_cast_fp16)[name = string("v_head_507_cast_fp16")];
            tensor<int32, [1]> p_head_507_axes_1 = const()[name = string("p_head_507_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_507_cast_fp16 = expand_dims(axes = p_head_507_axes_1, x = p_head_505_cast_fp16)[name = string("p_head_507_cast_fp16")];
            bool var_20202_transpose_x_3 = const()[name = string("op_20202_transpose_x_3"), val = bool(false)];
            bool var_20202_transpose_y_3 = const()[name = string("op_20202_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_20202_cast_fp16 = matmul(transpose_x = var_20202_transpose_x_3, transpose_y = var_20202_transpose_y_3, x = q_u_253_cast_fp16, y = k_head_507_cast_fp16)[name = string("op_20202_cast_fp16")];
            fp16 var_20203_to_fp16 = const()[name = string("op_20203_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_253_cast_fp16 = mul(x = var_20202_cast_fp16, y = var_20203_to_fp16)[name = string("scores_content_253_cast_fp16")];
            bool x_1329_transpose_x_3 = const()[name = string("x_1329_transpose_x_3"), val = bool(false)];
            bool x_1329_transpose_y_3 = const()[name = string("x_1329_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1329_cast_fp16 = matmul(transpose_x = x_1329_transpose_x_3, transpose_y = x_1329_transpose_y_3, x = q_v_253_cast_fp16, y = p_head_507_cast_fp16)[name = string("x_1329_cast_fp16")];
            tensor<int32, [8]> x_1331_pad_1 = const()[name = string("x_1331_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1331_mode_1 = const()[name = string("x_1331_mode_1"), val = string("constant")];
            fp16 const_2155_to_fp16 = const()[name = string("const_2155_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1331_cast_fp16 = pad(constant_val = const_2155_to_fp16, mode = x_1331_mode_1, pad = x_1331_pad_1, x = x_1329_cast_fp16)[name = string("x_1331_cast_fp16")];
            tensor<int32, [4]> var_20217 = const()[name = string("op_20217"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1333_cast_fp16 = reshape(shape = var_20217, x = x_1331_cast_fp16)[name = string("x_1333_cast_fp16")];
            tensor<int32, [4]> var_20221_begin_1 = const()[name = string("op_20221_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_20221_end_1 = const()[name = string("op_20221_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_20221_end_mask_1 = const()[name = string("op_20221_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_20221_cast_fp16 = slice_by_index(begin = var_20221_begin_1, end = var_20221_end_1, end_mask = var_20221_end_mask_1, x = x_1333_cast_fp16)[name = string("op_20221_cast_fp16")];
            tensor<int32, [4]> var_20223 = const()[name = string("op_20223"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_20224_cast_fp16 = reshape(shape = var_20223, x = var_20221_cast_fp16)[name = string("op_20224_cast_fp16")];
            tensor<int32, [4]> var_20229_begin_1 = const()[name = string("op_20229_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_20229_end_1 = const()[name = string("op_20229_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_20229_end_mask_1 = const()[name = string("op_20229_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_20229_cast_fp16 = slice_by_index(begin = var_20229_begin_1, end = var_20229_end_1, end_mask = var_20229_end_mask_1, x = var_20224_cast_fp16)[name = string("op_20229_cast_fp16")];
            fp16 var_20230_to_fp16 = const()[name = string("op_20230_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_253_cast_fp16 = mul(x = var_20229_cast_fp16, y = var_20230_to_fp16)[name = string("scores_pos_253_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_253_cast_fp16 = add(x = scores_content_253_cast_fp16, y = scores_pos_253_cast_fp16)[name = string("logits_253_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_20233_cast_fp16 = softmax(axis = var_19232, x = logits_253_cast_fp16)[name = string("op_20233_cast_fp16")];
            bool var_20235_transpose_x_1 = const()[name = string("op_20235_transpose_x_1"), val = bool(false)];
            bool var_20235_transpose_y_1 = const()[name = string("op_20235_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_20235_cast_fp16 = matmul(transpose_x = var_20235_transpose_x_1, transpose_y = var_20235_transpose_y_1, x = var_20233_cast_fp16, y = v_head_507_cast_fp16)[name = string("op_20235_cast_fp16")];
            tensor<int32, [1]> var_20236_axes_1 = const()[name = string("op_20236_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_20236_cast_fp16 = squeeze(axes = var_20236_axes_1, x = var_20235_cast_fp16)[name = string("op_20236_cast_fp16")];
            string dense_output_1273_pad_type_1 = const()[name = string("dense_output_1273_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1273_strides_1 = const()[name = string("dense_output_1273_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1273_pad_1 = const()[name = string("dense_output_1273_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1273_dilations_1 = const()[name = string("dense_output_1273_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1273_groups_1 = const()[name = string("dense_output_1273_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458335296))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458466432))))[name = string("layers_15_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1273_cast_fp16 = conv(dilations = dense_output_1273_dilations_1, groups = dense_output_1273_groups_1, pad = dense_output_1273_pad_1, pad_type = dense_output_1273_pad_type_1, strides = dense_output_1273_strides_1, weight = layers_15_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1273_cast_fp16")];
            string sparse_output_1273_pad_type_1 = const()[name = string("sparse_output_1273_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1273_strides_1 = const()[name = string("sparse_output_1273_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1273_pad_1 = const()[name = string("sparse_output_1273_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1273_dilations_1 = const()[name = string("sparse_output_1273_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1273_groups_1 = const()[name = string("sparse_output_1273_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458469696))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458467008))))[name = string("layers_15_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1273_cast_fp16 = conv(dilations = sparse_output_1273_dilations_1, groups = sparse_output_1273_groups_1, pad = sparse_output_1273_pad_1, pad_type = sparse_output_1273_pad_type_1, strides = sparse_output_1273_strides_1, weight = layers_15_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1273_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20251_cast_fp16 = add(x = dense_output_1273_cast_fp16, y = sparse_output_1273_cast_fp16)[name = string("op_20251_cast_fp16")];
            tensor<int32, [4]> var_20252 = const()[name = string("op_20252"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20254 = const()[name = string("op_20254"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20253_cast_fp16 = transpose(perm = var_20252, x = var_20251_cast_fp16)[name = string("transpose_337")];
            tensor<fp16, [1, 51, 128]> q_head_255_cast_fp16 = reshape(shape = var_20254, x = var_20253_cast_fp16)[name = string("q_head_255_cast_fp16")];
            string dense_output_1275_pad_type_1 = const()[name = string("dense_output_1275_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1275_strides_1 = const()[name = string("dense_output_1275_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1275_pad_1 = const()[name = string("dense_output_1275_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1275_dilations_1 = const()[name = string("dense_output_1275_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1275_groups_1 = const()[name = string("dense_output_1275_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458486144))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458617280))))[name = string("layers_15_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1275_cast_fp16 = conv(dilations = dense_output_1275_dilations_1, groups = dense_output_1275_groups_1, pad = dense_output_1275_pad_1, pad_type = dense_output_1275_pad_type_1, strides = dense_output_1275_strides_1, weight = layers_15_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1275_cast_fp16")];
            string sparse_output_1275_pad_type_1 = const()[name = string("sparse_output_1275_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1275_strides_1 = const()[name = string("sparse_output_1275_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1275_pad_1 = const()[name = string("sparse_output_1275_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1275_dilations_1 = const()[name = string("sparse_output_1275_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1275_groups_1 = const()[name = string("sparse_output_1275_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458620544))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458617856))))[name = string("layers_15_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1275_cast_fp16 = conv(dilations = sparse_output_1275_dilations_1, groups = sparse_output_1275_groups_1, pad = sparse_output_1275_pad_1, pad_type = sparse_output_1275_pad_type_1, strides = sparse_output_1275_strides_1, weight = layers_15_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1275_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20270_cast_fp16 = add(x = dense_output_1275_cast_fp16, y = sparse_output_1275_cast_fp16)[name = string("op_20270_cast_fp16")];
            tensor<int32, [4]> var_20271 = const()[name = string("op_20271"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20273 = const()[name = string("op_20273"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20272_cast_fp16 = transpose(perm = var_20271, x = var_20270_cast_fp16)[name = string("transpose_336")];
            tensor<fp16, [1, 51, 128]> k_head_509_cast_fp16 = reshape(shape = var_20273, x = var_20272_cast_fp16)[name = string("k_head_509_cast_fp16")];
            string dense_output_1277_pad_type_1 = const()[name = string("dense_output_1277_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1277_strides_1 = const()[name = string("dense_output_1277_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1277_pad_1 = const()[name = string("dense_output_1277_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1277_dilations_1 = const()[name = string("dense_output_1277_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1277_groups_1 = const()[name = string("dense_output_1277_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458636992))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458768128))))[name = string("layers_15_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1277_cast_fp16 = conv(dilations = dense_output_1277_dilations_1, groups = dense_output_1277_groups_1, pad = dense_output_1277_pad_1, pad_type = dense_output_1277_pad_type_1, strides = dense_output_1277_strides_1, weight = layers_15_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("dense_output_1277_cast_fp16")];
            string sparse_output_1277_pad_type_1 = const()[name = string("sparse_output_1277_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1277_strides_1 = const()[name = string("sparse_output_1277_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1277_pad_1 = const()[name = string("sparse_output_1277_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1277_dilations_1 = const()[name = string("sparse_output_1277_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1277_groups_1 = const()[name = string("sparse_output_1277_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458771392))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458768704))))[name = string("layers_15_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1277_cast_fp16 = conv(dilations = sparse_output_1277_dilations_1, groups = sparse_output_1277_groups_1, pad = sparse_output_1277_pad_1, pad_type = sparse_output_1277_pad_type_1, strides = sparse_output_1277_strides_1, weight = layers_15_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_719_cast_fp16)[name = string("sparse_output_1277_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20289_cast_fp16 = add(x = dense_output_1277_cast_fp16, y = sparse_output_1277_cast_fp16)[name = string("op_20289_cast_fp16")];
            tensor<int32, [4]> var_20290 = const()[name = string("op_20290"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20292 = const()[name = string("op_20292"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20291_cast_fp16 = transpose(perm = var_20290, x = var_20289_cast_fp16)[name = string("transpose_335")];
            tensor<fp16, [1, 51, 128]> v_head_509_cast_fp16 = reshape(shape = var_20292, x = var_20291_cast_fp16)[name = string("v_head_509_cast_fp16")];
            string dense_output_1279_pad_type_1 = const()[name = string("dense_output_1279_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1279_strides_1 = const()[name = string("dense_output_1279_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1279_pad_1 = const()[name = string("dense_output_1279_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1279_dilations_1 = const()[name = string("dense_output_1279_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1279_groups_1 = const()[name = string("dense_output_1279_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458787840))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458918976))))[name = string("layers_15_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1279_cast_fp16 = conv(dilations = dense_output_1279_dilations_1, groups = dense_output_1279_groups_1, pad = dense_output_1279_pad_1, pad_type = dense_output_1279_pad_type_1, strides = dense_output_1279_strides_1, weight = layers_15_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1279_cast_fp16")];
            string sparse_output_1279_pad_type_1 = const()[name = string("sparse_output_1279_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1279_strides_1 = const()[name = string("sparse_output_1279_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1279_pad_1 = const()[name = string("sparse_output_1279_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1279_dilations_1 = const()[name = string("sparse_output_1279_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1279_groups_1 = const()[name = string("sparse_output_1279_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_15_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458922240))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458919552))))[name = string("layers_15_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1279_cast_fp16 = conv(dilations = sparse_output_1279_dilations_1, groups = sparse_output_1279_groups_1, pad = sparse_output_1279_pad_1, pad_type = sparse_output_1279_pad_type_1, strides = sparse_output_1279_strides_1, weight = layers_15_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1279_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_20308_cast_fp16 = add(x = dense_output_1279_cast_fp16, y = sparse_output_1279_cast_fp16)[name = string("op_20308_cast_fp16")];
            tensor<int32, [4]> var_20309 = const()[name = string("op_20309"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20311 = const()[name = string("op_20311"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_20310_cast_fp16 = transpose(perm = var_20309, x = var_20308_cast_fp16)[name = string("transpose_334")];
            tensor<fp16, [1, 101, 128]> p_head_509_cast_fp16 = reshape(shape = var_20311, x = var_20310_cast_fp16)[name = string("p_head_509_cast_fp16")];
            tensor<fp16, [128]> var_20313_to_fp16 = const()[name = string("op_20313_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458938688)))];
            tensor<fp16, [1, 51, 128]> var_20314_cast_fp16 = add(x = q_head_255_cast_fp16, y = var_20313_to_fp16)[name = string("op_20314_cast_fp16")];
            tensor<int32, [1]> q_u_255_axes_1 = const()[name = string("q_u_255_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_255_cast_fp16 = expand_dims(axes = q_u_255_axes_1, x = var_20314_cast_fp16)[name = string("q_u_255_cast_fp16")];
            tensor<fp16, [128]> var_20316_to_fp16 = const()[name = string("op_20316_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458939008)))];
            tensor<fp16, [1, 51, 128]> var_20317_cast_fp16 = add(x = q_head_255_cast_fp16, y = var_20316_to_fp16)[name = string("op_20317_cast_fp16")];
            tensor<int32, [1]> q_v_255_axes_1 = const()[name = string("q_v_255_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_255_cast_fp16 = expand_dims(axes = q_v_255_axes_1, x = var_20317_cast_fp16)[name = string("q_v_255_cast_fp16")];
            tensor<int32, [1]> k_head_511_axes_1 = const()[name = string("k_head_511_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_511_cast_fp16 = expand_dims(axes = k_head_511_axes_1, x = k_head_509_cast_fp16)[name = string("k_head_511_cast_fp16")];
            tensor<int32, [1]> v_head_511_axes_1 = const()[name = string("v_head_511_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_511_cast_fp16 = expand_dims(axes = v_head_511_axes_1, x = v_head_509_cast_fp16)[name = string("v_head_511_cast_fp16")];
            tensor<int32, [1]> p_head_511_axes_1 = const()[name = string("p_head_511_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_511_cast_fp16 = expand_dims(axes = p_head_511_axes_1, x = p_head_509_cast_fp16)[name = string("p_head_511_cast_fp16")];
            bool var_20323_transpose_x_3 = const()[name = string("op_20323_transpose_x_3"), val = bool(false)];
            bool var_20323_transpose_y_3 = const()[name = string("op_20323_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_20323_cast_fp16 = matmul(transpose_x = var_20323_transpose_x_3, transpose_y = var_20323_transpose_y_3, x = q_u_255_cast_fp16, y = k_head_511_cast_fp16)[name = string("op_20323_cast_fp16")];
            fp16 var_20324_to_fp16 = const()[name = string("op_20324_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_255_cast_fp16 = mul(x = var_20323_cast_fp16, y = var_20324_to_fp16)[name = string("scores_content_255_cast_fp16")];
            bool x_1337_transpose_x_3 = const()[name = string("x_1337_transpose_x_3"), val = bool(false)];
            bool x_1337_transpose_y_3 = const()[name = string("x_1337_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1337_cast_fp16 = matmul(transpose_x = x_1337_transpose_x_3, transpose_y = x_1337_transpose_y_3, x = q_v_255_cast_fp16, y = p_head_511_cast_fp16)[name = string("x_1337_cast_fp16")];
            tensor<int32, [8]> x_1339_pad_1 = const()[name = string("x_1339_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1339_mode_1 = const()[name = string("x_1339_mode_1"), val = string("constant")];
            fp16 const_2161_to_fp16 = const()[name = string("const_2161_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1339_cast_fp16 = pad(constant_val = const_2161_to_fp16, mode = x_1339_mode_1, pad = x_1339_pad_1, x = x_1337_cast_fp16)[name = string("x_1339_cast_fp16")];
            tensor<int32, [4]> var_20338 = const()[name = string("op_20338"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1341_cast_fp16 = reshape(shape = var_20338, x = x_1339_cast_fp16)[name = string("x_1341_cast_fp16")];
            tensor<int32, [4]> var_20342_begin_1 = const()[name = string("op_20342_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_20342_end_1 = const()[name = string("op_20342_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_20342_end_mask_1 = const()[name = string("op_20342_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_20342_cast_fp16 = slice_by_index(begin = var_20342_begin_1, end = var_20342_end_1, end_mask = var_20342_end_mask_1, x = x_1341_cast_fp16)[name = string("op_20342_cast_fp16")];
            tensor<int32, [4]> var_20344 = const()[name = string("op_20344"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_20345_cast_fp16 = reshape(shape = var_20344, x = var_20342_cast_fp16)[name = string("op_20345_cast_fp16")];
            tensor<int32, [4]> var_20350_begin_1 = const()[name = string("op_20350_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_20350_end_1 = const()[name = string("op_20350_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_20350_end_mask_1 = const()[name = string("op_20350_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_20350_cast_fp16 = slice_by_index(begin = var_20350_begin_1, end = var_20350_end_1, end_mask = var_20350_end_mask_1, x = var_20345_cast_fp16)[name = string("op_20350_cast_fp16")];
            fp16 var_20351_to_fp16 = const()[name = string("op_20351_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_255_cast_fp16 = mul(x = var_20350_cast_fp16, y = var_20351_to_fp16)[name = string("scores_pos_255_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_255_cast_fp16 = add(x = scores_content_255_cast_fp16, y = scores_pos_255_cast_fp16)[name = string("logits_255_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_20354_cast_fp16 = softmax(axis = var_19232, x = logits_255_cast_fp16)[name = string("op_20354_cast_fp16")];
            bool var_20356_transpose_x_1 = const()[name = string("op_20356_transpose_x_1"), val = bool(false)];
            bool var_20356_transpose_y_1 = const()[name = string("op_20356_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_20356_cast_fp16 = matmul(transpose_x = var_20356_transpose_x_1, transpose_y = var_20356_transpose_y_1, x = var_20354_cast_fp16, y = v_head_511_cast_fp16)[name = string("op_20356_cast_fp16")];
            tensor<int32, [1]> o_head_31_axes_1 = const()[name = string("o_head_31_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_31_cast_fp16 = squeeze(axes = o_head_31_axes_1, x = var_20356_cast_fp16)[name = string("o_head_31_cast_fp16")];
            bool out_31_interleave_1 = const()[name = string("out_31_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_31_cast_fp16 = concat(axis = var_19232, interleave = out_31_interleave_1, values = (var_19510_cast_fp16, var_19631_cast_fp16, var_19752_cast_fp16, var_19873_cast_fp16, var_19994_cast_fp16, var_20115_cast_fp16, var_20236_cast_fp16, o_head_31_cast_fp16))[name = string("out_31_cast_fp16")];
            tensor<int32, [3]> var_20360_perm_1 = const()[name = string("op_20360_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_727_axes_1 = const()[name = string("input_727_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_20360_cast_fp16 = transpose(perm = var_20360_perm_1, x = out_31_cast_fp16)[name = string("transpose_333")];
            tensor<fp16, [1, 1024, 51, 1]> input_727_cast_fp16 = expand_dims(axes = input_727_axes_1, x = var_20360_cast_fp16)[name = string("input_727_cast_fp16")];
            string dense_output_1281_pad_type_1 = const()[name = string("dense_output_1281_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1281_strides_1 = const()[name = string("dense_output_1281_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1281_pad_1 = const()[name = string("dense_output_1281_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1281_dilations_1 = const()[name = string("dense_output_1281_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1281_groups_1 = const()[name = string("dense_output_1281_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458939328))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459987968))))[name = string("layers_15_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1281_cast_fp16 = conv(dilations = dense_output_1281_dilations_1, groups = dense_output_1281_groups_1, pad = dense_output_1281_pad_1, pad_type = dense_output_1281_pad_type_1, strides = dense_output_1281_strides_1, weight = layers_15_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_727_cast_fp16)[name = string("dense_output_1281_cast_fp16")];
            string sparse_output_1281_pad_type_1 = const()[name = string("sparse_output_1281_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1281_strides_1 = const()[name = string("sparse_output_1281_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1281_pad_1 = const()[name = string("sparse_output_1281_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1281_dilations_1 = const()[name = string("sparse_output_1281_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1281_groups_1 = const()[name = string("sparse_output_1281_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460009600))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459988544))))[name = string("layers_15_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1281_cast_fp16 = conv(dilations = sparse_output_1281_dilations_1, groups = sparse_output_1281_groups_1, pad = sparse_output_1281_pad_1, pad_type = sparse_output_1281_pad_type_1, strides = sparse_output_1281_strides_1, weight = layers_15_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_727_cast_fp16)[name = string("sparse_output_1281_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_31_cast_fp16 = add(x = dense_output_1281_cast_fp16, y = sparse_output_1281_cast_fp16)[name = string("out_conv_31_cast_fp16")];
            tensor<int32, [1]> var_20377_axes_1 = const()[name = string("op_20377_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_20377_cast_fp16 = squeeze(axes = var_20377_axes_1, x = out_conv_31_cast_fp16)[name = string("op_20377_cast_fp16")];
            tensor<int32, [3]> var_20378_perm_1 = const()[name = string("op_20378_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_20378_cast_fp16 = transpose(perm = var_20378_perm_1, x = var_20377_cast_fp16)[name = string("transpose_332")];
            tensor<fp16, [1, 51, 1024]> input_729_cast_fp16 = add(x = input_717_cast_fp16, y = var_20378_cast_fp16)[name = string("input_729_cast_fp16")];
            tensor<int32, [1]> x_1345_axes_1 = const()[name = string("x_1345_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_15_norm_conv_weight_to_fp16 = const()[name = string("layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460140736)))];
            tensor<fp16, [1024]> layers_15_norm_conv_bias_to_fp16 = const()[name = string("layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460142848)))];
            tensor<fp16, [1, 51, 1024]> x_1345_cast_fp16 = layer_norm(axes = x_1345_axes_1, beta = layers_15_norm_conv_bias_to_fp16, epsilon = var_19247_to_fp16, gamma = layers_15_norm_conv_weight_to_fp16, x = input_729_cast_fp16)[name = string("x_1345_cast_fp16")];
            tensor<int32, [3]> var_20388_perm_1 = const()[name = string("op_20388_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_731_axes_1 = const()[name = string("input_731_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_20388_cast_fp16 = transpose(perm = var_20388_perm_1, x = x_1345_cast_fp16)[name = string("transpose_331")];
            tensor<fp16, [1, 1024, 51, 1]> input_731_cast_fp16 = expand_dims(axes = input_731_axes_1, x = var_20388_cast_fp16)[name = string("input_731_cast_fp16")];
            string dense_output_1283_pad_type_1 = const()[name = string("dense_output_1283_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1283_strides_1 = const()[name = string("dense_output_1283_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1283_pad_1 = const()[name = string("dense_output_1283_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1283_dilations_1 = const()[name = string("dense_output_1283_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1283_groups_1 = const()[name = string("dense_output_1283_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_15_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460144960))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462242176))))[name = string("layers_15_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1283_cast_fp16 = conv(dilations = dense_output_1283_dilations_1, groups = dense_output_1283_groups_1, pad = dense_output_1283_pad_1, pad_type = dense_output_1283_pad_type_1, strides = dense_output_1283_strides_1, weight = layers_15_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_731_cast_fp16)[name = string("dense_output_1283_cast_fp16")];
            string sparse_output_1283_pad_type_1 = const()[name = string("sparse_output_1283_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1283_strides_1 = const()[name = string("sparse_output_1283_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1283_pad_1 = const()[name = string("sparse_output_1283_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1283_dilations_1 = const()[name = string("sparse_output_1283_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1283_groups_1 = const()[name = string("sparse_output_1283_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_15_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462284800))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462242752))))[name = string("layers_15_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1283_cast_fp16 = conv(dilations = sparse_output_1283_dilations_1, groups = sparse_output_1283_groups_1, pad = sparse_output_1283_pad_1, pad_type = sparse_output_1283_pad_type_1, strides = sparse_output_1283_strides_1, weight = layers_15_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_731_cast_fp16)[name = string("sparse_output_1283_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_733_cast_fp16 = add(x = dense_output_1283_cast_fp16, y = sparse_output_1283_cast_fp16)[name = string("input_733_cast_fp16")];
            int32 input_735_split_num_splits_1 = const()[name = string("input_735_split_num_splits_1"), val = int32(2)];
            int32 input_735_split_axis_1 = const()[name = string("input_735_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_735_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_735_split_cast_fp16_1 = split(axis = input_735_split_axis_1, num_splits = input_735_split_num_splits_1, x = input_733_cast_fp16)[name = string("input_735_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_735_split_1_sigmoid_cast_fp16 = sigmoid(x = input_735_split_cast_fp16_1)[name = string("input_735_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_735_cast_fp16 = mul(x = input_735_split_cast_fp16_0, y = input_735_split_1_sigmoid_cast_fp16)[name = string("input_735_cast_fp16")];
            tensor<int32, [8]> input_737_pad_1 = const()[name = string("input_737_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_737_mode_1 = const()[name = string("input_737_mode_1"), val = string("constant")];
            fp16 const_2163_to_fp16 = const()[name = string("const_2163_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_737_cast_fp16 = pad(constant_val = const_2163_to_fp16, mode = input_737_mode_1, pad = input_737_pad_1, x = input_735_cast_fp16)[name = string("input_737_cast_fp16")];
            string dense_output_1285_pad_type_1 = const()[name = string("dense_output_1285_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1285_strides_1 = const()[name = string("dense_output_1285_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1285_pad_1 = const()[name = string("dense_output_1285_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1285_dilations_1 = const()[name = string("dense_output_1285_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1285_groups_1 = const()[name = string("dense_output_1285_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1285_cast_fp16 = conv(dilations = dense_output_1285_dilations_1, groups = dense_output_1285_groups_1, pad = dense_output_1285_pad_1, pad_type = dense_output_1285_pad_type_1, strides = dense_output_1285_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_737_cast_fp16)[name = string("dense_output_1285_cast_fp16")];
            string sparse_output_1285_pad_type_1 = const()[name = string("sparse_output_1285_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1285_strides_1 = const()[name = string("sparse_output_1285_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1285_pad_1 = const()[name = string("sparse_output_1285_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1285_dilations_1 = const()[name = string("sparse_output_1285_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1285_groups_1 = const()[name = string("sparse_output_1285_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_15_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462547008))))[name = string("layers_15_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1285_cast_fp16 = conv(dilations = sparse_output_1285_dilations_1, groups = sparse_output_1285_groups_1, pad = sparse_output_1285_pad_1, pad_type = sparse_output_1285_pad_type_1, strides = sparse_output_1285_strides_1, weight = layers_15_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_737_cast_fp16)[name = string("sparse_output_1285_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_739_cast_fp16 = add(x = dense_output_1285_cast_fp16, y = sparse_output_1285_cast_fp16)[name = string("input_739_cast_fp16")];
            tensor<fp16, [1024]> layers_15_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_15_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462565504)))];
            tensor<fp16, [1024]> layers_15_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_15_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462567616)))];
            tensor<fp16, [1024]> layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462569728)))];
            tensor<fp16, [1024]> layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462571840)))];
            tensor<fp16, [1, 1024, 51, 1]> input_741_cast_fp16 = batch_norm(beta = layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_19247_to_fp16, gamma = layers_15_conv_batch_norm_weight_to_fp16, mean = layers_15_conv_batch_norm_running_mean_to_fp16, variance = layers_15_conv_batch_norm_running_var_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_743_cast_fp16 = silu(x = input_741_cast_fp16)[name = string("input_743_cast_fp16")];
            string dense_output_1287_pad_type_1 = const()[name = string("dense_output_1287_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1287_strides_1 = const()[name = string("dense_output_1287_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1287_pad_1 = const()[name = string("dense_output_1287_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1287_dilations_1 = const()[name = string("dense_output_1287_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1287_groups_1 = const()[name = string("dense_output_1287_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462573952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463622592))))[name = string("layers_15_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1287_cast_fp16 = conv(dilations = dense_output_1287_dilations_1, groups = dense_output_1287_groups_1, pad = dense_output_1287_pad_1, pad_type = dense_output_1287_pad_type_1, strides = dense_output_1287_strides_1, weight = layers_15_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_743_cast_fp16)[name = string("dense_output_1287_cast_fp16")];
            string sparse_output_1287_pad_type_1 = const()[name = string("sparse_output_1287_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1287_strides_1 = const()[name = string("sparse_output_1287_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1287_pad_1 = const()[name = string("sparse_output_1287_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1287_dilations_1 = const()[name = string("sparse_output_1287_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1287_groups_1 = const()[name = string("sparse_output_1287_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_15_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463644224))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463623168))))[name = string("layers_15_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1287_cast_fp16 = conv(dilations = sparse_output_1287_dilations_1, groups = sparse_output_1287_groups_1, pad = sparse_output_1287_pad_1, pad_type = sparse_output_1287_pad_type_1, strides = sparse_output_1287_strides_1, weight = layers_15_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_743_cast_fp16)[name = string("sparse_output_1287_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1347_cast_fp16 = add(x = dense_output_1287_cast_fp16, y = sparse_output_1287_cast_fp16)[name = string("x_1347_cast_fp16")];
            tensor<int32, [1]> var_20444_axes_1 = const()[name = string("op_20444_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_20444_cast_fp16 = squeeze(axes = var_20444_axes_1, x = x_1347_cast_fp16)[name = string("op_20444_cast_fp16")];
            tensor<int32, [3]> var_20445_perm_1 = const()[name = string("op_20445_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_20445_cast_fp16 = transpose(perm = var_20445_perm_1, x = var_20444_cast_fp16)[name = string("transpose_330")];
            tensor<fp16, [1, 51, 1024]> input_745_cast_fp16 = add(x = input_729_cast_fp16, y = var_20445_cast_fp16)[name = string("input_745_cast_fp16")];
            tensor<int32, [1]> x_1349_axes_1 = const()[name = string("x_1349_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463775360)))];
            tensor<fp16, [1024]> layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463777472)))];
            tensor<fp16, [1, 51, 1024]> x_1349_cast_fp16 = layer_norm(axes = x_1349_axes_1, beta = layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_19247_to_fp16, gamma = layers_15_norm_feed_forward2_weight_to_fp16, x = input_745_cast_fp16)[name = string("x_1349_cast_fp16")];
            tensor<int32, [4]> var_20455 = const()[name = string("op_20455"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1351_cast_fp16 = reshape(shape = var_20455, x = x_1349_cast_fp16)[name = string("x_1351_cast_fp16")];
            tensor<int32, [4]> input_747_perm_1 = const()[name = string("input_747_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1289_pad_type_1 = const()[name = string("dense_output_1289_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1289_strides_1 = const()[name = string("dense_output_1289_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1289_pad_1 = const()[name = string("dense_output_1289_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1289_dilations_1 = const()[name = string("dense_output_1289_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1289_groups_1 = const()[name = string("dense_output_1289_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_15_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463779584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467973952))))[name = string("layers_15_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_747_cast_fp16 = transpose(perm = input_747_perm_1, x = x_1351_cast_fp16)[name = string("transpose_329")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1289_cast_fp16 = conv(dilations = dense_output_1289_dilations_1, groups = dense_output_1289_groups_1, pad = dense_output_1289_pad_1, pad_type = dense_output_1289_pad_type_1, strides = dense_output_1289_strides_1, weight = layers_15_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_747_cast_fp16)[name = string("dense_output_1289_cast_fp16")];
            string sparse_output_1289_pad_type_1 = const()[name = string("sparse_output_1289_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1289_strides_1 = const()[name = string("sparse_output_1289_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1289_pad_1 = const()[name = string("sparse_output_1289_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1289_dilations_1 = const()[name = string("sparse_output_1289_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1289_groups_1 = const()[name = string("sparse_output_1289_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_15_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468058496))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467974528))))[name = string("layers_15_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1289_cast_fp16 = conv(dilations = sparse_output_1289_dilations_1, groups = sparse_output_1289_groups_1, pad = sparse_output_1289_pad_1, pad_type = sparse_output_1289_pad_type_1, strides = sparse_output_1289_strides_1, weight = layers_15_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_747_cast_fp16)[name = string("sparse_output_1289_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_749_cast_fp16 = add(x = dense_output_1289_cast_fp16, y = sparse_output_1289_cast_fp16)[name = string("input_749_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_751_cast_fp16 = silu(x = input_749_cast_fp16)[name = string("input_751_cast_fp16")];
            string dense_output_1291_pad_type_1 = const()[name = string("dense_output_1291_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1291_strides_1 = const()[name = string("dense_output_1291_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1291_pad_1 = const()[name = string("dense_output_1291_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1291_dilations_1 = const()[name = string("dense_output_1291_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1291_groups_1 = const()[name = string("dense_output_1291_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_15_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468582848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(472777216))))[name = string("layers_15_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1291_cast_fp16 = conv(dilations = dense_output_1291_dilations_1, groups = dense_output_1291_groups_1, pad = dense_output_1291_pad_1, pad_type = dense_output_1291_pad_type_1, strides = dense_output_1291_strides_1, weight = layers_15_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_751_cast_fp16)[name = string("dense_output_1291_cast_fp16")];
            string sparse_output_1291_pad_type_1 = const()[name = string("sparse_output_1291_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1291_strides_1 = const()[name = string("sparse_output_1291_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1291_pad_1 = const()[name = string("sparse_output_1291_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1291_dilations_1 = const()[name = string("sparse_output_1291_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1291_groups_1 = const()[name = string("sparse_output_1291_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_15_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(472861760))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(472777792))))[name = string("layers_15_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1291_cast_fp16 = conv(dilations = sparse_output_1291_dilations_1, groups = sparse_output_1291_groups_1, pad = sparse_output_1291_pad_1, pad_type = sparse_output_1291_pad_type_1, strides = sparse_output_1291_strides_1, weight = layers_15_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_751_cast_fp16)[name = string("sparse_output_1291_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1353_cast_fp16 = add(x = dense_output_1291_cast_fp16, y = sparse_output_1291_cast_fp16)[name = string("x_1353_cast_fp16")];
            tensor<int32, [4]> x_1355_perm_1 = const()[name = string("x_1355_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_20490 = const()[name = string("op_20490"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1355_cast_fp16 = transpose(perm = x_1355_perm_1, x = x_1353_cast_fp16)[name = string("transpose_328")];
            tensor<fp16, [1, 51, 1024]> var_20491_cast_fp16 = reshape(shape = var_20490, x = x_1355_cast_fp16)[name = string("op_20491_cast_fp16")];
            fp16 var_20492_to_fp16 = const()[name = string("op_20492_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_20493_cast_fp16 = mul(x = var_20491_cast_fp16, y = var_20492_to_fp16)[name = string("op_20493_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_753_cast_fp16 = add(x = input_745_cast_fp16, y = var_20493_cast_fp16)[name = string("input_753_cast_fp16")];
            tensor<int32, [1]> input_755_axes_1 = const()[name = string("input_755_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_15_norm_out_weight_to_fp16 = const()[name = string("layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473386112)))];
            tensor<fp16, [1024]> layers_15_norm_out_bias_to_fp16 = const()[name = string("layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473388224)))];
            tensor<fp16, [1, 51, 1024]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_1, beta = layers_15_norm_out_bias_to_fp16, epsilon = var_19247_to_fp16, gamma = layers_15_norm_out_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")];
            int32 var_20501 = const()[name = string("op_20501"), val = int32(-1)];
            tensor<int32, [1]> x_1357_axes_1 = const()[name = string("x_1357_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473390336)))];
            tensor<fp16, [1024]> layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473392448)))];
            fp16 var_20516_to_fp16 = const()[name = string("op_20516_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1357_cast_fp16 = layer_norm(axes = x_1357_axes_1, beta = layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_20516_to_fp16, gamma = layers_16_norm_feed_forward1_weight_to_fp16, x = input_755_cast_fp16)[name = string("x_1357_cast_fp16")];
            tensor<int32, [4]> var_20535 = const()[name = string("op_20535"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1359_cast_fp16 = reshape(shape = var_20535, x = x_1357_cast_fp16)[name = string("x_1359_cast_fp16")];
            tensor<int32, [4]> input_757_perm_1 = const()[name = string("input_757_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1293_pad_type_1 = const()[name = string("dense_output_1293_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1293_strides_1 = const()[name = string("dense_output_1293_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1293_pad_1 = const()[name = string("dense_output_1293_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1293_dilations_1 = const()[name = string("dense_output_1293_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1293_groups_1 = const()[name = string("dense_output_1293_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_16_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473394560))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477588928))))[name = string("layers_16_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_757_cast_fp16 = transpose(perm = input_757_perm_1, x = x_1359_cast_fp16)[name = string("transpose_327")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1293_cast_fp16 = conv(dilations = dense_output_1293_dilations_1, groups = dense_output_1293_groups_1, pad = dense_output_1293_pad_1, pad_type = dense_output_1293_pad_type_1, strides = dense_output_1293_strides_1, weight = layers_16_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_757_cast_fp16)[name = string("dense_output_1293_cast_fp16")];
            string sparse_output_1293_pad_type_1 = const()[name = string("sparse_output_1293_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1293_strides_1 = const()[name = string("sparse_output_1293_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1293_pad_1 = const()[name = string("sparse_output_1293_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1293_dilations_1 = const()[name = string("sparse_output_1293_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1293_groups_1 = const()[name = string("sparse_output_1293_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_16_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477673472))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477589504))))[name = string("layers_16_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1293_cast_fp16 = conv(dilations = sparse_output_1293_dilations_1, groups = sparse_output_1293_groups_1, pad = sparse_output_1293_pad_1, pad_type = sparse_output_1293_pad_type_1, strides = sparse_output_1293_strides_1, weight = layers_16_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_757_cast_fp16)[name = string("sparse_output_1293_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_759_cast_fp16 = add(x = dense_output_1293_cast_fp16, y = sparse_output_1293_cast_fp16)[name = string("input_759_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_761_cast_fp16 = silu(x = input_759_cast_fp16)[name = string("input_761_cast_fp16")];
            string dense_output_1295_pad_type_1 = const()[name = string("dense_output_1295_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1295_strides_1 = const()[name = string("dense_output_1295_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1295_pad_1 = const()[name = string("dense_output_1295_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1295_dilations_1 = const()[name = string("dense_output_1295_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1295_groups_1 = const()[name = string("dense_output_1295_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_16_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478197824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482392192))))[name = string("layers_16_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1295_cast_fp16 = conv(dilations = dense_output_1295_dilations_1, groups = dense_output_1295_groups_1, pad = dense_output_1295_pad_1, pad_type = dense_output_1295_pad_type_1, strides = dense_output_1295_strides_1, weight = layers_16_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_761_cast_fp16)[name = string("dense_output_1295_cast_fp16")];
            string sparse_output_1295_pad_type_1 = const()[name = string("sparse_output_1295_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1295_strides_1 = const()[name = string("sparse_output_1295_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1295_pad_1 = const()[name = string("sparse_output_1295_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1295_dilations_1 = const()[name = string("sparse_output_1295_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1295_groups_1 = const()[name = string("sparse_output_1295_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_16_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482476736))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482392768))))[name = string("layers_16_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1295_cast_fp16 = conv(dilations = sparse_output_1295_dilations_1, groups = sparse_output_1295_groups_1, pad = sparse_output_1295_pad_1, pad_type = sparse_output_1295_pad_type_1, strides = sparse_output_1295_strides_1, weight = layers_16_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_761_cast_fp16)[name = string("sparse_output_1295_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1361_cast_fp16 = add(x = dense_output_1295_cast_fp16, y = sparse_output_1295_cast_fp16)[name = string("x_1361_cast_fp16")];
            tensor<int32, [4]> x_1363_perm_1 = const()[name = string("x_1363_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_20570 = const()[name = string("op_20570"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1363_cast_fp16 = transpose(perm = x_1363_perm_1, x = x_1361_cast_fp16)[name = string("transpose_326")];
            tensor<fp16, [1, 51, 1024]> var_20571_cast_fp16 = reshape(shape = var_20570, x = x_1363_cast_fp16)[name = string("op_20571_cast_fp16")];
            fp16 var_20572_to_fp16 = const()[name = string("op_20572_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_20573_cast_fp16 = mul(x = var_20571_cast_fp16, y = var_20572_to_fp16)[name = string("op_20573_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_763_cast_fp16 = add(x = input_755_cast_fp16, y = var_20573_cast_fp16)[name = string("input_763_cast_fp16")];
            tensor<int32, [1]> q_33_axes_1 = const()[name = string("q_33_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_16_norm_self_att_weight_to_fp16 = const()[name = string("layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483001088)))];
            tensor<fp16, [1024]> layers_16_norm_self_att_bias_to_fp16 = const()[name = string("layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483003200)))];
            tensor<fp16, [1, 51, 1024]> q_33_cast_fp16 = layer_norm(axes = q_33_axes_1, beta = layers_16_norm_self_att_bias_to_fp16, epsilon = var_20516_to_fp16, gamma = layers_16_norm_self_att_weight_to_fp16, x = input_763_cast_fp16)[name = string("q_33_cast_fp16")];
            tensor<int32, [3]> var_20647 = const()[name = string("op_20647"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_765_axes_1 = const()[name = string("input_765_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_20648_cast_fp16 = transpose(perm = var_20647, x = q_33_cast_fp16)[name = string("transpose_325")];
            tensor<fp16, [1, 1024, 51, 1]> input_765_cast_fp16 = expand_dims(axes = input_765_axes_1, x = var_20648_cast_fp16)[name = string("input_765_cast_fp16")];
            string dense_output_1297_pad_type_1 = const()[name = string("dense_output_1297_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1297_strides_1 = const()[name = string("dense_output_1297_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1297_pad_1 = const()[name = string("dense_output_1297_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1297_dilations_1 = const()[name = string("dense_output_1297_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1297_groups_1 = const()[name = string("dense_output_1297_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483005312))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483136448))))[name = string("layers_16_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1297_cast_fp16 = conv(dilations = dense_output_1297_dilations_1, groups = dense_output_1297_groups_1, pad = dense_output_1297_pad_1, pad_type = dense_output_1297_pad_type_1, strides = dense_output_1297_strides_1, weight = layers_16_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1297_cast_fp16")];
            string sparse_output_1297_pad_type_1 = const()[name = string("sparse_output_1297_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1297_strides_1 = const()[name = string("sparse_output_1297_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1297_pad_1 = const()[name = string("sparse_output_1297_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1297_dilations_1 = const()[name = string("sparse_output_1297_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1297_groups_1 = const()[name = string("sparse_output_1297_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483139712))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483137024))))[name = string("layers_16_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1297_cast_fp16 = conv(dilations = sparse_output_1297_dilations_1, groups = sparse_output_1297_groups_1, pad = sparse_output_1297_pad_1, pad_type = sparse_output_1297_pad_type_1, strides = sparse_output_1297_strides_1, weight = layers_16_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1297_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20673_cast_fp16 = add(x = dense_output_1297_cast_fp16, y = sparse_output_1297_cast_fp16)[name = string("op_20673_cast_fp16")];
            tensor<int32, [4]> var_20674 = const()[name = string("op_20674"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20676 = const()[name = string("op_20676"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20675_cast_fp16 = transpose(perm = var_20674, x = var_20673_cast_fp16)[name = string("transpose_324")];
            tensor<fp16, [1, 51, 128]> q_head_257_cast_fp16 = reshape(shape = var_20676, x = var_20675_cast_fp16)[name = string("q_head_257_cast_fp16")];
            string dense_output_1299_pad_type_1 = const()[name = string("dense_output_1299_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1299_strides_1 = const()[name = string("dense_output_1299_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1299_pad_1 = const()[name = string("dense_output_1299_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1299_dilations_1 = const()[name = string("dense_output_1299_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1299_groups_1 = const()[name = string("dense_output_1299_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483156160))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483287296))))[name = string("layers_16_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1299_cast_fp16 = conv(dilations = dense_output_1299_dilations_1, groups = dense_output_1299_groups_1, pad = dense_output_1299_pad_1, pad_type = dense_output_1299_pad_type_1, strides = dense_output_1299_strides_1, weight = layers_16_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1299_cast_fp16")];
            string sparse_output_1299_pad_type_1 = const()[name = string("sparse_output_1299_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1299_strides_1 = const()[name = string("sparse_output_1299_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1299_pad_1 = const()[name = string("sparse_output_1299_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1299_dilations_1 = const()[name = string("sparse_output_1299_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1299_groups_1 = const()[name = string("sparse_output_1299_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483290560))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483287872))))[name = string("layers_16_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1299_cast_fp16 = conv(dilations = sparse_output_1299_dilations_1, groups = sparse_output_1299_groups_1, pad = sparse_output_1299_pad_1, pad_type = sparse_output_1299_pad_type_1, strides = sparse_output_1299_strides_1, weight = layers_16_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1299_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20692_cast_fp16 = add(x = dense_output_1299_cast_fp16, y = sparse_output_1299_cast_fp16)[name = string("op_20692_cast_fp16")];
            tensor<int32, [4]> var_20693 = const()[name = string("op_20693"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20695 = const()[name = string("op_20695"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20694_cast_fp16 = transpose(perm = var_20693, x = var_20692_cast_fp16)[name = string("transpose_323")];
            tensor<fp16, [1, 51, 128]> k_head_513_cast_fp16 = reshape(shape = var_20695, x = var_20694_cast_fp16)[name = string("k_head_513_cast_fp16")];
            string dense_output_1301_pad_type_1 = const()[name = string("dense_output_1301_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1301_strides_1 = const()[name = string("dense_output_1301_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1301_pad_1 = const()[name = string("dense_output_1301_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1301_dilations_1 = const()[name = string("dense_output_1301_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1301_groups_1 = const()[name = string("dense_output_1301_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483307008))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483438144))))[name = string("layers_16_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1301_cast_fp16 = conv(dilations = dense_output_1301_dilations_1, groups = dense_output_1301_groups_1, pad = dense_output_1301_pad_1, pad_type = dense_output_1301_pad_type_1, strides = dense_output_1301_strides_1, weight = layers_16_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1301_cast_fp16")];
            string sparse_output_1301_pad_type_1 = const()[name = string("sparse_output_1301_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1301_strides_1 = const()[name = string("sparse_output_1301_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1301_pad_1 = const()[name = string("sparse_output_1301_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1301_dilations_1 = const()[name = string("sparse_output_1301_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1301_groups_1 = const()[name = string("sparse_output_1301_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483441408))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483438720))))[name = string("layers_16_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1301_cast_fp16 = conv(dilations = sparse_output_1301_dilations_1, groups = sparse_output_1301_groups_1, pad = sparse_output_1301_pad_1, pad_type = sparse_output_1301_pad_type_1, strides = sparse_output_1301_strides_1, weight = layers_16_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1301_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20711_cast_fp16 = add(x = dense_output_1301_cast_fp16, y = sparse_output_1301_cast_fp16)[name = string("op_20711_cast_fp16")];
            tensor<int32, [4]> var_20712 = const()[name = string("op_20712"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20714 = const()[name = string("op_20714"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20713_cast_fp16 = transpose(perm = var_20712, x = var_20711_cast_fp16)[name = string("transpose_322")];
            tensor<fp16, [1, 51, 128]> v_head_513_cast_fp16 = reshape(shape = var_20714, x = var_20713_cast_fp16)[name = string("v_head_513_cast_fp16")];
            string dense_output_1303_pad_type_1 = const()[name = string("dense_output_1303_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1303_strides_1 = const()[name = string("dense_output_1303_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1303_pad_1 = const()[name = string("dense_output_1303_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1303_dilations_1 = const()[name = string("dense_output_1303_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1303_groups_1 = const()[name = string("dense_output_1303_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483457856))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483588992))))[name = string("layers_16_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1303_cast_fp16 = conv(dilations = dense_output_1303_dilations_1, groups = dense_output_1303_groups_1, pad = dense_output_1303_pad_1, pad_type = dense_output_1303_pad_type_1, strides = dense_output_1303_strides_1, weight = layers_16_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1303_cast_fp16")];
            string sparse_output_1303_pad_type_1 = const()[name = string("sparse_output_1303_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1303_strides_1 = const()[name = string("sparse_output_1303_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1303_pad_1 = const()[name = string("sparse_output_1303_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1303_dilations_1 = const()[name = string("sparse_output_1303_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1303_groups_1 = const()[name = string("sparse_output_1303_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483592256))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483589568))))[name = string("layers_16_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1303_cast_fp16 = conv(dilations = sparse_output_1303_dilations_1, groups = sparse_output_1303_groups_1, pad = sparse_output_1303_pad_1, pad_type = sparse_output_1303_pad_type_1, strides = sparse_output_1303_strides_1, weight = layers_16_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1303_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_20730_cast_fp16 = add(x = dense_output_1303_cast_fp16, y = sparse_output_1303_cast_fp16)[name = string("op_20730_cast_fp16")];
            tensor<int32, [4]> var_20731 = const()[name = string("op_20731"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20733 = const()[name = string("op_20733"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_20732_cast_fp16 = transpose(perm = var_20731, x = var_20730_cast_fp16)[name = string("transpose_321")];
            tensor<fp16, [1, 101, 128]> p_head_513_cast_fp16 = reshape(shape = var_20733, x = var_20732_cast_fp16)[name = string("p_head_513_cast_fp16")];
            tensor<fp16, [128]> var_20735_to_fp16 = const()[name = string("op_20735_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483608704)))];
            tensor<fp16, [1, 51, 128]> var_20736_cast_fp16 = add(x = q_head_257_cast_fp16, y = var_20735_to_fp16)[name = string("op_20736_cast_fp16")];
            tensor<int32, [1]> q_u_257_axes_1 = const()[name = string("q_u_257_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_257_cast_fp16 = expand_dims(axes = q_u_257_axes_1, x = var_20736_cast_fp16)[name = string("q_u_257_cast_fp16")];
            tensor<fp16, [128]> var_20738_to_fp16 = const()[name = string("op_20738_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483609024)))];
            tensor<fp16, [1, 51, 128]> var_20739_cast_fp16 = add(x = q_head_257_cast_fp16, y = var_20738_to_fp16)[name = string("op_20739_cast_fp16")];
            tensor<int32, [1]> q_v_257_axes_1 = const()[name = string("q_v_257_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_257_cast_fp16 = expand_dims(axes = q_v_257_axes_1, x = var_20739_cast_fp16)[name = string("q_v_257_cast_fp16")];
            tensor<int32, [1]> k_head_515_axes_1 = const()[name = string("k_head_515_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_515_cast_fp16 = expand_dims(axes = k_head_515_axes_1, x = k_head_513_cast_fp16)[name = string("k_head_515_cast_fp16")];
            tensor<int32, [1]> v_head_515_axes_1 = const()[name = string("v_head_515_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_515_cast_fp16 = expand_dims(axes = v_head_515_axes_1, x = v_head_513_cast_fp16)[name = string("v_head_515_cast_fp16")];
            tensor<int32, [1]> p_head_515_axes_1 = const()[name = string("p_head_515_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_515_cast_fp16 = expand_dims(axes = p_head_515_axes_1, x = p_head_513_cast_fp16)[name = string("p_head_515_cast_fp16")];
            bool var_20745_transpose_x_3 = const()[name = string("op_20745_transpose_x_3"), val = bool(false)];
            bool var_20745_transpose_y_3 = const()[name = string("op_20745_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_20745_cast_fp16 = matmul(transpose_x = var_20745_transpose_x_3, transpose_y = var_20745_transpose_y_3, x = q_u_257_cast_fp16, y = k_head_515_cast_fp16)[name = string("op_20745_cast_fp16")];
            fp16 var_20746_to_fp16 = const()[name = string("op_20746_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_257_cast_fp16 = mul(x = var_20745_cast_fp16, y = var_20746_to_fp16)[name = string("scores_content_257_cast_fp16")];
            bool x_1365_transpose_x_3 = const()[name = string("x_1365_transpose_x_3"), val = bool(false)];
            bool x_1365_transpose_y_3 = const()[name = string("x_1365_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1365_cast_fp16 = matmul(transpose_x = x_1365_transpose_x_3, transpose_y = x_1365_transpose_y_3, x = q_v_257_cast_fp16, y = p_head_515_cast_fp16)[name = string("x_1365_cast_fp16")];
            tensor<int32, [8]> x_1367_pad_1 = const()[name = string("x_1367_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1367_mode_1 = const()[name = string("x_1367_mode_1"), val = string("constant")];
            fp16 const_2173_to_fp16 = const()[name = string("const_2173_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1367_cast_fp16 = pad(constant_val = const_2173_to_fp16, mode = x_1367_mode_1, pad = x_1367_pad_1, x = x_1365_cast_fp16)[name = string("x_1367_cast_fp16")];
            tensor<int32, [4]> var_20760 = const()[name = string("op_20760"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1369_cast_fp16 = reshape(shape = var_20760, x = x_1367_cast_fp16)[name = string("x_1369_cast_fp16")];
            tensor<int32, [4]> var_20764_begin_1 = const()[name = string("op_20764_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_20764_end_1 = const()[name = string("op_20764_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_20764_end_mask_1 = const()[name = string("op_20764_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_20764_cast_fp16 = slice_by_index(begin = var_20764_begin_1, end = var_20764_end_1, end_mask = var_20764_end_mask_1, x = x_1369_cast_fp16)[name = string("op_20764_cast_fp16")];
            tensor<int32, [4]> var_20766 = const()[name = string("op_20766"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_20767_cast_fp16 = reshape(shape = var_20766, x = var_20764_cast_fp16)[name = string("op_20767_cast_fp16")];
            tensor<int32, [4]> var_20772_begin_1 = const()[name = string("op_20772_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_20772_end_1 = const()[name = string("op_20772_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_20772_end_mask_1 = const()[name = string("op_20772_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_20772_cast_fp16 = slice_by_index(begin = var_20772_begin_1, end = var_20772_end_1, end_mask = var_20772_end_mask_1, x = var_20767_cast_fp16)[name = string("op_20772_cast_fp16")];
            fp16 var_20773_to_fp16 = const()[name = string("op_20773_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_257_cast_fp16 = mul(x = var_20772_cast_fp16, y = var_20773_to_fp16)[name = string("scores_pos_257_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_257_cast_fp16 = add(x = scores_content_257_cast_fp16, y = scores_pos_257_cast_fp16)[name = string("logits_257_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_20776_cast_fp16 = softmax(axis = var_20501, x = logits_257_cast_fp16)[name = string("op_20776_cast_fp16")];
            bool var_20778_transpose_x_1 = const()[name = string("op_20778_transpose_x_1"), val = bool(false)];
            bool var_20778_transpose_y_1 = const()[name = string("op_20778_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_20778_cast_fp16 = matmul(transpose_x = var_20778_transpose_x_1, transpose_y = var_20778_transpose_y_1, x = var_20776_cast_fp16, y = v_head_515_cast_fp16)[name = string("op_20778_cast_fp16")];
            tensor<int32, [1]> var_20779_axes_1 = const()[name = string("op_20779_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_20779_cast_fp16 = squeeze(axes = var_20779_axes_1, x = var_20778_cast_fp16)[name = string("op_20779_cast_fp16")];
            string dense_output_1305_pad_type_1 = const()[name = string("dense_output_1305_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1305_strides_1 = const()[name = string("dense_output_1305_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1305_pad_1 = const()[name = string("dense_output_1305_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1305_dilations_1 = const()[name = string("dense_output_1305_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1305_groups_1 = const()[name = string("dense_output_1305_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483609344))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483740480))))[name = string("layers_16_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1305_cast_fp16 = conv(dilations = dense_output_1305_dilations_1, groups = dense_output_1305_groups_1, pad = dense_output_1305_pad_1, pad_type = dense_output_1305_pad_type_1, strides = dense_output_1305_strides_1, weight = layers_16_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1305_cast_fp16")];
            string sparse_output_1305_pad_type_1 = const()[name = string("sparse_output_1305_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1305_strides_1 = const()[name = string("sparse_output_1305_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1305_pad_1 = const()[name = string("sparse_output_1305_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1305_dilations_1 = const()[name = string("sparse_output_1305_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1305_groups_1 = const()[name = string("sparse_output_1305_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483743744))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483741056))))[name = string("layers_16_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1305_cast_fp16 = conv(dilations = sparse_output_1305_dilations_1, groups = sparse_output_1305_groups_1, pad = sparse_output_1305_pad_1, pad_type = sparse_output_1305_pad_type_1, strides = sparse_output_1305_strides_1, weight = layers_16_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1305_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20794_cast_fp16 = add(x = dense_output_1305_cast_fp16, y = sparse_output_1305_cast_fp16)[name = string("op_20794_cast_fp16")];
            tensor<int32, [4]> var_20795 = const()[name = string("op_20795"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20797 = const()[name = string("op_20797"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20796_cast_fp16 = transpose(perm = var_20795, x = var_20794_cast_fp16)[name = string("transpose_320")];
            tensor<fp16, [1, 51, 128]> q_head_259_cast_fp16 = reshape(shape = var_20797, x = var_20796_cast_fp16)[name = string("q_head_259_cast_fp16")];
            string dense_output_1307_pad_type_1 = const()[name = string("dense_output_1307_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1307_strides_1 = const()[name = string("dense_output_1307_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1307_pad_1 = const()[name = string("dense_output_1307_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1307_dilations_1 = const()[name = string("dense_output_1307_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1307_groups_1 = const()[name = string("dense_output_1307_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483760192))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483891328))))[name = string("layers_16_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1307_cast_fp16 = conv(dilations = dense_output_1307_dilations_1, groups = dense_output_1307_groups_1, pad = dense_output_1307_pad_1, pad_type = dense_output_1307_pad_type_1, strides = dense_output_1307_strides_1, weight = layers_16_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1307_cast_fp16")];
            string sparse_output_1307_pad_type_1 = const()[name = string("sparse_output_1307_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1307_strides_1 = const()[name = string("sparse_output_1307_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1307_pad_1 = const()[name = string("sparse_output_1307_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1307_dilations_1 = const()[name = string("sparse_output_1307_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1307_groups_1 = const()[name = string("sparse_output_1307_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483894592))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483891904))))[name = string("layers_16_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1307_cast_fp16 = conv(dilations = sparse_output_1307_dilations_1, groups = sparse_output_1307_groups_1, pad = sparse_output_1307_pad_1, pad_type = sparse_output_1307_pad_type_1, strides = sparse_output_1307_strides_1, weight = layers_16_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1307_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20813_cast_fp16 = add(x = dense_output_1307_cast_fp16, y = sparse_output_1307_cast_fp16)[name = string("op_20813_cast_fp16")];
            tensor<int32, [4]> var_20814 = const()[name = string("op_20814"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20816 = const()[name = string("op_20816"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20815_cast_fp16 = transpose(perm = var_20814, x = var_20813_cast_fp16)[name = string("transpose_319")];
            tensor<fp16, [1, 51, 128]> k_head_517_cast_fp16 = reshape(shape = var_20816, x = var_20815_cast_fp16)[name = string("k_head_517_cast_fp16")];
            string dense_output_1309_pad_type_1 = const()[name = string("dense_output_1309_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1309_strides_1 = const()[name = string("dense_output_1309_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1309_pad_1 = const()[name = string("dense_output_1309_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1309_dilations_1 = const()[name = string("dense_output_1309_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1309_groups_1 = const()[name = string("dense_output_1309_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483911040))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484042176))))[name = string("layers_16_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1309_cast_fp16 = conv(dilations = dense_output_1309_dilations_1, groups = dense_output_1309_groups_1, pad = dense_output_1309_pad_1, pad_type = dense_output_1309_pad_type_1, strides = dense_output_1309_strides_1, weight = layers_16_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1309_cast_fp16")];
            string sparse_output_1309_pad_type_1 = const()[name = string("sparse_output_1309_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1309_strides_1 = const()[name = string("sparse_output_1309_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1309_pad_1 = const()[name = string("sparse_output_1309_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1309_dilations_1 = const()[name = string("sparse_output_1309_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1309_groups_1 = const()[name = string("sparse_output_1309_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484045440))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484042752))))[name = string("layers_16_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1309_cast_fp16 = conv(dilations = sparse_output_1309_dilations_1, groups = sparse_output_1309_groups_1, pad = sparse_output_1309_pad_1, pad_type = sparse_output_1309_pad_type_1, strides = sparse_output_1309_strides_1, weight = layers_16_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1309_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20832_cast_fp16 = add(x = dense_output_1309_cast_fp16, y = sparse_output_1309_cast_fp16)[name = string("op_20832_cast_fp16")];
            tensor<int32, [4]> var_20833 = const()[name = string("op_20833"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20835 = const()[name = string("op_20835"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20834_cast_fp16 = transpose(perm = var_20833, x = var_20832_cast_fp16)[name = string("transpose_318")];
            tensor<fp16, [1, 51, 128]> v_head_517_cast_fp16 = reshape(shape = var_20835, x = var_20834_cast_fp16)[name = string("v_head_517_cast_fp16")];
            string dense_output_1311_pad_type_1 = const()[name = string("dense_output_1311_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1311_strides_1 = const()[name = string("dense_output_1311_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1311_pad_1 = const()[name = string("dense_output_1311_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1311_dilations_1 = const()[name = string("dense_output_1311_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1311_groups_1 = const()[name = string("dense_output_1311_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484061888))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484193024))))[name = string("layers_16_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1311_cast_fp16 = conv(dilations = dense_output_1311_dilations_1, groups = dense_output_1311_groups_1, pad = dense_output_1311_pad_1, pad_type = dense_output_1311_pad_type_1, strides = dense_output_1311_strides_1, weight = layers_16_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1311_cast_fp16")];
            string sparse_output_1311_pad_type_1 = const()[name = string("sparse_output_1311_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1311_strides_1 = const()[name = string("sparse_output_1311_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1311_pad_1 = const()[name = string("sparse_output_1311_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1311_dilations_1 = const()[name = string("sparse_output_1311_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1311_groups_1 = const()[name = string("sparse_output_1311_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484196288))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484193600))))[name = string("layers_16_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1311_cast_fp16 = conv(dilations = sparse_output_1311_dilations_1, groups = sparse_output_1311_groups_1, pad = sparse_output_1311_pad_1, pad_type = sparse_output_1311_pad_type_1, strides = sparse_output_1311_strides_1, weight = layers_16_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1311_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_20851_cast_fp16 = add(x = dense_output_1311_cast_fp16, y = sparse_output_1311_cast_fp16)[name = string("op_20851_cast_fp16")];
            tensor<int32, [4]> var_20852 = const()[name = string("op_20852"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20854 = const()[name = string("op_20854"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_20853_cast_fp16 = transpose(perm = var_20852, x = var_20851_cast_fp16)[name = string("transpose_317")];
            tensor<fp16, [1, 101, 128]> p_head_517_cast_fp16 = reshape(shape = var_20854, x = var_20853_cast_fp16)[name = string("p_head_517_cast_fp16")];
            tensor<fp16, [128]> var_20856_to_fp16 = const()[name = string("op_20856_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484212736)))];
            tensor<fp16, [1, 51, 128]> var_20857_cast_fp16 = add(x = q_head_259_cast_fp16, y = var_20856_to_fp16)[name = string("op_20857_cast_fp16")];
            tensor<int32, [1]> q_u_259_axes_1 = const()[name = string("q_u_259_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_259_cast_fp16 = expand_dims(axes = q_u_259_axes_1, x = var_20857_cast_fp16)[name = string("q_u_259_cast_fp16")];
            tensor<fp16, [128]> var_20859_to_fp16 = const()[name = string("op_20859_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484213056)))];
            tensor<fp16, [1, 51, 128]> var_20860_cast_fp16 = add(x = q_head_259_cast_fp16, y = var_20859_to_fp16)[name = string("op_20860_cast_fp16")];
            tensor<int32, [1]> q_v_259_axes_1 = const()[name = string("q_v_259_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_259_cast_fp16 = expand_dims(axes = q_v_259_axes_1, x = var_20860_cast_fp16)[name = string("q_v_259_cast_fp16")];
            tensor<int32, [1]> k_head_519_axes_1 = const()[name = string("k_head_519_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_519_cast_fp16 = expand_dims(axes = k_head_519_axes_1, x = k_head_517_cast_fp16)[name = string("k_head_519_cast_fp16")];
            tensor<int32, [1]> v_head_519_axes_1 = const()[name = string("v_head_519_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_519_cast_fp16 = expand_dims(axes = v_head_519_axes_1, x = v_head_517_cast_fp16)[name = string("v_head_519_cast_fp16")];
            tensor<int32, [1]> p_head_519_axes_1 = const()[name = string("p_head_519_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_519_cast_fp16 = expand_dims(axes = p_head_519_axes_1, x = p_head_517_cast_fp16)[name = string("p_head_519_cast_fp16")];
            bool var_20866_transpose_x_3 = const()[name = string("op_20866_transpose_x_3"), val = bool(false)];
            bool var_20866_transpose_y_3 = const()[name = string("op_20866_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_20866_cast_fp16 = matmul(transpose_x = var_20866_transpose_x_3, transpose_y = var_20866_transpose_y_3, x = q_u_259_cast_fp16, y = k_head_519_cast_fp16)[name = string("op_20866_cast_fp16")];
            fp16 var_20867_to_fp16 = const()[name = string("op_20867_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_259_cast_fp16 = mul(x = var_20866_cast_fp16, y = var_20867_to_fp16)[name = string("scores_content_259_cast_fp16")];
            bool x_1373_transpose_x_3 = const()[name = string("x_1373_transpose_x_3"), val = bool(false)];
            bool x_1373_transpose_y_3 = const()[name = string("x_1373_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1373_cast_fp16 = matmul(transpose_x = x_1373_transpose_x_3, transpose_y = x_1373_transpose_y_3, x = q_v_259_cast_fp16, y = p_head_519_cast_fp16)[name = string("x_1373_cast_fp16")];
            tensor<int32, [8]> x_1375_pad_1 = const()[name = string("x_1375_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1375_mode_1 = const()[name = string("x_1375_mode_1"), val = string("constant")];
            fp16 const_2179_to_fp16 = const()[name = string("const_2179_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1375_cast_fp16 = pad(constant_val = const_2179_to_fp16, mode = x_1375_mode_1, pad = x_1375_pad_1, x = x_1373_cast_fp16)[name = string("x_1375_cast_fp16")];
            tensor<int32, [4]> var_20881 = const()[name = string("op_20881"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1377_cast_fp16 = reshape(shape = var_20881, x = x_1375_cast_fp16)[name = string("x_1377_cast_fp16")];
            tensor<int32, [4]> var_20885_begin_1 = const()[name = string("op_20885_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_20885_end_1 = const()[name = string("op_20885_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_20885_end_mask_1 = const()[name = string("op_20885_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_20885_cast_fp16 = slice_by_index(begin = var_20885_begin_1, end = var_20885_end_1, end_mask = var_20885_end_mask_1, x = x_1377_cast_fp16)[name = string("op_20885_cast_fp16")];
            tensor<int32, [4]> var_20887 = const()[name = string("op_20887"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_20888_cast_fp16 = reshape(shape = var_20887, x = var_20885_cast_fp16)[name = string("op_20888_cast_fp16")];
            tensor<int32, [4]> var_20893_begin_1 = const()[name = string("op_20893_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_20893_end_1 = const()[name = string("op_20893_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_20893_end_mask_1 = const()[name = string("op_20893_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_20893_cast_fp16 = slice_by_index(begin = var_20893_begin_1, end = var_20893_end_1, end_mask = var_20893_end_mask_1, x = var_20888_cast_fp16)[name = string("op_20893_cast_fp16")];
            fp16 var_20894_to_fp16 = const()[name = string("op_20894_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_259_cast_fp16 = mul(x = var_20893_cast_fp16, y = var_20894_to_fp16)[name = string("scores_pos_259_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_259_cast_fp16 = add(x = scores_content_259_cast_fp16, y = scores_pos_259_cast_fp16)[name = string("logits_259_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_20897_cast_fp16 = softmax(axis = var_20501, x = logits_259_cast_fp16)[name = string("op_20897_cast_fp16")];
            bool var_20899_transpose_x_1 = const()[name = string("op_20899_transpose_x_1"), val = bool(false)];
            bool var_20899_transpose_y_1 = const()[name = string("op_20899_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_20899_cast_fp16 = matmul(transpose_x = var_20899_transpose_x_1, transpose_y = var_20899_transpose_y_1, x = var_20897_cast_fp16, y = v_head_519_cast_fp16)[name = string("op_20899_cast_fp16")];
            tensor<int32, [1]> var_20900_axes_1 = const()[name = string("op_20900_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_20900_cast_fp16 = squeeze(axes = var_20900_axes_1, x = var_20899_cast_fp16)[name = string("op_20900_cast_fp16")];
            string dense_output_1313_pad_type_1 = const()[name = string("dense_output_1313_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1313_strides_1 = const()[name = string("dense_output_1313_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1313_pad_1 = const()[name = string("dense_output_1313_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1313_dilations_1 = const()[name = string("dense_output_1313_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1313_groups_1 = const()[name = string("dense_output_1313_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484213376))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484344512))))[name = string("layers_16_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1313_cast_fp16 = conv(dilations = dense_output_1313_dilations_1, groups = dense_output_1313_groups_1, pad = dense_output_1313_pad_1, pad_type = dense_output_1313_pad_type_1, strides = dense_output_1313_strides_1, weight = layers_16_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1313_cast_fp16")];
            string sparse_output_1313_pad_type_1 = const()[name = string("sparse_output_1313_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1313_strides_1 = const()[name = string("sparse_output_1313_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1313_pad_1 = const()[name = string("sparse_output_1313_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1313_dilations_1 = const()[name = string("sparse_output_1313_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1313_groups_1 = const()[name = string("sparse_output_1313_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484347776))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484345088))))[name = string("layers_16_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1313_cast_fp16 = conv(dilations = sparse_output_1313_dilations_1, groups = sparse_output_1313_groups_1, pad = sparse_output_1313_pad_1, pad_type = sparse_output_1313_pad_type_1, strides = sparse_output_1313_strides_1, weight = layers_16_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1313_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20915_cast_fp16 = add(x = dense_output_1313_cast_fp16, y = sparse_output_1313_cast_fp16)[name = string("op_20915_cast_fp16")];
            tensor<int32, [4]> var_20916 = const()[name = string("op_20916"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20918 = const()[name = string("op_20918"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20917_cast_fp16 = transpose(perm = var_20916, x = var_20915_cast_fp16)[name = string("transpose_316")];
            tensor<fp16, [1, 51, 128]> q_head_261_cast_fp16 = reshape(shape = var_20918, x = var_20917_cast_fp16)[name = string("q_head_261_cast_fp16")];
            string dense_output_1315_pad_type_1 = const()[name = string("dense_output_1315_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1315_strides_1 = const()[name = string("dense_output_1315_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1315_pad_1 = const()[name = string("dense_output_1315_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1315_dilations_1 = const()[name = string("dense_output_1315_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1315_groups_1 = const()[name = string("dense_output_1315_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484364224))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484495360))))[name = string("layers_16_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1315_cast_fp16 = conv(dilations = dense_output_1315_dilations_1, groups = dense_output_1315_groups_1, pad = dense_output_1315_pad_1, pad_type = dense_output_1315_pad_type_1, strides = dense_output_1315_strides_1, weight = layers_16_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1315_cast_fp16")];
            string sparse_output_1315_pad_type_1 = const()[name = string("sparse_output_1315_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1315_strides_1 = const()[name = string("sparse_output_1315_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1315_pad_1 = const()[name = string("sparse_output_1315_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1315_dilations_1 = const()[name = string("sparse_output_1315_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1315_groups_1 = const()[name = string("sparse_output_1315_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484498624))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484495936))))[name = string("layers_16_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1315_cast_fp16 = conv(dilations = sparse_output_1315_dilations_1, groups = sparse_output_1315_groups_1, pad = sparse_output_1315_pad_1, pad_type = sparse_output_1315_pad_type_1, strides = sparse_output_1315_strides_1, weight = layers_16_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1315_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20934_cast_fp16 = add(x = dense_output_1315_cast_fp16, y = sparse_output_1315_cast_fp16)[name = string("op_20934_cast_fp16")];
            tensor<int32, [4]> var_20935 = const()[name = string("op_20935"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20937 = const()[name = string("op_20937"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20936_cast_fp16 = transpose(perm = var_20935, x = var_20934_cast_fp16)[name = string("transpose_315")];
            tensor<fp16, [1, 51, 128]> k_head_521_cast_fp16 = reshape(shape = var_20937, x = var_20936_cast_fp16)[name = string("k_head_521_cast_fp16")];
            string dense_output_1317_pad_type_1 = const()[name = string("dense_output_1317_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1317_strides_1 = const()[name = string("dense_output_1317_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1317_pad_1 = const()[name = string("dense_output_1317_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1317_dilations_1 = const()[name = string("dense_output_1317_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1317_groups_1 = const()[name = string("dense_output_1317_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484515072))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484646208))))[name = string("layers_16_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1317_cast_fp16 = conv(dilations = dense_output_1317_dilations_1, groups = dense_output_1317_groups_1, pad = dense_output_1317_pad_1, pad_type = dense_output_1317_pad_type_1, strides = dense_output_1317_strides_1, weight = layers_16_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1317_cast_fp16")];
            string sparse_output_1317_pad_type_1 = const()[name = string("sparse_output_1317_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1317_strides_1 = const()[name = string("sparse_output_1317_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1317_pad_1 = const()[name = string("sparse_output_1317_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1317_dilations_1 = const()[name = string("sparse_output_1317_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1317_groups_1 = const()[name = string("sparse_output_1317_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484649472))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484646784))))[name = string("layers_16_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1317_cast_fp16 = conv(dilations = sparse_output_1317_dilations_1, groups = sparse_output_1317_groups_1, pad = sparse_output_1317_pad_1, pad_type = sparse_output_1317_pad_type_1, strides = sparse_output_1317_strides_1, weight = layers_16_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1317_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_20953_cast_fp16 = add(x = dense_output_1317_cast_fp16, y = sparse_output_1317_cast_fp16)[name = string("op_20953_cast_fp16")];
            tensor<int32, [4]> var_20954 = const()[name = string("op_20954"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20956 = const()[name = string("op_20956"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_20955_cast_fp16 = transpose(perm = var_20954, x = var_20953_cast_fp16)[name = string("transpose_314")];
            tensor<fp16, [1, 51, 128]> v_head_521_cast_fp16 = reshape(shape = var_20956, x = var_20955_cast_fp16)[name = string("v_head_521_cast_fp16")];
            string dense_output_1319_pad_type_1 = const()[name = string("dense_output_1319_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1319_strides_1 = const()[name = string("dense_output_1319_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1319_pad_1 = const()[name = string("dense_output_1319_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1319_dilations_1 = const()[name = string("dense_output_1319_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1319_groups_1 = const()[name = string("dense_output_1319_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484665920))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484797056))))[name = string("layers_16_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1319_cast_fp16 = conv(dilations = dense_output_1319_dilations_1, groups = dense_output_1319_groups_1, pad = dense_output_1319_pad_1, pad_type = dense_output_1319_pad_type_1, strides = dense_output_1319_strides_1, weight = layers_16_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1319_cast_fp16")];
            string sparse_output_1319_pad_type_1 = const()[name = string("sparse_output_1319_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1319_strides_1 = const()[name = string("sparse_output_1319_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1319_pad_1 = const()[name = string("sparse_output_1319_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1319_dilations_1 = const()[name = string("sparse_output_1319_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1319_groups_1 = const()[name = string("sparse_output_1319_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484800320))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484797632))))[name = string("layers_16_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1319_cast_fp16 = conv(dilations = sparse_output_1319_dilations_1, groups = sparse_output_1319_groups_1, pad = sparse_output_1319_pad_1, pad_type = sparse_output_1319_pad_type_1, strides = sparse_output_1319_strides_1, weight = layers_16_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1319_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_20972_cast_fp16 = add(x = dense_output_1319_cast_fp16, y = sparse_output_1319_cast_fp16)[name = string("op_20972_cast_fp16")];
            tensor<int32, [4]> var_20973 = const()[name = string("op_20973"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_20975 = const()[name = string("op_20975"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_20974_cast_fp16 = transpose(perm = var_20973, x = var_20972_cast_fp16)[name = string("transpose_313")];
            tensor<fp16, [1, 101, 128]> p_head_521_cast_fp16 = reshape(shape = var_20975, x = var_20974_cast_fp16)[name = string("p_head_521_cast_fp16")];
            tensor<fp16, [128]> var_20977_to_fp16 = const()[name = string("op_20977_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484816768)))];
            tensor<fp16, [1, 51, 128]> var_20978_cast_fp16 = add(x = q_head_261_cast_fp16, y = var_20977_to_fp16)[name = string("op_20978_cast_fp16")];
            tensor<int32, [1]> q_u_261_axes_1 = const()[name = string("q_u_261_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_261_cast_fp16 = expand_dims(axes = q_u_261_axes_1, x = var_20978_cast_fp16)[name = string("q_u_261_cast_fp16")];
            tensor<fp16, [128]> var_20980_to_fp16 = const()[name = string("op_20980_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484817088)))];
            tensor<fp16, [1, 51, 128]> var_20981_cast_fp16 = add(x = q_head_261_cast_fp16, y = var_20980_to_fp16)[name = string("op_20981_cast_fp16")];
            tensor<int32, [1]> q_v_261_axes_1 = const()[name = string("q_v_261_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_261_cast_fp16 = expand_dims(axes = q_v_261_axes_1, x = var_20981_cast_fp16)[name = string("q_v_261_cast_fp16")];
            tensor<int32, [1]> k_head_523_axes_1 = const()[name = string("k_head_523_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_523_cast_fp16 = expand_dims(axes = k_head_523_axes_1, x = k_head_521_cast_fp16)[name = string("k_head_523_cast_fp16")];
            tensor<int32, [1]> v_head_523_axes_1 = const()[name = string("v_head_523_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_523_cast_fp16 = expand_dims(axes = v_head_523_axes_1, x = v_head_521_cast_fp16)[name = string("v_head_523_cast_fp16")];
            tensor<int32, [1]> p_head_523_axes_1 = const()[name = string("p_head_523_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_523_cast_fp16 = expand_dims(axes = p_head_523_axes_1, x = p_head_521_cast_fp16)[name = string("p_head_523_cast_fp16")];
            bool var_20987_transpose_x_3 = const()[name = string("op_20987_transpose_x_3"), val = bool(false)];
            bool var_20987_transpose_y_3 = const()[name = string("op_20987_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_20987_cast_fp16 = matmul(transpose_x = var_20987_transpose_x_3, transpose_y = var_20987_transpose_y_3, x = q_u_261_cast_fp16, y = k_head_523_cast_fp16)[name = string("op_20987_cast_fp16")];
            fp16 var_20988_to_fp16 = const()[name = string("op_20988_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_261_cast_fp16 = mul(x = var_20987_cast_fp16, y = var_20988_to_fp16)[name = string("scores_content_261_cast_fp16")];
            bool x_1381_transpose_x_3 = const()[name = string("x_1381_transpose_x_3"), val = bool(false)];
            bool x_1381_transpose_y_3 = const()[name = string("x_1381_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1381_cast_fp16 = matmul(transpose_x = x_1381_transpose_x_3, transpose_y = x_1381_transpose_y_3, x = q_v_261_cast_fp16, y = p_head_523_cast_fp16)[name = string("x_1381_cast_fp16")];
            tensor<int32, [8]> x_1383_pad_1 = const()[name = string("x_1383_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1383_mode_1 = const()[name = string("x_1383_mode_1"), val = string("constant")];
            fp16 const_2185_to_fp16 = const()[name = string("const_2185_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1383_cast_fp16 = pad(constant_val = const_2185_to_fp16, mode = x_1383_mode_1, pad = x_1383_pad_1, x = x_1381_cast_fp16)[name = string("x_1383_cast_fp16")];
            tensor<int32, [4]> var_21002 = const()[name = string("op_21002"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1385_cast_fp16 = reshape(shape = var_21002, x = x_1383_cast_fp16)[name = string("x_1385_cast_fp16")];
            tensor<int32, [4]> var_21006_begin_1 = const()[name = string("op_21006_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_21006_end_1 = const()[name = string("op_21006_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_21006_end_mask_1 = const()[name = string("op_21006_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_21006_cast_fp16 = slice_by_index(begin = var_21006_begin_1, end = var_21006_end_1, end_mask = var_21006_end_mask_1, x = x_1385_cast_fp16)[name = string("op_21006_cast_fp16")];
            tensor<int32, [4]> var_21008 = const()[name = string("op_21008"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_21009_cast_fp16 = reshape(shape = var_21008, x = var_21006_cast_fp16)[name = string("op_21009_cast_fp16")];
            tensor<int32, [4]> var_21014_begin_1 = const()[name = string("op_21014_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_21014_end_1 = const()[name = string("op_21014_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_21014_end_mask_1 = const()[name = string("op_21014_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_21014_cast_fp16 = slice_by_index(begin = var_21014_begin_1, end = var_21014_end_1, end_mask = var_21014_end_mask_1, x = var_21009_cast_fp16)[name = string("op_21014_cast_fp16")];
            fp16 var_21015_to_fp16 = const()[name = string("op_21015_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_261_cast_fp16 = mul(x = var_21014_cast_fp16, y = var_21015_to_fp16)[name = string("scores_pos_261_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_261_cast_fp16 = add(x = scores_content_261_cast_fp16, y = scores_pos_261_cast_fp16)[name = string("logits_261_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_21018_cast_fp16 = softmax(axis = var_20501, x = logits_261_cast_fp16)[name = string("op_21018_cast_fp16")];
            bool var_21020_transpose_x_1 = const()[name = string("op_21020_transpose_x_1"), val = bool(false)];
            bool var_21020_transpose_y_1 = const()[name = string("op_21020_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_21020_cast_fp16 = matmul(transpose_x = var_21020_transpose_x_1, transpose_y = var_21020_transpose_y_1, x = var_21018_cast_fp16, y = v_head_523_cast_fp16)[name = string("op_21020_cast_fp16")];
            tensor<int32, [1]> var_21021_axes_1 = const()[name = string("op_21021_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_21021_cast_fp16 = squeeze(axes = var_21021_axes_1, x = var_21020_cast_fp16)[name = string("op_21021_cast_fp16")];
            string dense_output_1321_pad_type_1 = const()[name = string("dense_output_1321_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1321_strides_1 = const()[name = string("dense_output_1321_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1321_pad_1 = const()[name = string("dense_output_1321_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1321_dilations_1 = const()[name = string("dense_output_1321_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1321_groups_1 = const()[name = string("dense_output_1321_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484817408))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484948544))))[name = string("layers_16_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1321_cast_fp16 = conv(dilations = dense_output_1321_dilations_1, groups = dense_output_1321_groups_1, pad = dense_output_1321_pad_1, pad_type = dense_output_1321_pad_type_1, strides = dense_output_1321_strides_1, weight = layers_16_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1321_cast_fp16")];
            string sparse_output_1321_pad_type_1 = const()[name = string("sparse_output_1321_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1321_strides_1 = const()[name = string("sparse_output_1321_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1321_pad_1 = const()[name = string("sparse_output_1321_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1321_dilations_1 = const()[name = string("sparse_output_1321_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1321_groups_1 = const()[name = string("sparse_output_1321_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484951808))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484949120))))[name = string("layers_16_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1321_cast_fp16 = conv(dilations = sparse_output_1321_dilations_1, groups = sparse_output_1321_groups_1, pad = sparse_output_1321_pad_1, pad_type = sparse_output_1321_pad_type_1, strides = sparse_output_1321_strides_1, weight = layers_16_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1321_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21036_cast_fp16 = add(x = dense_output_1321_cast_fp16, y = sparse_output_1321_cast_fp16)[name = string("op_21036_cast_fp16")];
            tensor<int32, [4]> var_21037 = const()[name = string("op_21037"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21039 = const()[name = string("op_21039"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21038_cast_fp16 = transpose(perm = var_21037, x = var_21036_cast_fp16)[name = string("transpose_312")];
            tensor<fp16, [1, 51, 128]> q_head_263_cast_fp16 = reshape(shape = var_21039, x = var_21038_cast_fp16)[name = string("q_head_263_cast_fp16")];
            string dense_output_1323_pad_type_1 = const()[name = string("dense_output_1323_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1323_strides_1 = const()[name = string("dense_output_1323_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1323_pad_1 = const()[name = string("dense_output_1323_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1323_dilations_1 = const()[name = string("dense_output_1323_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1323_groups_1 = const()[name = string("dense_output_1323_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484968256))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485099392))))[name = string("layers_16_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1323_cast_fp16 = conv(dilations = dense_output_1323_dilations_1, groups = dense_output_1323_groups_1, pad = dense_output_1323_pad_1, pad_type = dense_output_1323_pad_type_1, strides = dense_output_1323_strides_1, weight = layers_16_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1323_cast_fp16")];
            string sparse_output_1323_pad_type_1 = const()[name = string("sparse_output_1323_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1323_strides_1 = const()[name = string("sparse_output_1323_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1323_pad_1 = const()[name = string("sparse_output_1323_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1323_dilations_1 = const()[name = string("sparse_output_1323_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1323_groups_1 = const()[name = string("sparse_output_1323_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485102656))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485099968))))[name = string("layers_16_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1323_cast_fp16 = conv(dilations = sparse_output_1323_dilations_1, groups = sparse_output_1323_groups_1, pad = sparse_output_1323_pad_1, pad_type = sparse_output_1323_pad_type_1, strides = sparse_output_1323_strides_1, weight = layers_16_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1323_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21055_cast_fp16 = add(x = dense_output_1323_cast_fp16, y = sparse_output_1323_cast_fp16)[name = string("op_21055_cast_fp16")];
            tensor<int32, [4]> var_21056 = const()[name = string("op_21056"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21058 = const()[name = string("op_21058"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21057_cast_fp16 = transpose(perm = var_21056, x = var_21055_cast_fp16)[name = string("transpose_311")];
            tensor<fp16, [1, 51, 128]> k_head_525_cast_fp16 = reshape(shape = var_21058, x = var_21057_cast_fp16)[name = string("k_head_525_cast_fp16")];
            string dense_output_1325_pad_type_1 = const()[name = string("dense_output_1325_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1325_strides_1 = const()[name = string("dense_output_1325_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1325_pad_1 = const()[name = string("dense_output_1325_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1325_dilations_1 = const()[name = string("dense_output_1325_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1325_groups_1 = const()[name = string("dense_output_1325_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485119104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485250240))))[name = string("layers_16_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1325_cast_fp16 = conv(dilations = dense_output_1325_dilations_1, groups = dense_output_1325_groups_1, pad = dense_output_1325_pad_1, pad_type = dense_output_1325_pad_type_1, strides = dense_output_1325_strides_1, weight = layers_16_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1325_cast_fp16")];
            string sparse_output_1325_pad_type_1 = const()[name = string("sparse_output_1325_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1325_strides_1 = const()[name = string("sparse_output_1325_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1325_pad_1 = const()[name = string("sparse_output_1325_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1325_dilations_1 = const()[name = string("sparse_output_1325_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1325_groups_1 = const()[name = string("sparse_output_1325_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485253504))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485250816))))[name = string("layers_16_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1325_cast_fp16 = conv(dilations = sparse_output_1325_dilations_1, groups = sparse_output_1325_groups_1, pad = sparse_output_1325_pad_1, pad_type = sparse_output_1325_pad_type_1, strides = sparse_output_1325_strides_1, weight = layers_16_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1325_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21074_cast_fp16 = add(x = dense_output_1325_cast_fp16, y = sparse_output_1325_cast_fp16)[name = string("op_21074_cast_fp16")];
            tensor<int32, [4]> var_21075 = const()[name = string("op_21075"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21077 = const()[name = string("op_21077"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21076_cast_fp16 = transpose(perm = var_21075, x = var_21074_cast_fp16)[name = string("transpose_310")];
            tensor<fp16, [1, 51, 128]> v_head_525_cast_fp16 = reshape(shape = var_21077, x = var_21076_cast_fp16)[name = string("v_head_525_cast_fp16")];
            string dense_output_1327_pad_type_1 = const()[name = string("dense_output_1327_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1327_strides_1 = const()[name = string("dense_output_1327_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1327_pad_1 = const()[name = string("dense_output_1327_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1327_dilations_1 = const()[name = string("dense_output_1327_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1327_groups_1 = const()[name = string("dense_output_1327_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485269952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485401088))))[name = string("layers_16_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1327_cast_fp16 = conv(dilations = dense_output_1327_dilations_1, groups = dense_output_1327_groups_1, pad = dense_output_1327_pad_1, pad_type = dense_output_1327_pad_type_1, strides = dense_output_1327_strides_1, weight = layers_16_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1327_cast_fp16")];
            string sparse_output_1327_pad_type_1 = const()[name = string("sparse_output_1327_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1327_strides_1 = const()[name = string("sparse_output_1327_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1327_pad_1 = const()[name = string("sparse_output_1327_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1327_dilations_1 = const()[name = string("sparse_output_1327_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1327_groups_1 = const()[name = string("sparse_output_1327_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485404352))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485401664))))[name = string("layers_16_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1327_cast_fp16 = conv(dilations = sparse_output_1327_dilations_1, groups = sparse_output_1327_groups_1, pad = sparse_output_1327_pad_1, pad_type = sparse_output_1327_pad_type_1, strides = sparse_output_1327_strides_1, weight = layers_16_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1327_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_21093_cast_fp16 = add(x = dense_output_1327_cast_fp16, y = sparse_output_1327_cast_fp16)[name = string("op_21093_cast_fp16")];
            tensor<int32, [4]> var_21094 = const()[name = string("op_21094"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21096 = const()[name = string("op_21096"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_21095_cast_fp16 = transpose(perm = var_21094, x = var_21093_cast_fp16)[name = string("transpose_309")];
            tensor<fp16, [1, 101, 128]> p_head_525_cast_fp16 = reshape(shape = var_21096, x = var_21095_cast_fp16)[name = string("p_head_525_cast_fp16")];
            tensor<fp16, [128]> var_21098_to_fp16 = const()[name = string("op_21098_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485420800)))];
            tensor<fp16, [1, 51, 128]> var_21099_cast_fp16 = add(x = q_head_263_cast_fp16, y = var_21098_to_fp16)[name = string("op_21099_cast_fp16")];
            tensor<int32, [1]> q_u_263_axes_1 = const()[name = string("q_u_263_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_263_cast_fp16 = expand_dims(axes = q_u_263_axes_1, x = var_21099_cast_fp16)[name = string("q_u_263_cast_fp16")];
            tensor<fp16, [128]> var_21101_to_fp16 = const()[name = string("op_21101_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485421120)))];
            tensor<fp16, [1, 51, 128]> var_21102_cast_fp16 = add(x = q_head_263_cast_fp16, y = var_21101_to_fp16)[name = string("op_21102_cast_fp16")];
            tensor<int32, [1]> q_v_263_axes_1 = const()[name = string("q_v_263_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_263_cast_fp16 = expand_dims(axes = q_v_263_axes_1, x = var_21102_cast_fp16)[name = string("q_v_263_cast_fp16")];
            tensor<int32, [1]> k_head_527_axes_1 = const()[name = string("k_head_527_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_527_cast_fp16 = expand_dims(axes = k_head_527_axes_1, x = k_head_525_cast_fp16)[name = string("k_head_527_cast_fp16")];
            tensor<int32, [1]> v_head_527_axes_1 = const()[name = string("v_head_527_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_527_cast_fp16 = expand_dims(axes = v_head_527_axes_1, x = v_head_525_cast_fp16)[name = string("v_head_527_cast_fp16")];
            tensor<int32, [1]> p_head_527_axes_1 = const()[name = string("p_head_527_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_527_cast_fp16 = expand_dims(axes = p_head_527_axes_1, x = p_head_525_cast_fp16)[name = string("p_head_527_cast_fp16")];
            bool var_21108_transpose_x_3 = const()[name = string("op_21108_transpose_x_3"), val = bool(false)];
            bool var_21108_transpose_y_3 = const()[name = string("op_21108_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_21108_cast_fp16 = matmul(transpose_x = var_21108_transpose_x_3, transpose_y = var_21108_transpose_y_3, x = q_u_263_cast_fp16, y = k_head_527_cast_fp16)[name = string("op_21108_cast_fp16")];
            fp16 var_21109_to_fp16 = const()[name = string("op_21109_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_263_cast_fp16 = mul(x = var_21108_cast_fp16, y = var_21109_to_fp16)[name = string("scores_content_263_cast_fp16")];
            bool x_1389_transpose_x_3 = const()[name = string("x_1389_transpose_x_3"), val = bool(false)];
            bool x_1389_transpose_y_3 = const()[name = string("x_1389_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1389_cast_fp16 = matmul(transpose_x = x_1389_transpose_x_3, transpose_y = x_1389_transpose_y_3, x = q_v_263_cast_fp16, y = p_head_527_cast_fp16)[name = string("x_1389_cast_fp16")];
            tensor<int32, [8]> x_1391_pad_1 = const()[name = string("x_1391_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1391_mode_1 = const()[name = string("x_1391_mode_1"), val = string("constant")];
            fp16 const_2191_to_fp16 = const()[name = string("const_2191_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1391_cast_fp16 = pad(constant_val = const_2191_to_fp16, mode = x_1391_mode_1, pad = x_1391_pad_1, x = x_1389_cast_fp16)[name = string("x_1391_cast_fp16")];
            tensor<int32, [4]> var_21123 = const()[name = string("op_21123"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1393_cast_fp16 = reshape(shape = var_21123, x = x_1391_cast_fp16)[name = string("x_1393_cast_fp16")];
            tensor<int32, [4]> var_21127_begin_1 = const()[name = string("op_21127_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_21127_end_1 = const()[name = string("op_21127_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_21127_end_mask_1 = const()[name = string("op_21127_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_21127_cast_fp16 = slice_by_index(begin = var_21127_begin_1, end = var_21127_end_1, end_mask = var_21127_end_mask_1, x = x_1393_cast_fp16)[name = string("op_21127_cast_fp16")];
            tensor<int32, [4]> var_21129 = const()[name = string("op_21129"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_21130_cast_fp16 = reshape(shape = var_21129, x = var_21127_cast_fp16)[name = string("op_21130_cast_fp16")];
            tensor<int32, [4]> var_21135_begin_1 = const()[name = string("op_21135_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_21135_end_1 = const()[name = string("op_21135_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_21135_end_mask_1 = const()[name = string("op_21135_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_21135_cast_fp16 = slice_by_index(begin = var_21135_begin_1, end = var_21135_end_1, end_mask = var_21135_end_mask_1, x = var_21130_cast_fp16)[name = string("op_21135_cast_fp16")];
            fp16 var_21136_to_fp16 = const()[name = string("op_21136_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_263_cast_fp16 = mul(x = var_21135_cast_fp16, y = var_21136_to_fp16)[name = string("scores_pos_263_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_263_cast_fp16 = add(x = scores_content_263_cast_fp16, y = scores_pos_263_cast_fp16)[name = string("logits_263_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_21139_cast_fp16 = softmax(axis = var_20501, x = logits_263_cast_fp16)[name = string("op_21139_cast_fp16")];
            bool var_21141_transpose_x_1 = const()[name = string("op_21141_transpose_x_1"), val = bool(false)];
            bool var_21141_transpose_y_1 = const()[name = string("op_21141_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_21141_cast_fp16 = matmul(transpose_x = var_21141_transpose_x_1, transpose_y = var_21141_transpose_y_1, x = var_21139_cast_fp16, y = v_head_527_cast_fp16)[name = string("op_21141_cast_fp16")];
            tensor<int32, [1]> var_21142_axes_1 = const()[name = string("op_21142_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_21142_cast_fp16 = squeeze(axes = var_21142_axes_1, x = var_21141_cast_fp16)[name = string("op_21142_cast_fp16")];
            string dense_output_1329_pad_type_1 = const()[name = string("dense_output_1329_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1329_strides_1 = const()[name = string("dense_output_1329_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1329_pad_1 = const()[name = string("dense_output_1329_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1329_dilations_1 = const()[name = string("dense_output_1329_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1329_groups_1 = const()[name = string("dense_output_1329_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485421440))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485552576))))[name = string("layers_16_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1329_cast_fp16 = conv(dilations = dense_output_1329_dilations_1, groups = dense_output_1329_groups_1, pad = dense_output_1329_pad_1, pad_type = dense_output_1329_pad_type_1, strides = dense_output_1329_strides_1, weight = layers_16_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1329_cast_fp16")];
            string sparse_output_1329_pad_type_1 = const()[name = string("sparse_output_1329_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1329_strides_1 = const()[name = string("sparse_output_1329_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1329_pad_1 = const()[name = string("sparse_output_1329_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1329_dilations_1 = const()[name = string("sparse_output_1329_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1329_groups_1 = const()[name = string("sparse_output_1329_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485555840))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485553152))))[name = string("layers_16_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1329_cast_fp16 = conv(dilations = sparse_output_1329_dilations_1, groups = sparse_output_1329_groups_1, pad = sparse_output_1329_pad_1, pad_type = sparse_output_1329_pad_type_1, strides = sparse_output_1329_strides_1, weight = layers_16_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1329_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21157_cast_fp16 = add(x = dense_output_1329_cast_fp16, y = sparse_output_1329_cast_fp16)[name = string("op_21157_cast_fp16")];
            tensor<int32, [4]> var_21158 = const()[name = string("op_21158"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21160 = const()[name = string("op_21160"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21159_cast_fp16 = transpose(perm = var_21158, x = var_21157_cast_fp16)[name = string("transpose_308")];
            tensor<fp16, [1, 51, 128]> q_head_265_cast_fp16 = reshape(shape = var_21160, x = var_21159_cast_fp16)[name = string("q_head_265_cast_fp16")];
            string dense_output_1331_pad_type_1 = const()[name = string("dense_output_1331_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1331_strides_1 = const()[name = string("dense_output_1331_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1331_pad_1 = const()[name = string("dense_output_1331_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1331_dilations_1 = const()[name = string("dense_output_1331_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1331_groups_1 = const()[name = string("dense_output_1331_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485572288))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485703424))))[name = string("layers_16_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1331_cast_fp16 = conv(dilations = dense_output_1331_dilations_1, groups = dense_output_1331_groups_1, pad = dense_output_1331_pad_1, pad_type = dense_output_1331_pad_type_1, strides = dense_output_1331_strides_1, weight = layers_16_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1331_cast_fp16")];
            string sparse_output_1331_pad_type_1 = const()[name = string("sparse_output_1331_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1331_strides_1 = const()[name = string("sparse_output_1331_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1331_pad_1 = const()[name = string("sparse_output_1331_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1331_dilations_1 = const()[name = string("sparse_output_1331_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1331_groups_1 = const()[name = string("sparse_output_1331_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485706688))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485704000))))[name = string("layers_16_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1331_cast_fp16 = conv(dilations = sparse_output_1331_dilations_1, groups = sparse_output_1331_groups_1, pad = sparse_output_1331_pad_1, pad_type = sparse_output_1331_pad_type_1, strides = sparse_output_1331_strides_1, weight = layers_16_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1331_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21176_cast_fp16 = add(x = dense_output_1331_cast_fp16, y = sparse_output_1331_cast_fp16)[name = string("op_21176_cast_fp16")];
            tensor<int32, [4]> var_21177 = const()[name = string("op_21177"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21179 = const()[name = string("op_21179"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21178_cast_fp16 = transpose(perm = var_21177, x = var_21176_cast_fp16)[name = string("transpose_307")];
            tensor<fp16, [1, 51, 128]> k_head_529_cast_fp16 = reshape(shape = var_21179, x = var_21178_cast_fp16)[name = string("k_head_529_cast_fp16")];
            string dense_output_1333_pad_type_1 = const()[name = string("dense_output_1333_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1333_strides_1 = const()[name = string("dense_output_1333_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1333_pad_1 = const()[name = string("dense_output_1333_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1333_dilations_1 = const()[name = string("dense_output_1333_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1333_groups_1 = const()[name = string("dense_output_1333_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485723136))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485854272))))[name = string("layers_16_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1333_cast_fp16 = conv(dilations = dense_output_1333_dilations_1, groups = dense_output_1333_groups_1, pad = dense_output_1333_pad_1, pad_type = dense_output_1333_pad_type_1, strides = dense_output_1333_strides_1, weight = layers_16_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1333_cast_fp16")];
            string sparse_output_1333_pad_type_1 = const()[name = string("sparse_output_1333_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1333_strides_1 = const()[name = string("sparse_output_1333_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1333_pad_1 = const()[name = string("sparse_output_1333_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1333_dilations_1 = const()[name = string("sparse_output_1333_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1333_groups_1 = const()[name = string("sparse_output_1333_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485857536))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485854848))))[name = string("layers_16_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1333_cast_fp16 = conv(dilations = sparse_output_1333_dilations_1, groups = sparse_output_1333_groups_1, pad = sparse_output_1333_pad_1, pad_type = sparse_output_1333_pad_type_1, strides = sparse_output_1333_strides_1, weight = layers_16_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1333_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21195_cast_fp16 = add(x = dense_output_1333_cast_fp16, y = sparse_output_1333_cast_fp16)[name = string("op_21195_cast_fp16")];
            tensor<int32, [4]> var_21196 = const()[name = string("op_21196"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21198 = const()[name = string("op_21198"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21197_cast_fp16 = transpose(perm = var_21196, x = var_21195_cast_fp16)[name = string("transpose_306")];
            tensor<fp16, [1, 51, 128]> v_head_529_cast_fp16 = reshape(shape = var_21198, x = var_21197_cast_fp16)[name = string("v_head_529_cast_fp16")];
            string dense_output_1335_pad_type_1 = const()[name = string("dense_output_1335_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1335_strides_1 = const()[name = string("dense_output_1335_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1335_pad_1 = const()[name = string("dense_output_1335_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1335_dilations_1 = const()[name = string("dense_output_1335_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1335_groups_1 = const()[name = string("dense_output_1335_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485873984))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486005120))))[name = string("layers_16_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1335_cast_fp16 = conv(dilations = dense_output_1335_dilations_1, groups = dense_output_1335_groups_1, pad = dense_output_1335_pad_1, pad_type = dense_output_1335_pad_type_1, strides = dense_output_1335_strides_1, weight = layers_16_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1335_cast_fp16")];
            string sparse_output_1335_pad_type_1 = const()[name = string("sparse_output_1335_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1335_strides_1 = const()[name = string("sparse_output_1335_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1335_pad_1 = const()[name = string("sparse_output_1335_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1335_dilations_1 = const()[name = string("sparse_output_1335_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1335_groups_1 = const()[name = string("sparse_output_1335_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486008384))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486005696))))[name = string("layers_16_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1335_cast_fp16 = conv(dilations = sparse_output_1335_dilations_1, groups = sparse_output_1335_groups_1, pad = sparse_output_1335_pad_1, pad_type = sparse_output_1335_pad_type_1, strides = sparse_output_1335_strides_1, weight = layers_16_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1335_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_21214_cast_fp16 = add(x = dense_output_1335_cast_fp16, y = sparse_output_1335_cast_fp16)[name = string("op_21214_cast_fp16")];
            tensor<int32, [4]> var_21215 = const()[name = string("op_21215"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21217 = const()[name = string("op_21217"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_21216_cast_fp16 = transpose(perm = var_21215, x = var_21214_cast_fp16)[name = string("transpose_305")];
            tensor<fp16, [1, 101, 128]> p_head_529_cast_fp16 = reshape(shape = var_21217, x = var_21216_cast_fp16)[name = string("p_head_529_cast_fp16")];
            tensor<fp16, [128]> var_21219_to_fp16 = const()[name = string("op_21219_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486024832)))];
            tensor<fp16, [1, 51, 128]> var_21220_cast_fp16 = add(x = q_head_265_cast_fp16, y = var_21219_to_fp16)[name = string("op_21220_cast_fp16")];
            tensor<int32, [1]> q_u_265_axes_1 = const()[name = string("q_u_265_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_265_cast_fp16 = expand_dims(axes = q_u_265_axes_1, x = var_21220_cast_fp16)[name = string("q_u_265_cast_fp16")];
            tensor<fp16, [128]> var_21222_to_fp16 = const()[name = string("op_21222_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486025152)))];
            tensor<fp16, [1, 51, 128]> var_21223_cast_fp16 = add(x = q_head_265_cast_fp16, y = var_21222_to_fp16)[name = string("op_21223_cast_fp16")];
            tensor<int32, [1]> q_v_265_axes_1 = const()[name = string("q_v_265_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_265_cast_fp16 = expand_dims(axes = q_v_265_axes_1, x = var_21223_cast_fp16)[name = string("q_v_265_cast_fp16")];
            tensor<int32, [1]> k_head_531_axes_1 = const()[name = string("k_head_531_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_531_cast_fp16 = expand_dims(axes = k_head_531_axes_1, x = k_head_529_cast_fp16)[name = string("k_head_531_cast_fp16")];
            tensor<int32, [1]> v_head_531_axes_1 = const()[name = string("v_head_531_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_531_cast_fp16 = expand_dims(axes = v_head_531_axes_1, x = v_head_529_cast_fp16)[name = string("v_head_531_cast_fp16")];
            tensor<int32, [1]> p_head_531_axes_1 = const()[name = string("p_head_531_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_531_cast_fp16 = expand_dims(axes = p_head_531_axes_1, x = p_head_529_cast_fp16)[name = string("p_head_531_cast_fp16")];
            bool var_21229_transpose_x_3 = const()[name = string("op_21229_transpose_x_3"), val = bool(false)];
            bool var_21229_transpose_y_3 = const()[name = string("op_21229_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_21229_cast_fp16 = matmul(transpose_x = var_21229_transpose_x_3, transpose_y = var_21229_transpose_y_3, x = q_u_265_cast_fp16, y = k_head_531_cast_fp16)[name = string("op_21229_cast_fp16")];
            fp16 var_21230_to_fp16 = const()[name = string("op_21230_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_265_cast_fp16 = mul(x = var_21229_cast_fp16, y = var_21230_to_fp16)[name = string("scores_content_265_cast_fp16")];
            bool x_1397_transpose_x_3 = const()[name = string("x_1397_transpose_x_3"), val = bool(false)];
            bool x_1397_transpose_y_3 = const()[name = string("x_1397_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1397_cast_fp16 = matmul(transpose_x = x_1397_transpose_x_3, transpose_y = x_1397_transpose_y_3, x = q_v_265_cast_fp16, y = p_head_531_cast_fp16)[name = string("x_1397_cast_fp16")];
            tensor<int32, [8]> x_1399_pad_1 = const()[name = string("x_1399_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1399_mode_1 = const()[name = string("x_1399_mode_1"), val = string("constant")];
            fp16 const_2197_to_fp16 = const()[name = string("const_2197_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1399_cast_fp16 = pad(constant_val = const_2197_to_fp16, mode = x_1399_mode_1, pad = x_1399_pad_1, x = x_1397_cast_fp16)[name = string("x_1399_cast_fp16")];
            tensor<int32, [4]> var_21244 = const()[name = string("op_21244"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1401_cast_fp16 = reshape(shape = var_21244, x = x_1399_cast_fp16)[name = string("x_1401_cast_fp16")];
            tensor<int32, [4]> var_21248_begin_1 = const()[name = string("op_21248_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_21248_end_1 = const()[name = string("op_21248_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_21248_end_mask_1 = const()[name = string("op_21248_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_21248_cast_fp16 = slice_by_index(begin = var_21248_begin_1, end = var_21248_end_1, end_mask = var_21248_end_mask_1, x = x_1401_cast_fp16)[name = string("op_21248_cast_fp16")];
            tensor<int32, [4]> var_21250 = const()[name = string("op_21250"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_21251_cast_fp16 = reshape(shape = var_21250, x = var_21248_cast_fp16)[name = string("op_21251_cast_fp16")];
            tensor<int32, [4]> var_21256_begin_1 = const()[name = string("op_21256_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_21256_end_1 = const()[name = string("op_21256_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_21256_end_mask_1 = const()[name = string("op_21256_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_21256_cast_fp16 = slice_by_index(begin = var_21256_begin_1, end = var_21256_end_1, end_mask = var_21256_end_mask_1, x = var_21251_cast_fp16)[name = string("op_21256_cast_fp16")];
            fp16 var_21257_to_fp16 = const()[name = string("op_21257_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_265_cast_fp16 = mul(x = var_21256_cast_fp16, y = var_21257_to_fp16)[name = string("scores_pos_265_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_265_cast_fp16 = add(x = scores_content_265_cast_fp16, y = scores_pos_265_cast_fp16)[name = string("logits_265_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_21260_cast_fp16 = softmax(axis = var_20501, x = logits_265_cast_fp16)[name = string("op_21260_cast_fp16")];
            bool var_21262_transpose_x_1 = const()[name = string("op_21262_transpose_x_1"), val = bool(false)];
            bool var_21262_transpose_y_1 = const()[name = string("op_21262_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_21262_cast_fp16 = matmul(transpose_x = var_21262_transpose_x_1, transpose_y = var_21262_transpose_y_1, x = var_21260_cast_fp16, y = v_head_531_cast_fp16)[name = string("op_21262_cast_fp16")];
            tensor<int32, [1]> var_21263_axes_1 = const()[name = string("op_21263_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_21263_cast_fp16 = squeeze(axes = var_21263_axes_1, x = var_21262_cast_fp16)[name = string("op_21263_cast_fp16")];
            string dense_output_1337_pad_type_1 = const()[name = string("dense_output_1337_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1337_strides_1 = const()[name = string("dense_output_1337_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1337_pad_1 = const()[name = string("dense_output_1337_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1337_dilations_1 = const()[name = string("dense_output_1337_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1337_groups_1 = const()[name = string("dense_output_1337_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486025472))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486156608))))[name = string("layers_16_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1337_cast_fp16 = conv(dilations = dense_output_1337_dilations_1, groups = dense_output_1337_groups_1, pad = dense_output_1337_pad_1, pad_type = dense_output_1337_pad_type_1, strides = dense_output_1337_strides_1, weight = layers_16_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1337_cast_fp16")];
            string sparse_output_1337_pad_type_1 = const()[name = string("sparse_output_1337_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1337_strides_1 = const()[name = string("sparse_output_1337_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1337_pad_1 = const()[name = string("sparse_output_1337_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1337_dilations_1 = const()[name = string("sparse_output_1337_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1337_groups_1 = const()[name = string("sparse_output_1337_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486159872))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486157184))))[name = string("layers_16_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1337_cast_fp16 = conv(dilations = sparse_output_1337_dilations_1, groups = sparse_output_1337_groups_1, pad = sparse_output_1337_pad_1, pad_type = sparse_output_1337_pad_type_1, strides = sparse_output_1337_strides_1, weight = layers_16_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1337_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21278_cast_fp16 = add(x = dense_output_1337_cast_fp16, y = sparse_output_1337_cast_fp16)[name = string("op_21278_cast_fp16")];
            tensor<int32, [4]> var_21279 = const()[name = string("op_21279"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21281 = const()[name = string("op_21281"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21280_cast_fp16 = transpose(perm = var_21279, x = var_21278_cast_fp16)[name = string("transpose_304")];
            tensor<fp16, [1, 51, 128]> q_head_267_cast_fp16 = reshape(shape = var_21281, x = var_21280_cast_fp16)[name = string("q_head_267_cast_fp16")];
            string dense_output_1339_pad_type_1 = const()[name = string("dense_output_1339_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1339_strides_1 = const()[name = string("dense_output_1339_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1339_pad_1 = const()[name = string("dense_output_1339_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1339_dilations_1 = const()[name = string("dense_output_1339_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1339_groups_1 = const()[name = string("dense_output_1339_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486176320))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486307456))))[name = string("layers_16_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1339_cast_fp16 = conv(dilations = dense_output_1339_dilations_1, groups = dense_output_1339_groups_1, pad = dense_output_1339_pad_1, pad_type = dense_output_1339_pad_type_1, strides = dense_output_1339_strides_1, weight = layers_16_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1339_cast_fp16")];
            string sparse_output_1339_pad_type_1 = const()[name = string("sparse_output_1339_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1339_strides_1 = const()[name = string("sparse_output_1339_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1339_pad_1 = const()[name = string("sparse_output_1339_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1339_dilations_1 = const()[name = string("sparse_output_1339_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1339_groups_1 = const()[name = string("sparse_output_1339_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486310720))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486308032))))[name = string("layers_16_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1339_cast_fp16 = conv(dilations = sparse_output_1339_dilations_1, groups = sparse_output_1339_groups_1, pad = sparse_output_1339_pad_1, pad_type = sparse_output_1339_pad_type_1, strides = sparse_output_1339_strides_1, weight = layers_16_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1339_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21297_cast_fp16 = add(x = dense_output_1339_cast_fp16, y = sparse_output_1339_cast_fp16)[name = string("op_21297_cast_fp16")];
            tensor<int32, [4]> var_21298 = const()[name = string("op_21298"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21300 = const()[name = string("op_21300"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21299_cast_fp16 = transpose(perm = var_21298, x = var_21297_cast_fp16)[name = string("transpose_303")];
            tensor<fp16, [1, 51, 128]> k_head_533_cast_fp16 = reshape(shape = var_21300, x = var_21299_cast_fp16)[name = string("k_head_533_cast_fp16")];
            string dense_output_1341_pad_type_1 = const()[name = string("dense_output_1341_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1341_strides_1 = const()[name = string("dense_output_1341_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1341_pad_1 = const()[name = string("dense_output_1341_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1341_dilations_1 = const()[name = string("dense_output_1341_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1341_groups_1 = const()[name = string("dense_output_1341_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486327168))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486458304))))[name = string("layers_16_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1341_cast_fp16 = conv(dilations = dense_output_1341_dilations_1, groups = dense_output_1341_groups_1, pad = dense_output_1341_pad_1, pad_type = dense_output_1341_pad_type_1, strides = dense_output_1341_strides_1, weight = layers_16_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1341_cast_fp16")];
            string sparse_output_1341_pad_type_1 = const()[name = string("sparse_output_1341_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1341_strides_1 = const()[name = string("sparse_output_1341_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1341_pad_1 = const()[name = string("sparse_output_1341_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1341_dilations_1 = const()[name = string("sparse_output_1341_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1341_groups_1 = const()[name = string("sparse_output_1341_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486461568))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486458880))))[name = string("layers_16_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1341_cast_fp16 = conv(dilations = sparse_output_1341_dilations_1, groups = sparse_output_1341_groups_1, pad = sparse_output_1341_pad_1, pad_type = sparse_output_1341_pad_type_1, strides = sparse_output_1341_strides_1, weight = layers_16_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1341_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21316_cast_fp16 = add(x = dense_output_1341_cast_fp16, y = sparse_output_1341_cast_fp16)[name = string("op_21316_cast_fp16")];
            tensor<int32, [4]> var_21317 = const()[name = string("op_21317"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21319 = const()[name = string("op_21319"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21318_cast_fp16 = transpose(perm = var_21317, x = var_21316_cast_fp16)[name = string("transpose_302")];
            tensor<fp16, [1, 51, 128]> v_head_533_cast_fp16 = reshape(shape = var_21319, x = var_21318_cast_fp16)[name = string("v_head_533_cast_fp16")];
            string dense_output_1343_pad_type_1 = const()[name = string("dense_output_1343_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1343_strides_1 = const()[name = string("dense_output_1343_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1343_pad_1 = const()[name = string("dense_output_1343_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1343_dilations_1 = const()[name = string("dense_output_1343_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1343_groups_1 = const()[name = string("dense_output_1343_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486478016))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486609152))))[name = string("layers_16_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1343_cast_fp16 = conv(dilations = dense_output_1343_dilations_1, groups = dense_output_1343_groups_1, pad = dense_output_1343_pad_1, pad_type = dense_output_1343_pad_type_1, strides = dense_output_1343_strides_1, weight = layers_16_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1343_cast_fp16")];
            string sparse_output_1343_pad_type_1 = const()[name = string("sparse_output_1343_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1343_strides_1 = const()[name = string("sparse_output_1343_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1343_pad_1 = const()[name = string("sparse_output_1343_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1343_dilations_1 = const()[name = string("sparse_output_1343_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1343_groups_1 = const()[name = string("sparse_output_1343_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486612416))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486609728))))[name = string("layers_16_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1343_cast_fp16 = conv(dilations = sparse_output_1343_dilations_1, groups = sparse_output_1343_groups_1, pad = sparse_output_1343_pad_1, pad_type = sparse_output_1343_pad_type_1, strides = sparse_output_1343_strides_1, weight = layers_16_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1343_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_21335_cast_fp16 = add(x = dense_output_1343_cast_fp16, y = sparse_output_1343_cast_fp16)[name = string("op_21335_cast_fp16")];
            tensor<int32, [4]> var_21336 = const()[name = string("op_21336"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21338 = const()[name = string("op_21338"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_21337_cast_fp16 = transpose(perm = var_21336, x = var_21335_cast_fp16)[name = string("transpose_301")];
            tensor<fp16, [1, 101, 128]> p_head_533_cast_fp16 = reshape(shape = var_21338, x = var_21337_cast_fp16)[name = string("p_head_533_cast_fp16")];
            tensor<fp16, [128]> var_21340_to_fp16 = const()[name = string("op_21340_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486628864)))];
            tensor<fp16, [1, 51, 128]> var_21341_cast_fp16 = add(x = q_head_267_cast_fp16, y = var_21340_to_fp16)[name = string("op_21341_cast_fp16")];
            tensor<int32, [1]> q_u_267_axes_1 = const()[name = string("q_u_267_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_267_cast_fp16 = expand_dims(axes = q_u_267_axes_1, x = var_21341_cast_fp16)[name = string("q_u_267_cast_fp16")];
            tensor<fp16, [128]> var_21343_to_fp16 = const()[name = string("op_21343_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486629184)))];
            tensor<fp16, [1, 51, 128]> var_21344_cast_fp16 = add(x = q_head_267_cast_fp16, y = var_21343_to_fp16)[name = string("op_21344_cast_fp16")];
            tensor<int32, [1]> q_v_267_axes_1 = const()[name = string("q_v_267_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_267_cast_fp16 = expand_dims(axes = q_v_267_axes_1, x = var_21344_cast_fp16)[name = string("q_v_267_cast_fp16")];
            tensor<int32, [1]> k_head_535_axes_1 = const()[name = string("k_head_535_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_535_cast_fp16 = expand_dims(axes = k_head_535_axes_1, x = k_head_533_cast_fp16)[name = string("k_head_535_cast_fp16")];
            tensor<int32, [1]> v_head_535_axes_1 = const()[name = string("v_head_535_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_535_cast_fp16 = expand_dims(axes = v_head_535_axes_1, x = v_head_533_cast_fp16)[name = string("v_head_535_cast_fp16")];
            tensor<int32, [1]> p_head_535_axes_1 = const()[name = string("p_head_535_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_535_cast_fp16 = expand_dims(axes = p_head_535_axes_1, x = p_head_533_cast_fp16)[name = string("p_head_535_cast_fp16")];
            bool var_21350_transpose_x_3 = const()[name = string("op_21350_transpose_x_3"), val = bool(false)];
            bool var_21350_transpose_y_3 = const()[name = string("op_21350_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_21350_cast_fp16 = matmul(transpose_x = var_21350_transpose_x_3, transpose_y = var_21350_transpose_y_3, x = q_u_267_cast_fp16, y = k_head_535_cast_fp16)[name = string("op_21350_cast_fp16")];
            fp16 var_21351_to_fp16 = const()[name = string("op_21351_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_267_cast_fp16 = mul(x = var_21350_cast_fp16, y = var_21351_to_fp16)[name = string("scores_content_267_cast_fp16")];
            bool x_1405_transpose_x_3 = const()[name = string("x_1405_transpose_x_3"), val = bool(false)];
            bool x_1405_transpose_y_3 = const()[name = string("x_1405_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1405_cast_fp16 = matmul(transpose_x = x_1405_transpose_x_3, transpose_y = x_1405_transpose_y_3, x = q_v_267_cast_fp16, y = p_head_535_cast_fp16)[name = string("x_1405_cast_fp16")];
            tensor<int32, [8]> x_1407_pad_1 = const()[name = string("x_1407_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1407_mode_1 = const()[name = string("x_1407_mode_1"), val = string("constant")];
            fp16 const_2203_to_fp16 = const()[name = string("const_2203_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1407_cast_fp16 = pad(constant_val = const_2203_to_fp16, mode = x_1407_mode_1, pad = x_1407_pad_1, x = x_1405_cast_fp16)[name = string("x_1407_cast_fp16")];
            tensor<int32, [4]> var_21365 = const()[name = string("op_21365"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1409_cast_fp16 = reshape(shape = var_21365, x = x_1407_cast_fp16)[name = string("x_1409_cast_fp16")];
            tensor<int32, [4]> var_21369_begin_1 = const()[name = string("op_21369_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_21369_end_1 = const()[name = string("op_21369_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_21369_end_mask_1 = const()[name = string("op_21369_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_21369_cast_fp16 = slice_by_index(begin = var_21369_begin_1, end = var_21369_end_1, end_mask = var_21369_end_mask_1, x = x_1409_cast_fp16)[name = string("op_21369_cast_fp16")];
            tensor<int32, [4]> var_21371 = const()[name = string("op_21371"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_21372_cast_fp16 = reshape(shape = var_21371, x = var_21369_cast_fp16)[name = string("op_21372_cast_fp16")];
            tensor<int32, [4]> var_21377_begin_1 = const()[name = string("op_21377_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_21377_end_1 = const()[name = string("op_21377_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_21377_end_mask_1 = const()[name = string("op_21377_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_21377_cast_fp16 = slice_by_index(begin = var_21377_begin_1, end = var_21377_end_1, end_mask = var_21377_end_mask_1, x = var_21372_cast_fp16)[name = string("op_21377_cast_fp16")];
            fp16 var_21378_to_fp16 = const()[name = string("op_21378_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_267_cast_fp16 = mul(x = var_21377_cast_fp16, y = var_21378_to_fp16)[name = string("scores_pos_267_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_267_cast_fp16 = add(x = scores_content_267_cast_fp16, y = scores_pos_267_cast_fp16)[name = string("logits_267_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_21381_cast_fp16 = softmax(axis = var_20501, x = logits_267_cast_fp16)[name = string("op_21381_cast_fp16")];
            bool var_21383_transpose_x_1 = const()[name = string("op_21383_transpose_x_1"), val = bool(false)];
            bool var_21383_transpose_y_1 = const()[name = string("op_21383_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_21383_cast_fp16 = matmul(transpose_x = var_21383_transpose_x_1, transpose_y = var_21383_transpose_y_1, x = var_21381_cast_fp16, y = v_head_535_cast_fp16)[name = string("op_21383_cast_fp16")];
            tensor<int32, [1]> var_21384_axes_1 = const()[name = string("op_21384_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_21384_cast_fp16 = squeeze(axes = var_21384_axes_1, x = var_21383_cast_fp16)[name = string("op_21384_cast_fp16")];
            string dense_output_1345_pad_type_1 = const()[name = string("dense_output_1345_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1345_strides_1 = const()[name = string("dense_output_1345_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1345_pad_1 = const()[name = string("dense_output_1345_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1345_dilations_1 = const()[name = string("dense_output_1345_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1345_groups_1 = const()[name = string("dense_output_1345_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486629504))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486760640))))[name = string("layers_16_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1345_cast_fp16 = conv(dilations = dense_output_1345_dilations_1, groups = dense_output_1345_groups_1, pad = dense_output_1345_pad_1, pad_type = dense_output_1345_pad_type_1, strides = dense_output_1345_strides_1, weight = layers_16_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1345_cast_fp16")];
            string sparse_output_1345_pad_type_1 = const()[name = string("sparse_output_1345_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1345_strides_1 = const()[name = string("sparse_output_1345_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1345_pad_1 = const()[name = string("sparse_output_1345_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1345_dilations_1 = const()[name = string("sparse_output_1345_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1345_groups_1 = const()[name = string("sparse_output_1345_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486763904))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486761216))))[name = string("layers_16_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1345_cast_fp16 = conv(dilations = sparse_output_1345_dilations_1, groups = sparse_output_1345_groups_1, pad = sparse_output_1345_pad_1, pad_type = sparse_output_1345_pad_type_1, strides = sparse_output_1345_strides_1, weight = layers_16_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1345_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21399_cast_fp16 = add(x = dense_output_1345_cast_fp16, y = sparse_output_1345_cast_fp16)[name = string("op_21399_cast_fp16")];
            tensor<int32, [4]> var_21400 = const()[name = string("op_21400"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21402 = const()[name = string("op_21402"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21401_cast_fp16 = transpose(perm = var_21400, x = var_21399_cast_fp16)[name = string("transpose_300")];
            tensor<fp16, [1, 51, 128]> q_head_269_cast_fp16 = reshape(shape = var_21402, x = var_21401_cast_fp16)[name = string("q_head_269_cast_fp16")];
            string dense_output_1347_pad_type_1 = const()[name = string("dense_output_1347_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1347_strides_1 = const()[name = string("dense_output_1347_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1347_pad_1 = const()[name = string("dense_output_1347_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1347_dilations_1 = const()[name = string("dense_output_1347_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1347_groups_1 = const()[name = string("dense_output_1347_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486780352))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486911488))))[name = string("layers_16_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1347_cast_fp16 = conv(dilations = dense_output_1347_dilations_1, groups = dense_output_1347_groups_1, pad = dense_output_1347_pad_1, pad_type = dense_output_1347_pad_type_1, strides = dense_output_1347_strides_1, weight = layers_16_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1347_cast_fp16")];
            string sparse_output_1347_pad_type_1 = const()[name = string("sparse_output_1347_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1347_strides_1 = const()[name = string("sparse_output_1347_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1347_pad_1 = const()[name = string("sparse_output_1347_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1347_dilations_1 = const()[name = string("sparse_output_1347_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1347_groups_1 = const()[name = string("sparse_output_1347_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486914752))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486912064))))[name = string("layers_16_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1347_cast_fp16 = conv(dilations = sparse_output_1347_dilations_1, groups = sparse_output_1347_groups_1, pad = sparse_output_1347_pad_1, pad_type = sparse_output_1347_pad_type_1, strides = sparse_output_1347_strides_1, weight = layers_16_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1347_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21418_cast_fp16 = add(x = dense_output_1347_cast_fp16, y = sparse_output_1347_cast_fp16)[name = string("op_21418_cast_fp16")];
            tensor<int32, [4]> var_21419 = const()[name = string("op_21419"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21421 = const()[name = string("op_21421"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21420_cast_fp16 = transpose(perm = var_21419, x = var_21418_cast_fp16)[name = string("transpose_299")];
            tensor<fp16, [1, 51, 128]> k_head_537_cast_fp16 = reshape(shape = var_21421, x = var_21420_cast_fp16)[name = string("k_head_537_cast_fp16")];
            string dense_output_1349_pad_type_1 = const()[name = string("dense_output_1349_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1349_strides_1 = const()[name = string("dense_output_1349_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1349_pad_1 = const()[name = string("dense_output_1349_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1349_dilations_1 = const()[name = string("dense_output_1349_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1349_groups_1 = const()[name = string("dense_output_1349_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486931200))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487062336))))[name = string("layers_16_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1349_cast_fp16 = conv(dilations = dense_output_1349_dilations_1, groups = dense_output_1349_groups_1, pad = dense_output_1349_pad_1, pad_type = dense_output_1349_pad_type_1, strides = dense_output_1349_strides_1, weight = layers_16_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1349_cast_fp16")];
            string sparse_output_1349_pad_type_1 = const()[name = string("sparse_output_1349_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1349_strides_1 = const()[name = string("sparse_output_1349_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1349_pad_1 = const()[name = string("sparse_output_1349_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1349_dilations_1 = const()[name = string("sparse_output_1349_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1349_groups_1 = const()[name = string("sparse_output_1349_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487065600))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487062912))))[name = string("layers_16_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1349_cast_fp16 = conv(dilations = sparse_output_1349_dilations_1, groups = sparse_output_1349_groups_1, pad = sparse_output_1349_pad_1, pad_type = sparse_output_1349_pad_type_1, strides = sparse_output_1349_strides_1, weight = layers_16_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1349_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21437_cast_fp16 = add(x = dense_output_1349_cast_fp16, y = sparse_output_1349_cast_fp16)[name = string("op_21437_cast_fp16")];
            tensor<int32, [4]> var_21438 = const()[name = string("op_21438"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21440 = const()[name = string("op_21440"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21439_cast_fp16 = transpose(perm = var_21438, x = var_21437_cast_fp16)[name = string("transpose_298")];
            tensor<fp16, [1, 51, 128]> v_head_537_cast_fp16 = reshape(shape = var_21440, x = var_21439_cast_fp16)[name = string("v_head_537_cast_fp16")];
            string dense_output_1351_pad_type_1 = const()[name = string("dense_output_1351_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1351_strides_1 = const()[name = string("dense_output_1351_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1351_pad_1 = const()[name = string("dense_output_1351_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1351_dilations_1 = const()[name = string("dense_output_1351_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1351_groups_1 = const()[name = string("dense_output_1351_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487082048))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487213184))))[name = string("layers_16_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1351_cast_fp16 = conv(dilations = dense_output_1351_dilations_1, groups = dense_output_1351_groups_1, pad = dense_output_1351_pad_1, pad_type = dense_output_1351_pad_type_1, strides = dense_output_1351_strides_1, weight = layers_16_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1351_cast_fp16")];
            string sparse_output_1351_pad_type_1 = const()[name = string("sparse_output_1351_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1351_strides_1 = const()[name = string("sparse_output_1351_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1351_pad_1 = const()[name = string("sparse_output_1351_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1351_dilations_1 = const()[name = string("sparse_output_1351_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1351_groups_1 = const()[name = string("sparse_output_1351_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487216448))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487213760))))[name = string("layers_16_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1351_cast_fp16 = conv(dilations = sparse_output_1351_dilations_1, groups = sparse_output_1351_groups_1, pad = sparse_output_1351_pad_1, pad_type = sparse_output_1351_pad_type_1, strides = sparse_output_1351_strides_1, weight = layers_16_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1351_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_21456_cast_fp16 = add(x = dense_output_1351_cast_fp16, y = sparse_output_1351_cast_fp16)[name = string("op_21456_cast_fp16")];
            tensor<int32, [4]> var_21457 = const()[name = string("op_21457"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21459 = const()[name = string("op_21459"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_21458_cast_fp16 = transpose(perm = var_21457, x = var_21456_cast_fp16)[name = string("transpose_297")];
            tensor<fp16, [1, 101, 128]> p_head_537_cast_fp16 = reshape(shape = var_21459, x = var_21458_cast_fp16)[name = string("p_head_537_cast_fp16")];
            tensor<fp16, [128]> var_21461_to_fp16 = const()[name = string("op_21461_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487232896)))];
            tensor<fp16, [1, 51, 128]> var_21462_cast_fp16 = add(x = q_head_269_cast_fp16, y = var_21461_to_fp16)[name = string("op_21462_cast_fp16")];
            tensor<int32, [1]> q_u_269_axes_1 = const()[name = string("q_u_269_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_269_cast_fp16 = expand_dims(axes = q_u_269_axes_1, x = var_21462_cast_fp16)[name = string("q_u_269_cast_fp16")];
            tensor<fp16, [128]> var_21464_to_fp16 = const()[name = string("op_21464_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487233216)))];
            tensor<fp16, [1, 51, 128]> var_21465_cast_fp16 = add(x = q_head_269_cast_fp16, y = var_21464_to_fp16)[name = string("op_21465_cast_fp16")];
            tensor<int32, [1]> q_v_269_axes_1 = const()[name = string("q_v_269_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_269_cast_fp16 = expand_dims(axes = q_v_269_axes_1, x = var_21465_cast_fp16)[name = string("q_v_269_cast_fp16")];
            tensor<int32, [1]> k_head_539_axes_1 = const()[name = string("k_head_539_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_539_cast_fp16 = expand_dims(axes = k_head_539_axes_1, x = k_head_537_cast_fp16)[name = string("k_head_539_cast_fp16")];
            tensor<int32, [1]> v_head_539_axes_1 = const()[name = string("v_head_539_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_539_cast_fp16 = expand_dims(axes = v_head_539_axes_1, x = v_head_537_cast_fp16)[name = string("v_head_539_cast_fp16")];
            tensor<int32, [1]> p_head_539_axes_1 = const()[name = string("p_head_539_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_539_cast_fp16 = expand_dims(axes = p_head_539_axes_1, x = p_head_537_cast_fp16)[name = string("p_head_539_cast_fp16")];
            bool var_21471_transpose_x_3 = const()[name = string("op_21471_transpose_x_3"), val = bool(false)];
            bool var_21471_transpose_y_3 = const()[name = string("op_21471_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_21471_cast_fp16 = matmul(transpose_x = var_21471_transpose_x_3, transpose_y = var_21471_transpose_y_3, x = q_u_269_cast_fp16, y = k_head_539_cast_fp16)[name = string("op_21471_cast_fp16")];
            fp16 var_21472_to_fp16 = const()[name = string("op_21472_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_269_cast_fp16 = mul(x = var_21471_cast_fp16, y = var_21472_to_fp16)[name = string("scores_content_269_cast_fp16")];
            bool x_1413_transpose_x_3 = const()[name = string("x_1413_transpose_x_3"), val = bool(false)];
            bool x_1413_transpose_y_3 = const()[name = string("x_1413_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1413_cast_fp16 = matmul(transpose_x = x_1413_transpose_x_3, transpose_y = x_1413_transpose_y_3, x = q_v_269_cast_fp16, y = p_head_539_cast_fp16)[name = string("x_1413_cast_fp16")];
            tensor<int32, [8]> x_1415_pad_1 = const()[name = string("x_1415_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1415_mode_1 = const()[name = string("x_1415_mode_1"), val = string("constant")];
            fp16 const_2209_to_fp16 = const()[name = string("const_2209_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1415_cast_fp16 = pad(constant_val = const_2209_to_fp16, mode = x_1415_mode_1, pad = x_1415_pad_1, x = x_1413_cast_fp16)[name = string("x_1415_cast_fp16")];
            tensor<int32, [4]> var_21486 = const()[name = string("op_21486"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1417_cast_fp16 = reshape(shape = var_21486, x = x_1415_cast_fp16)[name = string("x_1417_cast_fp16")];
            tensor<int32, [4]> var_21490_begin_1 = const()[name = string("op_21490_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_21490_end_1 = const()[name = string("op_21490_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_21490_end_mask_1 = const()[name = string("op_21490_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_21490_cast_fp16 = slice_by_index(begin = var_21490_begin_1, end = var_21490_end_1, end_mask = var_21490_end_mask_1, x = x_1417_cast_fp16)[name = string("op_21490_cast_fp16")];
            tensor<int32, [4]> var_21492 = const()[name = string("op_21492"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_21493_cast_fp16 = reshape(shape = var_21492, x = var_21490_cast_fp16)[name = string("op_21493_cast_fp16")];
            tensor<int32, [4]> var_21498_begin_1 = const()[name = string("op_21498_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_21498_end_1 = const()[name = string("op_21498_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_21498_end_mask_1 = const()[name = string("op_21498_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_21498_cast_fp16 = slice_by_index(begin = var_21498_begin_1, end = var_21498_end_1, end_mask = var_21498_end_mask_1, x = var_21493_cast_fp16)[name = string("op_21498_cast_fp16")];
            fp16 var_21499_to_fp16 = const()[name = string("op_21499_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_269_cast_fp16 = mul(x = var_21498_cast_fp16, y = var_21499_to_fp16)[name = string("scores_pos_269_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_269_cast_fp16 = add(x = scores_content_269_cast_fp16, y = scores_pos_269_cast_fp16)[name = string("logits_269_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_21502_cast_fp16 = softmax(axis = var_20501, x = logits_269_cast_fp16)[name = string("op_21502_cast_fp16")];
            bool var_21504_transpose_x_1 = const()[name = string("op_21504_transpose_x_1"), val = bool(false)];
            bool var_21504_transpose_y_1 = const()[name = string("op_21504_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_21504_cast_fp16 = matmul(transpose_x = var_21504_transpose_x_1, transpose_y = var_21504_transpose_y_1, x = var_21502_cast_fp16, y = v_head_539_cast_fp16)[name = string("op_21504_cast_fp16")];
            tensor<int32, [1]> var_21505_axes_1 = const()[name = string("op_21505_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_21505_cast_fp16 = squeeze(axes = var_21505_axes_1, x = var_21504_cast_fp16)[name = string("op_21505_cast_fp16")];
            string dense_output_1353_pad_type_1 = const()[name = string("dense_output_1353_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1353_strides_1 = const()[name = string("dense_output_1353_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1353_pad_1 = const()[name = string("dense_output_1353_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1353_dilations_1 = const()[name = string("dense_output_1353_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1353_groups_1 = const()[name = string("dense_output_1353_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487233536))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487364672))))[name = string("layers_16_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1353_cast_fp16 = conv(dilations = dense_output_1353_dilations_1, groups = dense_output_1353_groups_1, pad = dense_output_1353_pad_1, pad_type = dense_output_1353_pad_type_1, strides = dense_output_1353_strides_1, weight = layers_16_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1353_cast_fp16")];
            string sparse_output_1353_pad_type_1 = const()[name = string("sparse_output_1353_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1353_strides_1 = const()[name = string("sparse_output_1353_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1353_pad_1 = const()[name = string("sparse_output_1353_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1353_dilations_1 = const()[name = string("sparse_output_1353_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1353_groups_1 = const()[name = string("sparse_output_1353_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487367936))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487365248))))[name = string("layers_16_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1353_cast_fp16 = conv(dilations = sparse_output_1353_dilations_1, groups = sparse_output_1353_groups_1, pad = sparse_output_1353_pad_1, pad_type = sparse_output_1353_pad_type_1, strides = sparse_output_1353_strides_1, weight = layers_16_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1353_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21520_cast_fp16 = add(x = dense_output_1353_cast_fp16, y = sparse_output_1353_cast_fp16)[name = string("op_21520_cast_fp16")];
            tensor<int32, [4]> var_21521 = const()[name = string("op_21521"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21523 = const()[name = string("op_21523"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21522_cast_fp16 = transpose(perm = var_21521, x = var_21520_cast_fp16)[name = string("transpose_296")];
            tensor<fp16, [1, 51, 128]> q_head_271_cast_fp16 = reshape(shape = var_21523, x = var_21522_cast_fp16)[name = string("q_head_271_cast_fp16")];
            string dense_output_1355_pad_type_1 = const()[name = string("dense_output_1355_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1355_strides_1 = const()[name = string("dense_output_1355_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1355_pad_1 = const()[name = string("dense_output_1355_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1355_dilations_1 = const()[name = string("dense_output_1355_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1355_groups_1 = const()[name = string("dense_output_1355_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487384384))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487515520))))[name = string("layers_16_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1355_cast_fp16 = conv(dilations = dense_output_1355_dilations_1, groups = dense_output_1355_groups_1, pad = dense_output_1355_pad_1, pad_type = dense_output_1355_pad_type_1, strides = dense_output_1355_strides_1, weight = layers_16_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1355_cast_fp16")];
            string sparse_output_1355_pad_type_1 = const()[name = string("sparse_output_1355_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1355_strides_1 = const()[name = string("sparse_output_1355_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1355_pad_1 = const()[name = string("sparse_output_1355_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1355_dilations_1 = const()[name = string("sparse_output_1355_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1355_groups_1 = const()[name = string("sparse_output_1355_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487518784))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487516096))))[name = string("layers_16_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1355_cast_fp16 = conv(dilations = sparse_output_1355_dilations_1, groups = sparse_output_1355_groups_1, pad = sparse_output_1355_pad_1, pad_type = sparse_output_1355_pad_type_1, strides = sparse_output_1355_strides_1, weight = layers_16_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1355_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21539_cast_fp16 = add(x = dense_output_1355_cast_fp16, y = sparse_output_1355_cast_fp16)[name = string("op_21539_cast_fp16")];
            tensor<int32, [4]> var_21540 = const()[name = string("op_21540"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21542 = const()[name = string("op_21542"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21541_cast_fp16 = transpose(perm = var_21540, x = var_21539_cast_fp16)[name = string("transpose_295")];
            tensor<fp16, [1, 51, 128]> k_head_541_cast_fp16 = reshape(shape = var_21542, x = var_21541_cast_fp16)[name = string("k_head_541_cast_fp16")];
            string dense_output_1357_pad_type_1 = const()[name = string("dense_output_1357_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1357_strides_1 = const()[name = string("dense_output_1357_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1357_pad_1 = const()[name = string("dense_output_1357_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1357_dilations_1 = const()[name = string("dense_output_1357_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1357_groups_1 = const()[name = string("dense_output_1357_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487535232))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487666368))))[name = string("layers_16_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1357_cast_fp16 = conv(dilations = dense_output_1357_dilations_1, groups = dense_output_1357_groups_1, pad = dense_output_1357_pad_1, pad_type = dense_output_1357_pad_type_1, strides = dense_output_1357_strides_1, weight = layers_16_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_765_cast_fp16)[name = string("dense_output_1357_cast_fp16")];
            string sparse_output_1357_pad_type_1 = const()[name = string("sparse_output_1357_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1357_strides_1 = const()[name = string("sparse_output_1357_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1357_pad_1 = const()[name = string("sparse_output_1357_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1357_dilations_1 = const()[name = string("sparse_output_1357_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1357_groups_1 = const()[name = string("sparse_output_1357_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487669632))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487666944))))[name = string("layers_16_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1357_cast_fp16 = conv(dilations = sparse_output_1357_dilations_1, groups = sparse_output_1357_groups_1, pad = sparse_output_1357_pad_1, pad_type = sparse_output_1357_pad_type_1, strides = sparse_output_1357_strides_1, weight = layers_16_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_765_cast_fp16)[name = string("sparse_output_1357_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21558_cast_fp16 = add(x = dense_output_1357_cast_fp16, y = sparse_output_1357_cast_fp16)[name = string("op_21558_cast_fp16")];
            tensor<int32, [4]> var_21559 = const()[name = string("op_21559"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21561 = const()[name = string("op_21561"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21560_cast_fp16 = transpose(perm = var_21559, x = var_21558_cast_fp16)[name = string("transpose_294")];
            tensor<fp16, [1, 51, 128]> v_head_541_cast_fp16 = reshape(shape = var_21561, x = var_21560_cast_fp16)[name = string("v_head_541_cast_fp16")];
            string dense_output_1359_pad_type_1 = const()[name = string("dense_output_1359_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1359_strides_1 = const()[name = string("dense_output_1359_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1359_pad_1 = const()[name = string("dense_output_1359_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1359_dilations_1 = const()[name = string("dense_output_1359_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1359_groups_1 = const()[name = string("dense_output_1359_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487686080))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487817216))))[name = string("layers_16_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1359_cast_fp16 = conv(dilations = dense_output_1359_dilations_1, groups = dense_output_1359_groups_1, pad = dense_output_1359_pad_1, pad_type = dense_output_1359_pad_type_1, strides = dense_output_1359_strides_1, weight = layers_16_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1359_cast_fp16")];
            string sparse_output_1359_pad_type_1 = const()[name = string("sparse_output_1359_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1359_strides_1 = const()[name = string("sparse_output_1359_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1359_pad_1 = const()[name = string("sparse_output_1359_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1359_dilations_1 = const()[name = string("sparse_output_1359_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1359_groups_1 = const()[name = string("sparse_output_1359_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_16_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487820480))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487817792))))[name = string("layers_16_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1359_cast_fp16 = conv(dilations = sparse_output_1359_dilations_1, groups = sparse_output_1359_groups_1, pad = sparse_output_1359_pad_1, pad_type = sparse_output_1359_pad_type_1, strides = sparse_output_1359_strides_1, weight = layers_16_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1359_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_21577_cast_fp16 = add(x = dense_output_1359_cast_fp16, y = sparse_output_1359_cast_fp16)[name = string("op_21577_cast_fp16")];
            tensor<int32, [4]> var_21578 = const()[name = string("op_21578"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21580 = const()[name = string("op_21580"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_21579_cast_fp16 = transpose(perm = var_21578, x = var_21577_cast_fp16)[name = string("transpose_293")];
            tensor<fp16, [1, 101, 128]> p_head_541_cast_fp16 = reshape(shape = var_21580, x = var_21579_cast_fp16)[name = string("p_head_541_cast_fp16")];
            tensor<fp16, [128]> var_21582_to_fp16 = const()[name = string("op_21582_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487836928)))];
            tensor<fp16, [1, 51, 128]> var_21583_cast_fp16 = add(x = q_head_271_cast_fp16, y = var_21582_to_fp16)[name = string("op_21583_cast_fp16")];
            tensor<int32, [1]> q_u_271_axes_1 = const()[name = string("q_u_271_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_271_cast_fp16 = expand_dims(axes = q_u_271_axes_1, x = var_21583_cast_fp16)[name = string("q_u_271_cast_fp16")];
            tensor<fp16, [128]> var_21585_to_fp16 = const()[name = string("op_21585_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487837248)))];
            tensor<fp16, [1, 51, 128]> var_21586_cast_fp16 = add(x = q_head_271_cast_fp16, y = var_21585_to_fp16)[name = string("op_21586_cast_fp16")];
            tensor<int32, [1]> q_v_271_axes_1 = const()[name = string("q_v_271_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_271_cast_fp16 = expand_dims(axes = q_v_271_axes_1, x = var_21586_cast_fp16)[name = string("q_v_271_cast_fp16")];
            tensor<int32, [1]> k_head_543_axes_1 = const()[name = string("k_head_543_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_543_cast_fp16 = expand_dims(axes = k_head_543_axes_1, x = k_head_541_cast_fp16)[name = string("k_head_543_cast_fp16")];
            tensor<int32, [1]> v_head_543_axes_1 = const()[name = string("v_head_543_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_543_cast_fp16 = expand_dims(axes = v_head_543_axes_1, x = v_head_541_cast_fp16)[name = string("v_head_543_cast_fp16")];
            tensor<int32, [1]> p_head_543_axes_1 = const()[name = string("p_head_543_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_543_cast_fp16 = expand_dims(axes = p_head_543_axes_1, x = p_head_541_cast_fp16)[name = string("p_head_543_cast_fp16")];
            bool var_21592_transpose_x_3 = const()[name = string("op_21592_transpose_x_3"), val = bool(false)];
            bool var_21592_transpose_y_3 = const()[name = string("op_21592_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_21592_cast_fp16 = matmul(transpose_x = var_21592_transpose_x_3, transpose_y = var_21592_transpose_y_3, x = q_u_271_cast_fp16, y = k_head_543_cast_fp16)[name = string("op_21592_cast_fp16")];
            fp16 var_21593_to_fp16 = const()[name = string("op_21593_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_271_cast_fp16 = mul(x = var_21592_cast_fp16, y = var_21593_to_fp16)[name = string("scores_content_271_cast_fp16")];
            bool x_1421_transpose_x_3 = const()[name = string("x_1421_transpose_x_3"), val = bool(false)];
            bool x_1421_transpose_y_3 = const()[name = string("x_1421_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1421_cast_fp16 = matmul(transpose_x = x_1421_transpose_x_3, transpose_y = x_1421_transpose_y_3, x = q_v_271_cast_fp16, y = p_head_543_cast_fp16)[name = string("x_1421_cast_fp16")];
            tensor<int32, [8]> x_1423_pad_1 = const()[name = string("x_1423_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1423_mode_1 = const()[name = string("x_1423_mode_1"), val = string("constant")];
            fp16 const_2215_to_fp16 = const()[name = string("const_2215_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1423_cast_fp16 = pad(constant_val = const_2215_to_fp16, mode = x_1423_mode_1, pad = x_1423_pad_1, x = x_1421_cast_fp16)[name = string("x_1423_cast_fp16")];
            tensor<int32, [4]> var_21607 = const()[name = string("op_21607"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1425_cast_fp16 = reshape(shape = var_21607, x = x_1423_cast_fp16)[name = string("x_1425_cast_fp16")];
            tensor<int32, [4]> var_21611_begin_1 = const()[name = string("op_21611_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_21611_end_1 = const()[name = string("op_21611_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_21611_end_mask_1 = const()[name = string("op_21611_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_21611_cast_fp16 = slice_by_index(begin = var_21611_begin_1, end = var_21611_end_1, end_mask = var_21611_end_mask_1, x = x_1425_cast_fp16)[name = string("op_21611_cast_fp16")];
            tensor<int32, [4]> var_21613 = const()[name = string("op_21613"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_21614_cast_fp16 = reshape(shape = var_21613, x = var_21611_cast_fp16)[name = string("op_21614_cast_fp16")];
            tensor<int32, [4]> var_21619_begin_1 = const()[name = string("op_21619_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_21619_end_1 = const()[name = string("op_21619_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_21619_end_mask_1 = const()[name = string("op_21619_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_21619_cast_fp16 = slice_by_index(begin = var_21619_begin_1, end = var_21619_end_1, end_mask = var_21619_end_mask_1, x = var_21614_cast_fp16)[name = string("op_21619_cast_fp16")];
            fp16 var_21620_to_fp16 = const()[name = string("op_21620_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_271_cast_fp16 = mul(x = var_21619_cast_fp16, y = var_21620_to_fp16)[name = string("scores_pos_271_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_271_cast_fp16 = add(x = scores_content_271_cast_fp16, y = scores_pos_271_cast_fp16)[name = string("logits_271_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_21623_cast_fp16 = softmax(axis = var_20501, x = logits_271_cast_fp16)[name = string("op_21623_cast_fp16")];
            bool var_21625_transpose_x_1 = const()[name = string("op_21625_transpose_x_1"), val = bool(false)];
            bool var_21625_transpose_y_1 = const()[name = string("op_21625_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_21625_cast_fp16 = matmul(transpose_x = var_21625_transpose_x_1, transpose_y = var_21625_transpose_y_1, x = var_21623_cast_fp16, y = v_head_543_cast_fp16)[name = string("op_21625_cast_fp16")];
            tensor<int32, [1]> o_head_33_axes_1 = const()[name = string("o_head_33_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_33_cast_fp16 = squeeze(axes = o_head_33_axes_1, x = var_21625_cast_fp16)[name = string("o_head_33_cast_fp16")];
            bool out_33_interleave_1 = const()[name = string("out_33_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_33_cast_fp16 = concat(axis = var_20501, interleave = out_33_interleave_1, values = (var_20779_cast_fp16, var_20900_cast_fp16, var_21021_cast_fp16, var_21142_cast_fp16, var_21263_cast_fp16, var_21384_cast_fp16, var_21505_cast_fp16, o_head_33_cast_fp16))[name = string("out_33_cast_fp16")];
            tensor<int32, [3]> var_21629_perm_1 = const()[name = string("op_21629_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_773_axes_1 = const()[name = string("input_773_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_21629_cast_fp16 = transpose(perm = var_21629_perm_1, x = out_33_cast_fp16)[name = string("transpose_292")];
            tensor<fp16, [1, 1024, 51, 1]> input_773_cast_fp16 = expand_dims(axes = input_773_axes_1, x = var_21629_cast_fp16)[name = string("input_773_cast_fp16")];
            string dense_output_1361_pad_type_1 = const()[name = string("dense_output_1361_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1361_strides_1 = const()[name = string("dense_output_1361_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1361_pad_1 = const()[name = string("dense_output_1361_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1361_dilations_1 = const()[name = string("dense_output_1361_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1361_groups_1 = const()[name = string("dense_output_1361_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487837568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488886208))))[name = string("layers_16_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1361_cast_fp16 = conv(dilations = dense_output_1361_dilations_1, groups = dense_output_1361_groups_1, pad = dense_output_1361_pad_1, pad_type = dense_output_1361_pad_type_1, strides = dense_output_1361_strides_1, weight = layers_16_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_773_cast_fp16)[name = string("dense_output_1361_cast_fp16")];
            string sparse_output_1361_pad_type_1 = const()[name = string("sparse_output_1361_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1361_strides_1 = const()[name = string("sparse_output_1361_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1361_pad_1 = const()[name = string("sparse_output_1361_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1361_dilations_1 = const()[name = string("sparse_output_1361_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1361_groups_1 = const()[name = string("sparse_output_1361_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488907840))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488886784))))[name = string("layers_16_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1361_cast_fp16 = conv(dilations = sparse_output_1361_dilations_1, groups = sparse_output_1361_groups_1, pad = sparse_output_1361_pad_1, pad_type = sparse_output_1361_pad_type_1, strides = sparse_output_1361_strides_1, weight = layers_16_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_773_cast_fp16)[name = string("sparse_output_1361_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_33_cast_fp16 = add(x = dense_output_1361_cast_fp16, y = sparse_output_1361_cast_fp16)[name = string("out_conv_33_cast_fp16")];
            tensor<int32, [1]> var_21646_axes_1 = const()[name = string("op_21646_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_21646_cast_fp16 = squeeze(axes = var_21646_axes_1, x = out_conv_33_cast_fp16)[name = string("op_21646_cast_fp16")];
            tensor<int32, [3]> var_21647_perm_1 = const()[name = string("op_21647_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_21647_cast_fp16 = transpose(perm = var_21647_perm_1, x = var_21646_cast_fp16)[name = string("transpose_291")];
            tensor<fp16, [1, 51, 1024]> input_775_cast_fp16 = add(x = input_763_cast_fp16, y = var_21647_cast_fp16)[name = string("input_775_cast_fp16")];
            tensor<int32, [1]> x_1429_axes_1 = const()[name = string("x_1429_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_16_norm_conv_weight_to_fp16 = const()[name = string("layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489038976)))];
            tensor<fp16, [1024]> layers_16_norm_conv_bias_to_fp16 = const()[name = string("layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489041088)))];
            tensor<fp16, [1, 51, 1024]> x_1429_cast_fp16 = layer_norm(axes = x_1429_axes_1, beta = layers_16_norm_conv_bias_to_fp16, epsilon = var_20516_to_fp16, gamma = layers_16_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_1429_cast_fp16")];
            tensor<int32, [3]> var_21657_perm_1 = const()[name = string("op_21657_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_777_axes_1 = const()[name = string("input_777_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_21657_cast_fp16 = transpose(perm = var_21657_perm_1, x = x_1429_cast_fp16)[name = string("transpose_290")];
            tensor<fp16, [1, 1024, 51, 1]> input_777_cast_fp16 = expand_dims(axes = input_777_axes_1, x = var_21657_cast_fp16)[name = string("input_777_cast_fp16")];
            string dense_output_1363_pad_type_1 = const()[name = string("dense_output_1363_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1363_strides_1 = const()[name = string("dense_output_1363_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1363_pad_1 = const()[name = string("dense_output_1363_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1363_dilations_1 = const()[name = string("dense_output_1363_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1363_groups_1 = const()[name = string("dense_output_1363_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_16_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489043200))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491140416))))[name = string("layers_16_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1363_cast_fp16 = conv(dilations = dense_output_1363_dilations_1, groups = dense_output_1363_groups_1, pad = dense_output_1363_pad_1, pad_type = dense_output_1363_pad_type_1, strides = dense_output_1363_strides_1, weight = layers_16_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_777_cast_fp16)[name = string("dense_output_1363_cast_fp16")];
            string sparse_output_1363_pad_type_1 = const()[name = string("sparse_output_1363_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1363_strides_1 = const()[name = string("sparse_output_1363_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1363_pad_1 = const()[name = string("sparse_output_1363_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1363_dilations_1 = const()[name = string("sparse_output_1363_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1363_groups_1 = const()[name = string("sparse_output_1363_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_16_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491183040))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491140992))))[name = string("layers_16_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1363_cast_fp16 = conv(dilations = sparse_output_1363_dilations_1, groups = sparse_output_1363_groups_1, pad = sparse_output_1363_pad_1, pad_type = sparse_output_1363_pad_type_1, strides = sparse_output_1363_strides_1, weight = layers_16_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_777_cast_fp16)[name = string("sparse_output_1363_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_779_cast_fp16 = add(x = dense_output_1363_cast_fp16, y = sparse_output_1363_cast_fp16)[name = string("input_779_cast_fp16")];
            int32 input_781_split_num_splits_1 = const()[name = string("input_781_split_num_splits_1"), val = int32(2)];
            int32 input_781_split_axis_1 = const()[name = string("input_781_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_781_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_781_split_cast_fp16_1 = split(axis = input_781_split_axis_1, num_splits = input_781_split_num_splits_1, x = input_779_cast_fp16)[name = string("input_781_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_781_split_1_sigmoid_cast_fp16 = sigmoid(x = input_781_split_cast_fp16_1)[name = string("input_781_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_781_cast_fp16 = mul(x = input_781_split_cast_fp16_0, y = input_781_split_1_sigmoid_cast_fp16)[name = string("input_781_cast_fp16")];
            tensor<int32, [8]> input_783_pad_1 = const()[name = string("input_783_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_783_mode_1 = const()[name = string("input_783_mode_1"), val = string("constant")];
            fp16 const_2217_to_fp16 = const()[name = string("const_2217_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_783_cast_fp16 = pad(constant_val = const_2217_to_fp16, mode = input_783_mode_1, pad = input_783_pad_1, x = input_781_cast_fp16)[name = string("input_783_cast_fp16")];
            string dense_output_1365_pad_type_1 = const()[name = string("dense_output_1365_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1365_strides_1 = const()[name = string("dense_output_1365_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1365_pad_1 = const()[name = string("dense_output_1365_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1365_dilations_1 = const()[name = string("dense_output_1365_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1365_groups_1 = const()[name = string("dense_output_1365_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1365_cast_fp16 = conv(dilations = dense_output_1365_dilations_1, groups = dense_output_1365_groups_1, pad = dense_output_1365_pad_1, pad_type = dense_output_1365_pad_type_1, strides = dense_output_1365_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_783_cast_fp16)[name = string("dense_output_1365_cast_fp16")];
            string sparse_output_1365_pad_type_1 = const()[name = string("sparse_output_1365_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1365_strides_1 = const()[name = string("sparse_output_1365_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1365_pad_1 = const()[name = string("sparse_output_1365_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1365_dilations_1 = const()[name = string("sparse_output_1365_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1365_groups_1 = const()[name = string("sparse_output_1365_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_16_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491445248))))[name = string("layers_16_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1365_cast_fp16 = conv(dilations = sparse_output_1365_dilations_1, groups = sparse_output_1365_groups_1, pad = sparse_output_1365_pad_1, pad_type = sparse_output_1365_pad_type_1, strides = sparse_output_1365_strides_1, weight = layers_16_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_783_cast_fp16)[name = string("sparse_output_1365_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_785_cast_fp16 = add(x = dense_output_1365_cast_fp16, y = sparse_output_1365_cast_fp16)[name = string("input_785_cast_fp16")];
            tensor<fp16, [1024]> layers_16_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_16_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491463744)))];
            tensor<fp16, [1024]> layers_16_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_16_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491465856)))];
            tensor<fp16, [1024]> layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491467968)))];
            tensor<fp16, [1024]> layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491470080)))];
            tensor<fp16, [1, 1024, 51, 1]> input_787_cast_fp16 = batch_norm(beta = layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_20516_to_fp16, gamma = layers_16_conv_batch_norm_weight_to_fp16, mean = layers_16_conv_batch_norm_running_mean_to_fp16, variance = layers_16_conv_batch_norm_running_var_to_fp16, x = input_785_cast_fp16)[name = string("input_787_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_789_cast_fp16 = silu(x = input_787_cast_fp16)[name = string("input_789_cast_fp16")];
            string dense_output_1367_pad_type_1 = const()[name = string("dense_output_1367_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1367_strides_1 = const()[name = string("dense_output_1367_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1367_pad_1 = const()[name = string("dense_output_1367_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1367_dilations_1 = const()[name = string("dense_output_1367_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1367_groups_1 = const()[name = string("dense_output_1367_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491472192))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492520832))))[name = string("layers_16_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1367_cast_fp16 = conv(dilations = dense_output_1367_dilations_1, groups = dense_output_1367_groups_1, pad = dense_output_1367_pad_1, pad_type = dense_output_1367_pad_type_1, strides = dense_output_1367_strides_1, weight = layers_16_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_789_cast_fp16)[name = string("dense_output_1367_cast_fp16")];
            string sparse_output_1367_pad_type_1 = const()[name = string("sparse_output_1367_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1367_strides_1 = const()[name = string("sparse_output_1367_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1367_pad_1 = const()[name = string("sparse_output_1367_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1367_dilations_1 = const()[name = string("sparse_output_1367_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1367_groups_1 = const()[name = string("sparse_output_1367_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_16_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492542464))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492521408))))[name = string("layers_16_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1367_cast_fp16 = conv(dilations = sparse_output_1367_dilations_1, groups = sparse_output_1367_groups_1, pad = sparse_output_1367_pad_1, pad_type = sparse_output_1367_pad_type_1, strides = sparse_output_1367_strides_1, weight = layers_16_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_789_cast_fp16)[name = string("sparse_output_1367_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1431_cast_fp16 = add(x = dense_output_1367_cast_fp16, y = sparse_output_1367_cast_fp16)[name = string("x_1431_cast_fp16")];
            tensor<int32, [1]> var_21713_axes_1 = const()[name = string("op_21713_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_21713_cast_fp16 = squeeze(axes = var_21713_axes_1, x = x_1431_cast_fp16)[name = string("op_21713_cast_fp16")];
            tensor<int32, [3]> var_21714_perm_1 = const()[name = string("op_21714_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_21714_cast_fp16 = transpose(perm = var_21714_perm_1, x = var_21713_cast_fp16)[name = string("transpose_289")];
            tensor<fp16, [1, 51, 1024]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = var_21714_cast_fp16)[name = string("input_791_cast_fp16")];
            tensor<int32, [1]> x_1433_axes_1 = const()[name = string("x_1433_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492673600)))];
            tensor<fp16, [1024]> layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492675712)))];
            tensor<fp16, [1, 51, 1024]> x_1433_cast_fp16 = layer_norm(axes = x_1433_axes_1, beta = layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_20516_to_fp16, gamma = layers_16_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("x_1433_cast_fp16")];
            tensor<int32, [4]> var_21724 = const()[name = string("op_21724"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1435_cast_fp16 = reshape(shape = var_21724, x = x_1433_cast_fp16)[name = string("x_1435_cast_fp16")];
            tensor<int32, [4]> input_793_perm_1 = const()[name = string("input_793_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1369_pad_type_1 = const()[name = string("dense_output_1369_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1369_strides_1 = const()[name = string("dense_output_1369_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1369_pad_1 = const()[name = string("dense_output_1369_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1369_dilations_1 = const()[name = string("dense_output_1369_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1369_groups_1 = const()[name = string("dense_output_1369_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_16_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492677824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(496872192))))[name = string("layers_16_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_793_cast_fp16 = transpose(perm = input_793_perm_1, x = x_1435_cast_fp16)[name = string("transpose_288")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1369_cast_fp16 = conv(dilations = dense_output_1369_dilations_1, groups = dense_output_1369_groups_1, pad = dense_output_1369_pad_1, pad_type = dense_output_1369_pad_type_1, strides = dense_output_1369_strides_1, weight = layers_16_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("dense_output_1369_cast_fp16")];
            string sparse_output_1369_pad_type_1 = const()[name = string("sparse_output_1369_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1369_strides_1 = const()[name = string("sparse_output_1369_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1369_pad_1 = const()[name = string("sparse_output_1369_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1369_dilations_1 = const()[name = string("sparse_output_1369_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1369_groups_1 = const()[name = string("sparse_output_1369_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_16_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(496956736))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(496872768))))[name = string("layers_16_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1369_cast_fp16 = conv(dilations = sparse_output_1369_dilations_1, groups = sparse_output_1369_groups_1, pad = sparse_output_1369_pad_1, pad_type = sparse_output_1369_pad_type_1, strides = sparse_output_1369_strides_1, weight = layers_16_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_793_cast_fp16)[name = string("sparse_output_1369_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_795_cast_fp16 = add(x = dense_output_1369_cast_fp16, y = sparse_output_1369_cast_fp16)[name = string("input_795_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_797_cast_fp16 = silu(x = input_795_cast_fp16)[name = string("input_797_cast_fp16")];
            string dense_output_1371_pad_type_1 = const()[name = string("dense_output_1371_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1371_strides_1 = const()[name = string("dense_output_1371_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1371_pad_1 = const()[name = string("dense_output_1371_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1371_dilations_1 = const()[name = string("dense_output_1371_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1371_groups_1 = const()[name = string("dense_output_1371_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_16_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497481088))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501675456))))[name = string("layers_16_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1371_cast_fp16 = conv(dilations = dense_output_1371_dilations_1, groups = dense_output_1371_groups_1, pad = dense_output_1371_pad_1, pad_type = dense_output_1371_pad_type_1, strides = dense_output_1371_strides_1, weight = layers_16_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("dense_output_1371_cast_fp16")];
            string sparse_output_1371_pad_type_1 = const()[name = string("sparse_output_1371_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1371_strides_1 = const()[name = string("sparse_output_1371_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1371_pad_1 = const()[name = string("sparse_output_1371_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1371_dilations_1 = const()[name = string("sparse_output_1371_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1371_groups_1 = const()[name = string("sparse_output_1371_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_16_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501760000))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501676032))))[name = string("layers_16_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1371_cast_fp16 = conv(dilations = sparse_output_1371_dilations_1, groups = sparse_output_1371_groups_1, pad = sparse_output_1371_pad_1, pad_type = sparse_output_1371_pad_type_1, strides = sparse_output_1371_strides_1, weight = layers_16_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_797_cast_fp16)[name = string("sparse_output_1371_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1437_cast_fp16 = add(x = dense_output_1371_cast_fp16, y = sparse_output_1371_cast_fp16)[name = string("x_1437_cast_fp16")];
            tensor<int32, [4]> x_1439_perm_1 = const()[name = string("x_1439_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_21759 = const()[name = string("op_21759"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1439_cast_fp16 = transpose(perm = x_1439_perm_1, x = x_1437_cast_fp16)[name = string("transpose_287")];
            tensor<fp16, [1, 51, 1024]> var_21760_cast_fp16 = reshape(shape = var_21759, x = x_1439_cast_fp16)[name = string("op_21760_cast_fp16")];
            fp16 var_21761_to_fp16 = const()[name = string("op_21761_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_21762_cast_fp16 = mul(x = var_21760_cast_fp16, y = var_21761_to_fp16)[name = string("op_21762_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_799_cast_fp16 = add(x = input_791_cast_fp16, y = var_21762_cast_fp16)[name = string("input_799_cast_fp16")];
            tensor<int32, [1]> input_801_axes_1 = const()[name = string("input_801_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_16_norm_out_weight_to_fp16 = const()[name = string("layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502284352)))];
            tensor<fp16, [1024]> layers_16_norm_out_bias_to_fp16 = const()[name = string("layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502286464)))];
            tensor<fp16, [1, 51, 1024]> input_801_cast_fp16 = layer_norm(axes = input_801_axes_1, beta = layers_16_norm_out_bias_to_fp16, epsilon = var_20516_to_fp16, gamma = layers_16_norm_out_weight_to_fp16, x = input_799_cast_fp16)[name = string("input_801_cast_fp16")];
            int32 var_21770 = const()[name = string("op_21770"), val = int32(-1)];
            tensor<int32, [1]> x_1441_axes_1 = const()[name = string("x_1441_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502288576)))];
            tensor<fp16, [1024]> layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502290688)))];
            fp16 var_21785_to_fp16 = const()[name = string("op_21785_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1441_cast_fp16 = layer_norm(axes = x_1441_axes_1, beta = layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_21785_to_fp16, gamma = layers_17_norm_feed_forward1_weight_to_fp16, x = input_801_cast_fp16)[name = string("x_1441_cast_fp16")];
            tensor<int32, [4]> var_21804 = const()[name = string("op_21804"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1443_cast_fp16 = reshape(shape = var_21804, x = x_1441_cast_fp16)[name = string("x_1443_cast_fp16")];
            tensor<int32, [4]> input_803_perm_1 = const()[name = string("input_803_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1373_pad_type_1 = const()[name = string("dense_output_1373_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1373_strides_1 = const()[name = string("dense_output_1373_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1373_pad_1 = const()[name = string("dense_output_1373_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1373_dilations_1 = const()[name = string("dense_output_1373_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1373_groups_1 = const()[name = string("dense_output_1373_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_17_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502292800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506487168))))[name = string("layers_17_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_803_cast_fp16 = transpose(perm = input_803_perm_1, x = x_1443_cast_fp16)[name = string("transpose_286")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1373_cast_fp16 = conv(dilations = dense_output_1373_dilations_1, groups = dense_output_1373_groups_1, pad = dense_output_1373_pad_1, pad_type = dense_output_1373_pad_type_1, strides = dense_output_1373_strides_1, weight = layers_17_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_803_cast_fp16)[name = string("dense_output_1373_cast_fp16")];
            string sparse_output_1373_pad_type_1 = const()[name = string("sparse_output_1373_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1373_strides_1 = const()[name = string("sparse_output_1373_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1373_pad_1 = const()[name = string("sparse_output_1373_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1373_dilations_1 = const()[name = string("sparse_output_1373_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1373_groups_1 = const()[name = string("sparse_output_1373_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_17_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506571712))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506487744))))[name = string("layers_17_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1373_cast_fp16 = conv(dilations = sparse_output_1373_dilations_1, groups = sparse_output_1373_groups_1, pad = sparse_output_1373_pad_1, pad_type = sparse_output_1373_pad_type_1, strides = sparse_output_1373_strides_1, weight = layers_17_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_803_cast_fp16)[name = string("sparse_output_1373_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_805_cast_fp16 = add(x = dense_output_1373_cast_fp16, y = sparse_output_1373_cast_fp16)[name = string("input_805_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_807_cast_fp16 = silu(x = input_805_cast_fp16)[name = string("input_807_cast_fp16")];
            string dense_output_1375_pad_type_1 = const()[name = string("dense_output_1375_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1375_strides_1 = const()[name = string("dense_output_1375_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1375_pad_1 = const()[name = string("dense_output_1375_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1375_dilations_1 = const()[name = string("dense_output_1375_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1375_groups_1 = const()[name = string("dense_output_1375_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_17_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(507096064))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511290432))))[name = string("layers_17_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1375_cast_fp16 = conv(dilations = dense_output_1375_dilations_1, groups = dense_output_1375_groups_1, pad = dense_output_1375_pad_1, pad_type = dense_output_1375_pad_type_1, strides = dense_output_1375_strides_1, weight = layers_17_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("dense_output_1375_cast_fp16")];
            string sparse_output_1375_pad_type_1 = const()[name = string("sparse_output_1375_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1375_strides_1 = const()[name = string("sparse_output_1375_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1375_pad_1 = const()[name = string("sparse_output_1375_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1375_dilations_1 = const()[name = string("sparse_output_1375_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1375_groups_1 = const()[name = string("sparse_output_1375_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_17_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511374976))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511291008))))[name = string("layers_17_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1375_cast_fp16 = conv(dilations = sparse_output_1375_dilations_1, groups = sparse_output_1375_groups_1, pad = sparse_output_1375_pad_1, pad_type = sparse_output_1375_pad_type_1, strides = sparse_output_1375_strides_1, weight = layers_17_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_807_cast_fp16)[name = string("sparse_output_1375_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1445_cast_fp16 = add(x = dense_output_1375_cast_fp16, y = sparse_output_1375_cast_fp16)[name = string("x_1445_cast_fp16")];
            tensor<int32, [4]> x_1447_perm_1 = const()[name = string("x_1447_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_21839 = const()[name = string("op_21839"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1447_cast_fp16 = transpose(perm = x_1447_perm_1, x = x_1445_cast_fp16)[name = string("transpose_285")];
            tensor<fp16, [1, 51, 1024]> var_21840_cast_fp16 = reshape(shape = var_21839, x = x_1447_cast_fp16)[name = string("op_21840_cast_fp16")];
            fp16 var_21841_to_fp16 = const()[name = string("op_21841_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_21842_cast_fp16 = mul(x = var_21840_cast_fp16, y = var_21841_to_fp16)[name = string("op_21842_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_809_cast_fp16 = add(x = input_801_cast_fp16, y = var_21842_cast_fp16)[name = string("input_809_cast_fp16")];
            tensor<int32, [1]> q_35_axes_1 = const()[name = string("q_35_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_17_norm_self_att_weight_to_fp16 = const()[name = string("layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511899328)))];
            tensor<fp16, [1024]> layers_17_norm_self_att_bias_to_fp16 = const()[name = string("layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511901440)))];
            tensor<fp16, [1, 51, 1024]> q_35_cast_fp16 = layer_norm(axes = q_35_axes_1, beta = layers_17_norm_self_att_bias_to_fp16, epsilon = var_21785_to_fp16, gamma = layers_17_norm_self_att_weight_to_fp16, x = input_809_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<int32, [3]> var_21916 = const()[name = string("op_21916"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_811_axes_1 = const()[name = string("input_811_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_21917_cast_fp16 = transpose(perm = var_21916, x = q_35_cast_fp16)[name = string("transpose_284")];
            tensor<fp16, [1, 1024, 51, 1]> input_811_cast_fp16 = expand_dims(axes = input_811_axes_1, x = var_21917_cast_fp16)[name = string("input_811_cast_fp16")];
            string dense_output_1377_pad_type_1 = const()[name = string("dense_output_1377_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1377_strides_1 = const()[name = string("dense_output_1377_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1377_pad_1 = const()[name = string("dense_output_1377_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1377_dilations_1 = const()[name = string("dense_output_1377_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1377_groups_1 = const()[name = string("dense_output_1377_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511903552))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512034688))))[name = string("layers_17_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1377_cast_fp16 = conv(dilations = dense_output_1377_dilations_1, groups = dense_output_1377_groups_1, pad = dense_output_1377_pad_1, pad_type = dense_output_1377_pad_type_1, strides = dense_output_1377_strides_1, weight = layers_17_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1377_cast_fp16")];
            string sparse_output_1377_pad_type_1 = const()[name = string("sparse_output_1377_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1377_strides_1 = const()[name = string("sparse_output_1377_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1377_pad_1 = const()[name = string("sparse_output_1377_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1377_dilations_1 = const()[name = string("sparse_output_1377_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1377_groups_1 = const()[name = string("sparse_output_1377_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512037952))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512035264))))[name = string("layers_17_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1377_cast_fp16 = conv(dilations = sparse_output_1377_dilations_1, groups = sparse_output_1377_groups_1, pad = sparse_output_1377_pad_1, pad_type = sparse_output_1377_pad_type_1, strides = sparse_output_1377_strides_1, weight = layers_17_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1377_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21942_cast_fp16 = add(x = dense_output_1377_cast_fp16, y = sparse_output_1377_cast_fp16)[name = string("op_21942_cast_fp16")];
            tensor<int32, [4]> var_21943 = const()[name = string("op_21943"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21945 = const()[name = string("op_21945"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21944_cast_fp16 = transpose(perm = var_21943, x = var_21942_cast_fp16)[name = string("transpose_283")];
            tensor<fp16, [1, 51, 128]> q_head_273_cast_fp16 = reshape(shape = var_21945, x = var_21944_cast_fp16)[name = string("q_head_273_cast_fp16")];
            string dense_output_1379_pad_type_1 = const()[name = string("dense_output_1379_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1379_strides_1 = const()[name = string("dense_output_1379_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1379_pad_1 = const()[name = string("dense_output_1379_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1379_dilations_1 = const()[name = string("dense_output_1379_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1379_groups_1 = const()[name = string("dense_output_1379_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512054400))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512185536))))[name = string("layers_17_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1379_cast_fp16 = conv(dilations = dense_output_1379_dilations_1, groups = dense_output_1379_groups_1, pad = dense_output_1379_pad_1, pad_type = dense_output_1379_pad_type_1, strides = dense_output_1379_strides_1, weight = layers_17_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1379_cast_fp16")];
            string sparse_output_1379_pad_type_1 = const()[name = string("sparse_output_1379_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1379_strides_1 = const()[name = string("sparse_output_1379_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1379_pad_1 = const()[name = string("sparse_output_1379_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1379_dilations_1 = const()[name = string("sparse_output_1379_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1379_groups_1 = const()[name = string("sparse_output_1379_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512188800))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512186112))))[name = string("layers_17_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1379_cast_fp16 = conv(dilations = sparse_output_1379_dilations_1, groups = sparse_output_1379_groups_1, pad = sparse_output_1379_pad_1, pad_type = sparse_output_1379_pad_type_1, strides = sparse_output_1379_strides_1, weight = layers_17_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1379_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21961_cast_fp16 = add(x = dense_output_1379_cast_fp16, y = sparse_output_1379_cast_fp16)[name = string("op_21961_cast_fp16")];
            tensor<int32, [4]> var_21962 = const()[name = string("op_21962"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21964 = const()[name = string("op_21964"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21963_cast_fp16 = transpose(perm = var_21962, x = var_21961_cast_fp16)[name = string("transpose_282")];
            tensor<fp16, [1, 51, 128]> k_head_545_cast_fp16 = reshape(shape = var_21964, x = var_21963_cast_fp16)[name = string("k_head_545_cast_fp16")];
            string dense_output_1381_pad_type_1 = const()[name = string("dense_output_1381_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1381_strides_1 = const()[name = string("dense_output_1381_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1381_pad_1 = const()[name = string("dense_output_1381_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1381_dilations_1 = const()[name = string("dense_output_1381_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1381_groups_1 = const()[name = string("dense_output_1381_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512205248))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512336384))))[name = string("layers_17_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1381_cast_fp16 = conv(dilations = dense_output_1381_dilations_1, groups = dense_output_1381_groups_1, pad = dense_output_1381_pad_1, pad_type = dense_output_1381_pad_type_1, strides = dense_output_1381_strides_1, weight = layers_17_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1381_cast_fp16")];
            string sparse_output_1381_pad_type_1 = const()[name = string("sparse_output_1381_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1381_strides_1 = const()[name = string("sparse_output_1381_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1381_pad_1 = const()[name = string("sparse_output_1381_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1381_dilations_1 = const()[name = string("sparse_output_1381_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1381_groups_1 = const()[name = string("sparse_output_1381_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512339648))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512336960))))[name = string("layers_17_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1381_cast_fp16 = conv(dilations = sparse_output_1381_dilations_1, groups = sparse_output_1381_groups_1, pad = sparse_output_1381_pad_1, pad_type = sparse_output_1381_pad_type_1, strides = sparse_output_1381_strides_1, weight = layers_17_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1381_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_21980_cast_fp16 = add(x = dense_output_1381_cast_fp16, y = sparse_output_1381_cast_fp16)[name = string("op_21980_cast_fp16")];
            tensor<int32, [4]> var_21981 = const()[name = string("op_21981"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_21983 = const()[name = string("op_21983"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_21982_cast_fp16 = transpose(perm = var_21981, x = var_21980_cast_fp16)[name = string("transpose_281")];
            tensor<fp16, [1, 51, 128]> v_head_545_cast_fp16 = reshape(shape = var_21983, x = var_21982_cast_fp16)[name = string("v_head_545_cast_fp16")];
            string dense_output_1383_pad_type_1 = const()[name = string("dense_output_1383_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1383_strides_1 = const()[name = string("dense_output_1383_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1383_pad_1 = const()[name = string("dense_output_1383_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1383_dilations_1 = const()[name = string("dense_output_1383_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1383_groups_1 = const()[name = string("dense_output_1383_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512356096))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512487232))))[name = string("layers_17_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1383_cast_fp16 = conv(dilations = dense_output_1383_dilations_1, groups = dense_output_1383_groups_1, pad = dense_output_1383_pad_1, pad_type = dense_output_1383_pad_type_1, strides = dense_output_1383_strides_1, weight = layers_17_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1383_cast_fp16")];
            string sparse_output_1383_pad_type_1 = const()[name = string("sparse_output_1383_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1383_strides_1 = const()[name = string("sparse_output_1383_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1383_pad_1 = const()[name = string("sparse_output_1383_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1383_dilations_1 = const()[name = string("sparse_output_1383_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1383_groups_1 = const()[name = string("sparse_output_1383_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512490496))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512487808))))[name = string("layers_17_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1383_cast_fp16 = conv(dilations = sparse_output_1383_dilations_1, groups = sparse_output_1383_groups_1, pad = sparse_output_1383_pad_1, pad_type = sparse_output_1383_pad_type_1, strides = sparse_output_1383_strides_1, weight = layers_17_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1383_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_21999_cast_fp16 = add(x = dense_output_1383_cast_fp16, y = sparse_output_1383_cast_fp16)[name = string("op_21999_cast_fp16")];
            tensor<int32, [4]> var_22000 = const()[name = string("op_22000"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22002 = const()[name = string("op_22002"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_22001_cast_fp16 = transpose(perm = var_22000, x = var_21999_cast_fp16)[name = string("transpose_280")];
            tensor<fp16, [1, 101, 128]> p_head_545_cast_fp16 = reshape(shape = var_22002, x = var_22001_cast_fp16)[name = string("p_head_545_cast_fp16")];
            tensor<fp16, [128]> var_22004_to_fp16 = const()[name = string("op_22004_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512506944)))];
            tensor<fp16, [1, 51, 128]> var_22005_cast_fp16 = add(x = q_head_273_cast_fp16, y = var_22004_to_fp16)[name = string("op_22005_cast_fp16")];
            tensor<int32, [1]> q_u_273_axes_1 = const()[name = string("q_u_273_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_273_cast_fp16 = expand_dims(axes = q_u_273_axes_1, x = var_22005_cast_fp16)[name = string("q_u_273_cast_fp16")];
            tensor<fp16, [128]> var_22007_to_fp16 = const()[name = string("op_22007_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512507264)))];
            tensor<fp16, [1, 51, 128]> var_22008_cast_fp16 = add(x = q_head_273_cast_fp16, y = var_22007_to_fp16)[name = string("op_22008_cast_fp16")];
            tensor<int32, [1]> q_v_273_axes_1 = const()[name = string("q_v_273_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_273_cast_fp16 = expand_dims(axes = q_v_273_axes_1, x = var_22008_cast_fp16)[name = string("q_v_273_cast_fp16")];
            tensor<int32, [1]> k_head_547_axes_1 = const()[name = string("k_head_547_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_547_cast_fp16 = expand_dims(axes = k_head_547_axes_1, x = k_head_545_cast_fp16)[name = string("k_head_547_cast_fp16")];
            tensor<int32, [1]> v_head_547_axes_1 = const()[name = string("v_head_547_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_547_cast_fp16 = expand_dims(axes = v_head_547_axes_1, x = v_head_545_cast_fp16)[name = string("v_head_547_cast_fp16")];
            tensor<int32, [1]> p_head_547_axes_1 = const()[name = string("p_head_547_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_547_cast_fp16 = expand_dims(axes = p_head_547_axes_1, x = p_head_545_cast_fp16)[name = string("p_head_547_cast_fp16")];
            bool var_22014_transpose_x_3 = const()[name = string("op_22014_transpose_x_3"), val = bool(false)];
            bool var_22014_transpose_y_3 = const()[name = string("op_22014_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_22014_cast_fp16 = matmul(transpose_x = var_22014_transpose_x_3, transpose_y = var_22014_transpose_y_3, x = q_u_273_cast_fp16, y = k_head_547_cast_fp16)[name = string("op_22014_cast_fp16")];
            fp16 var_22015_to_fp16 = const()[name = string("op_22015_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_273_cast_fp16 = mul(x = var_22014_cast_fp16, y = var_22015_to_fp16)[name = string("scores_content_273_cast_fp16")];
            bool x_1449_transpose_x_3 = const()[name = string("x_1449_transpose_x_3"), val = bool(false)];
            bool x_1449_transpose_y_3 = const()[name = string("x_1449_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1449_cast_fp16 = matmul(transpose_x = x_1449_transpose_x_3, transpose_y = x_1449_transpose_y_3, x = q_v_273_cast_fp16, y = p_head_547_cast_fp16)[name = string("x_1449_cast_fp16")];
            tensor<int32, [8]> x_1451_pad_1 = const()[name = string("x_1451_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1451_mode_1 = const()[name = string("x_1451_mode_1"), val = string("constant")];
            fp16 const_2227_to_fp16 = const()[name = string("const_2227_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1451_cast_fp16 = pad(constant_val = const_2227_to_fp16, mode = x_1451_mode_1, pad = x_1451_pad_1, x = x_1449_cast_fp16)[name = string("x_1451_cast_fp16")];
            tensor<int32, [4]> var_22029 = const()[name = string("op_22029"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1453_cast_fp16 = reshape(shape = var_22029, x = x_1451_cast_fp16)[name = string("x_1453_cast_fp16")];
            tensor<int32, [4]> var_22033_begin_1 = const()[name = string("op_22033_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_22033_end_1 = const()[name = string("op_22033_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_22033_end_mask_1 = const()[name = string("op_22033_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_22033_cast_fp16 = slice_by_index(begin = var_22033_begin_1, end = var_22033_end_1, end_mask = var_22033_end_mask_1, x = x_1453_cast_fp16)[name = string("op_22033_cast_fp16")];
            tensor<int32, [4]> var_22035 = const()[name = string("op_22035"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_22036_cast_fp16 = reshape(shape = var_22035, x = var_22033_cast_fp16)[name = string("op_22036_cast_fp16")];
            tensor<int32, [4]> var_22041_begin_1 = const()[name = string("op_22041_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_22041_end_1 = const()[name = string("op_22041_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_22041_end_mask_1 = const()[name = string("op_22041_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_22041_cast_fp16 = slice_by_index(begin = var_22041_begin_1, end = var_22041_end_1, end_mask = var_22041_end_mask_1, x = var_22036_cast_fp16)[name = string("op_22041_cast_fp16")];
            fp16 var_22042_to_fp16 = const()[name = string("op_22042_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_273_cast_fp16 = mul(x = var_22041_cast_fp16, y = var_22042_to_fp16)[name = string("scores_pos_273_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_273_cast_fp16 = add(x = scores_content_273_cast_fp16, y = scores_pos_273_cast_fp16)[name = string("logits_273_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_22045_cast_fp16 = softmax(axis = var_21770, x = logits_273_cast_fp16)[name = string("op_22045_cast_fp16")];
            bool var_22047_transpose_x_1 = const()[name = string("op_22047_transpose_x_1"), val = bool(false)];
            bool var_22047_transpose_y_1 = const()[name = string("op_22047_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_22047_cast_fp16 = matmul(transpose_x = var_22047_transpose_x_1, transpose_y = var_22047_transpose_y_1, x = var_22045_cast_fp16, y = v_head_547_cast_fp16)[name = string("op_22047_cast_fp16")];
            tensor<int32, [1]> var_22048_axes_1 = const()[name = string("op_22048_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_22048_cast_fp16 = squeeze(axes = var_22048_axes_1, x = var_22047_cast_fp16)[name = string("op_22048_cast_fp16")];
            string dense_output_1385_pad_type_1 = const()[name = string("dense_output_1385_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1385_strides_1 = const()[name = string("dense_output_1385_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1385_pad_1 = const()[name = string("dense_output_1385_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1385_dilations_1 = const()[name = string("dense_output_1385_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1385_groups_1 = const()[name = string("dense_output_1385_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512507584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512638720))))[name = string("layers_17_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1385_cast_fp16 = conv(dilations = dense_output_1385_dilations_1, groups = dense_output_1385_groups_1, pad = dense_output_1385_pad_1, pad_type = dense_output_1385_pad_type_1, strides = dense_output_1385_strides_1, weight = layers_17_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1385_cast_fp16")];
            string sparse_output_1385_pad_type_1 = const()[name = string("sparse_output_1385_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1385_strides_1 = const()[name = string("sparse_output_1385_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1385_pad_1 = const()[name = string("sparse_output_1385_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1385_dilations_1 = const()[name = string("sparse_output_1385_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1385_groups_1 = const()[name = string("sparse_output_1385_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512641984))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512639296))))[name = string("layers_17_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1385_cast_fp16 = conv(dilations = sparse_output_1385_dilations_1, groups = sparse_output_1385_groups_1, pad = sparse_output_1385_pad_1, pad_type = sparse_output_1385_pad_type_1, strides = sparse_output_1385_strides_1, weight = layers_17_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1385_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22063_cast_fp16 = add(x = dense_output_1385_cast_fp16, y = sparse_output_1385_cast_fp16)[name = string("op_22063_cast_fp16")];
            tensor<int32, [4]> var_22064 = const()[name = string("op_22064"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22066 = const()[name = string("op_22066"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22065_cast_fp16 = transpose(perm = var_22064, x = var_22063_cast_fp16)[name = string("transpose_279")];
            tensor<fp16, [1, 51, 128]> q_head_275_cast_fp16 = reshape(shape = var_22066, x = var_22065_cast_fp16)[name = string("q_head_275_cast_fp16")];
            string dense_output_1387_pad_type_1 = const()[name = string("dense_output_1387_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1387_strides_1 = const()[name = string("dense_output_1387_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1387_pad_1 = const()[name = string("dense_output_1387_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1387_dilations_1 = const()[name = string("dense_output_1387_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1387_groups_1 = const()[name = string("dense_output_1387_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512658432))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512789568))))[name = string("layers_17_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1387_cast_fp16 = conv(dilations = dense_output_1387_dilations_1, groups = dense_output_1387_groups_1, pad = dense_output_1387_pad_1, pad_type = dense_output_1387_pad_type_1, strides = dense_output_1387_strides_1, weight = layers_17_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1387_cast_fp16")];
            string sparse_output_1387_pad_type_1 = const()[name = string("sparse_output_1387_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1387_strides_1 = const()[name = string("sparse_output_1387_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1387_pad_1 = const()[name = string("sparse_output_1387_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1387_dilations_1 = const()[name = string("sparse_output_1387_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1387_groups_1 = const()[name = string("sparse_output_1387_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512792832))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512790144))))[name = string("layers_17_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1387_cast_fp16 = conv(dilations = sparse_output_1387_dilations_1, groups = sparse_output_1387_groups_1, pad = sparse_output_1387_pad_1, pad_type = sparse_output_1387_pad_type_1, strides = sparse_output_1387_strides_1, weight = layers_17_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1387_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22082_cast_fp16 = add(x = dense_output_1387_cast_fp16, y = sparse_output_1387_cast_fp16)[name = string("op_22082_cast_fp16")];
            tensor<int32, [4]> var_22083 = const()[name = string("op_22083"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22085 = const()[name = string("op_22085"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22084_cast_fp16 = transpose(perm = var_22083, x = var_22082_cast_fp16)[name = string("transpose_278")];
            tensor<fp16, [1, 51, 128]> k_head_549_cast_fp16 = reshape(shape = var_22085, x = var_22084_cast_fp16)[name = string("k_head_549_cast_fp16")];
            string dense_output_1389_pad_type_1 = const()[name = string("dense_output_1389_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1389_strides_1 = const()[name = string("dense_output_1389_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1389_pad_1 = const()[name = string("dense_output_1389_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1389_dilations_1 = const()[name = string("dense_output_1389_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1389_groups_1 = const()[name = string("dense_output_1389_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512809280))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512940416))))[name = string("layers_17_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1389_cast_fp16 = conv(dilations = dense_output_1389_dilations_1, groups = dense_output_1389_groups_1, pad = dense_output_1389_pad_1, pad_type = dense_output_1389_pad_type_1, strides = dense_output_1389_strides_1, weight = layers_17_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1389_cast_fp16")];
            string sparse_output_1389_pad_type_1 = const()[name = string("sparse_output_1389_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1389_strides_1 = const()[name = string("sparse_output_1389_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1389_pad_1 = const()[name = string("sparse_output_1389_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1389_dilations_1 = const()[name = string("sparse_output_1389_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1389_groups_1 = const()[name = string("sparse_output_1389_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512943680))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512940992))))[name = string("layers_17_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1389_cast_fp16 = conv(dilations = sparse_output_1389_dilations_1, groups = sparse_output_1389_groups_1, pad = sparse_output_1389_pad_1, pad_type = sparse_output_1389_pad_type_1, strides = sparse_output_1389_strides_1, weight = layers_17_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1389_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22101_cast_fp16 = add(x = dense_output_1389_cast_fp16, y = sparse_output_1389_cast_fp16)[name = string("op_22101_cast_fp16")];
            tensor<int32, [4]> var_22102 = const()[name = string("op_22102"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22104 = const()[name = string("op_22104"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22103_cast_fp16 = transpose(perm = var_22102, x = var_22101_cast_fp16)[name = string("transpose_277")];
            tensor<fp16, [1, 51, 128]> v_head_549_cast_fp16 = reshape(shape = var_22104, x = var_22103_cast_fp16)[name = string("v_head_549_cast_fp16")];
            string dense_output_1391_pad_type_1 = const()[name = string("dense_output_1391_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1391_strides_1 = const()[name = string("dense_output_1391_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1391_pad_1 = const()[name = string("dense_output_1391_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1391_dilations_1 = const()[name = string("dense_output_1391_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1391_groups_1 = const()[name = string("dense_output_1391_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512960128))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513091264))))[name = string("layers_17_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1391_cast_fp16 = conv(dilations = dense_output_1391_dilations_1, groups = dense_output_1391_groups_1, pad = dense_output_1391_pad_1, pad_type = dense_output_1391_pad_type_1, strides = dense_output_1391_strides_1, weight = layers_17_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1391_cast_fp16")];
            string sparse_output_1391_pad_type_1 = const()[name = string("sparse_output_1391_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1391_strides_1 = const()[name = string("sparse_output_1391_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1391_pad_1 = const()[name = string("sparse_output_1391_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1391_dilations_1 = const()[name = string("sparse_output_1391_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1391_groups_1 = const()[name = string("sparse_output_1391_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513094528))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513091840))))[name = string("layers_17_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1391_cast_fp16 = conv(dilations = sparse_output_1391_dilations_1, groups = sparse_output_1391_groups_1, pad = sparse_output_1391_pad_1, pad_type = sparse_output_1391_pad_type_1, strides = sparse_output_1391_strides_1, weight = layers_17_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1391_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_22120_cast_fp16 = add(x = dense_output_1391_cast_fp16, y = sparse_output_1391_cast_fp16)[name = string("op_22120_cast_fp16")];
            tensor<int32, [4]> var_22121 = const()[name = string("op_22121"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22123 = const()[name = string("op_22123"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_22122_cast_fp16 = transpose(perm = var_22121, x = var_22120_cast_fp16)[name = string("transpose_276")];
            tensor<fp16, [1, 101, 128]> p_head_549_cast_fp16 = reshape(shape = var_22123, x = var_22122_cast_fp16)[name = string("p_head_549_cast_fp16")];
            tensor<fp16, [128]> var_22125_to_fp16 = const()[name = string("op_22125_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513110976)))];
            tensor<fp16, [1, 51, 128]> var_22126_cast_fp16 = add(x = q_head_275_cast_fp16, y = var_22125_to_fp16)[name = string("op_22126_cast_fp16")];
            tensor<int32, [1]> q_u_275_axes_1 = const()[name = string("q_u_275_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_275_cast_fp16 = expand_dims(axes = q_u_275_axes_1, x = var_22126_cast_fp16)[name = string("q_u_275_cast_fp16")];
            tensor<fp16, [128]> var_22128_to_fp16 = const()[name = string("op_22128_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513111296)))];
            tensor<fp16, [1, 51, 128]> var_22129_cast_fp16 = add(x = q_head_275_cast_fp16, y = var_22128_to_fp16)[name = string("op_22129_cast_fp16")];
            tensor<int32, [1]> q_v_275_axes_1 = const()[name = string("q_v_275_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_275_cast_fp16 = expand_dims(axes = q_v_275_axes_1, x = var_22129_cast_fp16)[name = string("q_v_275_cast_fp16")];
            tensor<int32, [1]> k_head_551_axes_1 = const()[name = string("k_head_551_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_551_cast_fp16 = expand_dims(axes = k_head_551_axes_1, x = k_head_549_cast_fp16)[name = string("k_head_551_cast_fp16")];
            tensor<int32, [1]> v_head_551_axes_1 = const()[name = string("v_head_551_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_551_cast_fp16 = expand_dims(axes = v_head_551_axes_1, x = v_head_549_cast_fp16)[name = string("v_head_551_cast_fp16")];
            tensor<int32, [1]> p_head_551_axes_1 = const()[name = string("p_head_551_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_551_cast_fp16 = expand_dims(axes = p_head_551_axes_1, x = p_head_549_cast_fp16)[name = string("p_head_551_cast_fp16")];
            bool var_22135_transpose_x_3 = const()[name = string("op_22135_transpose_x_3"), val = bool(false)];
            bool var_22135_transpose_y_3 = const()[name = string("op_22135_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_22135_cast_fp16 = matmul(transpose_x = var_22135_transpose_x_3, transpose_y = var_22135_transpose_y_3, x = q_u_275_cast_fp16, y = k_head_551_cast_fp16)[name = string("op_22135_cast_fp16")];
            fp16 var_22136_to_fp16 = const()[name = string("op_22136_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_275_cast_fp16 = mul(x = var_22135_cast_fp16, y = var_22136_to_fp16)[name = string("scores_content_275_cast_fp16")];
            bool x_1457_transpose_x_3 = const()[name = string("x_1457_transpose_x_3"), val = bool(false)];
            bool x_1457_transpose_y_3 = const()[name = string("x_1457_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1457_cast_fp16 = matmul(transpose_x = x_1457_transpose_x_3, transpose_y = x_1457_transpose_y_3, x = q_v_275_cast_fp16, y = p_head_551_cast_fp16)[name = string("x_1457_cast_fp16")];
            tensor<int32, [8]> x_1459_pad_1 = const()[name = string("x_1459_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1459_mode_1 = const()[name = string("x_1459_mode_1"), val = string("constant")];
            fp16 const_2233_to_fp16 = const()[name = string("const_2233_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1459_cast_fp16 = pad(constant_val = const_2233_to_fp16, mode = x_1459_mode_1, pad = x_1459_pad_1, x = x_1457_cast_fp16)[name = string("x_1459_cast_fp16")];
            tensor<int32, [4]> var_22150 = const()[name = string("op_22150"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1461_cast_fp16 = reshape(shape = var_22150, x = x_1459_cast_fp16)[name = string("x_1461_cast_fp16")];
            tensor<int32, [4]> var_22154_begin_1 = const()[name = string("op_22154_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_22154_end_1 = const()[name = string("op_22154_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_22154_end_mask_1 = const()[name = string("op_22154_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_22154_cast_fp16 = slice_by_index(begin = var_22154_begin_1, end = var_22154_end_1, end_mask = var_22154_end_mask_1, x = x_1461_cast_fp16)[name = string("op_22154_cast_fp16")];
            tensor<int32, [4]> var_22156 = const()[name = string("op_22156"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_22157_cast_fp16 = reshape(shape = var_22156, x = var_22154_cast_fp16)[name = string("op_22157_cast_fp16")];
            tensor<int32, [4]> var_22162_begin_1 = const()[name = string("op_22162_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_22162_end_1 = const()[name = string("op_22162_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_22162_end_mask_1 = const()[name = string("op_22162_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_22162_cast_fp16 = slice_by_index(begin = var_22162_begin_1, end = var_22162_end_1, end_mask = var_22162_end_mask_1, x = var_22157_cast_fp16)[name = string("op_22162_cast_fp16")];
            fp16 var_22163_to_fp16 = const()[name = string("op_22163_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_275_cast_fp16 = mul(x = var_22162_cast_fp16, y = var_22163_to_fp16)[name = string("scores_pos_275_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_275_cast_fp16 = add(x = scores_content_275_cast_fp16, y = scores_pos_275_cast_fp16)[name = string("logits_275_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_22166_cast_fp16 = softmax(axis = var_21770, x = logits_275_cast_fp16)[name = string("op_22166_cast_fp16")];
            bool var_22168_transpose_x_1 = const()[name = string("op_22168_transpose_x_1"), val = bool(false)];
            bool var_22168_transpose_y_1 = const()[name = string("op_22168_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_22168_cast_fp16 = matmul(transpose_x = var_22168_transpose_x_1, transpose_y = var_22168_transpose_y_1, x = var_22166_cast_fp16, y = v_head_551_cast_fp16)[name = string("op_22168_cast_fp16")];
            tensor<int32, [1]> var_22169_axes_1 = const()[name = string("op_22169_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_22169_cast_fp16 = squeeze(axes = var_22169_axes_1, x = var_22168_cast_fp16)[name = string("op_22169_cast_fp16")];
            string dense_output_1393_pad_type_1 = const()[name = string("dense_output_1393_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1393_strides_1 = const()[name = string("dense_output_1393_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1393_pad_1 = const()[name = string("dense_output_1393_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1393_dilations_1 = const()[name = string("dense_output_1393_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1393_groups_1 = const()[name = string("dense_output_1393_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513111616))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513242752))))[name = string("layers_17_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1393_cast_fp16 = conv(dilations = dense_output_1393_dilations_1, groups = dense_output_1393_groups_1, pad = dense_output_1393_pad_1, pad_type = dense_output_1393_pad_type_1, strides = dense_output_1393_strides_1, weight = layers_17_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1393_cast_fp16")];
            string sparse_output_1393_pad_type_1 = const()[name = string("sparse_output_1393_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1393_strides_1 = const()[name = string("sparse_output_1393_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1393_pad_1 = const()[name = string("sparse_output_1393_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1393_dilations_1 = const()[name = string("sparse_output_1393_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1393_groups_1 = const()[name = string("sparse_output_1393_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513246016))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513243328))))[name = string("layers_17_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1393_cast_fp16 = conv(dilations = sparse_output_1393_dilations_1, groups = sparse_output_1393_groups_1, pad = sparse_output_1393_pad_1, pad_type = sparse_output_1393_pad_type_1, strides = sparse_output_1393_strides_1, weight = layers_17_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1393_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22184_cast_fp16 = add(x = dense_output_1393_cast_fp16, y = sparse_output_1393_cast_fp16)[name = string("op_22184_cast_fp16")];
            tensor<int32, [4]> var_22185 = const()[name = string("op_22185"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22187 = const()[name = string("op_22187"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22186_cast_fp16 = transpose(perm = var_22185, x = var_22184_cast_fp16)[name = string("transpose_275")];
            tensor<fp16, [1, 51, 128]> q_head_277_cast_fp16 = reshape(shape = var_22187, x = var_22186_cast_fp16)[name = string("q_head_277_cast_fp16")];
            string dense_output_1395_pad_type_1 = const()[name = string("dense_output_1395_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1395_strides_1 = const()[name = string("dense_output_1395_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1395_pad_1 = const()[name = string("dense_output_1395_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1395_dilations_1 = const()[name = string("dense_output_1395_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1395_groups_1 = const()[name = string("dense_output_1395_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513262464))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513393600))))[name = string("layers_17_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1395_cast_fp16 = conv(dilations = dense_output_1395_dilations_1, groups = dense_output_1395_groups_1, pad = dense_output_1395_pad_1, pad_type = dense_output_1395_pad_type_1, strides = dense_output_1395_strides_1, weight = layers_17_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1395_cast_fp16")];
            string sparse_output_1395_pad_type_1 = const()[name = string("sparse_output_1395_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1395_strides_1 = const()[name = string("sparse_output_1395_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1395_pad_1 = const()[name = string("sparse_output_1395_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1395_dilations_1 = const()[name = string("sparse_output_1395_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1395_groups_1 = const()[name = string("sparse_output_1395_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513396864))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513394176))))[name = string("layers_17_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1395_cast_fp16 = conv(dilations = sparse_output_1395_dilations_1, groups = sparse_output_1395_groups_1, pad = sparse_output_1395_pad_1, pad_type = sparse_output_1395_pad_type_1, strides = sparse_output_1395_strides_1, weight = layers_17_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1395_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22203_cast_fp16 = add(x = dense_output_1395_cast_fp16, y = sparse_output_1395_cast_fp16)[name = string("op_22203_cast_fp16")];
            tensor<int32, [4]> var_22204 = const()[name = string("op_22204"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22206 = const()[name = string("op_22206"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22205_cast_fp16 = transpose(perm = var_22204, x = var_22203_cast_fp16)[name = string("transpose_274")];
            tensor<fp16, [1, 51, 128]> k_head_553_cast_fp16 = reshape(shape = var_22206, x = var_22205_cast_fp16)[name = string("k_head_553_cast_fp16")];
            string dense_output_1397_pad_type_1 = const()[name = string("dense_output_1397_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1397_strides_1 = const()[name = string("dense_output_1397_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1397_pad_1 = const()[name = string("dense_output_1397_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1397_dilations_1 = const()[name = string("dense_output_1397_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1397_groups_1 = const()[name = string("dense_output_1397_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513413312))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513544448))))[name = string("layers_17_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1397_cast_fp16 = conv(dilations = dense_output_1397_dilations_1, groups = dense_output_1397_groups_1, pad = dense_output_1397_pad_1, pad_type = dense_output_1397_pad_type_1, strides = dense_output_1397_strides_1, weight = layers_17_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1397_cast_fp16")];
            string sparse_output_1397_pad_type_1 = const()[name = string("sparse_output_1397_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1397_strides_1 = const()[name = string("sparse_output_1397_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1397_pad_1 = const()[name = string("sparse_output_1397_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1397_dilations_1 = const()[name = string("sparse_output_1397_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1397_groups_1 = const()[name = string("sparse_output_1397_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513547712))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513545024))))[name = string("layers_17_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1397_cast_fp16 = conv(dilations = sparse_output_1397_dilations_1, groups = sparse_output_1397_groups_1, pad = sparse_output_1397_pad_1, pad_type = sparse_output_1397_pad_type_1, strides = sparse_output_1397_strides_1, weight = layers_17_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1397_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22222_cast_fp16 = add(x = dense_output_1397_cast_fp16, y = sparse_output_1397_cast_fp16)[name = string("op_22222_cast_fp16")];
            tensor<int32, [4]> var_22223 = const()[name = string("op_22223"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22225 = const()[name = string("op_22225"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22224_cast_fp16 = transpose(perm = var_22223, x = var_22222_cast_fp16)[name = string("transpose_273")];
            tensor<fp16, [1, 51, 128]> v_head_553_cast_fp16 = reshape(shape = var_22225, x = var_22224_cast_fp16)[name = string("v_head_553_cast_fp16")];
            string dense_output_1399_pad_type_1 = const()[name = string("dense_output_1399_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1399_strides_1 = const()[name = string("dense_output_1399_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1399_pad_1 = const()[name = string("dense_output_1399_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1399_dilations_1 = const()[name = string("dense_output_1399_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1399_groups_1 = const()[name = string("dense_output_1399_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513564160))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513695296))))[name = string("layers_17_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1399_cast_fp16 = conv(dilations = dense_output_1399_dilations_1, groups = dense_output_1399_groups_1, pad = dense_output_1399_pad_1, pad_type = dense_output_1399_pad_type_1, strides = dense_output_1399_strides_1, weight = layers_17_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1399_cast_fp16")];
            string sparse_output_1399_pad_type_1 = const()[name = string("sparse_output_1399_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1399_strides_1 = const()[name = string("sparse_output_1399_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1399_pad_1 = const()[name = string("sparse_output_1399_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1399_dilations_1 = const()[name = string("sparse_output_1399_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1399_groups_1 = const()[name = string("sparse_output_1399_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513698560))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513695872))))[name = string("layers_17_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1399_cast_fp16 = conv(dilations = sparse_output_1399_dilations_1, groups = sparse_output_1399_groups_1, pad = sparse_output_1399_pad_1, pad_type = sparse_output_1399_pad_type_1, strides = sparse_output_1399_strides_1, weight = layers_17_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1399_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_22241_cast_fp16 = add(x = dense_output_1399_cast_fp16, y = sparse_output_1399_cast_fp16)[name = string("op_22241_cast_fp16")];
            tensor<int32, [4]> var_22242 = const()[name = string("op_22242"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22244 = const()[name = string("op_22244"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_22243_cast_fp16 = transpose(perm = var_22242, x = var_22241_cast_fp16)[name = string("transpose_272")];
            tensor<fp16, [1, 101, 128]> p_head_553_cast_fp16 = reshape(shape = var_22244, x = var_22243_cast_fp16)[name = string("p_head_553_cast_fp16")];
            tensor<fp16, [128]> var_22246_to_fp16 = const()[name = string("op_22246_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513715008)))];
            tensor<fp16, [1, 51, 128]> var_22247_cast_fp16 = add(x = q_head_277_cast_fp16, y = var_22246_to_fp16)[name = string("op_22247_cast_fp16")];
            tensor<int32, [1]> q_u_277_axes_1 = const()[name = string("q_u_277_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_277_cast_fp16 = expand_dims(axes = q_u_277_axes_1, x = var_22247_cast_fp16)[name = string("q_u_277_cast_fp16")];
            tensor<fp16, [128]> var_22249_to_fp16 = const()[name = string("op_22249_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513715328)))];
            tensor<fp16, [1, 51, 128]> var_22250_cast_fp16 = add(x = q_head_277_cast_fp16, y = var_22249_to_fp16)[name = string("op_22250_cast_fp16")];
            tensor<int32, [1]> q_v_277_axes_1 = const()[name = string("q_v_277_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_277_cast_fp16 = expand_dims(axes = q_v_277_axes_1, x = var_22250_cast_fp16)[name = string("q_v_277_cast_fp16")];
            tensor<int32, [1]> k_head_555_axes_1 = const()[name = string("k_head_555_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_555_cast_fp16 = expand_dims(axes = k_head_555_axes_1, x = k_head_553_cast_fp16)[name = string("k_head_555_cast_fp16")];
            tensor<int32, [1]> v_head_555_axes_1 = const()[name = string("v_head_555_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_555_cast_fp16 = expand_dims(axes = v_head_555_axes_1, x = v_head_553_cast_fp16)[name = string("v_head_555_cast_fp16")];
            tensor<int32, [1]> p_head_555_axes_1 = const()[name = string("p_head_555_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_555_cast_fp16 = expand_dims(axes = p_head_555_axes_1, x = p_head_553_cast_fp16)[name = string("p_head_555_cast_fp16")];
            bool var_22256_transpose_x_3 = const()[name = string("op_22256_transpose_x_3"), val = bool(false)];
            bool var_22256_transpose_y_3 = const()[name = string("op_22256_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_22256_cast_fp16 = matmul(transpose_x = var_22256_transpose_x_3, transpose_y = var_22256_transpose_y_3, x = q_u_277_cast_fp16, y = k_head_555_cast_fp16)[name = string("op_22256_cast_fp16")];
            fp16 var_22257_to_fp16 = const()[name = string("op_22257_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_277_cast_fp16 = mul(x = var_22256_cast_fp16, y = var_22257_to_fp16)[name = string("scores_content_277_cast_fp16")];
            bool x_1465_transpose_x_3 = const()[name = string("x_1465_transpose_x_3"), val = bool(false)];
            bool x_1465_transpose_y_3 = const()[name = string("x_1465_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1465_cast_fp16 = matmul(transpose_x = x_1465_transpose_x_3, transpose_y = x_1465_transpose_y_3, x = q_v_277_cast_fp16, y = p_head_555_cast_fp16)[name = string("x_1465_cast_fp16")];
            tensor<int32, [8]> x_1467_pad_1 = const()[name = string("x_1467_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1467_mode_1 = const()[name = string("x_1467_mode_1"), val = string("constant")];
            fp16 const_2239_to_fp16 = const()[name = string("const_2239_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1467_cast_fp16 = pad(constant_val = const_2239_to_fp16, mode = x_1467_mode_1, pad = x_1467_pad_1, x = x_1465_cast_fp16)[name = string("x_1467_cast_fp16")];
            tensor<int32, [4]> var_22271 = const()[name = string("op_22271"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1469_cast_fp16 = reshape(shape = var_22271, x = x_1467_cast_fp16)[name = string("x_1469_cast_fp16")];
            tensor<int32, [4]> var_22275_begin_1 = const()[name = string("op_22275_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_22275_end_1 = const()[name = string("op_22275_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_22275_end_mask_1 = const()[name = string("op_22275_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_22275_cast_fp16 = slice_by_index(begin = var_22275_begin_1, end = var_22275_end_1, end_mask = var_22275_end_mask_1, x = x_1469_cast_fp16)[name = string("op_22275_cast_fp16")];
            tensor<int32, [4]> var_22277 = const()[name = string("op_22277"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_22278_cast_fp16 = reshape(shape = var_22277, x = var_22275_cast_fp16)[name = string("op_22278_cast_fp16")];
            tensor<int32, [4]> var_22283_begin_1 = const()[name = string("op_22283_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_22283_end_1 = const()[name = string("op_22283_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_22283_end_mask_1 = const()[name = string("op_22283_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_22283_cast_fp16 = slice_by_index(begin = var_22283_begin_1, end = var_22283_end_1, end_mask = var_22283_end_mask_1, x = var_22278_cast_fp16)[name = string("op_22283_cast_fp16")];
            fp16 var_22284_to_fp16 = const()[name = string("op_22284_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_277_cast_fp16 = mul(x = var_22283_cast_fp16, y = var_22284_to_fp16)[name = string("scores_pos_277_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_277_cast_fp16 = add(x = scores_content_277_cast_fp16, y = scores_pos_277_cast_fp16)[name = string("logits_277_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_22287_cast_fp16 = softmax(axis = var_21770, x = logits_277_cast_fp16)[name = string("op_22287_cast_fp16")];
            bool var_22289_transpose_x_1 = const()[name = string("op_22289_transpose_x_1"), val = bool(false)];
            bool var_22289_transpose_y_1 = const()[name = string("op_22289_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_22289_cast_fp16 = matmul(transpose_x = var_22289_transpose_x_1, transpose_y = var_22289_transpose_y_1, x = var_22287_cast_fp16, y = v_head_555_cast_fp16)[name = string("op_22289_cast_fp16")];
            tensor<int32, [1]> var_22290_axes_1 = const()[name = string("op_22290_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_22290_cast_fp16 = squeeze(axes = var_22290_axes_1, x = var_22289_cast_fp16)[name = string("op_22290_cast_fp16")];
            string dense_output_1401_pad_type_1 = const()[name = string("dense_output_1401_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1401_strides_1 = const()[name = string("dense_output_1401_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1401_pad_1 = const()[name = string("dense_output_1401_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1401_dilations_1 = const()[name = string("dense_output_1401_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1401_groups_1 = const()[name = string("dense_output_1401_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513715648))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513846784))))[name = string("layers_17_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1401_cast_fp16 = conv(dilations = dense_output_1401_dilations_1, groups = dense_output_1401_groups_1, pad = dense_output_1401_pad_1, pad_type = dense_output_1401_pad_type_1, strides = dense_output_1401_strides_1, weight = layers_17_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1401_cast_fp16")];
            string sparse_output_1401_pad_type_1 = const()[name = string("sparse_output_1401_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1401_strides_1 = const()[name = string("sparse_output_1401_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1401_pad_1 = const()[name = string("sparse_output_1401_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1401_dilations_1 = const()[name = string("sparse_output_1401_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1401_groups_1 = const()[name = string("sparse_output_1401_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513850048))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513847360))))[name = string("layers_17_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1401_cast_fp16 = conv(dilations = sparse_output_1401_dilations_1, groups = sparse_output_1401_groups_1, pad = sparse_output_1401_pad_1, pad_type = sparse_output_1401_pad_type_1, strides = sparse_output_1401_strides_1, weight = layers_17_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1401_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22305_cast_fp16 = add(x = dense_output_1401_cast_fp16, y = sparse_output_1401_cast_fp16)[name = string("op_22305_cast_fp16")];
            tensor<int32, [4]> var_22306 = const()[name = string("op_22306"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22308 = const()[name = string("op_22308"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22307_cast_fp16 = transpose(perm = var_22306, x = var_22305_cast_fp16)[name = string("transpose_271")];
            tensor<fp16, [1, 51, 128]> q_head_279_cast_fp16 = reshape(shape = var_22308, x = var_22307_cast_fp16)[name = string("q_head_279_cast_fp16")];
            string dense_output_1403_pad_type_1 = const()[name = string("dense_output_1403_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1403_strides_1 = const()[name = string("dense_output_1403_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1403_pad_1 = const()[name = string("dense_output_1403_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1403_dilations_1 = const()[name = string("dense_output_1403_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1403_groups_1 = const()[name = string("dense_output_1403_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513866496))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513997632))))[name = string("layers_17_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1403_cast_fp16 = conv(dilations = dense_output_1403_dilations_1, groups = dense_output_1403_groups_1, pad = dense_output_1403_pad_1, pad_type = dense_output_1403_pad_type_1, strides = dense_output_1403_strides_1, weight = layers_17_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1403_cast_fp16")];
            string sparse_output_1403_pad_type_1 = const()[name = string("sparse_output_1403_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1403_strides_1 = const()[name = string("sparse_output_1403_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1403_pad_1 = const()[name = string("sparse_output_1403_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1403_dilations_1 = const()[name = string("sparse_output_1403_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1403_groups_1 = const()[name = string("sparse_output_1403_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514000896))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513998208))))[name = string("layers_17_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1403_cast_fp16 = conv(dilations = sparse_output_1403_dilations_1, groups = sparse_output_1403_groups_1, pad = sparse_output_1403_pad_1, pad_type = sparse_output_1403_pad_type_1, strides = sparse_output_1403_strides_1, weight = layers_17_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1403_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22324_cast_fp16 = add(x = dense_output_1403_cast_fp16, y = sparse_output_1403_cast_fp16)[name = string("op_22324_cast_fp16")];
            tensor<int32, [4]> var_22325 = const()[name = string("op_22325"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22327 = const()[name = string("op_22327"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22326_cast_fp16 = transpose(perm = var_22325, x = var_22324_cast_fp16)[name = string("transpose_270")];
            tensor<fp16, [1, 51, 128]> k_head_557_cast_fp16 = reshape(shape = var_22327, x = var_22326_cast_fp16)[name = string("k_head_557_cast_fp16")];
            string dense_output_1405_pad_type_1 = const()[name = string("dense_output_1405_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1405_strides_1 = const()[name = string("dense_output_1405_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1405_pad_1 = const()[name = string("dense_output_1405_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1405_dilations_1 = const()[name = string("dense_output_1405_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1405_groups_1 = const()[name = string("dense_output_1405_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514017344))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514148480))))[name = string("layers_17_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1405_cast_fp16 = conv(dilations = dense_output_1405_dilations_1, groups = dense_output_1405_groups_1, pad = dense_output_1405_pad_1, pad_type = dense_output_1405_pad_type_1, strides = dense_output_1405_strides_1, weight = layers_17_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1405_cast_fp16")];
            string sparse_output_1405_pad_type_1 = const()[name = string("sparse_output_1405_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1405_strides_1 = const()[name = string("sparse_output_1405_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1405_pad_1 = const()[name = string("sparse_output_1405_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1405_dilations_1 = const()[name = string("sparse_output_1405_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1405_groups_1 = const()[name = string("sparse_output_1405_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514151744))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514149056))))[name = string("layers_17_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1405_cast_fp16 = conv(dilations = sparse_output_1405_dilations_1, groups = sparse_output_1405_groups_1, pad = sparse_output_1405_pad_1, pad_type = sparse_output_1405_pad_type_1, strides = sparse_output_1405_strides_1, weight = layers_17_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1405_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22343_cast_fp16 = add(x = dense_output_1405_cast_fp16, y = sparse_output_1405_cast_fp16)[name = string("op_22343_cast_fp16")];
            tensor<int32, [4]> var_22344 = const()[name = string("op_22344"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22346 = const()[name = string("op_22346"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22345_cast_fp16 = transpose(perm = var_22344, x = var_22343_cast_fp16)[name = string("transpose_269")];
            tensor<fp16, [1, 51, 128]> v_head_557_cast_fp16 = reshape(shape = var_22346, x = var_22345_cast_fp16)[name = string("v_head_557_cast_fp16")];
            string dense_output_1407_pad_type_1 = const()[name = string("dense_output_1407_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1407_strides_1 = const()[name = string("dense_output_1407_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1407_pad_1 = const()[name = string("dense_output_1407_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1407_dilations_1 = const()[name = string("dense_output_1407_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1407_groups_1 = const()[name = string("dense_output_1407_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514168192))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514299328))))[name = string("layers_17_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1407_cast_fp16 = conv(dilations = dense_output_1407_dilations_1, groups = dense_output_1407_groups_1, pad = dense_output_1407_pad_1, pad_type = dense_output_1407_pad_type_1, strides = dense_output_1407_strides_1, weight = layers_17_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1407_cast_fp16")];
            string sparse_output_1407_pad_type_1 = const()[name = string("sparse_output_1407_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1407_strides_1 = const()[name = string("sparse_output_1407_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1407_pad_1 = const()[name = string("sparse_output_1407_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1407_dilations_1 = const()[name = string("sparse_output_1407_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1407_groups_1 = const()[name = string("sparse_output_1407_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514302592))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514299904))))[name = string("layers_17_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1407_cast_fp16 = conv(dilations = sparse_output_1407_dilations_1, groups = sparse_output_1407_groups_1, pad = sparse_output_1407_pad_1, pad_type = sparse_output_1407_pad_type_1, strides = sparse_output_1407_strides_1, weight = layers_17_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1407_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_22362_cast_fp16 = add(x = dense_output_1407_cast_fp16, y = sparse_output_1407_cast_fp16)[name = string("op_22362_cast_fp16")];
            tensor<int32, [4]> var_22363 = const()[name = string("op_22363"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22365 = const()[name = string("op_22365"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_22364_cast_fp16 = transpose(perm = var_22363, x = var_22362_cast_fp16)[name = string("transpose_268")];
            tensor<fp16, [1, 101, 128]> p_head_557_cast_fp16 = reshape(shape = var_22365, x = var_22364_cast_fp16)[name = string("p_head_557_cast_fp16")];
            tensor<fp16, [128]> var_22367_to_fp16 = const()[name = string("op_22367_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514319040)))];
            tensor<fp16, [1, 51, 128]> var_22368_cast_fp16 = add(x = q_head_279_cast_fp16, y = var_22367_to_fp16)[name = string("op_22368_cast_fp16")];
            tensor<int32, [1]> q_u_279_axes_1 = const()[name = string("q_u_279_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_279_cast_fp16 = expand_dims(axes = q_u_279_axes_1, x = var_22368_cast_fp16)[name = string("q_u_279_cast_fp16")];
            tensor<fp16, [128]> var_22370_to_fp16 = const()[name = string("op_22370_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514319360)))];
            tensor<fp16, [1, 51, 128]> var_22371_cast_fp16 = add(x = q_head_279_cast_fp16, y = var_22370_to_fp16)[name = string("op_22371_cast_fp16")];
            tensor<int32, [1]> q_v_279_axes_1 = const()[name = string("q_v_279_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_279_cast_fp16 = expand_dims(axes = q_v_279_axes_1, x = var_22371_cast_fp16)[name = string("q_v_279_cast_fp16")];
            tensor<int32, [1]> k_head_559_axes_1 = const()[name = string("k_head_559_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_559_cast_fp16 = expand_dims(axes = k_head_559_axes_1, x = k_head_557_cast_fp16)[name = string("k_head_559_cast_fp16")];
            tensor<int32, [1]> v_head_559_axes_1 = const()[name = string("v_head_559_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_559_cast_fp16 = expand_dims(axes = v_head_559_axes_1, x = v_head_557_cast_fp16)[name = string("v_head_559_cast_fp16")];
            tensor<int32, [1]> p_head_559_axes_1 = const()[name = string("p_head_559_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_559_cast_fp16 = expand_dims(axes = p_head_559_axes_1, x = p_head_557_cast_fp16)[name = string("p_head_559_cast_fp16")];
            bool var_22377_transpose_x_3 = const()[name = string("op_22377_transpose_x_3"), val = bool(false)];
            bool var_22377_transpose_y_3 = const()[name = string("op_22377_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_22377_cast_fp16 = matmul(transpose_x = var_22377_transpose_x_3, transpose_y = var_22377_transpose_y_3, x = q_u_279_cast_fp16, y = k_head_559_cast_fp16)[name = string("op_22377_cast_fp16")];
            fp16 var_22378_to_fp16 = const()[name = string("op_22378_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_279_cast_fp16 = mul(x = var_22377_cast_fp16, y = var_22378_to_fp16)[name = string("scores_content_279_cast_fp16")];
            bool x_1473_transpose_x_3 = const()[name = string("x_1473_transpose_x_3"), val = bool(false)];
            bool x_1473_transpose_y_3 = const()[name = string("x_1473_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1473_cast_fp16 = matmul(transpose_x = x_1473_transpose_x_3, transpose_y = x_1473_transpose_y_3, x = q_v_279_cast_fp16, y = p_head_559_cast_fp16)[name = string("x_1473_cast_fp16")];
            tensor<int32, [8]> x_1475_pad_1 = const()[name = string("x_1475_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1475_mode_1 = const()[name = string("x_1475_mode_1"), val = string("constant")];
            fp16 const_2245_to_fp16 = const()[name = string("const_2245_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1475_cast_fp16 = pad(constant_val = const_2245_to_fp16, mode = x_1475_mode_1, pad = x_1475_pad_1, x = x_1473_cast_fp16)[name = string("x_1475_cast_fp16")];
            tensor<int32, [4]> var_22392 = const()[name = string("op_22392"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1477_cast_fp16 = reshape(shape = var_22392, x = x_1475_cast_fp16)[name = string("x_1477_cast_fp16")];
            tensor<int32, [4]> var_22396_begin_1 = const()[name = string("op_22396_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_22396_end_1 = const()[name = string("op_22396_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_22396_end_mask_1 = const()[name = string("op_22396_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_22396_cast_fp16 = slice_by_index(begin = var_22396_begin_1, end = var_22396_end_1, end_mask = var_22396_end_mask_1, x = x_1477_cast_fp16)[name = string("op_22396_cast_fp16")];
            tensor<int32, [4]> var_22398 = const()[name = string("op_22398"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_22399_cast_fp16 = reshape(shape = var_22398, x = var_22396_cast_fp16)[name = string("op_22399_cast_fp16")];
            tensor<int32, [4]> var_22404_begin_1 = const()[name = string("op_22404_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_22404_end_1 = const()[name = string("op_22404_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_22404_end_mask_1 = const()[name = string("op_22404_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_22404_cast_fp16 = slice_by_index(begin = var_22404_begin_1, end = var_22404_end_1, end_mask = var_22404_end_mask_1, x = var_22399_cast_fp16)[name = string("op_22404_cast_fp16")];
            fp16 var_22405_to_fp16 = const()[name = string("op_22405_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_279_cast_fp16 = mul(x = var_22404_cast_fp16, y = var_22405_to_fp16)[name = string("scores_pos_279_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_279_cast_fp16 = add(x = scores_content_279_cast_fp16, y = scores_pos_279_cast_fp16)[name = string("logits_279_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_22408_cast_fp16 = softmax(axis = var_21770, x = logits_279_cast_fp16)[name = string("op_22408_cast_fp16")];
            bool var_22410_transpose_x_1 = const()[name = string("op_22410_transpose_x_1"), val = bool(false)];
            bool var_22410_transpose_y_1 = const()[name = string("op_22410_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_22410_cast_fp16 = matmul(transpose_x = var_22410_transpose_x_1, transpose_y = var_22410_transpose_y_1, x = var_22408_cast_fp16, y = v_head_559_cast_fp16)[name = string("op_22410_cast_fp16")];
            tensor<int32, [1]> var_22411_axes_1 = const()[name = string("op_22411_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_22411_cast_fp16 = squeeze(axes = var_22411_axes_1, x = var_22410_cast_fp16)[name = string("op_22411_cast_fp16")];
            string dense_output_1409_pad_type_1 = const()[name = string("dense_output_1409_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1409_strides_1 = const()[name = string("dense_output_1409_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1409_pad_1 = const()[name = string("dense_output_1409_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1409_dilations_1 = const()[name = string("dense_output_1409_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1409_groups_1 = const()[name = string("dense_output_1409_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514319680))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514450816))))[name = string("layers_17_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1409_cast_fp16 = conv(dilations = dense_output_1409_dilations_1, groups = dense_output_1409_groups_1, pad = dense_output_1409_pad_1, pad_type = dense_output_1409_pad_type_1, strides = dense_output_1409_strides_1, weight = layers_17_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1409_cast_fp16")];
            string sparse_output_1409_pad_type_1 = const()[name = string("sparse_output_1409_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1409_strides_1 = const()[name = string("sparse_output_1409_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1409_pad_1 = const()[name = string("sparse_output_1409_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1409_dilations_1 = const()[name = string("sparse_output_1409_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1409_groups_1 = const()[name = string("sparse_output_1409_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514454080))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514451392))))[name = string("layers_17_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1409_cast_fp16 = conv(dilations = sparse_output_1409_dilations_1, groups = sparse_output_1409_groups_1, pad = sparse_output_1409_pad_1, pad_type = sparse_output_1409_pad_type_1, strides = sparse_output_1409_strides_1, weight = layers_17_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1409_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22426_cast_fp16 = add(x = dense_output_1409_cast_fp16, y = sparse_output_1409_cast_fp16)[name = string("op_22426_cast_fp16")];
            tensor<int32, [4]> var_22427 = const()[name = string("op_22427"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22429 = const()[name = string("op_22429"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22428_cast_fp16 = transpose(perm = var_22427, x = var_22426_cast_fp16)[name = string("transpose_267")];
            tensor<fp16, [1, 51, 128]> q_head_281_cast_fp16 = reshape(shape = var_22429, x = var_22428_cast_fp16)[name = string("q_head_281_cast_fp16")];
            string dense_output_1411_pad_type_1 = const()[name = string("dense_output_1411_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1411_strides_1 = const()[name = string("dense_output_1411_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1411_pad_1 = const()[name = string("dense_output_1411_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1411_dilations_1 = const()[name = string("dense_output_1411_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1411_groups_1 = const()[name = string("dense_output_1411_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514470528))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514601664))))[name = string("layers_17_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1411_cast_fp16 = conv(dilations = dense_output_1411_dilations_1, groups = dense_output_1411_groups_1, pad = dense_output_1411_pad_1, pad_type = dense_output_1411_pad_type_1, strides = dense_output_1411_strides_1, weight = layers_17_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1411_cast_fp16")];
            string sparse_output_1411_pad_type_1 = const()[name = string("sparse_output_1411_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1411_strides_1 = const()[name = string("sparse_output_1411_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1411_pad_1 = const()[name = string("sparse_output_1411_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1411_dilations_1 = const()[name = string("sparse_output_1411_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1411_groups_1 = const()[name = string("sparse_output_1411_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514604928))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514602240))))[name = string("layers_17_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1411_cast_fp16 = conv(dilations = sparse_output_1411_dilations_1, groups = sparse_output_1411_groups_1, pad = sparse_output_1411_pad_1, pad_type = sparse_output_1411_pad_type_1, strides = sparse_output_1411_strides_1, weight = layers_17_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1411_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22445_cast_fp16 = add(x = dense_output_1411_cast_fp16, y = sparse_output_1411_cast_fp16)[name = string("op_22445_cast_fp16")];
            tensor<int32, [4]> var_22446 = const()[name = string("op_22446"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22448 = const()[name = string("op_22448"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22447_cast_fp16 = transpose(perm = var_22446, x = var_22445_cast_fp16)[name = string("transpose_266")];
            tensor<fp16, [1, 51, 128]> k_head_561_cast_fp16 = reshape(shape = var_22448, x = var_22447_cast_fp16)[name = string("k_head_561_cast_fp16")];
            string dense_output_1413_pad_type_1 = const()[name = string("dense_output_1413_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1413_strides_1 = const()[name = string("dense_output_1413_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1413_pad_1 = const()[name = string("dense_output_1413_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1413_dilations_1 = const()[name = string("dense_output_1413_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1413_groups_1 = const()[name = string("dense_output_1413_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514621376))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514752512))))[name = string("layers_17_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1413_cast_fp16 = conv(dilations = dense_output_1413_dilations_1, groups = dense_output_1413_groups_1, pad = dense_output_1413_pad_1, pad_type = dense_output_1413_pad_type_1, strides = dense_output_1413_strides_1, weight = layers_17_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1413_cast_fp16")];
            string sparse_output_1413_pad_type_1 = const()[name = string("sparse_output_1413_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1413_strides_1 = const()[name = string("sparse_output_1413_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1413_pad_1 = const()[name = string("sparse_output_1413_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1413_dilations_1 = const()[name = string("sparse_output_1413_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1413_groups_1 = const()[name = string("sparse_output_1413_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514755776))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514753088))))[name = string("layers_17_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1413_cast_fp16 = conv(dilations = sparse_output_1413_dilations_1, groups = sparse_output_1413_groups_1, pad = sparse_output_1413_pad_1, pad_type = sparse_output_1413_pad_type_1, strides = sparse_output_1413_strides_1, weight = layers_17_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1413_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22464_cast_fp16 = add(x = dense_output_1413_cast_fp16, y = sparse_output_1413_cast_fp16)[name = string("op_22464_cast_fp16")];
            tensor<int32, [4]> var_22465 = const()[name = string("op_22465"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22467 = const()[name = string("op_22467"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22466_cast_fp16 = transpose(perm = var_22465, x = var_22464_cast_fp16)[name = string("transpose_265")];
            tensor<fp16, [1, 51, 128]> v_head_561_cast_fp16 = reshape(shape = var_22467, x = var_22466_cast_fp16)[name = string("v_head_561_cast_fp16")];
            string dense_output_1415_pad_type_1 = const()[name = string("dense_output_1415_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1415_strides_1 = const()[name = string("dense_output_1415_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1415_pad_1 = const()[name = string("dense_output_1415_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1415_dilations_1 = const()[name = string("dense_output_1415_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1415_groups_1 = const()[name = string("dense_output_1415_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514772224))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514903360))))[name = string("layers_17_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1415_cast_fp16 = conv(dilations = dense_output_1415_dilations_1, groups = dense_output_1415_groups_1, pad = dense_output_1415_pad_1, pad_type = dense_output_1415_pad_type_1, strides = dense_output_1415_strides_1, weight = layers_17_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1415_cast_fp16")];
            string sparse_output_1415_pad_type_1 = const()[name = string("sparse_output_1415_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1415_strides_1 = const()[name = string("sparse_output_1415_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1415_pad_1 = const()[name = string("sparse_output_1415_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1415_dilations_1 = const()[name = string("sparse_output_1415_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1415_groups_1 = const()[name = string("sparse_output_1415_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514906624))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514903936))))[name = string("layers_17_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1415_cast_fp16 = conv(dilations = sparse_output_1415_dilations_1, groups = sparse_output_1415_groups_1, pad = sparse_output_1415_pad_1, pad_type = sparse_output_1415_pad_type_1, strides = sparse_output_1415_strides_1, weight = layers_17_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1415_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_22483_cast_fp16 = add(x = dense_output_1415_cast_fp16, y = sparse_output_1415_cast_fp16)[name = string("op_22483_cast_fp16")];
            tensor<int32, [4]> var_22484 = const()[name = string("op_22484"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22486 = const()[name = string("op_22486"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_22485_cast_fp16 = transpose(perm = var_22484, x = var_22483_cast_fp16)[name = string("transpose_264")];
            tensor<fp16, [1, 101, 128]> p_head_561_cast_fp16 = reshape(shape = var_22486, x = var_22485_cast_fp16)[name = string("p_head_561_cast_fp16")];
            tensor<fp16, [128]> var_22488_to_fp16 = const()[name = string("op_22488_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514923072)))];
            tensor<fp16, [1, 51, 128]> var_22489_cast_fp16 = add(x = q_head_281_cast_fp16, y = var_22488_to_fp16)[name = string("op_22489_cast_fp16")];
            tensor<int32, [1]> q_u_281_axes_1 = const()[name = string("q_u_281_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_281_cast_fp16 = expand_dims(axes = q_u_281_axes_1, x = var_22489_cast_fp16)[name = string("q_u_281_cast_fp16")];
            tensor<fp16, [128]> var_22491_to_fp16 = const()[name = string("op_22491_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514923392)))];
            tensor<fp16, [1, 51, 128]> var_22492_cast_fp16 = add(x = q_head_281_cast_fp16, y = var_22491_to_fp16)[name = string("op_22492_cast_fp16")];
            tensor<int32, [1]> q_v_281_axes_1 = const()[name = string("q_v_281_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_281_cast_fp16 = expand_dims(axes = q_v_281_axes_1, x = var_22492_cast_fp16)[name = string("q_v_281_cast_fp16")];
            tensor<int32, [1]> k_head_563_axes_1 = const()[name = string("k_head_563_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_563_cast_fp16 = expand_dims(axes = k_head_563_axes_1, x = k_head_561_cast_fp16)[name = string("k_head_563_cast_fp16")];
            tensor<int32, [1]> v_head_563_axes_1 = const()[name = string("v_head_563_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_563_cast_fp16 = expand_dims(axes = v_head_563_axes_1, x = v_head_561_cast_fp16)[name = string("v_head_563_cast_fp16")];
            tensor<int32, [1]> p_head_563_axes_1 = const()[name = string("p_head_563_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_563_cast_fp16 = expand_dims(axes = p_head_563_axes_1, x = p_head_561_cast_fp16)[name = string("p_head_563_cast_fp16")];
            bool var_22498_transpose_x_3 = const()[name = string("op_22498_transpose_x_3"), val = bool(false)];
            bool var_22498_transpose_y_3 = const()[name = string("op_22498_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_22498_cast_fp16 = matmul(transpose_x = var_22498_transpose_x_3, transpose_y = var_22498_transpose_y_3, x = q_u_281_cast_fp16, y = k_head_563_cast_fp16)[name = string("op_22498_cast_fp16")];
            fp16 var_22499_to_fp16 = const()[name = string("op_22499_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_281_cast_fp16 = mul(x = var_22498_cast_fp16, y = var_22499_to_fp16)[name = string("scores_content_281_cast_fp16")];
            bool x_1481_transpose_x_3 = const()[name = string("x_1481_transpose_x_3"), val = bool(false)];
            bool x_1481_transpose_y_3 = const()[name = string("x_1481_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1481_cast_fp16 = matmul(transpose_x = x_1481_transpose_x_3, transpose_y = x_1481_transpose_y_3, x = q_v_281_cast_fp16, y = p_head_563_cast_fp16)[name = string("x_1481_cast_fp16")];
            tensor<int32, [8]> x_1483_pad_1 = const()[name = string("x_1483_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1483_mode_1 = const()[name = string("x_1483_mode_1"), val = string("constant")];
            fp16 const_2251_to_fp16 = const()[name = string("const_2251_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1483_cast_fp16 = pad(constant_val = const_2251_to_fp16, mode = x_1483_mode_1, pad = x_1483_pad_1, x = x_1481_cast_fp16)[name = string("x_1483_cast_fp16")];
            tensor<int32, [4]> var_22513 = const()[name = string("op_22513"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1485_cast_fp16 = reshape(shape = var_22513, x = x_1483_cast_fp16)[name = string("x_1485_cast_fp16")];
            tensor<int32, [4]> var_22517_begin_1 = const()[name = string("op_22517_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_22517_end_1 = const()[name = string("op_22517_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_22517_end_mask_1 = const()[name = string("op_22517_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_22517_cast_fp16 = slice_by_index(begin = var_22517_begin_1, end = var_22517_end_1, end_mask = var_22517_end_mask_1, x = x_1485_cast_fp16)[name = string("op_22517_cast_fp16")];
            tensor<int32, [4]> var_22519 = const()[name = string("op_22519"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_22520_cast_fp16 = reshape(shape = var_22519, x = var_22517_cast_fp16)[name = string("op_22520_cast_fp16")];
            tensor<int32, [4]> var_22525_begin_1 = const()[name = string("op_22525_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_22525_end_1 = const()[name = string("op_22525_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_22525_end_mask_1 = const()[name = string("op_22525_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_22525_cast_fp16 = slice_by_index(begin = var_22525_begin_1, end = var_22525_end_1, end_mask = var_22525_end_mask_1, x = var_22520_cast_fp16)[name = string("op_22525_cast_fp16")];
            fp16 var_22526_to_fp16 = const()[name = string("op_22526_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_281_cast_fp16 = mul(x = var_22525_cast_fp16, y = var_22526_to_fp16)[name = string("scores_pos_281_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_281_cast_fp16 = add(x = scores_content_281_cast_fp16, y = scores_pos_281_cast_fp16)[name = string("logits_281_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_22529_cast_fp16 = softmax(axis = var_21770, x = logits_281_cast_fp16)[name = string("op_22529_cast_fp16")];
            bool var_22531_transpose_x_1 = const()[name = string("op_22531_transpose_x_1"), val = bool(false)];
            bool var_22531_transpose_y_1 = const()[name = string("op_22531_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_22531_cast_fp16 = matmul(transpose_x = var_22531_transpose_x_1, transpose_y = var_22531_transpose_y_1, x = var_22529_cast_fp16, y = v_head_563_cast_fp16)[name = string("op_22531_cast_fp16")];
            tensor<int32, [1]> var_22532_axes_1 = const()[name = string("op_22532_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_22532_cast_fp16 = squeeze(axes = var_22532_axes_1, x = var_22531_cast_fp16)[name = string("op_22532_cast_fp16")];
            string dense_output_1417_pad_type_1 = const()[name = string("dense_output_1417_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1417_strides_1 = const()[name = string("dense_output_1417_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1417_pad_1 = const()[name = string("dense_output_1417_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1417_dilations_1 = const()[name = string("dense_output_1417_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1417_groups_1 = const()[name = string("dense_output_1417_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(514923712))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515054848))))[name = string("layers_17_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1417_cast_fp16 = conv(dilations = dense_output_1417_dilations_1, groups = dense_output_1417_groups_1, pad = dense_output_1417_pad_1, pad_type = dense_output_1417_pad_type_1, strides = dense_output_1417_strides_1, weight = layers_17_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1417_cast_fp16")];
            string sparse_output_1417_pad_type_1 = const()[name = string("sparse_output_1417_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1417_strides_1 = const()[name = string("sparse_output_1417_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1417_pad_1 = const()[name = string("sparse_output_1417_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1417_dilations_1 = const()[name = string("sparse_output_1417_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1417_groups_1 = const()[name = string("sparse_output_1417_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515058112))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515055424))))[name = string("layers_17_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1417_cast_fp16 = conv(dilations = sparse_output_1417_dilations_1, groups = sparse_output_1417_groups_1, pad = sparse_output_1417_pad_1, pad_type = sparse_output_1417_pad_type_1, strides = sparse_output_1417_strides_1, weight = layers_17_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1417_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22547_cast_fp16 = add(x = dense_output_1417_cast_fp16, y = sparse_output_1417_cast_fp16)[name = string("op_22547_cast_fp16")];
            tensor<int32, [4]> var_22548 = const()[name = string("op_22548"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22550 = const()[name = string("op_22550"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22549_cast_fp16 = transpose(perm = var_22548, x = var_22547_cast_fp16)[name = string("transpose_263")];
            tensor<fp16, [1, 51, 128]> q_head_283_cast_fp16 = reshape(shape = var_22550, x = var_22549_cast_fp16)[name = string("q_head_283_cast_fp16")];
            string dense_output_1419_pad_type_1 = const()[name = string("dense_output_1419_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1419_strides_1 = const()[name = string("dense_output_1419_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1419_pad_1 = const()[name = string("dense_output_1419_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1419_dilations_1 = const()[name = string("dense_output_1419_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1419_groups_1 = const()[name = string("dense_output_1419_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515074560))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515205696))))[name = string("layers_17_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1419_cast_fp16 = conv(dilations = dense_output_1419_dilations_1, groups = dense_output_1419_groups_1, pad = dense_output_1419_pad_1, pad_type = dense_output_1419_pad_type_1, strides = dense_output_1419_strides_1, weight = layers_17_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1419_cast_fp16")];
            string sparse_output_1419_pad_type_1 = const()[name = string("sparse_output_1419_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1419_strides_1 = const()[name = string("sparse_output_1419_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1419_pad_1 = const()[name = string("sparse_output_1419_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1419_dilations_1 = const()[name = string("sparse_output_1419_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1419_groups_1 = const()[name = string("sparse_output_1419_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515208960))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515206272))))[name = string("layers_17_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1419_cast_fp16 = conv(dilations = sparse_output_1419_dilations_1, groups = sparse_output_1419_groups_1, pad = sparse_output_1419_pad_1, pad_type = sparse_output_1419_pad_type_1, strides = sparse_output_1419_strides_1, weight = layers_17_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1419_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22566_cast_fp16 = add(x = dense_output_1419_cast_fp16, y = sparse_output_1419_cast_fp16)[name = string("op_22566_cast_fp16")];
            tensor<int32, [4]> var_22567 = const()[name = string("op_22567"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22569 = const()[name = string("op_22569"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22568_cast_fp16 = transpose(perm = var_22567, x = var_22566_cast_fp16)[name = string("transpose_262")];
            tensor<fp16, [1, 51, 128]> k_head_565_cast_fp16 = reshape(shape = var_22569, x = var_22568_cast_fp16)[name = string("k_head_565_cast_fp16")];
            string dense_output_1421_pad_type_1 = const()[name = string("dense_output_1421_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1421_strides_1 = const()[name = string("dense_output_1421_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1421_pad_1 = const()[name = string("dense_output_1421_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1421_dilations_1 = const()[name = string("dense_output_1421_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1421_groups_1 = const()[name = string("dense_output_1421_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515225408))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515356544))))[name = string("layers_17_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1421_cast_fp16 = conv(dilations = dense_output_1421_dilations_1, groups = dense_output_1421_groups_1, pad = dense_output_1421_pad_1, pad_type = dense_output_1421_pad_type_1, strides = dense_output_1421_strides_1, weight = layers_17_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1421_cast_fp16")];
            string sparse_output_1421_pad_type_1 = const()[name = string("sparse_output_1421_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1421_strides_1 = const()[name = string("sparse_output_1421_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1421_pad_1 = const()[name = string("sparse_output_1421_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1421_dilations_1 = const()[name = string("sparse_output_1421_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1421_groups_1 = const()[name = string("sparse_output_1421_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515359808))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515357120))))[name = string("layers_17_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1421_cast_fp16 = conv(dilations = sparse_output_1421_dilations_1, groups = sparse_output_1421_groups_1, pad = sparse_output_1421_pad_1, pad_type = sparse_output_1421_pad_type_1, strides = sparse_output_1421_strides_1, weight = layers_17_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1421_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22585_cast_fp16 = add(x = dense_output_1421_cast_fp16, y = sparse_output_1421_cast_fp16)[name = string("op_22585_cast_fp16")];
            tensor<int32, [4]> var_22586 = const()[name = string("op_22586"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22588 = const()[name = string("op_22588"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22587_cast_fp16 = transpose(perm = var_22586, x = var_22585_cast_fp16)[name = string("transpose_261")];
            tensor<fp16, [1, 51, 128]> v_head_565_cast_fp16 = reshape(shape = var_22588, x = var_22587_cast_fp16)[name = string("v_head_565_cast_fp16")];
            string dense_output_1423_pad_type_1 = const()[name = string("dense_output_1423_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1423_strides_1 = const()[name = string("dense_output_1423_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1423_pad_1 = const()[name = string("dense_output_1423_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1423_dilations_1 = const()[name = string("dense_output_1423_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1423_groups_1 = const()[name = string("dense_output_1423_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515376256))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515507392))))[name = string("layers_17_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1423_cast_fp16 = conv(dilations = dense_output_1423_dilations_1, groups = dense_output_1423_groups_1, pad = dense_output_1423_pad_1, pad_type = dense_output_1423_pad_type_1, strides = dense_output_1423_strides_1, weight = layers_17_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1423_cast_fp16")];
            string sparse_output_1423_pad_type_1 = const()[name = string("sparse_output_1423_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1423_strides_1 = const()[name = string("sparse_output_1423_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1423_pad_1 = const()[name = string("sparse_output_1423_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1423_dilations_1 = const()[name = string("sparse_output_1423_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1423_groups_1 = const()[name = string("sparse_output_1423_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515510656))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515507968))))[name = string("layers_17_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1423_cast_fp16 = conv(dilations = sparse_output_1423_dilations_1, groups = sparse_output_1423_groups_1, pad = sparse_output_1423_pad_1, pad_type = sparse_output_1423_pad_type_1, strides = sparse_output_1423_strides_1, weight = layers_17_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1423_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_22604_cast_fp16 = add(x = dense_output_1423_cast_fp16, y = sparse_output_1423_cast_fp16)[name = string("op_22604_cast_fp16")];
            tensor<int32, [4]> var_22605 = const()[name = string("op_22605"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22607 = const()[name = string("op_22607"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_22606_cast_fp16 = transpose(perm = var_22605, x = var_22604_cast_fp16)[name = string("transpose_260")];
            tensor<fp16, [1, 101, 128]> p_head_565_cast_fp16 = reshape(shape = var_22607, x = var_22606_cast_fp16)[name = string("p_head_565_cast_fp16")];
            tensor<fp16, [128]> var_22609_to_fp16 = const()[name = string("op_22609_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515527104)))];
            tensor<fp16, [1, 51, 128]> var_22610_cast_fp16 = add(x = q_head_283_cast_fp16, y = var_22609_to_fp16)[name = string("op_22610_cast_fp16")];
            tensor<int32, [1]> q_u_283_axes_1 = const()[name = string("q_u_283_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_283_cast_fp16 = expand_dims(axes = q_u_283_axes_1, x = var_22610_cast_fp16)[name = string("q_u_283_cast_fp16")];
            tensor<fp16, [128]> var_22612_to_fp16 = const()[name = string("op_22612_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515527424)))];
            tensor<fp16, [1, 51, 128]> var_22613_cast_fp16 = add(x = q_head_283_cast_fp16, y = var_22612_to_fp16)[name = string("op_22613_cast_fp16")];
            tensor<int32, [1]> q_v_283_axes_1 = const()[name = string("q_v_283_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_283_cast_fp16 = expand_dims(axes = q_v_283_axes_1, x = var_22613_cast_fp16)[name = string("q_v_283_cast_fp16")];
            tensor<int32, [1]> k_head_567_axes_1 = const()[name = string("k_head_567_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_567_cast_fp16 = expand_dims(axes = k_head_567_axes_1, x = k_head_565_cast_fp16)[name = string("k_head_567_cast_fp16")];
            tensor<int32, [1]> v_head_567_axes_1 = const()[name = string("v_head_567_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_567_cast_fp16 = expand_dims(axes = v_head_567_axes_1, x = v_head_565_cast_fp16)[name = string("v_head_567_cast_fp16")];
            tensor<int32, [1]> p_head_567_axes_1 = const()[name = string("p_head_567_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_567_cast_fp16 = expand_dims(axes = p_head_567_axes_1, x = p_head_565_cast_fp16)[name = string("p_head_567_cast_fp16")];
            bool var_22619_transpose_x_3 = const()[name = string("op_22619_transpose_x_3"), val = bool(false)];
            bool var_22619_transpose_y_3 = const()[name = string("op_22619_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_22619_cast_fp16 = matmul(transpose_x = var_22619_transpose_x_3, transpose_y = var_22619_transpose_y_3, x = q_u_283_cast_fp16, y = k_head_567_cast_fp16)[name = string("op_22619_cast_fp16")];
            fp16 var_22620_to_fp16 = const()[name = string("op_22620_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_283_cast_fp16 = mul(x = var_22619_cast_fp16, y = var_22620_to_fp16)[name = string("scores_content_283_cast_fp16")];
            bool x_1489_transpose_x_3 = const()[name = string("x_1489_transpose_x_3"), val = bool(false)];
            bool x_1489_transpose_y_3 = const()[name = string("x_1489_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1489_cast_fp16 = matmul(transpose_x = x_1489_transpose_x_3, transpose_y = x_1489_transpose_y_3, x = q_v_283_cast_fp16, y = p_head_567_cast_fp16)[name = string("x_1489_cast_fp16")];
            tensor<int32, [8]> x_1491_pad_1 = const()[name = string("x_1491_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1491_mode_1 = const()[name = string("x_1491_mode_1"), val = string("constant")];
            fp16 const_2257_to_fp16 = const()[name = string("const_2257_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1491_cast_fp16 = pad(constant_val = const_2257_to_fp16, mode = x_1491_mode_1, pad = x_1491_pad_1, x = x_1489_cast_fp16)[name = string("x_1491_cast_fp16")];
            tensor<int32, [4]> var_22634 = const()[name = string("op_22634"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1493_cast_fp16 = reshape(shape = var_22634, x = x_1491_cast_fp16)[name = string("x_1493_cast_fp16")];
            tensor<int32, [4]> var_22638_begin_1 = const()[name = string("op_22638_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_22638_end_1 = const()[name = string("op_22638_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_22638_end_mask_1 = const()[name = string("op_22638_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_22638_cast_fp16 = slice_by_index(begin = var_22638_begin_1, end = var_22638_end_1, end_mask = var_22638_end_mask_1, x = x_1493_cast_fp16)[name = string("op_22638_cast_fp16")];
            tensor<int32, [4]> var_22640 = const()[name = string("op_22640"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_22641_cast_fp16 = reshape(shape = var_22640, x = var_22638_cast_fp16)[name = string("op_22641_cast_fp16")];
            tensor<int32, [4]> var_22646_begin_1 = const()[name = string("op_22646_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_22646_end_1 = const()[name = string("op_22646_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_22646_end_mask_1 = const()[name = string("op_22646_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_22646_cast_fp16 = slice_by_index(begin = var_22646_begin_1, end = var_22646_end_1, end_mask = var_22646_end_mask_1, x = var_22641_cast_fp16)[name = string("op_22646_cast_fp16")];
            fp16 var_22647_to_fp16 = const()[name = string("op_22647_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_283_cast_fp16 = mul(x = var_22646_cast_fp16, y = var_22647_to_fp16)[name = string("scores_pos_283_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_283_cast_fp16 = add(x = scores_content_283_cast_fp16, y = scores_pos_283_cast_fp16)[name = string("logits_283_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_22650_cast_fp16 = softmax(axis = var_21770, x = logits_283_cast_fp16)[name = string("op_22650_cast_fp16")];
            bool var_22652_transpose_x_1 = const()[name = string("op_22652_transpose_x_1"), val = bool(false)];
            bool var_22652_transpose_y_1 = const()[name = string("op_22652_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_22652_cast_fp16 = matmul(transpose_x = var_22652_transpose_x_1, transpose_y = var_22652_transpose_y_1, x = var_22650_cast_fp16, y = v_head_567_cast_fp16)[name = string("op_22652_cast_fp16")];
            tensor<int32, [1]> var_22653_axes_1 = const()[name = string("op_22653_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_22653_cast_fp16 = squeeze(axes = var_22653_axes_1, x = var_22652_cast_fp16)[name = string("op_22653_cast_fp16")];
            string dense_output_1425_pad_type_1 = const()[name = string("dense_output_1425_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1425_strides_1 = const()[name = string("dense_output_1425_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1425_pad_1 = const()[name = string("dense_output_1425_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1425_dilations_1 = const()[name = string("dense_output_1425_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1425_groups_1 = const()[name = string("dense_output_1425_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515527744))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515658880))))[name = string("layers_17_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1425_cast_fp16 = conv(dilations = dense_output_1425_dilations_1, groups = dense_output_1425_groups_1, pad = dense_output_1425_pad_1, pad_type = dense_output_1425_pad_type_1, strides = dense_output_1425_strides_1, weight = layers_17_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1425_cast_fp16")];
            string sparse_output_1425_pad_type_1 = const()[name = string("sparse_output_1425_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1425_strides_1 = const()[name = string("sparse_output_1425_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1425_pad_1 = const()[name = string("sparse_output_1425_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1425_dilations_1 = const()[name = string("sparse_output_1425_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1425_groups_1 = const()[name = string("sparse_output_1425_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515662144))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515659456))))[name = string("layers_17_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1425_cast_fp16 = conv(dilations = sparse_output_1425_dilations_1, groups = sparse_output_1425_groups_1, pad = sparse_output_1425_pad_1, pad_type = sparse_output_1425_pad_type_1, strides = sparse_output_1425_strides_1, weight = layers_17_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1425_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22668_cast_fp16 = add(x = dense_output_1425_cast_fp16, y = sparse_output_1425_cast_fp16)[name = string("op_22668_cast_fp16")];
            tensor<int32, [4]> var_22669 = const()[name = string("op_22669"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22671 = const()[name = string("op_22671"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22670_cast_fp16 = transpose(perm = var_22669, x = var_22668_cast_fp16)[name = string("transpose_259")];
            tensor<fp16, [1, 51, 128]> q_head_285_cast_fp16 = reshape(shape = var_22671, x = var_22670_cast_fp16)[name = string("q_head_285_cast_fp16")];
            string dense_output_1427_pad_type_1 = const()[name = string("dense_output_1427_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1427_strides_1 = const()[name = string("dense_output_1427_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1427_pad_1 = const()[name = string("dense_output_1427_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1427_dilations_1 = const()[name = string("dense_output_1427_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1427_groups_1 = const()[name = string("dense_output_1427_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515678592))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515809728))))[name = string("layers_17_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1427_cast_fp16 = conv(dilations = dense_output_1427_dilations_1, groups = dense_output_1427_groups_1, pad = dense_output_1427_pad_1, pad_type = dense_output_1427_pad_type_1, strides = dense_output_1427_strides_1, weight = layers_17_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1427_cast_fp16")];
            string sparse_output_1427_pad_type_1 = const()[name = string("sparse_output_1427_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1427_strides_1 = const()[name = string("sparse_output_1427_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1427_pad_1 = const()[name = string("sparse_output_1427_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1427_dilations_1 = const()[name = string("sparse_output_1427_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1427_groups_1 = const()[name = string("sparse_output_1427_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515812992))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515810304))))[name = string("layers_17_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1427_cast_fp16 = conv(dilations = sparse_output_1427_dilations_1, groups = sparse_output_1427_groups_1, pad = sparse_output_1427_pad_1, pad_type = sparse_output_1427_pad_type_1, strides = sparse_output_1427_strides_1, weight = layers_17_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1427_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22687_cast_fp16 = add(x = dense_output_1427_cast_fp16, y = sparse_output_1427_cast_fp16)[name = string("op_22687_cast_fp16")];
            tensor<int32, [4]> var_22688 = const()[name = string("op_22688"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22690 = const()[name = string("op_22690"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22689_cast_fp16 = transpose(perm = var_22688, x = var_22687_cast_fp16)[name = string("transpose_258")];
            tensor<fp16, [1, 51, 128]> k_head_569_cast_fp16 = reshape(shape = var_22690, x = var_22689_cast_fp16)[name = string("k_head_569_cast_fp16")];
            string dense_output_1429_pad_type_1 = const()[name = string("dense_output_1429_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1429_strides_1 = const()[name = string("dense_output_1429_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1429_pad_1 = const()[name = string("dense_output_1429_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1429_dilations_1 = const()[name = string("dense_output_1429_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1429_groups_1 = const()[name = string("dense_output_1429_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515829440))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515960576))))[name = string("layers_17_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1429_cast_fp16 = conv(dilations = dense_output_1429_dilations_1, groups = dense_output_1429_groups_1, pad = dense_output_1429_pad_1, pad_type = dense_output_1429_pad_type_1, strides = dense_output_1429_strides_1, weight = layers_17_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1429_cast_fp16")];
            string sparse_output_1429_pad_type_1 = const()[name = string("sparse_output_1429_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1429_strides_1 = const()[name = string("sparse_output_1429_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1429_pad_1 = const()[name = string("sparse_output_1429_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1429_dilations_1 = const()[name = string("sparse_output_1429_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1429_groups_1 = const()[name = string("sparse_output_1429_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515963840))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515961152))))[name = string("layers_17_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1429_cast_fp16 = conv(dilations = sparse_output_1429_dilations_1, groups = sparse_output_1429_groups_1, pad = sparse_output_1429_pad_1, pad_type = sparse_output_1429_pad_type_1, strides = sparse_output_1429_strides_1, weight = layers_17_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1429_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22706_cast_fp16 = add(x = dense_output_1429_cast_fp16, y = sparse_output_1429_cast_fp16)[name = string("op_22706_cast_fp16")];
            tensor<int32, [4]> var_22707 = const()[name = string("op_22707"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22709 = const()[name = string("op_22709"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22708_cast_fp16 = transpose(perm = var_22707, x = var_22706_cast_fp16)[name = string("transpose_257")];
            tensor<fp16, [1, 51, 128]> v_head_569_cast_fp16 = reshape(shape = var_22709, x = var_22708_cast_fp16)[name = string("v_head_569_cast_fp16")];
            string dense_output_1431_pad_type_1 = const()[name = string("dense_output_1431_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1431_strides_1 = const()[name = string("dense_output_1431_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1431_pad_1 = const()[name = string("dense_output_1431_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1431_dilations_1 = const()[name = string("dense_output_1431_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1431_groups_1 = const()[name = string("dense_output_1431_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515980288))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516111424))))[name = string("layers_17_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1431_cast_fp16 = conv(dilations = dense_output_1431_dilations_1, groups = dense_output_1431_groups_1, pad = dense_output_1431_pad_1, pad_type = dense_output_1431_pad_type_1, strides = dense_output_1431_strides_1, weight = layers_17_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1431_cast_fp16")];
            string sparse_output_1431_pad_type_1 = const()[name = string("sparse_output_1431_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1431_strides_1 = const()[name = string("sparse_output_1431_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1431_pad_1 = const()[name = string("sparse_output_1431_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1431_dilations_1 = const()[name = string("sparse_output_1431_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1431_groups_1 = const()[name = string("sparse_output_1431_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516114688))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516112000))))[name = string("layers_17_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1431_cast_fp16 = conv(dilations = sparse_output_1431_dilations_1, groups = sparse_output_1431_groups_1, pad = sparse_output_1431_pad_1, pad_type = sparse_output_1431_pad_type_1, strides = sparse_output_1431_strides_1, weight = layers_17_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1431_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_22725_cast_fp16 = add(x = dense_output_1431_cast_fp16, y = sparse_output_1431_cast_fp16)[name = string("op_22725_cast_fp16")];
            tensor<int32, [4]> var_22726 = const()[name = string("op_22726"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22728 = const()[name = string("op_22728"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_22727_cast_fp16 = transpose(perm = var_22726, x = var_22725_cast_fp16)[name = string("transpose_256")];
            tensor<fp16, [1, 101, 128]> p_head_569_cast_fp16 = reshape(shape = var_22728, x = var_22727_cast_fp16)[name = string("p_head_569_cast_fp16")];
            tensor<fp16, [128]> var_22730_to_fp16 = const()[name = string("op_22730_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516131136)))];
            tensor<fp16, [1, 51, 128]> var_22731_cast_fp16 = add(x = q_head_285_cast_fp16, y = var_22730_to_fp16)[name = string("op_22731_cast_fp16")];
            tensor<int32, [1]> q_u_285_axes_1 = const()[name = string("q_u_285_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_285_cast_fp16 = expand_dims(axes = q_u_285_axes_1, x = var_22731_cast_fp16)[name = string("q_u_285_cast_fp16")];
            tensor<fp16, [128]> var_22733_to_fp16 = const()[name = string("op_22733_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516131456)))];
            tensor<fp16, [1, 51, 128]> var_22734_cast_fp16 = add(x = q_head_285_cast_fp16, y = var_22733_to_fp16)[name = string("op_22734_cast_fp16")];
            tensor<int32, [1]> q_v_285_axes_1 = const()[name = string("q_v_285_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_285_cast_fp16 = expand_dims(axes = q_v_285_axes_1, x = var_22734_cast_fp16)[name = string("q_v_285_cast_fp16")];
            tensor<int32, [1]> k_head_571_axes_1 = const()[name = string("k_head_571_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_571_cast_fp16 = expand_dims(axes = k_head_571_axes_1, x = k_head_569_cast_fp16)[name = string("k_head_571_cast_fp16")];
            tensor<int32, [1]> v_head_571_axes_1 = const()[name = string("v_head_571_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_571_cast_fp16 = expand_dims(axes = v_head_571_axes_1, x = v_head_569_cast_fp16)[name = string("v_head_571_cast_fp16")];
            tensor<int32, [1]> p_head_571_axes_1 = const()[name = string("p_head_571_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_571_cast_fp16 = expand_dims(axes = p_head_571_axes_1, x = p_head_569_cast_fp16)[name = string("p_head_571_cast_fp16")];
            bool var_22740_transpose_x_3 = const()[name = string("op_22740_transpose_x_3"), val = bool(false)];
            bool var_22740_transpose_y_3 = const()[name = string("op_22740_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_22740_cast_fp16 = matmul(transpose_x = var_22740_transpose_x_3, transpose_y = var_22740_transpose_y_3, x = q_u_285_cast_fp16, y = k_head_571_cast_fp16)[name = string("op_22740_cast_fp16")];
            fp16 var_22741_to_fp16 = const()[name = string("op_22741_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_285_cast_fp16 = mul(x = var_22740_cast_fp16, y = var_22741_to_fp16)[name = string("scores_content_285_cast_fp16")];
            bool x_1497_transpose_x_3 = const()[name = string("x_1497_transpose_x_3"), val = bool(false)];
            bool x_1497_transpose_y_3 = const()[name = string("x_1497_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1497_cast_fp16 = matmul(transpose_x = x_1497_transpose_x_3, transpose_y = x_1497_transpose_y_3, x = q_v_285_cast_fp16, y = p_head_571_cast_fp16)[name = string("x_1497_cast_fp16")];
            tensor<int32, [8]> x_1499_pad_1 = const()[name = string("x_1499_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1499_mode_1 = const()[name = string("x_1499_mode_1"), val = string("constant")];
            fp16 const_2263_to_fp16 = const()[name = string("const_2263_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1499_cast_fp16 = pad(constant_val = const_2263_to_fp16, mode = x_1499_mode_1, pad = x_1499_pad_1, x = x_1497_cast_fp16)[name = string("x_1499_cast_fp16")];
            tensor<int32, [4]> var_22755 = const()[name = string("op_22755"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1501_cast_fp16 = reshape(shape = var_22755, x = x_1499_cast_fp16)[name = string("x_1501_cast_fp16")];
            tensor<int32, [4]> var_22759_begin_1 = const()[name = string("op_22759_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_22759_end_1 = const()[name = string("op_22759_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_22759_end_mask_1 = const()[name = string("op_22759_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_22759_cast_fp16 = slice_by_index(begin = var_22759_begin_1, end = var_22759_end_1, end_mask = var_22759_end_mask_1, x = x_1501_cast_fp16)[name = string("op_22759_cast_fp16")];
            tensor<int32, [4]> var_22761 = const()[name = string("op_22761"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_22762_cast_fp16 = reshape(shape = var_22761, x = var_22759_cast_fp16)[name = string("op_22762_cast_fp16")];
            tensor<int32, [4]> var_22767_begin_1 = const()[name = string("op_22767_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_22767_end_1 = const()[name = string("op_22767_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_22767_end_mask_1 = const()[name = string("op_22767_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_22767_cast_fp16 = slice_by_index(begin = var_22767_begin_1, end = var_22767_end_1, end_mask = var_22767_end_mask_1, x = var_22762_cast_fp16)[name = string("op_22767_cast_fp16")];
            fp16 var_22768_to_fp16 = const()[name = string("op_22768_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_285_cast_fp16 = mul(x = var_22767_cast_fp16, y = var_22768_to_fp16)[name = string("scores_pos_285_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_285_cast_fp16 = add(x = scores_content_285_cast_fp16, y = scores_pos_285_cast_fp16)[name = string("logits_285_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_22771_cast_fp16 = softmax(axis = var_21770, x = logits_285_cast_fp16)[name = string("op_22771_cast_fp16")];
            bool var_22773_transpose_x_1 = const()[name = string("op_22773_transpose_x_1"), val = bool(false)];
            bool var_22773_transpose_y_1 = const()[name = string("op_22773_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_22773_cast_fp16 = matmul(transpose_x = var_22773_transpose_x_1, transpose_y = var_22773_transpose_y_1, x = var_22771_cast_fp16, y = v_head_571_cast_fp16)[name = string("op_22773_cast_fp16")];
            tensor<int32, [1]> var_22774_axes_1 = const()[name = string("op_22774_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_22774_cast_fp16 = squeeze(axes = var_22774_axes_1, x = var_22773_cast_fp16)[name = string("op_22774_cast_fp16")];
            string dense_output_1433_pad_type_1 = const()[name = string("dense_output_1433_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1433_strides_1 = const()[name = string("dense_output_1433_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1433_pad_1 = const()[name = string("dense_output_1433_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1433_dilations_1 = const()[name = string("dense_output_1433_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1433_groups_1 = const()[name = string("dense_output_1433_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516131776))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516262912))))[name = string("layers_17_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1433_cast_fp16 = conv(dilations = dense_output_1433_dilations_1, groups = dense_output_1433_groups_1, pad = dense_output_1433_pad_1, pad_type = dense_output_1433_pad_type_1, strides = dense_output_1433_strides_1, weight = layers_17_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1433_cast_fp16")];
            string sparse_output_1433_pad_type_1 = const()[name = string("sparse_output_1433_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1433_strides_1 = const()[name = string("sparse_output_1433_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1433_pad_1 = const()[name = string("sparse_output_1433_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1433_dilations_1 = const()[name = string("sparse_output_1433_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1433_groups_1 = const()[name = string("sparse_output_1433_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516266176))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516263488))))[name = string("layers_17_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1433_cast_fp16 = conv(dilations = sparse_output_1433_dilations_1, groups = sparse_output_1433_groups_1, pad = sparse_output_1433_pad_1, pad_type = sparse_output_1433_pad_type_1, strides = sparse_output_1433_strides_1, weight = layers_17_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1433_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22789_cast_fp16 = add(x = dense_output_1433_cast_fp16, y = sparse_output_1433_cast_fp16)[name = string("op_22789_cast_fp16")];
            tensor<int32, [4]> var_22790 = const()[name = string("op_22790"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22792 = const()[name = string("op_22792"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22791_cast_fp16 = transpose(perm = var_22790, x = var_22789_cast_fp16)[name = string("transpose_255")];
            tensor<fp16, [1, 51, 128]> q_head_287_cast_fp16 = reshape(shape = var_22792, x = var_22791_cast_fp16)[name = string("q_head_287_cast_fp16")];
            string dense_output_1435_pad_type_1 = const()[name = string("dense_output_1435_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1435_strides_1 = const()[name = string("dense_output_1435_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1435_pad_1 = const()[name = string("dense_output_1435_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1435_dilations_1 = const()[name = string("dense_output_1435_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1435_groups_1 = const()[name = string("dense_output_1435_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516282624))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516413760))))[name = string("layers_17_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1435_cast_fp16 = conv(dilations = dense_output_1435_dilations_1, groups = dense_output_1435_groups_1, pad = dense_output_1435_pad_1, pad_type = dense_output_1435_pad_type_1, strides = dense_output_1435_strides_1, weight = layers_17_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1435_cast_fp16")];
            string sparse_output_1435_pad_type_1 = const()[name = string("sparse_output_1435_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1435_strides_1 = const()[name = string("sparse_output_1435_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1435_pad_1 = const()[name = string("sparse_output_1435_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1435_dilations_1 = const()[name = string("sparse_output_1435_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1435_groups_1 = const()[name = string("sparse_output_1435_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516417024))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516414336))))[name = string("layers_17_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1435_cast_fp16 = conv(dilations = sparse_output_1435_dilations_1, groups = sparse_output_1435_groups_1, pad = sparse_output_1435_pad_1, pad_type = sparse_output_1435_pad_type_1, strides = sparse_output_1435_strides_1, weight = layers_17_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1435_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22808_cast_fp16 = add(x = dense_output_1435_cast_fp16, y = sparse_output_1435_cast_fp16)[name = string("op_22808_cast_fp16")];
            tensor<int32, [4]> var_22809 = const()[name = string("op_22809"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22811 = const()[name = string("op_22811"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22810_cast_fp16 = transpose(perm = var_22809, x = var_22808_cast_fp16)[name = string("transpose_254")];
            tensor<fp16, [1, 51, 128]> k_head_573_cast_fp16 = reshape(shape = var_22811, x = var_22810_cast_fp16)[name = string("k_head_573_cast_fp16")];
            string dense_output_1437_pad_type_1 = const()[name = string("dense_output_1437_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1437_strides_1 = const()[name = string("dense_output_1437_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1437_pad_1 = const()[name = string("dense_output_1437_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1437_dilations_1 = const()[name = string("dense_output_1437_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1437_groups_1 = const()[name = string("dense_output_1437_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516433472))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516564608))))[name = string("layers_17_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1437_cast_fp16 = conv(dilations = dense_output_1437_dilations_1, groups = dense_output_1437_groups_1, pad = dense_output_1437_pad_1, pad_type = dense_output_1437_pad_type_1, strides = dense_output_1437_strides_1, weight = layers_17_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("dense_output_1437_cast_fp16")];
            string sparse_output_1437_pad_type_1 = const()[name = string("sparse_output_1437_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1437_strides_1 = const()[name = string("sparse_output_1437_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1437_pad_1 = const()[name = string("sparse_output_1437_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1437_dilations_1 = const()[name = string("sparse_output_1437_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1437_groups_1 = const()[name = string("sparse_output_1437_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516567872))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516565184))))[name = string("layers_17_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1437_cast_fp16 = conv(dilations = sparse_output_1437_dilations_1, groups = sparse_output_1437_groups_1, pad = sparse_output_1437_pad_1, pad_type = sparse_output_1437_pad_type_1, strides = sparse_output_1437_strides_1, weight = layers_17_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_811_cast_fp16)[name = string("sparse_output_1437_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_22827_cast_fp16 = add(x = dense_output_1437_cast_fp16, y = sparse_output_1437_cast_fp16)[name = string("op_22827_cast_fp16")];
            tensor<int32, [4]> var_22828 = const()[name = string("op_22828"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22830 = const()[name = string("op_22830"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_22829_cast_fp16 = transpose(perm = var_22828, x = var_22827_cast_fp16)[name = string("transpose_253")];
            tensor<fp16, [1, 51, 128]> v_head_573_cast_fp16 = reshape(shape = var_22830, x = var_22829_cast_fp16)[name = string("v_head_573_cast_fp16")];
            string dense_output_1439_pad_type_1 = const()[name = string("dense_output_1439_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1439_strides_1 = const()[name = string("dense_output_1439_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1439_pad_1 = const()[name = string("dense_output_1439_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1439_dilations_1 = const()[name = string("dense_output_1439_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1439_groups_1 = const()[name = string("dense_output_1439_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516584320))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516715456))))[name = string("layers_17_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1439_cast_fp16 = conv(dilations = dense_output_1439_dilations_1, groups = dense_output_1439_groups_1, pad = dense_output_1439_pad_1, pad_type = dense_output_1439_pad_type_1, strides = dense_output_1439_strides_1, weight = layers_17_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1439_cast_fp16")];
            string sparse_output_1439_pad_type_1 = const()[name = string("sparse_output_1439_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1439_strides_1 = const()[name = string("sparse_output_1439_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1439_pad_1 = const()[name = string("sparse_output_1439_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1439_dilations_1 = const()[name = string("sparse_output_1439_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1439_groups_1 = const()[name = string("sparse_output_1439_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_17_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516718720))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516716032))))[name = string("layers_17_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1439_cast_fp16 = conv(dilations = sparse_output_1439_dilations_1, groups = sparse_output_1439_groups_1, pad = sparse_output_1439_pad_1, pad_type = sparse_output_1439_pad_type_1, strides = sparse_output_1439_strides_1, weight = layers_17_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1439_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_22846_cast_fp16 = add(x = dense_output_1439_cast_fp16, y = sparse_output_1439_cast_fp16)[name = string("op_22846_cast_fp16")];
            tensor<int32, [4]> var_22847 = const()[name = string("op_22847"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_22849 = const()[name = string("op_22849"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_22848_cast_fp16 = transpose(perm = var_22847, x = var_22846_cast_fp16)[name = string("transpose_252")];
            tensor<fp16, [1, 101, 128]> p_head_573_cast_fp16 = reshape(shape = var_22849, x = var_22848_cast_fp16)[name = string("p_head_573_cast_fp16")];
            tensor<fp16, [128]> var_22851_to_fp16 = const()[name = string("op_22851_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516735168)))];
            tensor<fp16, [1, 51, 128]> var_22852_cast_fp16 = add(x = q_head_287_cast_fp16, y = var_22851_to_fp16)[name = string("op_22852_cast_fp16")];
            tensor<int32, [1]> q_u_287_axes_1 = const()[name = string("q_u_287_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_287_cast_fp16 = expand_dims(axes = q_u_287_axes_1, x = var_22852_cast_fp16)[name = string("q_u_287_cast_fp16")];
            tensor<fp16, [128]> var_22854_to_fp16 = const()[name = string("op_22854_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516735488)))];
            tensor<fp16, [1, 51, 128]> var_22855_cast_fp16 = add(x = q_head_287_cast_fp16, y = var_22854_to_fp16)[name = string("op_22855_cast_fp16")];
            tensor<int32, [1]> q_v_287_axes_1 = const()[name = string("q_v_287_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_287_cast_fp16 = expand_dims(axes = q_v_287_axes_1, x = var_22855_cast_fp16)[name = string("q_v_287_cast_fp16")];
            tensor<int32, [1]> k_head_575_axes_1 = const()[name = string("k_head_575_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_575_cast_fp16 = expand_dims(axes = k_head_575_axes_1, x = k_head_573_cast_fp16)[name = string("k_head_575_cast_fp16")];
            tensor<int32, [1]> v_head_575_axes_1 = const()[name = string("v_head_575_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_575_cast_fp16 = expand_dims(axes = v_head_575_axes_1, x = v_head_573_cast_fp16)[name = string("v_head_575_cast_fp16")];
            tensor<int32, [1]> p_head_575_axes_1 = const()[name = string("p_head_575_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_575_cast_fp16 = expand_dims(axes = p_head_575_axes_1, x = p_head_573_cast_fp16)[name = string("p_head_575_cast_fp16")];
            bool var_22861_transpose_x_3 = const()[name = string("op_22861_transpose_x_3"), val = bool(false)];
            bool var_22861_transpose_y_3 = const()[name = string("op_22861_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_22861_cast_fp16 = matmul(transpose_x = var_22861_transpose_x_3, transpose_y = var_22861_transpose_y_3, x = q_u_287_cast_fp16, y = k_head_575_cast_fp16)[name = string("op_22861_cast_fp16")];
            fp16 var_22862_to_fp16 = const()[name = string("op_22862_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_287_cast_fp16 = mul(x = var_22861_cast_fp16, y = var_22862_to_fp16)[name = string("scores_content_287_cast_fp16")];
            bool x_1505_transpose_x_3 = const()[name = string("x_1505_transpose_x_3"), val = bool(false)];
            bool x_1505_transpose_y_3 = const()[name = string("x_1505_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1505_cast_fp16 = matmul(transpose_x = x_1505_transpose_x_3, transpose_y = x_1505_transpose_y_3, x = q_v_287_cast_fp16, y = p_head_575_cast_fp16)[name = string("x_1505_cast_fp16")];
            tensor<int32, [8]> x_1507_pad_1 = const()[name = string("x_1507_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1507_mode_1 = const()[name = string("x_1507_mode_1"), val = string("constant")];
            fp16 const_2269_to_fp16 = const()[name = string("const_2269_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1507_cast_fp16 = pad(constant_val = const_2269_to_fp16, mode = x_1507_mode_1, pad = x_1507_pad_1, x = x_1505_cast_fp16)[name = string("x_1507_cast_fp16")];
            tensor<int32, [4]> var_22876 = const()[name = string("op_22876"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1509_cast_fp16 = reshape(shape = var_22876, x = x_1507_cast_fp16)[name = string("x_1509_cast_fp16")];
            tensor<int32, [4]> var_22880_begin_1 = const()[name = string("op_22880_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_22880_end_1 = const()[name = string("op_22880_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_22880_end_mask_1 = const()[name = string("op_22880_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_22880_cast_fp16 = slice_by_index(begin = var_22880_begin_1, end = var_22880_end_1, end_mask = var_22880_end_mask_1, x = x_1509_cast_fp16)[name = string("op_22880_cast_fp16")];
            tensor<int32, [4]> var_22882 = const()[name = string("op_22882"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_22883_cast_fp16 = reshape(shape = var_22882, x = var_22880_cast_fp16)[name = string("op_22883_cast_fp16")];
            tensor<int32, [4]> var_22888_begin_1 = const()[name = string("op_22888_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_22888_end_1 = const()[name = string("op_22888_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_22888_end_mask_1 = const()[name = string("op_22888_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_22888_cast_fp16 = slice_by_index(begin = var_22888_begin_1, end = var_22888_end_1, end_mask = var_22888_end_mask_1, x = var_22883_cast_fp16)[name = string("op_22888_cast_fp16")];
            fp16 var_22889_to_fp16 = const()[name = string("op_22889_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_287_cast_fp16 = mul(x = var_22888_cast_fp16, y = var_22889_to_fp16)[name = string("scores_pos_287_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_287_cast_fp16 = add(x = scores_content_287_cast_fp16, y = scores_pos_287_cast_fp16)[name = string("logits_287_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_22892_cast_fp16 = softmax(axis = var_21770, x = logits_287_cast_fp16)[name = string("op_22892_cast_fp16")];
            bool var_22894_transpose_x_1 = const()[name = string("op_22894_transpose_x_1"), val = bool(false)];
            bool var_22894_transpose_y_1 = const()[name = string("op_22894_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_22894_cast_fp16 = matmul(transpose_x = var_22894_transpose_x_1, transpose_y = var_22894_transpose_y_1, x = var_22892_cast_fp16, y = v_head_575_cast_fp16)[name = string("op_22894_cast_fp16")];
            tensor<int32, [1]> o_head_35_axes_1 = const()[name = string("o_head_35_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_35_cast_fp16 = squeeze(axes = o_head_35_axes_1, x = var_22894_cast_fp16)[name = string("o_head_35_cast_fp16")];
            bool out_35_interleave_1 = const()[name = string("out_35_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_35_cast_fp16 = concat(axis = var_21770, interleave = out_35_interleave_1, values = (var_22048_cast_fp16, var_22169_cast_fp16, var_22290_cast_fp16, var_22411_cast_fp16, var_22532_cast_fp16, var_22653_cast_fp16, var_22774_cast_fp16, o_head_35_cast_fp16))[name = string("out_35_cast_fp16")];
            tensor<int32, [3]> var_22898_perm_1 = const()[name = string("op_22898_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_819_axes_1 = const()[name = string("input_819_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_22898_cast_fp16 = transpose(perm = var_22898_perm_1, x = out_35_cast_fp16)[name = string("transpose_251")];
            tensor<fp16, [1, 1024, 51, 1]> input_819_cast_fp16 = expand_dims(axes = input_819_axes_1, x = var_22898_cast_fp16)[name = string("input_819_cast_fp16")];
            string dense_output_1441_pad_type_1 = const()[name = string("dense_output_1441_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1441_strides_1 = const()[name = string("dense_output_1441_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1441_pad_1 = const()[name = string("dense_output_1441_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1441_dilations_1 = const()[name = string("dense_output_1441_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1441_groups_1 = const()[name = string("dense_output_1441_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516735808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517784448))))[name = string("layers_17_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1441_cast_fp16 = conv(dilations = dense_output_1441_dilations_1, groups = dense_output_1441_groups_1, pad = dense_output_1441_pad_1, pad_type = dense_output_1441_pad_type_1, strides = dense_output_1441_strides_1, weight = layers_17_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("dense_output_1441_cast_fp16")];
            string sparse_output_1441_pad_type_1 = const()[name = string("sparse_output_1441_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1441_strides_1 = const()[name = string("sparse_output_1441_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1441_pad_1 = const()[name = string("sparse_output_1441_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1441_dilations_1 = const()[name = string("sparse_output_1441_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1441_groups_1 = const()[name = string("sparse_output_1441_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517806080))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517785024))))[name = string("layers_17_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1441_cast_fp16 = conv(dilations = sparse_output_1441_dilations_1, groups = sparse_output_1441_groups_1, pad = sparse_output_1441_pad_1, pad_type = sparse_output_1441_pad_type_1, strides = sparse_output_1441_strides_1, weight = layers_17_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_819_cast_fp16)[name = string("sparse_output_1441_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_35_cast_fp16 = add(x = dense_output_1441_cast_fp16, y = sparse_output_1441_cast_fp16)[name = string("out_conv_35_cast_fp16")];
            tensor<int32, [1]> var_22915_axes_1 = const()[name = string("op_22915_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_22915_cast_fp16 = squeeze(axes = var_22915_axes_1, x = out_conv_35_cast_fp16)[name = string("op_22915_cast_fp16")];
            tensor<int32, [3]> var_22916_perm_1 = const()[name = string("op_22916_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_22916_cast_fp16 = transpose(perm = var_22916_perm_1, x = var_22915_cast_fp16)[name = string("transpose_250")];
            tensor<fp16, [1, 51, 1024]> input_821_cast_fp16 = add(x = input_809_cast_fp16, y = var_22916_cast_fp16)[name = string("input_821_cast_fp16")];
            tensor<int32, [1]> x_1513_axes_1 = const()[name = string("x_1513_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_17_norm_conv_weight_to_fp16 = const()[name = string("layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517937216)))];
            tensor<fp16, [1024]> layers_17_norm_conv_bias_to_fp16 = const()[name = string("layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517939328)))];
            tensor<fp16, [1, 51, 1024]> x_1513_cast_fp16 = layer_norm(axes = x_1513_axes_1, beta = layers_17_norm_conv_bias_to_fp16, epsilon = var_21785_to_fp16, gamma = layers_17_norm_conv_weight_to_fp16, x = input_821_cast_fp16)[name = string("x_1513_cast_fp16")];
            tensor<int32, [3]> var_22926_perm_1 = const()[name = string("op_22926_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_823_axes_1 = const()[name = string("input_823_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_22926_cast_fp16 = transpose(perm = var_22926_perm_1, x = x_1513_cast_fp16)[name = string("transpose_249")];
            tensor<fp16, [1, 1024, 51, 1]> input_823_cast_fp16 = expand_dims(axes = input_823_axes_1, x = var_22926_cast_fp16)[name = string("input_823_cast_fp16")];
            string dense_output_1443_pad_type_1 = const()[name = string("dense_output_1443_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1443_strides_1 = const()[name = string("dense_output_1443_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1443_pad_1 = const()[name = string("dense_output_1443_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1443_dilations_1 = const()[name = string("dense_output_1443_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1443_groups_1 = const()[name = string("dense_output_1443_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_17_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517941440))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520038656))))[name = string("layers_17_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1443_cast_fp16 = conv(dilations = dense_output_1443_dilations_1, groups = dense_output_1443_groups_1, pad = dense_output_1443_pad_1, pad_type = dense_output_1443_pad_type_1, strides = dense_output_1443_strides_1, weight = layers_17_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("dense_output_1443_cast_fp16")];
            string sparse_output_1443_pad_type_1 = const()[name = string("sparse_output_1443_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1443_strides_1 = const()[name = string("sparse_output_1443_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1443_pad_1 = const()[name = string("sparse_output_1443_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1443_dilations_1 = const()[name = string("sparse_output_1443_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1443_groups_1 = const()[name = string("sparse_output_1443_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_17_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520081280))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520039232))))[name = string("layers_17_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1443_cast_fp16 = conv(dilations = sparse_output_1443_dilations_1, groups = sparse_output_1443_groups_1, pad = sparse_output_1443_pad_1, pad_type = sparse_output_1443_pad_type_1, strides = sparse_output_1443_strides_1, weight = layers_17_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_823_cast_fp16)[name = string("sparse_output_1443_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_825_cast_fp16 = add(x = dense_output_1443_cast_fp16, y = sparse_output_1443_cast_fp16)[name = string("input_825_cast_fp16")];
            int32 input_827_split_num_splits_1 = const()[name = string("input_827_split_num_splits_1"), val = int32(2)];
            int32 input_827_split_axis_1 = const()[name = string("input_827_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_827_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_827_split_cast_fp16_1 = split(axis = input_827_split_axis_1, num_splits = input_827_split_num_splits_1, x = input_825_cast_fp16)[name = string("input_827_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_827_split_1_sigmoid_cast_fp16 = sigmoid(x = input_827_split_cast_fp16_1)[name = string("input_827_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_827_cast_fp16 = mul(x = input_827_split_cast_fp16_0, y = input_827_split_1_sigmoid_cast_fp16)[name = string("input_827_cast_fp16")];
            tensor<int32, [8]> input_829_pad_1 = const()[name = string("input_829_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_829_mode_1 = const()[name = string("input_829_mode_1"), val = string("constant")];
            fp16 const_2271_to_fp16 = const()[name = string("const_2271_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_829_cast_fp16 = pad(constant_val = const_2271_to_fp16, mode = input_829_mode_1, pad = input_829_pad_1, x = input_827_cast_fp16)[name = string("input_829_cast_fp16")];
            string dense_output_1445_pad_type_1 = const()[name = string("dense_output_1445_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1445_strides_1 = const()[name = string("dense_output_1445_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1445_pad_1 = const()[name = string("dense_output_1445_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1445_dilations_1 = const()[name = string("dense_output_1445_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1445_groups_1 = const()[name = string("dense_output_1445_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1445_cast_fp16 = conv(dilations = dense_output_1445_dilations_1, groups = dense_output_1445_groups_1, pad = dense_output_1445_pad_1, pad_type = dense_output_1445_pad_type_1, strides = dense_output_1445_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_829_cast_fp16)[name = string("dense_output_1445_cast_fp16")];
            string sparse_output_1445_pad_type_1 = const()[name = string("sparse_output_1445_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1445_strides_1 = const()[name = string("sparse_output_1445_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1445_pad_1 = const()[name = string("sparse_output_1445_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1445_dilations_1 = const()[name = string("sparse_output_1445_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1445_groups_1 = const()[name = string("sparse_output_1445_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_17_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520343488))))[name = string("layers_17_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1445_cast_fp16 = conv(dilations = sparse_output_1445_dilations_1, groups = sparse_output_1445_groups_1, pad = sparse_output_1445_pad_1, pad_type = sparse_output_1445_pad_type_1, strides = sparse_output_1445_strides_1, weight = layers_17_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_829_cast_fp16)[name = string("sparse_output_1445_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_831_cast_fp16 = add(x = dense_output_1445_cast_fp16, y = sparse_output_1445_cast_fp16)[name = string("input_831_cast_fp16")];
            tensor<fp16, [1024]> layers_17_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_17_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520361984)))];
            tensor<fp16, [1024]> layers_17_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_17_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520364096)))];
            tensor<fp16, [1024]> layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_17_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520366208)))];
            tensor<fp16, [1024]> layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_17_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520368320)))];
            tensor<fp16, [1, 1024, 51, 1]> input_833_cast_fp16 = batch_norm(beta = layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_21785_to_fp16, gamma = layers_17_conv_batch_norm_weight_to_fp16, mean = layers_17_conv_batch_norm_running_mean_to_fp16, variance = layers_17_conv_batch_norm_running_var_to_fp16, x = input_831_cast_fp16)[name = string("input_833_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_835_cast_fp16 = silu(x = input_833_cast_fp16)[name = string("input_835_cast_fp16")];
            string dense_output_1447_pad_type_1 = const()[name = string("dense_output_1447_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1447_strides_1 = const()[name = string("dense_output_1447_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1447_pad_1 = const()[name = string("dense_output_1447_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1447_dilations_1 = const()[name = string("dense_output_1447_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1447_groups_1 = const()[name = string("dense_output_1447_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520370432))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(521419072))))[name = string("layers_17_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1447_cast_fp16 = conv(dilations = dense_output_1447_dilations_1, groups = dense_output_1447_groups_1, pad = dense_output_1447_pad_1, pad_type = dense_output_1447_pad_type_1, strides = dense_output_1447_strides_1, weight = layers_17_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_835_cast_fp16)[name = string("dense_output_1447_cast_fp16")];
            string sparse_output_1447_pad_type_1 = const()[name = string("sparse_output_1447_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1447_strides_1 = const()[name = string("sparse_output_1447_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1447_pad_1 = const()[name = string("sparse_output_1447_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1447_dilations_1 = const()[name = string("sparse_output_1447_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1447_groups_1 = const()[name = string("sparse_output_1447_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_17_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(521440704))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(521419648))))[name = string("layers_17_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1447_cast_fp16 = conv(dilations = sparse_output_1447_dilations_1, groups = sparse_output_1447_groups_1, pad = sparse_output_1447_pad_1, pad_type = sparse_output_1447_pad_type_1, strides = sparse_output_1447_strides_1, weight = layers_17_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_835_cast_fp16)[name = string("sparse_output_1447_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1515_cast_fp16 = add(x = dense_output_1447_cast_fp16, y = sparse_output_1447_cast_fp16)[name = string("x_1515_cast_fp16")];
            tensor<int32, [1]> var_22982_axes_1 = const()[name = string("op_22982_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_22982_cast_fp16 = squeeze(axes = var_22982_axes_1, x = x_1515_cast_fp16)[name = string("op_22982_cast_fp16")];
            tensor<int32, [3]> var_22983_perm_1 = const()[name = string("op_22983_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_22983_cast_fp16 = transpose(perm = var_22983_perm_1, x = var_22982_cast_fp16)[name = string("transpose_248")];
            tensor<fp16, [1, 51, 1024]> input_837_cast_fp16 = add(x = input_821_cast_fp16, y = var_22983_cast_fp16)[name = string("input_837_cast_fp16")];
            tensor<int32, [1]> x_1517_axes_1 = const()[name = string("x_1517_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(521571840)))];
            tensor<fp16, [1024]> layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(521573952)))];
            tensor<fp16, [1, 51, 1024]> x_1517_cast_fp16 = layer_norm(axes = x_1517_axes_1, beta = layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_21785_to_fp16, gamma = layers_17_norm_feed_forward2_weight_to_fp16, x = input_837_cast_fp16)[name = string("x_1517_cast_fp16")];
            tensor<int32, [4]> var_22993 = const()[name = string("op_22993"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1519_cast_fp16 = reshape(shape = var_22993, x = x_1517_cast_fp16)[name = string("x_1519_cast_fp16")];
            tensor<int32, [4]> input_839_perm_1 = const()[name = string("input_839_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1449_pad_type_1 = const()[name = string("dense_output_1449_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1449_strides_1 = const()[name = string("dense_output_1449_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1449_pad_1 = const()[name = string("dense_output_1449_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1449_dilations_1 = const()[name = string("dense_output_1449_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1449_groups_1 = const()[name = string("dense_output_1449_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_17_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(521576064))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525770432))))[name = string("layers_17_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_839_cast_fp16 = transpose(perm = input_839_perm_1, x = x_1519_cast_fp16)[name = string("transpose_247")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1449_cast_fp16 = conv(dilations = dense_output_1449_dilations_1, groups = dense_output_1449_groups_1, pad = dense_output_1449_pad_1, pad_type = dense_output_1449_pad_type_1, strides = dense_output_1449_strides_1, weight = layers_17_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_839_cast_fp16)[name = string("dense_output_1449_cast_fp16")];
            string sparse_output_1449_pad_type_1 = const()[name = string("sparse_output_1449_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1449_strides_1 = const()[name = string("sparse_output_1449_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1449_pad_1 = const()[name = string("sparse_output_1449_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1449_dilations_1 = const()[name = string("sparse_output_1449_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1449_groups_1 = const()[name = string("sparse_output_1449_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_17_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525854976))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525771008))))[name = string("layers_17_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1449_cast_fp16 = conv(dilations = sparse_output_1449_dilations_1, groups = sparse_output_1449_groups_1, pad = sparse_output_1449_pad_1, pad_type = sparse_output_1449_pad_type_1, strides = sparse_output_1449_strides_1, weight = layers_17_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_839_cast_fp16)[name = string("sparse_output_1449_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_841_cast_fp16 = add(x = dense_output_1449_cast_fp16, y = sparse_output_1449_cast_fp16)[name = string("input_841_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_843_cast_fp16 = silu(x = input_841_cast_fp16)[name = string("input_843_cast_fp16")];
            string dense_output_1451_pad_type_1 = const()[name = string("dense_output_1451_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1451_strides_1 = const()[name = string("dense_output_1451_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1451_pad_1 = const()[name = string("dense_output_1451_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1451_dilations_1 = const()[name = string("dense_output_1451_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1451_groups_1 = const()[name = string("dense_output_1451_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_17_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526379328))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530573696))))[name = string("layers_17_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1451_cast_fp16 = conv(dilations = dense_output_1451_dilations_1, groups = dense_output_1451_groups_1, pad = dense_output_1451_pad_1, pad_type = dense_output_1451_pad_type_1, strides = dense_output_1451_strides_1, weight = layers_17_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_843_cast_fp16)[name = string("dense_output_1451_cast_fp16")];
            string sparse_output_1451_pad_type_1 = const()[name = string("sparse_output_1451_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1451_strides_1 = const()[name = string("sparse_output_1451_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1451_pad_1 = const()[name = string("sparse_output_1451_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1451_dilations_1 = const()[name = string("sparse_output_1451_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1451_groups_1 = const()[name = string("sparse_output_1451_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_17_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530658240))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530574272))))[name = string("layers_17_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1451_cast_fp16 = conv(dilations = sparse_output_1451_dilations_1, groups = sparse_output_1451_groups_1, pad = sparse_output_1451_pad_1, pad_type = sparse_output_1451_pad_type_1, strides = sparse_output_1451_strides_1, weight = layers_17_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_843_cast_fp16)[name = string("sparse_output_1451_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1521_cast_fp16 = add(x = dense_output_1451_cast_fp16, y = sparse_output_1451_cast_fp16)[name = string("x_1521_cast_fp16")];
            tensor<int32, [4]> x_1523_perm_1 = const()[name = string("x_1523_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_23028 = const()[name = string("op_23028"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1523_cast_fp16 = transpose(perm = x_1523_perm_1, x = x_1521_cast_fp16)[name = string("transpose_246")];
            tensor<fp16, [1, 51, 1024]> var_23029_cast_fp16 = reshape(shape = var_23028, x = x_1523_cast_fp16)[name = string("op_23029_cast_fp16")];
            fp16 var_23030_to_fp16 = const()[name = string("op_23030_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_23031_cast_fp16 = mul(x = var_23029_cast_fp16, y = var_23030_to_fp16)[name = string("op_23031_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_845_cast_fp16 = add(x = input_837_cast_fp16, y = var_23031_cast_fp16)[name = string("input_845_cast_fp16")];
            tensor<int32, [1]> input_847_axes_1 = const()[name = string("input_847_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_17_norm_out_weight_to_fp16 = const()[name = string("layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531182592)))];
            tensor<fp16, [1024]> layers_17_norm_out_bias_to_fp16 = const()[name = string("layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531184704)))];
            tensor<fp16, [1, 51, 1024]> input_847_cast_fp16 = layer_norm(axes = input_847_axes_1, beta = layers_17_norm_out_bias_to_fp16, epsilon = var_21785_to_fp16, gamma = layers_17_norm_out_weight_to_fp16, x = input_845_cast_fp16)[name = string("input_847_cast_fp16")];
            int32 var_23039 = const()[name = string("op_23039"), val = int32(-1)];
            tensor<int32, [1]> x_1525_axes_1 = const()[name = string("x_1525_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531186816)))];
            tensor<fp16, [1024]> layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531188928)))];
            fp16 var_23054_to_fp16 = const()[name = string("op_23054_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1525_cast_fp16 = layer_norm(axes = x_1525_axes_1, beta = layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_23054_to_fp16, gamma = layers_18_norm_feed_forward1_weight_to_fp16, x = input_847_cast_fp16)[name = string("x_1525_cast_fp16")];
            tensor<int32, [4]> var_23073 = const()[name = string("op_23073"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1527_cast_fp16 = reshape(shape = var_23073, x = x_1525_cast_fp16)[name = string("x_1527_cast_fp16")];
            tensor<int32, [4]> input_849_perm_1 = const()[name = string("input_849_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1453_pad_type_1 = const()[name = string("dense_output_1453_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1453_strides_1 = const()[name = string("dense_output_1453_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1453_pad_1 = const()[name = string("dense_output_1453_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1453_dilations_1 = const()[name = string("dense_output_1453_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1453_groups_1 = const()[name = string("dense_output_1453_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_18_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531191040))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535385408))))[name = string("layers_18_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_849_cast_fp16 = transpose(perm = input_849_perm_1, x = x_1527_cast_fp16)[name = string("transpose_245")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1453_cast_fp16 = conv(dilations = dense_output_1453_dilations_1, groups = dense_output_1453_groups_1, pad = dense_output_1453_pad_1, pad_type = dense_output_1453_pad_type_1, strides = dense_output_1453_strides_1, weight = layers_18_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("dense_output_1453_cast_fp16")];
            string sparse_output_1453_pad_type_1 = const()[name = string("sparse_output_1453_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1453_strides_1 = const()[name = string("sparse_output_1453_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1453_pad_1 = const()[name = string("sparse_output_1453_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1453_dilations_1 = const()[name = string("sparse_output_1453_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1453_groups_1 = const()[name = string("sparse_output_1453_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_18_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535469952))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535385984))))[name = string("layers_18_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1453_cast_fp16 = conv(dilations = sparse_output_1453_dilations_1, groups = sparse_output_1453_groups_1, pad = sparse_output_1453_pad_1, pad_type = sparse_output_1453_pad_type_1, strides = sparse_output_1453_strides_1, weight = layers_18_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_849_cast_fp16)[name = string("sparse_output_1453_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_851_cast_fp16 = add(x = dense_output_1453_cast_fp16, y = sparse_output_1453_cast_fp16)[name = string("input_851_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_853_cast_fp16 = silu(x = input_851_cast_fp16)[name = string("input_853_cast_fp16")];
            string dense_output_1455_pad_type_1 = const()[name = string("dense_output_1455_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1455_strides_1 = const()[name = string("dense_output_1455_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1455_pad_1 = const()[name = string("dense_output_1455_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1455_dilations_1 = const()[name = string("dense_output_1455_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1455_groups_1 = const()[name = string("dense_output_1455_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_18_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535994304))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540188672))))[name = string("layers_18_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1455_cast_fp16 = conv(dilations = dense_output_1455_dilations_1, groups = dense_output_1455_groups_1, pad = dense_output_1455_pad_1, pad_type = dense_output_1455_pad_type_1, strides = dense_output_1455_strides_1, weight = layers_18_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_853_cast_fp16)[name = string("dense_output_1455_cast_fp16")];
            string sparse_output_1455_pad_type_1 = const()[name = string("sparse_output_1455_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1455_strides_1 = const()[name = string("sparse_output_1455_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1455_pad_1 = const()[name = string("sparse_output_1455_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1455_dilations_1 = const()[name = string("sparse_output_1455_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1455_groups_1 = const()[name = string("sparse_output_1455_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_18_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540273216))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540189248))))[name = string("layers_18_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1455_cast_fp16 = conv(dilations = sparse_output_1455_dilations_1, groups = sparse_output_1455_groups_1, pad = sparse_output_1455_pad_1, pad_type = sparse_output_1455_pad_type_1, strides = sparse_output_1455_strides_1, weight = layers_18_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_853_cast_fp16)[name = string("sparse_output_1455_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1529_cast_fp16 = add(x = dense_output_1455_cast_fp16, y = sparse_output_1455_cast_fp16)[name = string("x_1529_cast_fp16")];
            tensor<int32, [4]> x_1531_perm_1 = const()[name = string("x_1531_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_23108 = const()[name = string("op_23108"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1531_cast_fp16 = transpose(perm = x_1531_perm_1, x = x_1529_cast_fp16)[name = string("transpose_244")];
            tensor<fp16, [1, 51, 1024]> var_23109_cast_fp16 = reshape(shape = var_23108, x = x_1531_cast_fp16)[name = string("op_23109_cast_fp16")];
            fp16 var_23110_to_fp16 = const()[name = string("op_23110_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_23111_cast_fp16 = mul(x = var_23109_cast_fp16, y = var_23110_to_fp16)[name = string("op_23111_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_855_cast_fp16 = add(x = input_847_cast_fp16, y = var_23111_cast_fp16)[name = string("input_855_cast_fp16")];
            tensor<int32, [1]> q_37_axes_1 = const()[name = string("q_37_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_18_norm_self_att_weight_to_fp16 = const()[name = string("layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540797568)))];
            tensor<fp16, [1024]> layers_18_norm_self_att_bias_to_fp16 = const()[name = string("layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540799680)))];
            tensor<fp16, [1, 51, 1024]> q_37_cast_fp16 = layer_norm(axes = q_37_axes_1, beta = layers_18_norm_self_att_bias_to_fp16, epsilon = var_23054_to_fp16, gamma = layers_18_norm_self_att_weight_to_fp16, x = input_855_cast_fp16)[name = string("q_37_cast_fp16")];
            tensor<int32, [3]> var_23185 = const()[name = string("op_23185"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_857_axes_1 = const()[name = string("input_857_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_23186_cast_fp16 = transpose(perm = var_23185, x = q_37_cast_fp16)[name = string("transpose_243")];
            tensor<fp16, [1, 1024, 51, 1]> input_857_cast_fp16 = expand_dims(axes = input_857_axes_1, x = var_23186_cast_fp16)[name = string("input_857_cast_fp16")];
            string dense_output_1457_pad_type_1 = const()[name = string("dense_output_1457_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1457_strides_1 = const()[name = string("dense_output_1457_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1457_pad_1 = const()[name = string("dense_output_1457_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1457_dilations_1 = const()[name = string("dense_output_1457_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1457_groups_1 = const()[name = string("dense_output_1457_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540801792))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540932928))))[name = string("layers_18_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1457_cast_fp16 = conv(dilations = dense_output_1457_dilations_1, groups = dense_output_1457_groups_1, pad = dense_output_1457_pad_1, pad_type = dense_output_1457_pad_type_1, strides = dense_output_1457_strides_1, weight = layers_18_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1457_cast_fp16")];
            string sparse_output_1457_pad_type_1 = const()[name = string("sparse_output_1457_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1457_strides_1 = const()[name = string("sparse_output_1457_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1457_pad_1 = const()[name = string("sparse_output_1457_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1457_dilations_1 = const()[name = string("sparse_output_1457_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1457_groups_1 = const()[name = string("sparse_output_1457_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540936192))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540933504))))[name = string("layers_18_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1457_cast_fp16 = conv(dilations = sparse_output_1457_dilations_1, groups = sparse_output_1457_groups_1, pad = sparse_output_1457_pad_1, pad_type = sparse_output_1457_pad_type_1, strides = sparse_output_1457_strides_1, weight = layers_18_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1457_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23211_cast_fp16 = add(x = dense_output_1457_cast_fp16, y = sparse_output_1457_cast_fp16)[name = string("op_23211_cast_fp16")];
            tensor<int32, [4]> var_23212 = const()[name = string("op_23212"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23214 = const()[name = string("op_23214"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23213_cast_fp16 = transpose(perm = var_23212, x = var_23211_cast_fp16)[name = string("transpose_242")];
            tensor<fp16, [1, 51, 128]> q_head_289_cast_fp16 = reshape(shape = var_23214, x = var_23213_cast_fp16)[name = string("q_head_289_cast_fp16")];
            string dense_output_1459_pad_type_1 = const()[name = string("dense_output_1459_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1459_strides_1 = const()[name = string("dense_output_1459_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1459_pad_1 = const()[name = string("dense_output_1459_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1459_dilations_1 = const()[name = string("dense_output_1459_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1459_groups_1 = const()[name = string("dense_output_1459_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540952640))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541083776))))[name = string("layers_18_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1459_cast_fp16 = conv(dilations = dense_output_1459_dilations_1, groups = dense_output_1459_groups_1, pad = dense_output_1459_pad_1, pad_type = dense_output_1459_pad_type_1, strides = dense_output_1459_strides_1, weight = layers_18_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1459_cast_fp16")];
            string sparse_output_1459_pad_type_1 = const()[name = string("sparse_output_1459_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1459_strides_1 = const()[name = string("sparse_output_1459_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1459_pad_1 = const()[name = string("sparse_output_1459_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1459_dilations_1 = const()[name = string("sparse_output_1459_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1459_groups_1 = const()[name = string("sparse_output_1459_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541087040))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541084352))))[name = string("layers_18_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1459_cast_fp16 = conv(dilations = sparse_output_1459_dilations_1, groups = sparse_output_1459_groups_1, pad = sparse_output_1459_pad_1, pad_type = sparse_output_1459_pad_type_1, strides = sparse_output_1459_strides_1, weight = layers_18_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1459_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23230_cast_fp16 = add(x = dense_output_1459_cast_fp16, y = sparse_output_1459_cast_fp16)[name = string("op_23230_cast_fp16")];
            tensor<int32, [4]> var_23231 = const()[name = string("op_23231"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23233 = const()[name = string("op_23233"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23232_cast_fp16 = transpose(perm = var_23231, x = var_23230_cast_fp16)[name = string("transpose_241")];
            tensor<fp16, [1, 51, 128]> k_head_577_cast_fp16 = reshape(shape = var_23233, x = var_23232_cast_fp16)[name = string("k_head_577_cast_fp16")];
            string dense_output_1461_pad_type_1 = const()[name = string("dense_output_1461_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1461_strides_1 = const()[name = string("dense_output_1461_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1461_pad_1 = const()[name = string("dense_output_1461_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1461_dilations_1 = const()[name = string("dense_output_1461_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1461_groups_1 = const()[name = string("dense_output_1461_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541103488))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541234624))))[name = string("layers_18_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1461_cast_fp16 = conv(dilations = dense_output_1461_dilations_1, groups = dense_output_1461_groups_1, pad = dense_output_1461_pad_1, pad_type = dense_output_1461_pad_type_1, strides = dense_output_1461_strides_1, weight = layers_18_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1461_cast_fp16")];
            string sparse_output_1461_pad_type_1 = const()[name = string("sparse_output_1461_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1461_strides_1 = const()[name = string("sparse_output_1461_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1461_pad_1 = const()[name = string("sparse_output_1461_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1461_dilations_1 = const()[name = string("sparse_output_1461_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1461_groups_1 = const()[name = string("sparse_output_1461_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541237888))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541235200))))[name = string("layers_18_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1461_cast_fp16 = conv(dilations = sparse_output_1461_dilations_1, groups = sparse_output_1461_groups_1, pad = sparse_output_1461_pad_1, pad_type = sparse_output_1461_pad_type_1, strides = sparse_output_1461_strides_1, weight = layers_18_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1461_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23249_cast_fp16 = add(x = dense_output_1461_cast_fp16, y = sparse_output_1461_cast_fp16)[name = string("op_23249_cast_fp16")];
            tensor<int32, [4]> var_23250 = const()[name = string("op_23250"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23252 = const()[name = string("op_23252"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23251_cast_fp16 = transpose(perm = var_23250, x = var_23249_cast_fp16)[name = string("transpose_240")];
            tensor<fp16, [1, 51, 128]> v_head_577_cast_fp16 = reshape(shape = var_23252, x = var_23251_cast_fp16)[name = string("v_head_577_cast_fp16")];
            string dense_output_1463_pad_type_1 = const()[name = string("dense_output_1463_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1463_strides_1 = const()[name = string("dense_output_1463_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1463_pad_1 = const()[name = string("dense_output_1463_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1463_dilations_1 = const()[name = string("dense_output_1463_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1463_groups_1 = const()[name = string("dense_output_1463_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541254336))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541385472))))[name = string("layers_18_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1463_cast_fp16 = conv(dilations = dense_output_1463_dilations_1, groups = dense_output_1463_groups_1, pad = dense_output_1463_pad_1, pad_type = dense_output_1463_pad_type_1, strides = dense_output_1463_strides_1, weight = layers_18_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1463_cast_fp16")];
            string sparse_output_1463_pad_type_1 = const()[name = string("sparse_output_1463_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1463_strides_1 = const()[name = string("sparse_output_1463_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1463_pad_1 = const()[name = string("sparse_output_1463_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1463_dilations_1 = const()[name = string("sparse_output_1463_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1463_groups_1 = const()[name = string("sparse_output_1463_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541388736))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541386048))))[name = string("layers_18_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1463_cast_fp16 = conv(dilations = sparse_output_1463_dilations_1, groups = sparse_output_1463_groups_1, pad = sparse_output_1463_pad_1, pad_type = sparse_output_1463_pad_type_1, strides = sparse_output_1463_strides_1, weight = layers_18_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1463_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_23268_cast_fp16 = add(x = dense_output_1463_cast_fp16, y = sparse_output_1463_cast_fp16)[name = string("op_23268_cast_fp16")];
            tensor<int32, [4]> var_23269 = const()[name = string("op_23269"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23271 = const()[name = string("op_23271"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_23270_cast_fp16 = transpose(perm = var_23269, x = var_23268_cast_fp16)[name = string("transpose_239")];
            tensor<fp16, [1, 101, 128]> p_head_577_cast_fp16 = reshape(shape = var_23271, x = var_23270_cast_fp16)[name = string("p_head_577_cast_fp16")];
            tensor<fp16, [128]> var_23273_to_fp16 = const()[name = string("op_23273_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541405184)))];
            tensor<fp16, [1, 51, 128]> var_23274_cast_fp16 = add(x = q_head_289_cast_fp16, y = var_23273_to_fp16)[name = string("op_23274_cast_fp16")];
            tensor<int32, [1]> q_u_289_axes_1 = const()[name = string("q_u_289_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_289_cast_fp16 = expand_dims(axes = q_u_289_axes_1, x = var_23274_cast_fp16)[name = string("q_u_289_cast_fp16")];
            tensor<fp16, [128]> var_23276_to_fp16 = const()[name = string("op_23276_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541405504)))];
            tensor<fp16, [1, 51, 128]> var_23277_cast_fp16 = add(x = q_head_289_cast_fp16, y = var_23276_to_fp16)[name = string("op_23277_cast_fp16")];
            tensor<int32, [1]> q_v_289_axes_1 = const()[name = string("q_v_289_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_289_cast_fp16 = expand_dims(axes = q_v_289_axes_1, x = var_23277_cast_fp16)[name = string("q_v_289_cast_fp16")];
            tensor<int32, [1]> k_head_579_axes_1 = const()[name = string("k_head_579_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_579_cast_fp16 = expand_dims(axes = k_head_579_axes_1, x = k_head_577_cast_fp16)[name = string("k_head_579_cast_fp16")];
            tensor<int32, [1]> v_head_579_axes_1 = const()[name = string("v_head_579_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_579_cast_fp16 = expand_dims(axes = v_head_579_axes_1, x = v_head_577_cast_fp16)[name = string("v_head_579_cast_fp16")];
            tensor<int32, [1]> p_head_579_axes_1 = const()[name = string("p_head_579_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_579_cast_fp16 = expand_dims(axes = p_head_579_axes_1, x = p_head_577_cast_fp16)[name = string("p_head_579_cast_fp16")];
            bool var_23283_transpose_x_3 = const()[name = string("op_23283_transpose_x_3"), val = bool(false)];
            bool var_23283_transpose_y_3 = const()[name = string("op_23283_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_23283_cast_fp16 = matmul(transpose_x = var_23283_transpose_x_3, transpose_y = var_23283_transpose_y_3, x = q_u_289_cast_fp16, y = k_head_579_cast_fp16)[name = string("op_23283_cast_fp16")];
            fp16 var_23284_to_fp16 = const()[name = string("op_23284_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_289_cast_fp16 = mul(x = var_23283_cast_fp16, y = var_23284_to_fp16)[name = string("scores_content_289_cast_fp16")];
            bool x_1533_transpose_x_3 = const()[name = string("x_1533_transpose_x_3"), val = bool(false)];
            bool x_1533_transpose_y_3 = const()[name = string("x_1533_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1533_cast_fp16 = matmul(transpose_x = x_1533_transpose_x_3, transpose_y = x_1533_transpose_y_3, x = q_v_289_cast_fp16, y = p_head_579_cast_fp16)[name = string("x_1533_cast_fp16")];
            tensor<int32, [8]> x_1535_pad_1 = const()[name = string("x_1535_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1535_mode_1 = const()[name = string("x_1535_mode_1"), val = string("constant")];
            fp16 const_2281_to_fp16 = const()[name = string("const_2281_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1535_cast_fp16 = pad(constant_val = const_2281_to_fp16, mode = x_1535_mode_1, pad = x_1535_pad_1, x = x_1533_cast_fp16)[name = string("x_1535_cast_fp16")];
            tensor<int32, [4]> var_23298 = const()[name = string("op_23298"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1537_cast_fp16 = reshape(shape = var_23298, x = x_1535_cast_fp16)[name = string("x_1537_cast_fp16")];
            tensor<int32, [4]> var_23302_begin_1 = const()[name = string("op_23302_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_23302_end_1 = const()[name = string("op_23302_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_23302_end_mask_1 = const()[name = string("op_23302_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_23302_cast_fp16 = slice_by_index(begin = var_23302_begin_1, end = var_23302_end_1, end_mask = var_23302_end_mask_1, x = x_1537_cast_fp16)[name = string("op_23302_cast_fp16")];
            tensor<int32, [4]> var_23304 = const()[name = string("op_23304"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_23305_cast_fp16 = reshape(shape = var_23304, x = var_23302_cast_fp16)[name = string("op_23305_cast_fp16")];
            tensor<int32, [4]> var_23310_begin_1 = const()[name = string("op_23310_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_23310_end_1 = const()[name = string("op_23310_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_23310_end_mask_1 = const()[name = string("op_23310_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_23310_cast_fp16 = slice_by_index(begin = var_23310_begin_1, end = var_23310_end_1, end_mask = var_23310_end_mask_1, x = var_23305_cast_fp16)[name = string("op_23310_cast_fp16")];
            fp16 var_23311_to_fp16 = const()[name = string("op_23311_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_289_cast_fp16 = mul(x = var_23310_cast_fp16, y = var_23311_to_fp16)[name = string("scores_pos_289_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_289_cast_fp16 = add(x = scores_content_289_cast_fp16, y = scores_pos_289_cast_fp16)[name = string("logits_289_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_23314_cast_fp16 = softmax(axis = var_23039, x = logits_289_cast_fp16)[name = string("op_23314_cast_fp16")];
            bool var_23316_transpose_x_1 = const()[name = string("op_23316_transpose_x_1"), val = bool(false)];
            bool var_23316_transpose_y_1 = const()[name = string("op_23316_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_23316_cast_fp16 = matmul(transpose_x = var_23316_transpose_x_1, transpose_y = var_23316_transpose_y_1, x = var_23314_cast_fp16, y = v_head_579_cast_fp16)[name = string("op_23316_cast_fp16")];
            tensor<int32, [1]> var_23317_axes_1 = const()[name = string("op_23317_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_23317_cast_fp16 = squeeze(axes = var_23317_axes_1, x = var_23316_cast_fp16)[name = string("op_23317_cast_fp16")];
            string dense_output_1465_pad_type_1 = const()[name = string("dense_output_1465_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1465_strides_1 = const()[name = string("dense_output_1465_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1465_pad_1 = const()[name = string("dense_output_1465_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1465_dilations_1 = const()[name = string("dense_output_1465_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1465_groups_1 = const()[name = string("dense_output_1465_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541405824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541536960))))[name = string("layers_18_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1465_cast_fp16 = conv(dilations = dense_output_1465_dilations_1, groups = dense_output_1465_groups_1, pad = dense_output_1465_pad_1, pad_type = dense_output_1465_pad_type_1, strides = dense_output_1465_strides_1, weight = layers_18_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1465_cast_fp16")];
            string sparse_output_1465_pad_type_1 = const()[name = string("sparse_output_1465_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1465_strides_1 = const()[name = string("sparse_output_1465_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1465_pad_1 = const()[name = string("sparse_output_1465_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1465_dilations_1 = const()[name = string("sparse_output_1465_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1465_groups_1 = const()[name = string("sparse_output_1465_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541540224))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541537536))))[name = string("layers_18_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1465_cast_fp16 = conv(dilations = sparse_output_1465_dilations_1, groups = sparse_output_1465_groups_1, pad = sparse_output_1465_pad_1, pad_type = sparse_output_1465_pad_type_1, strides = sparse_output_1465_strides_1, weight = layers_18_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1465_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23332_cast_fp16 = add(x = dense_output_1465_cast_fp16, y = sparse_output_1465_cast_fp16)[name = string("op_23332_cast_fp16")];
            tensor<int32, [4]> var_23333 = const()[name = string("op_23333"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23335 = const()[name = string("op_23335"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23334_cast_fp16 = transpose(perm = var_23333, x = var_23332_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 51, 128]> q_head_291_cast_fp16 = reshape(shape = var_23335, x = var_23334_cast_fp16)[name = string("q_head_291_cast_fp16")];
            string dense_output_1467_pad_type_1 = const()[name = string("dense_output_1467_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1467_strides_1 = const()[name = string("dense_output_1467_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1467_pad_1 = const()[name = string("dense_output_1467_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1467_dilations_1 = const()[name = string("dense_output_1467_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1467_groups_1 = const()[name = string("dense_output_1467_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541556672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541687808))))[name = string("layers_18_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1467_cast_fp16 = conv(dilations = dense_output_1467_dilations_1, groups = dense_output_1467_groups_1, pad = dense_output_1467_pad_1, pad_type = dense_output_1467_pad_type_1, strides = dense_output_1467_strides_1, weight = layers_18_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1467_cast_fp16")];
            string sparse_output_1467_pad_type_1 = const()[name = string("sparse_output_1467_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1467_strides_1 = const()[name = string("sparse_output_1467_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1467_pad_1 = const()[name = string("sparse_output_1467_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1467_dilations_1 = const()[name = string("sparse_output_1467_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1467_groups_1 = const()[name = string("sparse_output_1467_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541691072))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541688384))))[name = string("layers_18_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1467_cast_fp16 = conv(dilations = sparse_output_1467_dilations_1, groups = sparse_output_1467_groups_1, pad = sparse_output_1467_pad_1, pad_type = sparse_output_1467_pad_type_1, strides = sparse_output_1467_strides_1, weight = layers_18_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1467_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23351_cast_fp16 = add(x = dense_output_1467_cast_fp16, y = sparse_output_1467_cast_fp16)[name = string("op_23351_cast_fp16")];
            tensor<int32, [4]> var_23352 = const()[name = string("op_23352"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23354 = const()[name = string("op_23354"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23353_cast_fp16 = transpose(perm = var_23352, x = var_23351_cast_fp16)[name = string("transpose_237")];
            tensor<fp16, [1, 51, 128]> k_head_581_cast_fp16 = reshape(shape = var_23354, x = var_23353_cast_fp16)[name = string("k_head_581_cast_fp16")];
            string dense_output_1469_pad_type_1 = const()[name = string("dense_output_1469_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1469_strides_1 = const()[name = string("dense_output_1469_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1469_pad_1 = const()[name = string("dense_output_1469_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1469_dilations_1 = const()[name = string("dense_output_1469_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1469_groups_1 = const()[name = string("dense_output_1469_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541707520))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541838656))))[name = string("layers_18_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1469_cast_fp16 = conv(dilations = dense_output_1469_dilations_1, groups = dense_output_1469_groups_1, pad = dense_output_1469_pad_1, pad_type = dense_output_1469_pad_type_1, strides = dense_output_1469_strides_1, weight = layers_18_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1469_cast_fp16")];
            string sparse_output_1469_pad_type_1 = const()[name = string("sparse_output_1469_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1469_strides_1 = const()[name = string("sparse_output_1469_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1469_pad_1 = const()[name = string("sparse_output_1469_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1469_dilations_1 = const()[name = string("sparse_output_1469_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1469_groups_1 = const()[name = string("sparse_output_1469_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541841920))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541839232))))[name = string("layers_18_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1469_cast_fp16 = conv(dilations = sparse_output_1469_dilations_1, groups = sparse_output_1469_groups_1, pad = sparse_output_1469_pad_1, pad_type = sparse_output_1469_pad_type_1, strides = sparse_output_1469_strides_1, weight = layers_18_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1469_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23370_cast_fp16 = add(x = dense_output_1469_cast_fp16, y = sparse_output_1469_cast_fp16)[name = string("op_23370_cast_fp16")];
            tensor<int32, [4]> var_23371 = const()[name = string("op_23371"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23373 = const()[name = string("op_23373"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23372_cast_fp16 = transpose(perm = var_23371, x = var_23370_cast_fp16)[name = string("transpose_236")];
            tensor<fp16, [1, 51, 128]> v_head_581_cast_fp16 = reshape(shape = var_23373, x = var_23372_cast_fp16)[name = string("v_head_581_cast_fp16")];
            string dense_output_1471_pad_type_1 = const()[name = string("dense_output_1471_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1471_strides_1 = const()[name = string("dense_output_1471_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1471_pad_1 = const()[name = string("dense_output_1471_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1471_dilations_1 = const()[name = string("dense_output_1471_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1471_groups_1 = const()[name = string("dense_output_1471_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541858368))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541989504))))[name = string("layers_18_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1471_cast_fp16 = conv(dilations = dense_output_1471_dilations_1, groups = dense_output_1471_groups_1, pad = dense_output_1471_pad_1, pad_type = dense_output_1471_pad_type_1, strides = dense_output_1471_strides_1, weight = layers_18_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1471_cast_fp16")];
            string sparse_output_1471_pad_type_1 = const()[name = string("sparse_output_1471_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1471_strides_1 = const()[name = string("sparse_output_1471_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1471_pad_1 = const()[name = string("sparse_output_1471_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1471_dilations_1 = const()[name = string("sparse_output_1471_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1471_groups_1 = const()[name = string("sparse_output_1471_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541992768))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541990080))))[name = string("layers_18_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1471_cast_fp16 = conv(dilations = sparse_output_1471_dilations_1, groups = sparse_output_1471_groups_1, pad = sparse_output_1471_pad_1, pad_type = sparse_output_1471_pad_type_1, strides = sparse_output_1471_strides_1, weight = layers_18_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1471_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_23389_cast_fp16 = add(x = dense_output_1471_cast_fp16, y = sparse_output_1471_cast_fp16)[name = string("op_23389_cast_fp16")];
            tensor<int32, [4]> var_23390 = const()[name = string("op_23390"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23392 = const()[name = string("op_23392"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_23391_cast_fp16 = transpose(perm = var_23390, x = var_23389_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [1, 101, 128]> p_head_581_cast_fp16 = reshape(shape = var_23392, x = var_23391_cast_fp16)[name = string("p_head_581_cast_fp16")];
            tensor<fp16, [128]> var_23394_to_fp16 = const()[name = string("op_23394_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542009216)))];
            tensor<fp16, [1, 51, 128]> var_23395_cast_fp16 = add(x = q_head_291_cast_fp16, y = var_23394_to_fp16)[name = string("op_23395_cast_fp16")];
            tensor<int32, [1]> q_u_291_axes_1 = const()[name = string("q_u_291_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_291_cast_fp16 = expand_dims(axes = q_u_291_axes_1, x = var_23395_cast_fp16)[name = string("q_u_291_cast_fp16")];
            tensor<fp16, [128]> var_23397_to_fp16 = const()[name = string("op_23397_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542009536)))];
            tensor<fp16, [1, 51, 128]> var_23398_cast_fp16 = add(x = q_head_291_cast_fp16, y = var_23397_to_fp16)[name = string("op_23398_cast_fp16")];
            tensor<int32, [1]> q_v_291_axes_1 = const()[name = string("q_v_291_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_291_cast_fp16 = expand_dims(axes = q_v_291_axes_1, x = var_23398_cast_fp16)[name = string("q_v_291_cast_fp16")];
            tensor<int32, [1]> k_head_583_axes_1 = const()[name = string("k_head_583_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_583_cast_fp16 = expand_dims(axes = k_head_583_axes_1, x = k_head_581_cast_fp16)[name = string("k_head_583_cast_fp16")];
            tensor<int32, [1]> v_head_583_axes_1 = const()[name = string("v_head_583_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_583_cast_fp16 = expand_dims(axes = v_head_583_axes_1, x = v_head_581_cast_fp16)[name = string("v_head_583_cast_fp16")];
            tensor<int32, [1]> p_head_583_axes_1 = const()[name = string("p_head_583_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_583_cast_fp16 = expand_dims(axes = p_head_583_axes_1, x = p_head_581_cast_fp16)[name = string("p_head_583_cast_fp16")];
            bool var_23404_transpose_x_3 = const()[name = string("op_23404_transpose_x_3"), val = bool(false)];
            bool var_23404_transpose_y_3 = const()[name = string("op_23404_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_23404_cast_fp16 = matmul(transpose_x = var_23404_transpose_x_3, transpose_y = var_23404_transpose_y_3, x = q_u_291_cast_fp16, y = k_head_583_cast_fp16)[name = string("op_23404_cast_fp16")];
            fp16 var_23405_to_fp16 = const()[name = string("op_23405_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_291_cast_fp16 = mul(x = var_23404_cast_fp16, y = var_23405_to_fp16)[name = string("scores_content_291_cast_fp16")];
            bool x_1541_transpose_x_3 = const()[name = string("x_1541_transpose_x_3"), val = bool(false)];
            bool x_1541_transpose_y_3 = const()[name = string("x_1541_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1541_cast_fp16 = matmul(transpose_x = x_1541_transpose_x_3, transpose_y = x_1541_transpose_y_3, x = q_v_291_cast_fp16, y = p_head_583_cast_fp16)[name = string("x_1541_cast_fp16")];
            tensor<int32, [8]> x_1543_pad_1 = const()[name = string("x_1543_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1543_mode_1 = const()[name = string("x_1543_mode_1"), val = string("constant")];
            fp16 const_2287_to_fp16 = const()[name = string("const_2287_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1543_cast_fp16 = pad(constant_val = const_2287_to_fp16, mode = x_1543_mode_1, pad = x_1543_pad_1, x = x_1541_cast_fp16)[name = string("x_1543_cast_fp16")];
            tensor<int32, [4]> var_23419 = const()[name = string("op_23419"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1545_cast_fp16 = reshape(shape = var_23419, x = x_1543_cast_fp16)[name = string("x_1545_cast_fp16")];
            tensor<int32, [4]> var_23423_begin_1 = const()[name = string("op_23423_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_23423_end_1 = const()[name = string("op_23423_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_23423_end_mask_1 = const()[name = string("op_23423_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_23423_cast_fp16 = slice_by_index(begin = var_23423_begin_1, end = var_23423_end_1, end_mask = var_23423_end_mask_1, x = x_1545_cast_fp16)[name = string("op_23423_cast_fp16")];
            tensor<int32, [4]> var_23425 = const()[name = string("op_23425"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_23426_cast_fp16 = reshape(shape = var_23425, x = var_23423_cast_fp16)[name = string("op_23426_cast_fp16")];
            tensor<int32, [4]> var_23431_begin_1 = const()[name = string("op_23431_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_23431_end_1 = const()[name = string("op_23431_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_23431_end_mask_1 = const()[name = string("op_23431_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_23431_cast_fp16 = slice_by_index(begin = var_23431_begin_1, end = var_23431_end_1, end_mask = var_23431_end_mask_1, x = var_23426_cast_fp16)[name = string("op_23431_cast_fp16")];
            fp16 var_23432_to_fp16 = const()[name = string("op_23432_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_291_cast_fp16 = mul(x = var_23431_cast_fp16, y = var_23432_to_fp16)[name = string("scores_pos_291_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_291_cast_fp16 = add(x = scores_content_291_cast_fp16, y = scores_pos_291_cast_fp16)[name = string("logits_291_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_23435_cast_fp16 = softmax(axis = var_23039, x = logits_291_cast_fp16)[name = string("op_23435_cast_fp16")];
            bool var_23437_transpose_x_1 = const()[name = string("op_23437_transpose_x_1"), val = bool(false)];
            bool var_23437_transpose_y_1 = const()[name = string("op_23437_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_23437_cast_fp16 = matmul(transpose_x = var_23437_transpose_x_1, transpose_y = var_23437_transpose_y_1, x = var_23435_cast_fp16, y = v_head_583_cast_fp16)[name = string("op_23437_cast_fp16")];
            tensor<int32, [1]> var_23438_axes_1 = const()[name = string("op_23438_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_23438_cast_fp16 = squeeze(axes = var_23438_axes_1, x = var_23437_cast_fp16)[name = string("op_23438_cast_fp16")];
            string dense_output_1473_pad_type_1 = const()[name = string("dense_output_1473_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1473_strides_1 = const()[name = string("dense_output_1473_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1473_pad_1 = const()[name = string("dense_output_1473_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1473_dilations_1 = const()[name = string("dense_output_1473_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1473_groups_1 = const()[name = string("dense_output_1473_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542009856))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542140992))))[name = string("layers_18_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1473_cast_fp16 = conv(dilations = dense_output_1473_dilations_1, groups = dense_output_1473_groups_1, pad = dense_output_1473_pad_1, pad_type = dense_output_1473_pad_type_1, strides = dense_output_1473_strides_1, weight = layers_18_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1473_cast_fp16")];
            string sparse_output_1473_pad_type_1 = const()[name = string("sparse_output_1473_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1473_strides_1 = const()[name = string("sparse_output_1473_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1473_pad_1 = const()[name = string("sparse_output_1473_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1473_dilations_1 = const()[name = string("sparse_output_1473_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1473_groups_1 = const()[name = string("sparse_output_1473_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542144256))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542141568))))[name = string("layers_18_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1473_cast_fp16 = conv(dilations = sparse_output_1473_dilations_1, groups = sparse_output_1473_groups_1, pad = sparse_output_1473_pad_1, pad_type = sparse_output_1473_pad_type_1, strides = sparse_output_1473_strides_1, weight = layers_18_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1473_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23453_cast_fp16 = add(x = dense_output_1473_cast_fp16, y = sparse_output_1473_cast_fp16)[name = string("op_23453_cast_fp16")];
            tensor<int32, [4]> var_23454 = const()[name = string("op_23454"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23456 = const()[name = string("op_23456"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23455_cast_fp16 = transpose(perm = var_23454, x = var_23453_cast_fp16)[name = string("transpose_234")];
            tensor<fp16, [1, 51, 128]> q_head_293_cast_fp16 = reshape(shape = var_23456, x = var_23455_cast_fp16)[name = string("q_head_293_cast_fp16")];
            string dense_output_1475_pad_type_1 = const()[name = string("dense_output_1475_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1475_strides_1 = const()[name = string("dense_output_1475_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1475_pad_1 = const()[name = string("dense_output_1475_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1475_dilations_1 = const()[name = string("dense_output_1475_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1475_groups_1 = const()[name = string("dense_output_1475_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542160704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542291840))))[name = string("layers_18_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1475_cast_fp16 = conv(dilations = dense_output_1475_dilations_1, groups = dense_output_1475_groups_1, pad = dense_output_1475_pad_1, pad_type = dense_output_1475_pad_type_1, strides = dense_output_1475_strides_1, weight = layers_18_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1475_cast_fp16")];
            string sparse_output_1475_pad_type_1 = const()[name = string("sparse_output_1475_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1475_strides_1 = const()[name = string("sparse_output_1475_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1475_pad_1 = const()[name = string("sparse_output_1475_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1475_dilations_1 = const()[name = string("sparse_output_1475_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1475_groups_1 = const()[name = string("sparse_output_1475_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542295104))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542292416))))[name = string("layers_18_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1475_cast_fp16 = conv(dilations = sparse_output_1475_dilations_1, groups = sparse_output_1475_groups_1, pad = sparse_output_1475_pad_1, pad_type = sparse_output_1475_pad_type_1, strides = sparse_output_1475_strides_1, weight = layers_18_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1475_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23472_cast_fp16 = add(x = dense_output_1475_cast_fp16, y = sparse_output_1475_cast_fp16)[name = string("op_23472_cast_fp16")];
            tensor<int32, [4]> var_23473 = const()[name = string("op_23473"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23475 = const()[name = string("op_23475"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23474_cast_fp16 = transpose(perm = var_23473, x = var_23472_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [1, 51, 128]> k_head_585_cast_fp16 = reshape(shape = var_23475, x = var_23474_cast_fp16)[name = string("k_head_585_cast_fp16")];
            string dense_output_1477_pad_type_1 = const()[name = string("dense_output_1477_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1477_strides_1 = const()[name = string("dense_output_1477_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1477_pad_1 = const()[name = string("dense_output_1477_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1477_dilations_1 = const()[name = string("dense_output_1477_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1477_groups_1 = const()[name = string("dense_output_1477_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542311552))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542442688))))[name = string("layers_18_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1477_cast_fp16 = conv(dilations = dense_output_1477_dilations_1, groups = dense_output_1477_groups_1, pad = dense_output_1477_pad_1, pad_type = dense_output_1477_pad_type_1, strides = dense_output_1477_strides_1, weight = layers_18_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1477_cast_fp16")];
            string sparse_output_1477_pad_type_1 = const()[name = string("sparse_output_1477_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1477_strides_1 = const()[name = string("sparse_output_1477_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1477_pad_1 = const()[name = string("sparse_output_1477_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1477_dilations_1 = const()[name = string("sparse_output_1477_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1477_groups_1 = const()[name = string("sparse_output_1477_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542445952))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542443264))))[name = string("layers_18_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1477_cast_fp16 = conv(dilations = sparse_output_1477_dilations_1, groups = sparse_output_1477_groups_1, pad = sparse_output_1477_pad_1, pad_type = sparse_output_1477_pad_type_1, strides = sparse_output_1477_strides_1, weight = layers_18_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1477_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23491_cast_fp16 = add(x = dense_output_1477_cast_fp16, y = sparse_output_1477_cast_fp16)[name = string("op_23491_cast_fp16")];
            tensor<int32, [4]> var_23492 = const()[name = string("op_23492"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23494 = const()[name = string("op_23494"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23493_cast_fp16 = transpose(perm = var_23492, x = var_23491_cast_fp16)[name = string("transpose_232")];
            tensor<fp16, [1, 51, 128]> v_head_585_cast_fp16 = reshape(shape = var_23494, x = var_23493_cast_fp16)[name = string("v_head_585_cast_fp16")];
            string dense_output_1479_pad_type_1 = const()[name = string("dense_output_1479_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1479_strides_1 = const()[name = string("dense_output_1479_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1479_pad_1 = const()[name = string("dense_output_1479_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1479_dilations_1 = const()[name = string("dense_output_1479_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1479_groups_1 = const()[name = string("dense_output_1479_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542462400))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542593536))))[name = string("layers_18_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1479_cast_fp16 = conv(dilations = dense_output_1479_dilations_1, groups = dense_output_1479_groups_1, pad = dense_output_1479_pad_1, pad_type = dense_output_1479_pad_type_1, strides = dense_output_1479_strides_1, weight = layers_18_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1479_cast_fp16")];
            string sparse_output_1479_pad_type_1 = const()[name = string("sparse_output_1479_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1479_strides_1 = const()[name = string("sparse_output_1479_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1479_pad_1 = const()[name = string("sparse_output_1479_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1479_dilations_1 = const()[name = string("sparse_output_1479_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1479_groups_1 = const()[name = string("sparse_output_1479_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542596800))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542594112))))[name = string("layers_18_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1479_cast_fp16 = conv(dilations = sparse_output_1479_dilations_1, groups = sparse_output_1479_groups_1, pad = sparse_output_1479_pad_1, pad_type = sparse_output_1479_pad_type_1, strides = sparse_output_1479_strides_1, weight = layers_18_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1479_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_23510_cast_fp16 = add(x = dense_output_1479_cast_fp16, y = sparse_output_1479_cast_fp16)[name = string("op_23510_cast_fp16")];
            tensor<int32, [4]> var_23511 = const()[name = string("op_23511"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23513 = const()[name = string("op_23513"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_23512_cast_fp16 = transpose(perm = var_23511, x = var_23510_cast_fp16)[name = string("transpose_231")];
            tensor<fp16, [1, 101, 128]> p_head_585_cast_fp16 = reshape(shape = var_23513, x = var_23512_cast_fp16)[name = string("p_head_585_cast_fp16")];
            tensor<fp16, [128]> var_23515_to_fp16 = const()[name = string("op_23515_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542613248)))];
            tensor<fp16, [1, 51, 128]> var_23516_cast_fp16 = add(x = q_head_293_cast_fp16, y = var_23515_to_fp16)[name = string("op_23516_cast_fp16")];
            tensor<int32, [1]> q_u_293_axes_1 = const()[name = string("q_u_293_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_293_cast_fp16 = expand_dims(axes = q_u_293_axes_1, x = var_23516_cast_fp16)[name = string("q_u_293_cast_fp16")];
            tensor<fp16, [128]> var_23518_to_fp16 = const()[name = string("op_23518_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542613568)))];
            tensor<fp16, [1, 51, 128]> var_23519_cast_fp16 = add(x = q_head_293_cast_fp16, y = var_23518_to_fp16)[name = string("op_23519_cast_fp16")];
            tensor<int32, [1]> q_v_293_axes_1 = const()[name = string("q_v_293_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_293_cast_fp16 = expand_dims(axes = q_v_293_axes_1, x = var_23519_cast_fp16)[name = string("q_v_293_cast_fp16")];
            tensor<int32, [1]> k_head_587_axes_1 = const()[name = string("k_head_587_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_587_cast_fp16 = expand_dims(axes = k_head_587_axes_1, x = k_head_585_cast_fp16)[name = string("k_head_587_cast_fp16")];
            tensor<int32, [1]> v_head_587_axes_1 = const()[name = string("v_head_587_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_587_cast_fp16 = expand_dims(axes = v_head_587_axes_1, x = v_head_585_cast_fp16)[name = string("v_head_587_cast_fp16")];
            tensor<int32, [1]> p_head_587_axes_1 = const()[name = string("p_head_587_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_587_cast_fp16 = expand_dims(axes = p_head_587_axes_1, x = p_head_585_cast_fp16)[name = string("p_head_587_cast_fp16")];
            bool var_23525_transpose_x_3 = const()[name = string("op_23525_transpose_x_3"), val = bool(false)];
            bool var_23525_transpose_y_3 = const()[name = string("op_23525_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_23525_cast_fp16 = matmul(transpose_x = var_23525_transpose_x_3, transpose_y = var_23525_transpose_y_3, x = q_u_293_cast_fp16, y = k_head_587_cast_fp16)[name = string("op_23525_cast_fp16")];
            fp16 var_23526_to_fp16 = const()[name = string("op_23526_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_293_cast_fp16 = mul(x = var_23525_cast_fp16, y = var_23526_to_fp16)[name = string("scores_content_293_cast_fp16")];
            bool x_1549_transpose_x_3 = const()[name = string("x_1549_transpose_x_3"), val = bool(false)];
            bool x_1549_transpose_y_3 = const()[name = string("x_1549_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1549_cast_fp16 = matmul(transpose_x = x_1549_transpose_x_3, transpose_y = x_1549_transpose_y_3, x = q_v_293_cast_fp16, y = p_head_587_cast_fp16)[name = string("x_1549_cast_fp16")];
            tensor<int32, [8]> x_1551_pad_1 = const()[name = string("x_1551_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1551_mode_1 = const()[name = string("x_1551_mode_1"), val = string("constant")];
            fp16 const_2293_to_fp16 = const()[name = string("const_2293_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1551_cast_fp16 = pad(constant_val = const_2293_to_fp16, mode = x_1551_mode_1, pad = x_1551_pad_1, x = x_1549_cast_fp16)[name = string("x_1551_cast_fp16")];
            tensor<int32, [4]> var_23540 = const()[name = string("op_23540"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1553_cast_fp16 = reshape(shape = var_23540, x = x_1551_cast_fp16)[name = string("x_1553_cast_fp16")];
            tensor<int32, [4]> var_23544_begin_1 = const()[name = string("op_23544_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_23544_end_1 = const()[name = string("op_23544_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_23544_end_mask_1 = const()[name = string("op_23544_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_23544_cast_fp16 = slice_by_index(begin = var_23544_begin_1, end = var_23544_end_1, end_mask = var_23544_end_mask_1, x = x_1553_cast_fp16)[name = string("op_23544_cast_fp16")];
            tensor<int32, [4]> var_23546 = const()[name = string("op_23546"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_23547_cast_fp16 = reshape(shape = var_23546, x = var_23544_cast_fp16)[name = string("op_23547_cast_fp16")];
            tensor<int32, [4]> var_23552_begin_1 = const()[name = string("op_23552_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_23552_end_1 = const()[name = string("op_23552_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_23552_end_mask_1 = const()[name = string("op_23552_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_23552_cast_fp16 = slice_by_index(begin = var_23552_begin_1, end = var_23552_end_1, end_mask = var_23552_end_mask_1, x = var_23547_cast_fp16)[name = string("op_23552_cast_fp16")];
            fp16 var_23553_to_fp16 = const()[name = string("op_23553_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_293_cast_fp16 = mul(x = var_23552_cast_fp16, y = var_23553_to_fp16)[name = string("scores_pos_293_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_293_cast_fp16 = add(x = scores_content_293_cast_fp16, y = scores_pos_293_cast_fp16)[name = string("logits_293_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_23556_cast_fp16 = softmax(axis = var_23039, x = logits_293_cast_fp16)[name = string("op_23556_cast_fp16")];
            bool var_23558_transpose_x_1 = const()[name = string("op_23558_transpose_x_1"), val = bool(false)];
            bool var_23558_transpose_y_1 = const()[name = string("op_23558_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_23558_cast_fp16 = matmul(transpose_x = var_23558_transpose_x_1, transpose_y = var_23558_transpose_y_1, x = var_23556_cast_fp16, y = v_head_587_cast_fp16)[name = string("op_23558_cast_fp16")];
            tensor<int32, [1]> var_23559_axes_1 = const()[name = string("op_23559_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_23559_cast_fp16 = squeeze(axes = var_23559_axes_1, x = var_23558_cast_fp16)[name = string("op_23559_cast_fp16")];
            string dense_output_1481_pad_type_1 = const()[name = string("dense_output_1481_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1481_strides_1 = const()[name = string("dense_output_1481_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1481_pad_1 = const()[name = string("dense_output_1481_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1481_dilations_1 = const()[name = string("dense_output_1481_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1481_groups_1 = const()[name = string("dense_output_1481_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542613888))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542745024))))[name = string("layers_18_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1481_cast_fp16 = conv(dilations = dense_output_1481_dilations_1, groups = dense_output_1481_groups_1, pad = dense_output_1481_pad_1, pad_type = dense_output_1481_pad_type_1, strides = dense_output_1481_strides_1, weight = layers_18_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1481_cast_fp16")];
            string sparse_output_1481_pad_type_1 = const()[name = string("sparse_output_1481_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1481_strides_1 = const()[name = string("sparse_output_1481_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1481_pad_1 = const()[name = string("sparse_output_1481_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1481_dilations_1 = const()[name = string("sparse_output_1481_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1481_groups_1 = const()[name = string("sparse_output_1481_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542748288))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542745600))))[name = string("layers_18_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1481_cast_fp16 = conv(dilations = sparse_output_1481_dilations_1, groups = sparse_output_1481_groups_1, pad = sparse_output_1481_pad_1, pad_type = sparse_output_1481_pad_type_1, strides = sparse_output_1481_strides_1, weight = layers_18_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1481_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23574_cast_fp16 = add(x = dense_output_1481_cast_fp16, y = sparse_output_1481_cast_fp16)[name = string("op_23574_cast_fp16")];
            tensor<int32, [4]> var_23575 = const()[name = string("op_23575"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23577 = const()[name = string("op_23577"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23576_cast_fp16 = transpose(perm = var_23575, x = var_23574_cast_fp16)[name = string("transpose_230")];
            tensor<fp16, [1, 51, 128]> q_head_295_cast_fp16 = reshape(shape = var_23577, x = var_23576_cast_fp16)[name = string("q_head_295_cast_fp16")];
            string dense_output_1483_pad_type_1 = const()[name = string("dense_output_1483_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1483_strides_1 = const()[name = string("dense_output_1483_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1483_pad_1 = const()[name = string("dense_output_1483_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1483_dilations_1 = const()[name = string("dense_output_1483_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1483_groups_1 = const()[name = string("dense_output_1483_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542764736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542895872))))[name = string("layers_18_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1483_cast_fp16 = conv(dilations = dense_output_1483_dilations_1, groups = dense_output_1483_groups_1, pad = dense_output_1483_pad_1, pad_type = dense_output_1483_pad_type_1, strides = dense_output_1483_strides_1, weight = layers_18_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1483_cast_fp16")];
            string sparse_output_1483_pad_type_1 = const()[name = string("sparse_output_1483_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1483_strides_1 = const()[name = string("sparse_output_1483_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1483_pad_1 = const()[name = string("sparse_output_1483_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1483_dilations_1 = const()[name = string("sparse_output_1483_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1483_groups_1 = const()[name = string("sparse_output_1483_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542899136))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542896448))))[name = string("layers_18_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1483_cast_fp16 = conv(dilations = sparse_output_1483_dilations_1, groups = sparse_output_1483_groups_1, pad = sparse_output_1483_pad_1, pad_type = sparse_output_1483_pad_type_1, strides = sparse_output_1483_strides_1, weight = layers_18_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1483_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23593_cast_fp16 = add(x = dense_output_1483_cast_fp16, y = sparse_output_1483_cast_fp16)[name = string("op_23593_cast_fp16")];
            tensor<int32, [4]> var_23594 = const()[name = string("op_23594"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23596 = const()[name = string("op_23596"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23595_cast_fp16 = transpose(perm = var_23594, x = var_23593_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 51, 128]> k_head_589_cast_fp16 = reshape(shape = var_23596, x = var_23595_cast_fp16)[name = string("k_head_589_cast_fp16")];
            string dense_output_1485_pad_type_1 = const()[name = string("dense_output_1485_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1485_strides_1 = const()[name = string("dense_output_1485_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1485_pad_1 = const()[name = string("dense_output_1485_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1485_dilations_1 = const()[name = string("dense_output_1485_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1485_groups_1 = const()[name = string("dense_output_1485_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542915584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543046720))))[name = string("layers_18_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1485_cast_fp16 = conv(dilations = dense_output_1485_dilations_1, groups = dense_output_1485_groups_1, pad = dense_output_1485_pad_1, pad_type = dense_output_1485_pad_type_1, strides = dense_output_1485_strides_1, weight = layers_18_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1485_cast_fp16")];
            string sparse_output_1485_pad_type_1 = const()[name = string("sparse_output_1485_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1485_strides_1 = const()[name = string("sparse_output_1485_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1485_pad_1 = const()[name = string("sparse_output_1485_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1485_dilations_1 = const()[name = string("sparse_output_1485_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1485_groups_1 = const()[name = string("sparse_output_1485_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543049984))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543047296))))[name = string("layers_18_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1485_cast_fp16 = conv(dilations = sparse_output_1485_dilations_1, groups = sparse_output_1485_groups_1, pad = sparse_output_1485_pad_1, pad_type = sparse_output_1485_pad_type_1, strides = sparse_output_1485_strides_1, weight = layers_18_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1485_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23612_cast_fp16 = add(x = dense_output_1485_cast_fp16, y = sparse_output_1485_cast_fp16)[name = string("op_23612_cast_fp16")];
            tensor<int32, [4]> var_23613 = const()[name = string("op_23613"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23615 = const()[name = string("op_23615"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23614_cast_fp16 = transpose(perm = var_23613, x = var_23612_cast_fp16)[name = string("transpose_228")];
            tensor<fp16, [1, 51, 128]> v_head_589_cast_fp16 = reshape(shape = var_23615, x = var_23614_cast_fp16)[name = string("v_head_589_cast_fp16")];
            string dense_output_1487_pad_type_1 = const()[name = string("dense_output_1487_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1487_strides_1 = const()[name = string("dense_output_1487_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1487_pad_1 = const()[name = string("dense_output_1487_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1487_dilations_1 = const()[name = string("dense_output_1487_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1487_groups_1 = const()[name = string("dense_output_1487_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543066432))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543197568))))[name = string("layers_18_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1487_cast_fp16 = conv(dilations = dense_output_1487_dilations_1, groups = dense_output_1487_groups_1, pad = dense_output_1487_pad_1, pad_type = dense_output_1487_pad_type_1, strides = dense_output_1487_strides_1, weight = layers_18_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1487_cast_fp16")];
            string sparse_output_1487_pad_type_1 = const()[name = string("sparse_output_1487_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1487_strides_1 = const()[name = string("sparse_output_1487_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1487_pad_1 = const()[name = string("sparse_output_1487_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1487_dilations_1 = const()[name = string("sparse_output_1487_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1487_groups_1 = const()[name = string("sparse_output_1487_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543200832))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543198144))))[name = string("layers_18_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1487_cast_fp16 = conv(dilations = sparse_output_1487_dilations_1, groups = sparse_output_1487_groups_1, pad = sparse_output_1487_pad_1, pad_type = sparse_output_1487_pad_type_1, strides = sparse_output_1487_strides_1, weight = layers_18_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1487_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_23631_cast_fp16 = add(x = dense_output_1487_cast_fp16, y = sparse_output_1487_cast_fp16)[name = string("op_23631_cast_fp16")];
            tensor<int32, [4]> var_23632 = const()[name = string("op_23632"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23634 = const()[name = string("op_23634"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_23633_cast_fp16 = transpose(perm = var_23632, x = var_23631_cast_fp16)[name = string("transpose_227")];
            tensor<fp16, [1, 101, 128]> p_head_589_cast_fp16 = reshape(shape = var_23634, x = var_23633_cast_fp16)[name = string("p_head_589_cast_fp16")];
            tensor<fp16, [128]> var_23636_to_fp16 = const()[name = string("op_23636_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543217280)))];
            tensor<fp16, [1, 51, 128]> var_23637_cast_fp16 = add(x = q_head_295_cast_fp16, y = var_23636_to_fp16)[name = string("op_23637_cast_fp16")];
            tensor<int32, [1]> q_u_295_axes_1 = const()[name = string("q_u_295_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_295_cast_fp16 = expand_dims(axes = q_u_295_axes_1, x = var_23637_cast_fp16)[name = string("q_u_295_cast_fp16")];
            tensor<fp16, [128]> var_23639_to_fp16 = const()[name = string("op_23639_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543217600)))];
            tensor<fp16, [1, 51, 128]> var_23640_cast_fp16 = add(x = q_head_295_cast_fp16, y = var_23639_to_fp16)[name = string("op_23640_cast_fp16")];
            tensor<int32, [1]> q_v_295_axes_1 = const()[name = string("q_v_295_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_295_cast_fp16 = expand_dims(axes = q_v_295_axes_1, x = var_23640_cast_fp16)[name = string("q_v_295_cast_fp16")];
            tensor<int32, [1]> k_head_591_axes_1 = const()[name = string("k_head_591_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_591_cast_fp16 = expand_dims(axes = k_head_591_axes_1, x = k_head_589_cast_fp16)[name = string("k_head_591_cast_fp16")];
            tensor<int32, [1]> v_head_591_axes_1 = const()[name = string("v_head_591_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_591_cast_fp16 = expand_dims(axes = v_head_591_axes_1, x = v_head_589_cast_fp16)[name = string("v_head_591_cast_fp16")];
            tensor<int32, [1]> p_head_591_axes_1 = const()[name = string("p_head_591_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_591_cast_fp16 = expand_dims(axes = p_head_591_axes_1, x = p_head_589_cast_fp16)[name = string("p_head_591_cast_fp16")];
            bool var_23646_transpose_x_3 = const()[name = string("op_23646_transpose_x_3"), val = bool(false)];
            bool var_23646_transpose_y_3 = const()[name = string("op_23646_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_23646_cast_fp16 = matmul(transpose_x = var_23646_transpose_x_3, transpose_y = var_23646_transpose_y_3, x = q_u_295_cast_fp16, y = k_head_591_cast_fp16)[name = string("op_23646_cast_fp16")];
            fp16 var_23647_to_fp16 = const()[name = string("op_23647_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_295_cast_fp16 = mul(x = var_23646_cast_fp16, y = var_23647_to_fp16)[name = string("scores_content_295_cast_fp16")];
            bool x_1557_transpose_x_3 = const()[name = string("x_1557_transpose_x_3"), val = bool(false)];
            bool x_1557_transpose_y_3 = const()[name = string("x_1557_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1557_cast_fp16 = matmul(transpose_x = x_1557_transpose_x_3, transpose_y = x_1557_transpose_y_3, x = q_v_295_cast_fp16, y = p_head_591_cast_fp16)[name = string("x_1557_cast_fp16")];
            tensor<int32, [8]> x_1559_pad_1 = const()[name = string("x_1559_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1559_mode_1 = const()[name = string("x_1559_mode_1"), val = string("constant")];
            fp16 const_2299_to_fp16 = const()[name = string("const_2299_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1559_cast_fp16 = pad(constant_val = const_2299_to_fp16, mode = x_1559_mode_1, pad = x_1559_pad_1, x = x_1557_cast_fp16)[name = string("x_1559_cast_fp16")];
            tensor<int32, [4]> var_23661 = const()[name = string("op_23661"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1561_cast_fp16 = reshape(shape = var_23661, x = x_1559_cast_fp16)[name = string("x_1561_cast_fp16")];
            tensor<int32, [4]> var_23665_begin_1 = const()[name = string("op_23665_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_23665_end_1 = const()[name = string("op_23665_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_23665_end_mask_1 = const()[name = string("op_23665_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_23665_cast_fp16 = slice_by_index(begin = var_23665_begin_1, end = var_23665_end_1, end_mask = var_23665_end_mask_1, x = x_1561_cast_fp16)[name = string("op_23665_cast_fp16")];
            tensor<int32, [4]> var_23667 = const()[name = string("op_23667"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_23668_cast_fp16 = reshape(shape = var_23667, x = var_23665_cast_fp16)[name = string("op_23668_cast_fp16")];
            tensor<int32, [4]> var_23673_begin_1 = const()[name = string("op_23673_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_23673_end_1 = const()[name = string("op_23673_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_23673_end_mask_1 = const()[name = string("op_23673_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_23673_cast_fp16 = slice_by_index(begin = var_23673_begin_1, end = var_23673_end_1, end_mask = var_23673_end_mask_1, x = var_23668_cast_fp16)[name = string("op_23673_cast_fp16")];
            fp16 var_23674_to_fp16 = const()[name = string("op_23674_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_295_cast_fp16 = mul(x = var_23673_cast_fp16, y = var_23674_to_fp16)[name = string("scores_pos_295_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_295_cast_fp16 = add(x = scores_content_295_cast_fp16, y = scores_pos_295_cast_fp16)[name = string("logits_295_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_23677_cast_fp16 = softmax(axis = var_23039, x = logits_295_cast_fp16)[name = string("op_23677_cast_fp16")];
            bool var_23679_transpose_x_1 = const()[name = string("op_23679_transpose_x_1"), val = bool(false)];
            bool var_23679_transpose_y_1 = const()[name = string("op_23679_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_23679_cast_fp16 = matmul(transpose_x = var_23679_transpose_x_1, transpose_y = var_23679_transpose_y_1, x = var_23677_cast_fp16, y = v_head_591_cast_fp16)[name = string("op_23679_cast_fp16")];
            tensor<int32, [1]> var_23680_axes_1 = const()[name = string("op_23680_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_23680_cast_fp16 = squeeze(axes = var_23680_axes_1, x = var_23679_cast_fp16)[name = string("op_23680_cast_fp16")];
            string dense_output_1489_pad_type_1 = const()[name = string("dense_output_1489_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1489_strides_1 = const()[name = string("dense_output_1489_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1489_pad_1 = const()[name = string("dense_output_1489_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1489_dilations_1 = const()[name = string("dense_output_1489_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1489_groups_1 = const()[name = string("dense_output_1489_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543217920))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543349056))))[name = string("layers_18_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1489_cast_fp16 = conv(dilations = dense_output_1489_dilations_1, groups = dense_output_1489_groups_1, pad = dense_output_1489_pad_1, pad_type = dense_output_1489_pad_type_1, strides = dense_output_1489_strides_1, weight = layers_18_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1489_cast_fp16")];
            string sparse_output_1489_pad_type_1 = const()[name = string("sparse_output_1489_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1489_strides_1 = const()[name = string("sparse_output_1489_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1489_pad_1 = const()[name = string("sparse_output_1489_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1489_dilations_1 = const()[name = string("sparse_output_1489_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1489_groups_1 = const()[name = string("sparse_output_1489_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543352320))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543349632))))[name = string("layers_18_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1489_cast_fp16 = conv(dilations = sparse_output_1489_dilations_1, groups = sparse_output_1489_groups_1, pad = sparse_output_1489_pad_1, pad_type = sparse_output_1489_pad_type_1, strides = sparse_output_1489_strides_1, weight = layers_18_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1489_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23695_cast_fp16 = add(x = dense_output_1489_cast_fp16, y = sparse_output_1489_cast_fp16)[name = string("op_23695_cast_fp16")];
            tensor<int32, [4]> var_23696 = const()[name = string("op_23696"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23698 = const()[name = string("op_23698"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23697_cast_fp16 = transpose(perm = var_23696, x = var_23695_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [1, 51, 128]> q_head_297_cast_fp16 = reshape(shape = var_23698, x = var_23697_cast_fp16)[name = string("q_head_297_cast_fp16")];
            string dense_output_1491_pad_type_1 = const()[name = string("dense_output_1491_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1491_strides_1 = const()[name = string("dense_output_1491_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1491_pad_1 = const()[name = string("dense_output_1491_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1491_dilations_1 = const()[name = string("dense_output_1491_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1491_groups_1 = const()[name = string("dense_output_1491_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543368768))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543499904))))[name = string("layers_18_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1491_cast_fp16 = conv(dilations = dense_output_1491_dilations_1, groups = dense_output_1491_groups_1, pad = dense_output_1491_pad_1, pad_type = dense_output_1491_pad_type_1, strides = dense_output_1491_strides_1, weight = layers_18_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1491_cast_fp16")];
            string sparse_output_1491_pad_type_1 = const()[name = string("sparse_output_1491_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1491_strides_1 = const()[name = string("sparse_output_1491_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1491_pad_1 = const()[name = string("sparse_output_1491_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1491_dilations_1 = const()[name = string("sparse_output_1491_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1491_groups_1 = const()[name = string("sparse_output_1491_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543503168))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543500480))))[name = string("layers_18_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1491_cast_fp16 = conv(dilations = sparse_output_1491_dilations_1, groups = sparse_output_1491_groups_1, pad = sparse_output_1491_pad_1, pad_type = sparse_output_1491_pad_type_1, strides = sparse_output_1491_strides_1, weight = layers_18_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1491_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23714_cast_fp16 = add(x = dense_output_1491_cast_fp16, y = sparse_output_1491_cast_fp16)[name = string("op_23714_cast_fp16")];
            tensor<int32, [4]> var_23715 = const()[name = string("op_23715"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23717 = const()[name = string("op_23717"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23716_cast_fp16 = transpose(perm = var_23715, x = var_23714_cast_fp16)[name = string("transpose_225")];
            tensor<fp16, [1, 51, 128]> k_head_593_cast_fp16 = reshape(shape = var_23717, x = var_23716_cast_fp16)[name = string("k_head_593_cast_fp16")];
            string dense_output_1493_pad_type_1 = const()[name = string("dense_output_1493_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1493_strides_1 = const()[name = string("dense_output_1493_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1493_pad_1 = const()[name = string("dense_output_1493_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1493_dilations_1 = const()[name = string("dense_output_1493_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1493_groups_1 = const()[name = string("dense_output_1493_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543519616))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543650752))))[name = string("layers_18_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1493_cast_fp16 = conv(dilations = dense_output_1493_dilations_1, groups = dense_output_1493_groups_1, pad = dense_output_1493_pad_1, pad_type = dense_output_1493_pad_type_1, strides = dense_output_1493_strides_1, weight = layers_18_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1493_cast_fp16")];
            string sparse_output_1493_pad_type_1 = const()[name = string("sparse_output_1493_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1493_strides_1 = const()[name = string("sparse_output_1493_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1493_pad_1 = const()[name = string("sparse_output_1493_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1493_dilations_1 = const()[name = string("sparse_output_1493_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1493_groups_1 = const()[name = string("sparse_output_1493_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543654016))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543651328))))[name = string("layers_18_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1493_cast_fp16 = conv(dilations = sparse_output_1493_dilations_1, groups = sparse_output_1493_groups_1, pad = sparse_output_1493_pad_1, pad_type = sparse_output_1493_pad_type_1, strides = sparse_output_1493_strides_1, weight = layers_18_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1493_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23733_cast_fp16 = add(x = dense_output_1493_cast_fp16, y = sparse_output_1493_cast_fp16)[name = string("op_23733_cast_fp16")];
            tensor<int32, [4]> var_23734 = const()[name = string("op_23734"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23736 = const()[name = string("op_23736"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23735_cast_fp16 = transpose(perm = var_23734, x = var_23733_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [1, 51, 128]> v_head_593_cast_fp16 = reshape(shape = var_23736, x = var_23735_cast_fp16)[name = string("v_head_593_cast_fp16")];
            string dense_output_1495_pad_type_1 = const()[name = string("dense_output_1495_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1495_strides_1 = const()[name = string("dense_output_1495_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1495_pad_1 = const()[name = string("dense_output_1495_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1495_dilations_1 = const()[name = string("dense_output_1495_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1495_groups_1 = const()[name = string("dense_output_1495_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543670464))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543801600))))[name = string("layers_18_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1495_cast_fp16 = conv(dilations = dense_output_1495_dilations_1, groups = dense_output_1495_groups_1, pad = dense_output_1495_pad_1, pad_type = dense_output_1495_pad_type_1, strides = dense_output_1495_strides_1, weight = layers_18_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1495_cast_fp16")];
            string sparse_output_1495_pad_type_1 = const()[name = string("sparse_output_1495_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1495_strides_1 = const()[name = string("sparse_output_1495_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1495_pad_1 = const()[name = string("sparse_output_1495_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1495_dilations_1 = const()[name = string("sparse_output_1495_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1495_groups_1 = const()[name = string("sparse_output_1495_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543804864))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543802176))))[name = string("layers_18_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1495_cast_fp16 = conv(dilations = sparse_output_1495_dilations_1, groups = sparse_output_1495_groups_1, pad = sparse_output_1495_pad_1, pad_type = sparse_output_1495_pad_type_1, strides = sparse_output_1495_strides_1, weight = layers_18_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1495_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_23752_cast_fp16 = add(x = dense_output_1495_cast_fp16, y = sparse_output_1495_cast_fp16)[name = string("op_23752_cast_fp16")];
            tensor<int32, [4]> var_23753 = const()[name = string("op_23753"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23755 = const()[name = string("op_23755"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_23754_cast_fp16 = transpose(perm = var_23753, x = var_23752_cast_fp16)[name = string("transpose_223")];
            tensor<fp16, [1, 101, 128]> p_head_593_cast_fp16 = reshape(shape = var_23755, x = var_23754_cast_fp16)[name = string("p_head_593_cast_fp16")];
            tensor<fp16, [128]> var_23757_to_fp16 = const()[name = string("op_23757_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543821312)))];
            tensor<fp16, [1, 51, 128]> var_23758_cast_fp16 = add(x = q_head_297_cast_fp16, y = var_23757_to_fp16)[name = string("op_23758_cast_fp16")];
            tensor<int32, [1]> q_u_297_axes_1 = const()[name = string("q_u_297_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_297_cast_fp16 = expand_dims(axes = q_u_297_axes_1, x = var_23758_cast_fp16)[name = string("q_u_297_cast_fp16")];
            tensor<fp16, [128]> var_23760_to_fp16 = const()[name = string("op_23760_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543821632)))];
            tensor<fp16, [1, 51, 128]> var_23761_cast_fp16 = add(x = q_head_297_cast_fp16, y = var_23760_to_fp16)[name = string("op_23761_cast_fp16")];
            tensor<int32, [1]> q_v_297_axes_1 = const()[name = string("q_v_297_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_297_cast_fp16 = expand_dims(axes = q_v_297_axes_1, x = var_23761_cast_fp16)[name = string("q_v_297_cast_fp16")];
            tensor<int32, [1]> k_head_595_axes_1 = const()[name = string("k_head_595_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_595_cast_fp16 = expand_dims(axes = k_head_595_axes_1, x = k_head_593_cast_fp16)[name = string("k_head_595_cast_fp16")];
            tensor<int32, [1]> v_head_595_axes_1 = const()[name = string("v_head_595_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_595_cast_fp16 = expand_dims(axes = v_head_595_axes_1, x = v_head_593_cast_fp16)[name = string("v_head_595_cast_fp16")];
            tensor<int32, [1]> p_head_595_axes_1 = const()[name = string("p_head_595_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_595_cast_fp16 = expand_dims(axes = p_head_595_axes_1, x = p_head_593_cast_fp16)[name = string("p_head_595_cast_fp16")];
            bool var_23767_transpose_x_3 = const()[name = string("op_23767_transpose_x_3"), val = bool(false)];
            bool var_23767_transpose_y_3 = const()[name = string("op_23767_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_23767_cast_fp16 = matmul(transpose_x = var_23767_transpose_x_3, transpose_y = var_23767_transpose_y_3, x = q_u_297_cast_fp16, y = k_head_595_cast_fp16)[name = string("op_23767_cast_fp16")];
            fp16 var_23768_to_fp16 = const()[name = string("op_23768_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_297_cast_fp16 = mul(x = var_23767_cast_fp16, y = var_23768_to_fp16)[name = string("scores_content_297_cast_fp16")];
            bool x_1565_transpose_x_3 = const()[name = string("x_1565_transpose_x_3"), val = bool(false)];
            bool x_1565_transpose_y_3 = const()[name = string("x_1565_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1565_cast_fp16 = matmul(transpose_x = x_1565_transpose_x_3, transpose_y = x_1565_transpose_y_3, x = q_v_297_cast_fp16, y = p_head_595_cast_fp16)[name = string("x_1565_cast_fp16")];
            tensor<int32, [8]> x_1567_pad_1 = const()[name = string("x_1567_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1567_mode_1 = const()[name = string("x_1567_mode_1"), val = string("constant")];
            fp16 const_2305_to_fp16 = const()[name = string("const_2305_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1567_cast_fp16 = pad(constant_val = const_2305_to_fp16, mode = x_1567_mode_1, pad = x_1567_pad_1, x = x_1565_cast_fp16)[name = string("x_1567_cast_fp16")];
            tensor<int32, [4]> var_23782 = const()[name = string("op_23782"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1569_cast_fp16 = reshape(shape = var_23782, x = x_1567_cast_fp16)[name = string("x_1569_cast_fp16")];
            tensor<int32, [4]> var_23786_begin_1 = const()[name = string("op_23786_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_23786_end_1 = const()[name = string("op_23786_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_23786_end_mask_1 = const()[name = string("op_23786_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_23786_cast_fp16 = slice_by_index(begin = var_23786_begin_1, end = var_23786_end_1, end_mask = var_23786_end_mask_1, x = x_1569_cast_fp16)[name = string("op_23786_cast_fp16")];
            tensor<int32, [4]> var_23788 = const()[name = string("op_23788"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_23789_cast_fp16 = reshape(shape = var_23788, x = var_23786_cast_fp16)[name = string("op_23789_cast_fp16")];
            tensor<int32, [4]> var_23794_begin_1 = const()[name = string("op_23794_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_23794_end_1 = const()[name = string("op_23794_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_23794_end_mask_1 = const()[name = string("op_23794_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_23794_cast_fp16 = slice_by_index(begin = var_23794_begin_1, end = var_23794_end_1, end_mask = var_23794_end_mask_1, x = var_23789_cast_fp16)[name = string("op_23794_cast_fp16")];
            fp16 var_23795_to_fp16 = const()[name = string("op_23795_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_297_cast_fp16 = mul(x = var_23794_cast_fp16, y = var_23795_to_fp16)[name = string("scores_pos_297_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_297_cast_fp16 = add(x = scores_content_297_cast_fp16, y = scores_pos_297_cast_fp16)[name = string("logits_297_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_23798_cast_fp16 = softmax(axis = var_23039, x = logits_297_cast_fp16)[name = string("op_23798_cast_fp16")];
            bool var_23800_transpose_x_1 = const()[name = string("op_23800_transpose_x_1"), val = bool(false)];
            bool var_23800_transpose_y_1 = const()[name = string("op_23800_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_23800_cast_fp16 = matmul(transpose_x = var_23800_transpose_x_1, transpose_y = var_23800_transpose_y_1, x = var_23798_cast_fp16, y = v_head_595_cast_fp16)[name = string("op_23800_cast_fp16")];
            tensor<int32, [1]> var_23801_axes_1 = const()[name = string("op_23801_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_23801_cast_fp16 = squeeze(axes = var_23801_axes_1, x = var_23800_cast_fp16)[name = string("op_23801_cast_fp16")];
            string dense_output_1497_pad_type_1 = const()[name = string("dense_output_1497_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1497_strides_1 = const()[name = string("dense_output_1497_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1497_pad_1 = const()[name = string("dense_output_1497_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1497_dilations_1 = const()[name = string("dense_output_1497_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1497_groups_1 = const()[name = string("dense_output_1497_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543821952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543953088))))[name = string("layers_18_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1497_cast_fp16 = conv(dilations = dense_output_1497_dilations_1, groups = dense_output_1497_groups_1, pad = dense_output_1497_pad_1, pad_type = dense_output_1497_pad_type_1, strides = dense_output_1497_strides_1, weight = layers_18_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1497_cast_fp16")];
            string sparse_output_1497_pad_type_1 = const()[name = string("sparse_output_1497_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1497_strides_1 = const()[name = string("sparse_output_1497_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1497_pad_1 = const()[name = string("sparse_output_1497_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1497_dilations_1 = const()[name = string("sparse_output_1497_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1497_groups_1 = const()[name = string("sparse_output_1497_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543956352))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543953664))))[name = string("layers_18_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1497_cast_fp16 = conv(dilations = sparse_output_1497_dilations_1, groups = sparse_output_1497_groups_1, pad = sparse_output_1497_pad_1, pad_type = sparse_output_1497_pad_type_1, strides = sparse_output_1497_strides_1, weight = layers_18_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1497_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23816_cast_fp16 = add(x = dense_output_1497_cast_fp16, y = sparse_output_1497_cast_fp16)[name = string("op_23816_cast_fp16")];
            tensor<int32, [4]> var_23817 = const()[name = string("op_23817"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23819 = const()[name = string("op_23819"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23818_cast_fp16 = transpose(perm = var_23817, x = var_23816_cast_fp16)[name = string("transpose_222")];
            tensor<fp16, [1, 51, 128]> q_head_299_cast_fp16 = reshape(shape = var_23819, x = var_23818_cast_fp16)[name = string("q_head_299_cast_fp16")];
            string dense_output_1499_pad_type_1 = const()[name = string("dense_output_1499_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1499_strides_1 = const()[name = string("dense_output_1499_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1499_pad_1 = const()[name = string("dense_output_1499_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1499_dilations_1 = const()[name = string("dense_output_1499_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1499_groups_1 = const()[name = string("dense_output_1499_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543972800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544103936))))[name = string("layers_18_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1499_cast_fp16 = conv(dilations = dense_output_1499_dilations_1, groups = dense_output_1499_groups_1, pad = dense_output_1499_pad_1, pad_type = dense_output_1499_pad_type_1, strides = dense_output_1499_strides_1, weight = layers_18_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1499_cast_fp16")];
            string sparse_output_1499_pad_type_1 = const()[name = string("sparse_output_1499_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1499_strides_1 = const()[name = string("sparse_output_1499_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1499_pad_1 = const()[name = string("sparse_output_1499_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1499_dilations_1 = const()[name = string("sparse_output_1499_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1499_groups_1 = const()[name = string("sparse_output_1499_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544107200))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544104512))))[name = string("layers_18_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1499_cast_fp16 = conv(dilations = sparse_output_1499_dilations_1, groups = sparse_output_1499_groups_1, pad = sparse_output_1499_pad_1, pad_type = sparse_output_1499_pad_type_1, strides = sparse_output_1499_strides_1, weight = layers_18_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1499_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23835_cast_fp16 = add(x = dense_output_1499_cast_fp16, y = sparse_output_1499_cast_fp16)[name = string("op_23835_cast_fp16")];
            tensor<int32, [4]> var_23836 = const()[name = string("op_23836"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23838 = const()[name = string("op_23838"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23837_cast_fp16 = transpose(perm = var_23836, x = var_23835_cast_fp16)[name = string("transpose_221")];
            tensor<fp16, [1, 51, 128]> k_head_597_cast_fp16 = reshape(shape = var_23838, x = var_23837_cast_fp16)[name = string("k_head_597_cast_fp16")];
            string dense_output_1501_pad_type_1 = const()[name = string("dense_output_1501_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1501_strides_1 = const()[name = string("dense_output_1501_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1501_pad_1 = const()[name = string("dense_output_1501_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1501_dilations_1 = const()[name = string("dense_output_1501_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1501_groups_1 = const()[name = string("dense_output_1501_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544123648))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544254784))))[name = string("layers_18_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1501_cast_fp16 = conv(dilations = dense_output_1501_dilations_1, groups = dense_output_1501_groups_1, pad = dense_output_1501_pad_1, pad_type = dense_output_1501_pad_type_1, strides = dense_output_1501_strides_1, weight = layers_18_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1501_cast_fp16")];
            string sparse_output_1501_pad_type_1 = const()[name = string("sparse_output_1501_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1501_strides_1 = const()[name = string("sparse_output_1501_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1501_pad_1 = const()[name = string("sparse_output_1501_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1501_dilations_1 = const()[name = string("sparse_output_1501_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1501_groups_1 = const()[name = string("sparse_output_1501_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544258048))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544255360))))[name = string("layers_18_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1501_cast_fp16 = conv(dilations = sparse_output_1501_dilations_1, groups = sparse_output_1501_groups_1, pad = sparse_output_1501_pad_1, pad_type = sparse_output_1501_pad_type_1, strides = sparse_output_1501_strides_1, weight = layers_18_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1501_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23854_cast_fp16 = add(x = dense_output_1501_cast_fp16, y = sparse_output_1501_cast_fp16)[name = string("op_23854_cast_fp16")];
            tensor<int32, [4]> var_23855 = const()[name = string("op_23855"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23857 = const()[name = string("op_23857"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23856_cast_fp16 = transpose(perm = var_23855, x = var_23854_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 51, 128]> v_head_597_cast_fp16 = reshape(shape = var_23857, x = var_23856_cast_fp16)[name = string("v_head_597_cast_fp16")];
            string dense_output_1503_pad_type_1 = const()[name = string("dense_output_1503_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1503_strides_1 = const()[name = string("dense_output_1503_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1503_pad_1 = const()[name = string("dense_output_1503_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1503_dilations_1 = const()[name = string("dense_output_1503_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1503_groups_1 = const()[name = string("dense_output_1503_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544274496))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544405632))))[name = string("layers_18_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1503_cast_fp16 = conv(dilations = dense_output_1503_dilations_1, groups = dense_output_1503_groups_1, pad = dense_output_1503_pad_1, pad_type = dense_output_1503_pad_type_1, strides = dense_output_1503_strides_1, weight = layers_18_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1503_cast_fp16")];
            string sparse_output_1503_pad_type_1 = const()[name = string("sparse_output_1503_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1503_strides_1 = const()[name = string("sparse_output_1503_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1503_pad_1 = const()[name = string("sparse_output_1503_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1503_dilations_1 = const()[name = string("sparse_output_1503_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1503_groups_1 = const()[name = string("sparse_output_1503_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544408896))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544406208))))[name = string("layers_18_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1503_cast_fp16 = conv(dilations = sparse_output_1503_dilations_1, groups = sparse_output_1503_groups_1, pad = sparse_output_1503_pad_1, pad_type = sparse_output_1503_pad_type_1, strides = sparse_output_1503_strides_1, weight = layers_18_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1503_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_23873_cast_fp16 = add(x = dense_output_1503_cast_fp16, y = sparse_output_1503_cast_fp16)[name = string("op_23873_cast_fp16")];
            tensor<int32, [4]> var_23874 = const()[name = string("op_23874"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23876 = const()[name = string("op_23876"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_23875_cast_fp16 = transpose(perm = var_23874, x = var_23873_cast_fp16)[name = string("transpose_219")];
            tensor<fp16, [1, 101, 128]> p_head_597_cast_fp16 = reshape(shape = var_23876, x = var_23875_cast_fp16)[name = string("p_head_597_cast_fp16")];
            tensor<fp16, [128]> var_23878_to_fp16 = const()[name = string("op_23878_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544425344)))];
            tensor<fp16, [1, 51, 128]> var_23879_cast_fp16 = add(x = q_head_299_cast_fp16, y = var_23878_to_fp16)[name = string("op_23879_cast_fp16")];
            tensor<int32, [1]> q_u_299_axes_1 = const()[name = string("q_u_299_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_299_cast_fp16 = expand_dims(axes = q_u_299_axes_1, x = var_23879_cast_fp16)[name = string("q_u_299_cast_fp16")];
            tensor<fp16, [128]> var_23881_to_fp16 = const()[name = string("op_23881_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544425664)))];
            tensor<fp16, [1, 51, 128]> var_23882_cast_fp16 = add(x = q_head_299_cast_fp16, y = var_23881_to_fp16)[name = string("op_23882_cast_fp16")];
            tensor<int32, [1]> q_v_299_axes_1 = const()[name = string("q_v_299_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_299_cast_fp16 = expand_dims(axes = q_v_299_axes_1, x = var_23882_cast_fp16)[name = string("q_v_299_cast_fp16")];
            tensor<int32, [1]> k_head_599_axes_1 = const()[name = string("k_head_599_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_599_cast_fp16 = expand_dims(axes = k_head_599_axes_1, x = k_head_597_cast_fp16)[name = string("k_head_599_cast_fp16")];
            tensor<int32, [1]> v_head_599_axes_1 = const()[name = string("v_head_599_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_599_cast_fp16 = expand_dims(axes = v_head_599_axes_1, x = v_head_597_cast_fp16)[name = string("v_head_599_cast_fp16")];
            tensor<int32, [1]> p_head_599_axes_1 = const()[name = string("p_head_599_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_599_cast_fp16 = expand_dims(axes = p_head_599_axes_1, x = p_head_597_cast_fp16)[name = string("p_head_599_cast_fp16")];
            bool var_23888_transpose_x_3 = const()[name = string("op_23888_transpose_x_3"), val = bool(false)];
            bool var_23888_transpose_y_3 = const()[name = string("op_23888_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_23888_cast_fp16 = matmul(transpose_x = var_23888_transpose_x_3, transpose_y = var_23888_transpose_y_3, x = q_u_299_cast_fp16, y = k_head_599_cast_fp16)[name = string("op_23888_cast_fp16")];
            fp16 var_23889_to_fp16 = const()[name = string("op_23889_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_299_cast_fp16 = mul(x = var_23888_cast_fp16, y = var_23889_to_fp16)[name = string("scores_content_299_cast_fp16")];
            bool x_1573_transpose_x_3 = const()[name = string("x_1573_transpose_x_3"), val = bool(false)];
            bool x_1573_transpose_y_3 = const()[name = string("x_1573_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1573_cast_fp16 = matmul(transpose_x = x_1573_transpose_x_3, transpose_y = x_1573_transpose_y_3, x = q_v_299_cast_fp16, y = p_head_599_cast_fp16)[name = string("x_1573_cast_fp16")];
            tensor<int32, [8]> x_1575_pad_1 = const()[name = string("x_1575_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1575_mode_1 = const()[name = string("x_1575_mode_1"), val = string("constant")];
            fp16 const_2311_to_fp16 = const()[name = string("const_2311_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1575_cast_fp16 = pad(constant_val = const_2311_to_fp16, mode = x_1575_mode_1, pad = x_1575_pad_1, x = x_1573_cast_fp16)[name = string("x_1575_cast_fp16")];
            tensor<int32, [4]> var_23903 = const()[name = string("op_23903"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1577_cast_fp16 = reshape(shape = var_23903, x = x_1575_cast_fp16)[name = string("x_1577_cast_fp16")];
            tensor<int32, [4]> var_23907_begin_1 = const()[name = string("op_23907_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_23907_end_1 = const()[name = string("op_23907_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_23907_end_mask_1 = const()[name = string("op_23907_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_23907_cast_fp16 = slice_by_index(begin = var_23907_begin_1, end = var_23907_end_1, end_mask = var_23907_end_mask_1, x = x_1577_cast_fp16)[name = string("op_23907_cast_fp16")];
            tensor<int32, [4]> var_23909 = const()[name = string("op_23909"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_23910_cast_fp16 = reshape(shape = var_23909, x = var_23907_cast_fp16)[name = string("op_23910_cast_fp16")];
            tensor<int32, [4]> var_23915_begin_1 = const()[name = string("op_23915_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_23915_end_1 = const()[name = string("op_23915_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_23915_end_mask_1 = const()[name = string("op_23915_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_23915_cast_fp16 = slice_by_index(begin = var_23915_begin_1, end = var_23915_end_1, end_mask = var_23915_end_mask_1, x = var_23910_cast_fp16)[name = string("op_23915_cast_fp16")];
            fp16 var_23916_to_fp16 = const()[name = string("op_23916_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_299_cast_fp16 = mul(x = var_23915_cast_fp16, y = var_23916_to_fp16)[name = string("scores_pos_299_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_299_cast_fp16 = add(x = scores_content_299_cast_fp16, y = scores_pos_299_cast_fp16)[name = string("logits_299_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_23919_cast_fp16 = softmax(axis = var_23039, x = logits_299_cast_fp16)[name = string("op_23919_cast_fp16")];
            bool var_23921_transpose_x_1 = const()[name = string("op_23921_transpose_x_1"), val = bool(false)];
            bool var_23921_transpose_y_1 = const()[name = string("op_23921_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_23921_cast_fp16 = matmul(transpose_x = var_23921_transpose_x_1, transpose_y = var_23921_transpose_y_1, x = var_23919_cast_fp16, y = v_head_599_cast_fp16)[name = string("op_23921_cast_fp16")];
            tensor<int32, [1]> var_23922_axes_1 = const()[name = string("op_23922_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_23922_cast_fp16 = squeeze(axes = var_23922_axes_1, x = var_23921_cast_fp16)[name = string("op_23922_cast_fp16")];
            string dense_output_1505_pad_type_1 = const()[name = string("dense_output_1505_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1505_strides_1 = const()[name = string("dense_output_1505_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1505_pad_1 = const()[name = string("dense_output_1505_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1505_dilations_1 = const()[name = string("dense_output_1505_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1505_groups_1 = const()[name = string("dense_output_1505_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544425984))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544557120))))[name = string("layers_18_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1505_cast_fp16 = conv(dilations = dense_output_1505_dilations_1, groups = dense_output_1505_groups_1, pad = dense_output_1505_pad_1, pad_type = dense_output_1505_pad_type_1, strides = dense_output_1505_strides_1, weight = layers_18_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1505_cast_fp16")];
            string sparse_output_1505_pad_type_1 = const()[name = string("sparse_output_1505_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1505_strides_1 = const()[name = string("sparse_output_1505_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1505_pad_1 = const()[name = string("sparse_output_1505_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1505_dilations_1 = const()[name = string("sparse_output_1505_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1505_groups_1 = const()[name = string("sparse_output_1505_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544560384))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544557696))))[name = string("layers_18_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1505_cast_fp16 = conv(dilations = sparse_output_1505_dilations_1, groups = sparse_output_1505_groups_1, pad = sparse_output_1505_pad_1, pad_type = sparse_output_1505_pad_type_1, strides = sparse_output_1505_strides_1, weight = layers_18_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1505_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23937_cast_fp16 = add(x = dense_output_1505_cast_fp16, y = sparse_output_1505_cast_fp16)[name = string("op_23937_cast_fp16")];
            tensor<int32, [4]> var_23938 = const()[name = string("op_23938"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23940 = const()[name = string("op_23940"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23939_cast_fp16 = transpose(perm = var_23938, x = var_23937_cast_fp16)[name = string("transpose_218")];
            tensor<fp16, [1, 51, 128]> q_head_301_cast_fp16 = reshape(shape = var_23940, x = var_23939_cast_fp16)[name = string("q_head_301_cast_fp16")];
            string dense_output_1507_pad_type_1 = const()[name = string("dense_output_1507_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1507_strides_1 = const()[name = string("dense_output_1507_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1507_pad_1 = const()[name = string("dense_output_1507_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1507_dilations_1 = const()[name = string("dense_output_1507_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1507_groups_1 = const()[name = string("dense_output_1507_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544576832))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544707968))))[name = string("layers_18_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1507_cast_fp16 = conv(dilations = dense_output_1507_dilations_1, groups = dense_output_1507_groups_1, pad = dense_output_1507_pad_1, pad_type = dense_output_1507_pad_type_1, strides = dense_output_1507_strides_1, weight = layers_18_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1507_cast_fp16")];
            string sparse_output_1507_pad_type_1 = const()[name = string("sparse_output_1507_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1507_strides_1 = const()[name = string("sparse_output_1507_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1507_pad_1 = const()[name = string("sparse_output_1507_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1507_dilations_1 = const()[name = string("sparse_output_1507_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1507_groups_1 = const()[name = string("sparse_output_1507_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544711232))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544708544))))[name = string("layers_18_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1507_cast_fp16 = conv(dilations = sparse_output_1507_dilations_1, groups = sparse_output_1507_groups_1, pad = sparse_output_1507_pad_1, pad_type = sparse_output_1507_pad_type_1, strides = sparse_output_1507_strides_1, weight = layers_18_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1507_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23956_cast_fp16 = add(x = dense_output_1507_cast_fp16, y = sparse_output_1507_cast_fp16)[name = string("op_23956_cast_fp16")];
            tensor<int32, [4]> var_23957 = const()[name = string("op_23957"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23959 = const()[name = string("op_23959"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23958_cast_fp16 = transpose(perm = var_23957, x = var_23956_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 51, 128]> k_head_601_cast_fp16 = reshape(shape = var_23959, x = var_23958_cast_fp16)[name = string("k_head_601_cast_fp16")];
            string dense_output_1509_pad_type_1 = const()[name = string("dense_output_1509_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1509_strides_1 = const()[name = string("dense_output_1509_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1509_pad_1 = const()[name = string("dense_output_1509_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1509_dilations_1 = const()[name = string("dense_output_1509_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1509_groups_1 = const()[name = string("dense_output_1509_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544727680))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544858816))))[name = string("layers_18_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1509_cast_fp16 = conv(dilations = dense_output_1509_dilations_1, groups = dense_output_1509_groups_1, pad = dense_output_1509_pad_1, pad_type = dense_output_1509_pad_type_1, strides = dense_output_1509_strides_1, weight = layers_18_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1509_cast_fp16")];
            string sparse_output_1509_pad_type_1 = const()[name = string("sparse_output_1509_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1509_strides_1 = const()[name = string("sparse_output_1509_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1509_pad_1 = const()[name = string("sparse_output_1509_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1509_dilations_1 = const()[name = string("sparse_output_1509_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1509_groups_1 = const()[name = string("sparse_output_1509_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544862080))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544859392))))[name = string("layers_18_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1509_cast_fp16 = conv(dilations = sparse_output_1509_dilations_1, groups = sparse_output_1509_groups_1, pad = sparse_output_1509_pad_1, pad_type = sparse_output_1509_pad_type_1, strides = sparse_output_1509_strides_1, weight = layers_18_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1509_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_23975_cast_fp16 = add(x = dense_output_1509_cast_fp16, y = sparse_output_1509_cast_fp16)[name = string("op_23975_cast_fp16")];
            tensor<int32, [4]> var_23976 = const()[name = string("op_23976"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23978 = const()[name = string("op_23978"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_23977_cast_fp16 = transpose(perm = var_23976, x = var_23975_cast_fp16)[name = string("transpose_216")];
            tensor<fp16, [1, 51, 128]> v_head_601_cast_fp16 = reshape(shape = var_23978, x = var_23977_cast_fp16)[name = string("v_head_601_cast_fp16")];
            string dense_output_1511_pad_type_1 = const()[name = string("dense_output_1511_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1511_strides_1 = const()[name = string("dense_output_1511_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1511_pad_1 = const()[name = string("dense_output_1511_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1511_dilations_1 = const()[name = string("dense_output_1511_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1511_groups_1 = const()[name = string("dense_output_1511_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544878528))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545009664))))[name = string("layers_18_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1511_cast_fp16 = conv(dilations = dense_output_1511_dilations_1, groups = dense_output_1511_groups_1, pad = dense_output_1511_pad_1, pad_type = dense_output_1511_pad_type_1, strides = dense_output_1511_strides_1, weight = layers_18_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1511_cast_fp16")];
            string sparse_output_1511_pad_type_1 = const()[name = string("sparse_output_1511_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1511_strides_1 = const()[name = string("sparse_output_1511_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1511_pad_1 = const()[name = string("sparse_output_1511_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1511_dilations_1 = const()[name = string("sparse_output_1511_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1511_groups_1 = const()[name = string("sparse_output_1511_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545012928))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545010240))))[name = string("layers_18_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1511_cast_fp16 = conv(dilations = sparse_output_1511_dilations_1, groups = sparse_output_1511_groups_1, pad = sparse_output_1511_pad_1, pad_type = sparse_output_1511_pad_type_1, strides = sparse_output_1511_strides_1, weight = layers_18_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1511_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_23994_cast_fp16 = add(x = dense_output_1511_cast_fp16, y = sparse_output_1511_cast_fp16)[name = string("op_23994_cast_fp16")];
            tensor<int32, [4]> var_23995 = const()[name = string("op_23995"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_23997 = const()[name = string("op_23997"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_23996_cast_fp16 = transpose(perm = var_23995, x = var_23994_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [1, 101, 128]> p_head_601_cast_fp16 = reshape(shape = var_23997, x = var_23996_cast_fp16)[name = string("p_head_601_cast_fp16")];
            tensor<fp16, [128]> var_23999_to_fp16 = const()[name = string("op_23999_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545029376)))];
            tensor<fp16, [1, 51, 128]> var_24000_cast_fp16 = add(x = q_head_301_cast_fp16, y = var_23999_to_fp16)[name = string("op_24000_cast_fp16")];
            tensor<int32, [1]> q_u_301_axes_1 = const()[name = string("q_u_301_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_301_cast_fp16 = expand_dims(axes = q_u_301_axes_1, x = var_24000_cast_fp16)[name = string("q_u_301_cast_fp16")];
            tensor<fp16, [128]> var_24002_to_fp16 = const()[name = string("op_24002_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545029696)))];
            tensor<fp16, [1, 51, 128]> var_24003_cast_fp16 = add(x = q_head_301_cast_fp16, y = var_24002_to_fp16)[name = string("op_24003_cast_fp16")];
            tensor<int32, [1]> q_v_301_axes_1 = const()[name = string("q_v_301_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_301_cast_fp16 = expand_dims(axes = q_v_301_axes_1, x = var_24003_cast_fp16)[name = string("q_v_301_cast_fp16")];
            tensor<int32, [1]> k_head_603_axes_1 = const()[name = string("k_head_603_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_603_cast_fp16 = expand_dims(axes = k_head_603_axes_1, x = k_head_601_cast_fp16)[name = string("k_head_603_cast_fp16")];
            tensor<int32, [1]> v_head_603_axes_1 = const()[name = string("v_head_603_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_603_cast_fp16 = expand_dims(axes = v_head_603_axes_1, x = v_head_601_cast_fp16)[name = string("v_head_603_cast_fp16")];
            tensor<int32, [1]> p_head_603_axes_1 = const()[name = string("p_head_603_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_603_cast_fp16 = expand_dims(axes = p_head_603_axes_1, x = p_head_601_cast_fp16)[name = string("p_head_603_cast_fp16")];
            bool var_24009_transpose_x_3 = const()[name = string("op_24009_transpose_x_3"), val = bool(false)];
            bool var_24009_transpose_y_3 = const()[name = string("op_24009_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_24009_cast_fp16 = matmul(transpose_x = var_24009_transpose_x_3, transpose_y = var_24009_transpose_y_3, x = q_u_301_cast_fp16, y = k_head_603_cast_fp16)[name = string("op_24009_cast_fp16")];
            fp16 var_24010_to_fp16 = const()[name = string("op_24010_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_301_cast_fp16 = mul(x = var_24009_cast_fp16, y = var_24010_to_fp16)[name = string("scores_content_301_cast_fp16")];
            bool x_1581_transpose_x_3 = const()[name = string("x_1581_transpose_x_3"), val = bool(false)];
            bool x_1581_transpose_y_3 = const()[name = string("x_1581_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1581_cast_fp16 = matmul(transpose_x = x_1581_transpose_x_3, transpose_y = x_1581_transpose_y_3, x = q_v_301_cast_fp16, y = p_head_603_cast_fp16)[name = string("x_1581_cast_fp16")];
            tensor<int32, [8]> x_1583_pad_1 = const()[name = string("x_1583_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1583_mode_1 = const()[name = string("x_1583_mode_1"), val = string("constant")];
            fp16 const_2317_to_fp16 = const()[name = string("const_2317_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1583_cast_fp16 = pad(constant_val = const_2317_to_fp16, mode = x_1583_mode_1, pad = x_1583_pad_1, x = x_1581_cast_fp16)[name = string("x_1583_cast_fp16")];
            tensor<int32, [4]> var_24024 = const()[name = string("op_24024"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1585_cast_fp16 = reshape(shape = var_24024, x = x_1583_cast_fp16)[name = string("x_1585_cast_fp16")];
            tensor<int32, [4]> var_24028_begin_1 = const()[name = string("op_24028_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_24028_end_1 = const()[name = string("op_24028_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_24028_end_mask_1 = const()[name = string("op_24028_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_24028_cast_fp16 = slice_by_index(begin = var_24028_begin_1, end = var_24028_end_1, end_mask = var_24028_end_mask_1, x = x_1585_cast_fp16)[name = string("op_24028_cast_fp16")];
            tensor<int32, [4]> var_24030 = const()[name = string("op_24030"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_24031_cast_fp16 = reshape(shape = var_24030, x = var_24028_cast_fp16)[name = string("op_24031_cast_fp16")];
            tensor<int32, [4]> var_24036_begin_1 = const()[name = string("op_24036_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_24036_end_1 = const()[name = string("op_24036_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_24036_end_mask_1 = const()[name = string("op_24036_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_24036_cast_fp16 = slice_by_index(begin = var_24036_begin_1, end = var_24036_end_1, end_mask = var_24036_end_mask_1, x = var_24031_cast_fp16)[name = string("op_24036_cast_fp16")];
            fp16 var_24037_to_fp16 = const()[name = string("op_24037_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_301_cast_fp16 = mul(x = var_24036_cast_fp16, y = var_24037_to_fp16)[name = string("scores_pos_301_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_301_cast_fp16 = add(x = scores_content_301_cast_fp16, y = scores_pos_301_cast_fp16)[name = string("logits_301_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_24040_cast_fp16 = softmax(axis = var_23039, x = logits_301_cast_fp16)[name = string("op_24040_cast_fp16")];
            bool var_24042_transpose_x_1 = const()[name = string("op_24042_transpose_x_1"), val = bool(false)];
            bool var_24042_transpose_y_1 = const()[name = string("op_24042_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_24042_cast_fp16 = matmul(transpose_x = var_24042_transpose_x_1, transpose_y = var_24042_transpose_y_1, x = var_24040_cast_fp16, y = v_head_603_cast_fp16)[name = string("op_24042_cast_fp16")];
            tensor<int32, [1]> var_24043_axes_1 = const()[name = string("op_24043_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_24043_cast_fp16 = squeeze(axes = var_24043_axes_1, x = var_24042_cast_fp16)[name = string("op_24043_cast_fp16")];
            string dense_output_1513_pad_type_1 = const()[name = string("dense_output_1513_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1513_strides_1 = const()[name = string("dense_output_1513_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1513_pad_1 = const()[name = string("dense_output_1513_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1513_dilations_1 = const()[name = string("dense_output_1513_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1513_groups_1 = const()[name = string("dense_output_1513_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545030016))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545161152))))[name = string("layers_18_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1513_cast_fp16 = conv(dilations = dense_output_1513_dilations_1, groups = dense_output_1513_groups_1, pad = dense_output_1513_pad_1, pad_type = dense_output_1513_pad_type_1, strides = dense_output_1513_strides_1, weight = layers_18_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1513_cast_fp16")];
            string sparse_output_1513_pad_type_1 = const()[name = string("sparse_output_1513_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1513_strides_1 = const()[name = string("sparse_output_1513_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1513_pad_1 = const()[name = string("sparse_output_1513_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1513_dilations_1 = const()[name = string("sparse_output_1513_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1513_groups_1 = const()[name = string("sparse_output_1513_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545164416))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545161728))))[name = string("layers_18_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1513_cast_fp16 = conv(dilations = sparse_output_1513_dilations_1, groups = sparse_output_1513_groups_1, pad = sparse_output_1513_pad_1, pad_type = sparse_output_1513_pad_type_1, strides = sparse_output_1513_strides_1, weight = layers_18_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1513_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24058_cast_fp16 = add(x = dense_output_1513_cast_fp16, y = sparse_output_1513_cast_fp16)[name = string("op_24058_cast_fp16")];
            tensor<int32, [4]> var_24059 = const()[name = string("op_24059"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24061 = const()[name = string("op_24061"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24060_cast_fp16 = transpose(perm = var_24059, x = var_24058_cast_fp16)[name = string("transpose_214")];
            tensor<fp16, [1, 51, 128]> q_head_303_cast_fp16 = reshape(shape = var_24061, x = var_24060_cast_fp16)[name = string("q_head_303_cast_fp16")];
            string dense_output_1515_pad_type_1 = const()[name = string("dense_output_1515_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1515_strides_1 = const()[name = string("dense_output_1515_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1515_pad_1 = const()[name = string("dense_output_1515_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1515_dilations_1 = const()[name = string("dense_output_1515_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1515_groups_1 = const()[name = string("dense_output_1515_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545180864))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545312000))))[name = string("layers_18_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1515_cast_fp16 = conv(dilations = dense_output_1515_dilations_1, groups = dense_output_1515_groups_1, pad = dense_output_1515_pad_1, pad_type = dense_output_1515_pad_type_1, strides = dense_output_1515_strides_1, weight = layers_18_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1515_cast_fp16")];
            string sparse_output_1515_pad_type_1 = const()[name = string("sparse_output_1515_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1515_strides_1 = const()[name = string("sparse_output_1515_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1515_pad_1 = const()[name = string("sparse_output_1515_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1515_dilations_1 = const()[name = string("sparse_output_1515_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1515_groups_1 = const()[name = string("sparse_output_1515_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545315264))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545312576))))[name = string("layers_18_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1515_cast_fp16 = conv(dilations = sparse_output_1515_dilations_1, groups = sparse_output_1515_groups_1, pad = sparse_output_1515_pad_1, pad_type = sparse_output_1515_pad_type_1, strides = sparse_output_1515_strides_1, weight = layers_18_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1515_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24077_cast_fp16 = add(x = dense_output_1515_cast_fp16, y = sparse_output_1515_cast_fp16)[name = string("op_24077_cast_fp16")];
            tensor<int32, [4]> var_24078 = const()[name = string("op_24078"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24080 = const()[name = string("op_24080"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24079_cast_fp16 = transpose(perm = var_24078, x = var_24077_cast_fp16)[name = string("transpose_213")];
            tensor<fp16, [1, 51, 128]> k_head_605_cast_fp16 = reshape(shape = var_24080, x = var_24079_cast_fp16)[name = string("k_head_605_cast_fp16")];
            string dense_output_1517_pad_type_1 = const()[name = string("dense_output_1517_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1517_strides_1 = const()[name = string("dense_output_1517_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1517_pad_1 = const()[name = string("dense_output_1517_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1517_dilations_1 = const()[name = string("dense_output_1517_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1517_groups_1 = const()[name = string("dense_output_1517_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545331712))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545462848))))[name = string("layers_18_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1517_cast_fp16 = conv(dilations = dense_output_1517_dilations_1, groups = dense_output_1517_groups_1, pad = dense_output_1517_pad_1, pad_type = dense_output_1517_pad_type_1, strides = dense_output_1517_strides_1, weight = layers_18_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = string("dense_output_1517_cast_fp16")];
            string sparse_output_1517_pad_type_1 = const()[name = string("sparse_output_1517_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1517_strides_1 = const()[name = string("sparse_output_1517_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1517_pad_1 = const()[name = string("sparse_output_1517_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1517_dilations_1 = const()[name = string("sparse_output_1517_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1517_groups_1 = const()[name = string("sparse_output_1517_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545466112))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545463424))))[name = string("layers_18_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1517_cast_fp16 = conv(dilations = sparse_output_1517_dilations_1, groups = sparse_output_1517_groups_1, pad = sparse_output_1517_pad_1, pad_type = sparse_output_1517_pad_type_1, strides = sparse_output_1517_strides_1, weight = layers_18_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_857_cast_fp16)[name = string("sparse_output_1517_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24096_cast_fp16 = add(x = dense_output_1517_cast_fp16, y = sparse_output_1517_cast_fp16)[name = string("op_24096_cast_fp16")];
            tensor<int32, [4]> var_24097 = const()[name = string("op_24097"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24099 = const()[name = string("op_24099"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24098_cast_fp16 = transpose(perm = var_24097, x = var_24096_cast_fp16)[name = string("transpose_212")];
            tensor<fp16, [1, 51, 128]> v_head_605_cast_fp16 = reshape(shape = var_24099, x = var_24098_cast_fp16)[name = string("v_head_605_cast_fp16")];
            string dense_output_1519_pad_type_1 = const()[name = string("dense_output_1519_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1519_strides_1 = const()[name = string("dense_output_1519_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1519_pad_1 = const()[name = string("dense_output_1519_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1519_dilations_1 = const()[name = string("dense_output_1519_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1519_groups_1 = const()[name = string("dense_output_1519_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545482560))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545613696))))[name = string("layers_18_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1519_cast_fp16 = conv(dilations = dense_output_1519_dilations_1, groups = dense_output_1519_groups_1, pad = dense_output_1519_pad_1, pad_type = dense_output_1519_pad_type_1, strides = dense_output_1519_strides_1, weight = layers_18_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1519_cast_fp16")];
            string sparse_output_1519_pad_type_1 = const()[name = string("sparse_output_1519_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1519_strides_1 = const()[name = string("sparse_output_1519_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1519_pad_1 = const()[name = string("sparse_output_1519_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1519_dilations_1 = const()[name = string("sparse_output_1519_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1519_groups_1 = const()[name = string("sparse_output_1519_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_18_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545616960))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545614272))))[name = string("layers_18_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1519_cast_fp16 = conv(dilations = sparse_output_1519_dilations_1, groups = sparse_output_1519_groups_1, pad = sparse_output_1519_pad_1, pad_type = sparse_output_1519_pad_type_1, strides = sparse_output_1519_strides_1, weight = layers_18_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1519_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_24115_cast_fp16 = add(x = dense_output_1519_cast_fp16, y = sparse_output_1519_cast_fp16)[name = string("op_24115_cast_fp16")];
            tensor<int32, [4]> var_24116 = const()[name = string("op_24116"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24118 = const()[name = string("op_24118"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_24117_cast_fp16 = transpose(perm = var_24116, x = var_24115_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 101, 128]> p_head_605_cast_fp16 = reshape(shape = var_24118, x = var_24117_cast_fp16)[name = string("p_head_605_cast_fp16")];
            tensor<fp16, [128]> var_24120_to_fp16 = const()[name = string("op_24120_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545633408)))];
            tensor<fp16, [1, 51, 128]> var_24121_cast_fp16 = add(x = q_head_303_cast_fp16, y = var_24120_to_fp16)[name = string("op_24121_cast_fp16")];
            tensor<int32, [1]> q_u_303_axes_1 = const()[name = string("q_u_303_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_303_cast_fp16 = expand_dims(axes = q_u_303_axes_1, x = var_24121_cast_fp16)[name = string("q_u_303_cast_fp16")];
            tensor<fp16, [128]> var_24123_to_fp16 = const()[name = string("op_24123_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545633728)))];
            tensor<fp16, [1, 51, 128]> var_24124_cast_fp16 = add(x = q_head_303_cast_fp16, y = var_24123_to_fp16)[name = string("op_24124_cast_fp16")];
            tensor<int32, [1]> q_v_303_axes_1 = const()[name = string("q_v_303_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_303_cast_fp16 = expand_dims(axes = q_v_303_axes_1, x = var_24124_cast_fp16)[name = string("q_v_303_cast_fp16")];
            tensor<int32, [1]> k_head_607_axes_1 = const()[name = string("k_head_607_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_607_cast_fp16 = expand_dims(axes = k_head_607_axes_1, x = k_head_605_cast_fp16)[name = string("k_head_607_cast_fp16")];
            tensor<int32, [1]> v_head_607_axes_1 = const()[name = string("v_head_607_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_607_cast_fp16 = expand_dims(axes = v_head_607_axes_1, x = v_head_605_cast_fp16)[name = string("v_head_607_cast_fp16")];
            tensor<int32, [1]> p_head_607_axes_1 = const()[name = string("p_head_607_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_607_cast_fp16 = expand_dims(axes = p_head_607_axes_1, x = p_head_605_cast_fp16)[name = string("p_head_607_cast_fp16")];
            bool var_24130_transpose_x_3 = const()[name = string("op_24130_transpose_x_3"), val = bool(false)];
            bool var_24130_transpose_y_3 = const()[name = string("op_24130_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_24130_cast_fp16 = matmul(transpose_x = var_24130_transpose_x_3, transpose_y = var_24130_transpose_y_3, x = q_u_303_cast_fp16, y = k_head_607_cast_fp16)[name = string("op_24130_cast_fp16")];
            fp16 var_24131_to_fp16 = const()[name = string("op_24131_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_303_cast_fp16 = mul(x = var_24130_cast_fp16, y = var_24131_to_fp16)[name = string("scores_content_303_cast_fp16")];
            bool x_1589_transpose_x_3 = const()[name = string("x_1589_transpose_x_3"), val = bool(false)];
            bool x_1589_transpose_y_3 = const()[name = string("x_1589_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1589_cast_fp16 = matmul(transpose_x = x_1589_transpose_x_3, transpose_y = x_1589_transpose_y_3, x = q_v_303_cast_fp16, y = p_head_607_cast_fp16)[name = string("x_1589_cast_fp16")];
            tensor<int32, [8]> x_1591_pad_1 = const()[name = string("x_1591_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1591_mode_1 = const()[name = string("x_1591_mode_1"), val = string("constant")];
            fp16 const_2323_to_fp16 = const()[name = string("const_2323_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1591_cast_fp16 = pad(constant_val = const_2323_to_fp16, mode = x_1591_mode_1, pad = x_1591_pad_1, x = x_1589_cast_fp16)[name = string("x_1591_cast_fp16")];
            tensor<int32, [4]> var_24145 = const()[name = string("op_24145"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1593_cast_fp16 = reshape(shape = var_24145, x = x_1591_cast_fp16)[name = string("x_1593_cast_fp16")];
            tensor<int32, [4]> var_24149_begin_1 = const()[name = string("op_24149_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_24149_end_1 = const()[name = string("op_24149_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_24149_end_mask_1 = const()[name = string("op_24149_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_24149_cast_fp16 = slice_by_index(begin = var_24149_begin_1, end = var_24149_end_1, end_mask = var_24149_end_mask_1, x = x_1593_cast_fp16)[name = string("op_24149_cast_fp16")];
            tensor<int32, [4]> var_24151 = const()[name = string("op_24151"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_24152_cast_fp16 = reshape(shape = var_24151, x = var_24149_cast_fp16)[name = string("op_24152_cast_fp16")];
            tensor<int32, [4]> var_24157_begin_1 = const()[name = string("op_24157_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_24157_end_1 = const()[name = string("op_24157_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_24157_end_mask_1 = const()[name = string("op_24157_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_24157_cast_fp16 = slice_by_index(begin = var_24157_begin_1, end = var_24157_end_1, end_mask = var_24157_end_mask_1, x = var_24152_cast_fp16)[name = string("op_24157_cast_fp16")];
            fp16 var_24158_to_fp16 = const()[name = string("op_24158_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_303_cast_fp16 = mul(x = var_24157_cast_fp16, y = var_24158_to_fp16)[name = string("scores_pos_303_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_303_cast_fp16 = add(x = scores_content_303_cast_fp16, y = scores_pos_303_cast_fp16)[name = string("logits_303_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_24161_cast_fp16 = softmax(axis = var_23039, x = logits_303_cast_fp16)[name = string("op_24161_cast_fp16")];
            bool var_24163_transpose_x_1 = const()[name = string("op_24163_transpose_x_1"), val = bool(false)];
            bool var_24163_transpose_y_1 = const()[name = string("op_24163_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_24163_cast_fp16 = matmul(transpose_x = var_24163_transpose_x_1, transpose_y = var_24163_transpose_y_1, x = var_24161_cast_fp16, y = v_head_607_cast_fp16)[name = string("op_24163_cast_fp16")];
            tensor<int32, [1]> o_head_37_axes_1 = const()[name = string("o_head_37_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_37_cast_fp16 = squeeze(axes = o_head_37_axes_1, x = var_24163_cast_fp16)[name = string("o_head_37_cast_fp16")];
            bool out_37_interleave_1 = const()[name = string("out_37_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_37_cast_fp16 = concat(axis = var_23039, interleave = out_37_interleave_1, values = (var_23317_cast_fp16, var_23438_cast_fp16, var_23559_cast_fp16, var_23680_cast_fp16, var_23801_cast_fp16, var_23922_cast_fp16, var_24043_cast_fp16, o_head_37_cast_fp16))[name = string("out_37_cast_fp16")];
            tensor<int32, [3]> var_24167_perm_1 = const()[name = string("op_24167_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_865_axes_1 = const()[name = string("input_865_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_24167_cast_fp16 = transpose(perm = var_24167_perm_1, x = out_37_cast_fp16)[name = string("transpose_210")];
            tensor<fp16, [1, 1024, 51, 1]> input_865_cast_fp16 = expand_dims(axes = input_865_axes_1, x = var_24167_cast_fp16)[name = string("input_865_cast_fp16")];
            string dense_output_1521_pad_type_1 = const()[name = string("dense_output_1521_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1521_strides_1 = const()[name = string("dense_output_1521_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1521_pad_1 = const()[name = string("dense_output_1521_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1521_dilations_1 = const()[name = string("dense_output_1521_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1521_groups_1 = const()[name = string("dense_output_1521_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545634048))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546682688))))[name = string("layers_18_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1521_cast_fp16 = conv(dilations = dense_output_1521_dilations_1, groups = dense_output_1521_groups_1, pad = dense_output_1521_pad_1, pad_type = dense_output_1521_pad_type_1, strides = dense_output_1521_strides_1, weight = layers_18_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_865_cast_fp16)[name = string("dense_output_1521_cast_fp16")];
            string sparse_output_1521_pad_type_1 = const()[name = string("sparse_output_1521_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1521_strides_1 = const()[name = string("sparse_output_1521_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1521_pad_1 = const()[name = string("sparse_output_1521_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1521_dilations_1 = const()[name = string("sparse_output_1521_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1521_groups_1 = const()[name = string("sparse_output_1521_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546704320))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546683264))))[name = string("layers_18_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1521_cast_fp16 = conv(dilations = sparse_output_1521_dilations_1, groups = sparse_output_1521_groups_1, pad = sparse_output_1521_pad_1, pad_type = sparse_output_1521_pad_type_1, strides = sparse_output_1521_strides_1, weight = layers_18_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_865_cast_fp16)[name = string("sparse_output_1521_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_37_cast_fp16 = add(x = dense_output_1521_cast_fp16, y = sparse_output_1521_cast_fp16)[name = string("out_conv_37_cast_fp16")];
            tensor<int32, [1]> var_24184_axes_1 = const()[name = string("op_24184_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_24184_cast_fp16 = squeeze(axes = var_24184_axes_1, x = out_conv_37_cast_fp16)[name = string("op_24184_cast_fp16")];
            tensor<int32, [3]> var_24185_perm_1 = const()[name = string("op_24185_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_24185_cast_fp16 = transpose(perm = var_24185_perm_1, x = var_24184_cast_fp16)[name = string("transpose_209")];
            tensor<fp16, [1, 51, 1024]> input_867_cast_fp16 = add(x = input_855_cast_fp16, y = var_24185_cast_fp16)[name = string("input_867_cast_fp16")];
            tensor<int32, [1]> x_1597_axes_1 = const()[name = string("x_1597_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_18_norm_conv_weight_to_fp16 = const()[name = string("layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546835456)))];
            tensor<fp16, [1024]> layers_18_norm_conv_bias_to_fp16 = const()[name = string("layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546837568)))];
            tensor<fp16, [1, 51, 1024]> x_1597_cast_fp16 = layer_norm(axes = x_1597_axes_1, beta = layers_18_norm_conv_bias_to_fp16, epsilon = var_23054_to_fp16, gamma = layers_18_norm_conv_weight_to_fp16, x = input_867_cast_fp16)[name = string("x_1597_cast_fp16")];
            tensor<int32, [3]> var_24195_perm_1 = const()[name = string("op_24195_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_869_axes_1 = const()[name = string("input_869_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_24195_cast_fp16 = transpose(perm = var_24195_perm_1, x = x_1597_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 1024, 51, 1]> input_869_cast_fp16 = expand_dims(axes = input_869_axes_1, x = var_24195_cast_fp16)[name = string("input_869_cast_fp16")];
            string dense_output_1523_pad_type_1 = const()[name = string("dense_output_1523_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1523_strides_1 = const()[name = string("dense_output_1523_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1523_pad_1 = const()[name = string("dense_output_1523_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1523_dilations_1 = const()[name = string("dense_output_1523_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1523_groups_1 = const()[name = string("dense_output_1523_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_18_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546839680))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548936896))))[name = string("layers_18_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1523_cast_fp16 = conv(dilations = dense_output_1523_dilations_1, groups = dense_output_1523_groups_1, pad = dense_output_1523_pad_1, pad_type = dense_output_1523_pad_type_1, strides = dense_output_1523_strides_1, weight = layers_18_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_869_cast_fp16)[name = string("dense_output_1523_cast_fp16")];
            string sparse_output_1523_pad_type_1 = const()[name = string("sparse_output_1523_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1523_strides_1 = const()[name = string("sparse_output_1523_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1523_pad_1 = const()[name = string("sparse_output_1523_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1523_dilations_1 = const()[name = string("sparse_output_1523_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1523_groups_1 = const()[name = string("sparse_output_1523_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_18_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548979520))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548937472))))[name = string("layers_18_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1523_cast_fp16 = conv(dilations = sparse_output_1523_dilations_1, groups = sparse_output_1523_groups_1, pad = sparse_output_1523_pad_1, pad_type = sparse_output_1523_pad_type_1, strides = sparse_output_1523_strides_1, weight = layers_18_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_869_cast_fp16)[name = string("sparse_output_1523_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_871_cast_fp16 = add(x = dense_output_1523_cast_fp16, y = sparse_output_1523_cast_fp16)[name = string("input_871_cast_fp16")];
            int32 input_873_split_num_splits_1 = const()[name = string("input_873_split_num_splits_1"), val = int32(2)];
            int32 input_873_split_axis_1 = const()[name = string("input_873_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_873_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_873_split_cast_fp16_1 = split(axis = input_873_split_axis_1, num_splits = input_873_split_num_splits_1, x = input_871_cast_fp16)[name = string("input_873_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_873_split_1_sigmoid_cast_fp16 = sigmoid(x = input_873_split_cast_fp16_1)[name = string("input_873_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_873_cast_fp16 = mul(x = input_873_split_cast_fp16_0, y = input_873_split_1_sigmoid_cast_fp16)[name = string("input_873_cast_fp16")];
            tensor<int32, [8]> input_875_pad_1 = const()[name = string("input_875_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_875_mode_1 = const()[name = string("input_875_mode_1"), val = string("constant")];
            fp16 const_2325_to_fp16 = const()[name = string("const_2325_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_875_cast_fp16 = pad(constant_val = const_2325_to_fp16, mode = input_875_mode_1, pad = input_875_pad_1, x = input_873_cast_fp16)[name = string("input_875_cast_fp16")];
            string dense_output_1525_pad_type_1 = const()[name = string("dense_output_1525_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1525_strides_1 = const()[name = string("dense_output_1525_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1525_pad_1 = const()[name = string("dense_output_1525_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1525_dilations_1 = const()[name = string("dense_output_1525_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1525_groups_1 = const()[name = string("dense_output_1525_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1525_cast_fp16 = conv(dilations = dense_output_1525_dilations_1, groups = dense_output_1525_groups_1, pad = dense_output_1525_pad_1, pad_type = dense_output_1525_pad_type_1, strides = dense_output_1525_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_875_cast_fp16)[name = string("dense_output_1525_cast_fp16")];
            string sparse_output_1525_pad_type_1 = const()[name = string("sparse_output_1525_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1525_strides_1 = const()[name = string("sparse_output_1525_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1525_pad_1 = const()[name = string("sparse_output_1525_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1525_dilations_1 = const()[name = string("sparse_output_1525_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1525_groups_1 = const()[name = string("sparse_output_1525_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_18_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549241728))))[name = string("layers_18_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1525_cast_fp16 = conv(dilations = sparse_output_1525_dilations_1, groups = sparse_output_1525_groups_1, pad = sparse_output_1525_pad_1, pad_type = sparse_output_1525_pad_type_1, strides = sparse_output_1525_strides_1, weight = layers_18_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_875_cast_fp16)[name = string("sparse_output_1525_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_877_cast_fp16 = add(x = dense_output_1525_cast_fp16, y = sparse_output_1525_cast_fp16)[name = string("input_877_cast_fp16")];
            tensor<fp16, [1024]> layers_18_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_18_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549260224)))];
            tensor<fp16, [1024]> layers_18_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_18_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549262336)))];
            tensor<fp16, [1024]> layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_18_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549264448)))];
            tensor<fp16, [1024]> layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_18_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549266560)))];
            tensor<fp16, [1, 1024, 51, 1]> input_879_cast_fp16 = batch_norm(beta = layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_23054_to_fp16, gamma = layers_18_conv_batch_norm_weight_to_fp16, mean = layers_18_conv_batch_norm_running_mean_to_fp16, variance = layers_18_conv_batch_norm_running_var_to_fp16, x = input_877_cast_fp16)[name = string("input_879_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_881_cast_fp16 = silu(x = input_879_cast_fp16)[name = string("input_881_cast_fp16")];
            string dense_output_1527_pad_type_1 = const()[name = string("dense_output_1527_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1527_strides_1 = const()[name = string("dense_output_1527_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1527_pad_1 = const()[name = string("dense_output_1527_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1527_dilations_1 = const()[name = string("dense_output_1527_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1527_groups_1 = const()[name = string("dense_output_1527_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549268672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550317312))))[name = string("layers_18_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1527_cast_fp16 = conv(dilations = dense_output_1527_dilations_1, groups = dense_output_1527_groups_1, pad = dense_output_1527_pad_1, pad_type = dense_output_1527_pad_type_1, strides = dense_output_1527_strides_1, weight = layers_18_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_881_cast_fp16)[name = string("dense_output_1527_cast_fp16")];
            string sparse_output_1527_pad_type_1 = const()[name = string("sparse_output_1527_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1527_strides_1 = const()[name = string("sparse_output_1527_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1527_pad_1 = const()[name = string("sparse_output_1527_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1527_dilations_1 = const()[name = string("sparse_output_1527_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1527_groups_1 = const()[name = string("sparse_output_1527_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_18_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550338944))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550317888))))[name = string("layers_18_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1527_cast_fp16 = conv(dilations = sparse_output_1527_dilations_1, groups = sparse_output_1527_groups_1, pad = sparse_output_1527_pad_1, pad_type = sparse_output_1527_pad_type_1, strides = sparse_output_1527_strides_1, weight = layers_18_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_881_cast_fp16)[name = string("sparse_output_1527_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1599_cast_fp16 = add(x = dense_output_1527_cast_fp16, y = sparse_output_1527_cast_fp16)[name = string("x_1599_cast_fp16")];
            tensor<int32, [1]> var_24251_axes_1 = const()[name = string("op_24251_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_24251_cast_fp16 = squeeze(axes = var_24251_axes_1, x = x_1599_cast_fp16)[name = string("op_24251_cast_fp16")];
            tensor<int32, [3]> var_24252_perm_1 = const()[name = string("op_24252_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_24252_cast_fp16 = transpose(perm = var_24252_perm_1, x = var_24251_cast_fp16)[name = string("transpose_207")];
            tensor<fp16, [1, 51, 1024]> input_883_cast_fp16 = add(x = input_867_cast_fp16, y = var_24252_cast_fp16)[name = string("input_883_cast_fp16")];
            tensor<int32, [1]> x_1601_axes_1 = const()[name = string("x_1601_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550470080)))];
            tensor<fp16, [1024]> layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550472192)))];
            tensor<fp16, [1, 51, 1024]> x_1601_cast_fp16 = layer_norm(axes = x_1601_axes_1, beta = layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_23054_to_fp16, gamma = layers_18_norm_feed_forward2_weight_to_fp16, x = input_883_cast_fp16)[name = string("x_1601_cast_fp16")];
            tensor<int32, [4]> var_24262 = const()[name = string("op_24262"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1603_cast_fp16 = reshape(shape = var_24262, x = x_1601_cast_fp16)[name = string("x_1603_cast_fp16")];
            tensor<int32, [4]> input_885_perm_1 = const()[name = string("input_885_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1529_pad_type_1 = const()[name = string("dense_output_1529_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1529_strides_1 = const()[name = string("dense_output_1529_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1529_pad_1 = const()[name = string("dense_output_1529_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1529_dilations_1 = const()[name = string("dense_output_1529_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1529_groups_1 = const()[name = string("dense_output_1529_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_18_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550474304))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554668672))))[name = string("layers_18_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_885_cast_fp16 = transpose(perm = input_885_perm_1, x = x_1603_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1529_cast_fp16 = conv(dilations = dense_output_1529_dilations_1, groups = dense_output_1529_groups_1, pad = dense_output_1529_pad_1, pad_type = dense_output_1529_pad_type_1, strides = dense_output_1529_strides_1, weight = layers_18_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_885_cast_fp16)[name = string("dense_output_1529_cast_fp16")];
            string sparse_output_1529_pad_type_1 = const()[name = string("sparse_output_1529_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1529_strides_1 = const()[name = string("sparse_output_1529_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1529_pad_1 = const()[name = string("sparse_output_1529_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1529_dilations_1 = const()[name = string("sparse_output_1529_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1529_groups_1 = const()[name = string("sparse_output_1529_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_18_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554753216))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554669248))))[name = string("layers_18_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1529_cast_fp16 = conv(dilations = sparse_output_1529_dilations_1, groups = sparse_output_1529_groups_1, pad = sparse_output_1529_pad_1, pad_type = sparse_output_1529_pad_type_1, strides = sparse_output_1529_strides_1, weight = layers_18_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_885_cast_fp16)[name = string("sparse_output_1529_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_887_cast_fp16 = add(x = dense_output_1529_cast_fp16, y = sparse_output_1529_cast_fp16)[name = string("input_887_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_889_cast_fp16 = silu(x = input_887_cast_fp16)[name = string("input_889_cast_fp16")];
            string dense_output_1531_pad_type_1 = const()[name = string("dense_output_1531_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1531_strides_1 = const()[name = string("dense_output_1531_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1531_pad_1 = const()[name = string("dense_output_1531_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1531_dilations_1 = const()[name = string("dense_output_1531_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1531_groups_1 = const()[name = string("dense_output_1531_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_18_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555277568))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(559471936))))[name = string("layers_18_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1531_cast_fp16 = conv(dilations = dense_output_1531_dilations_1, groups = dense_output_1531_groups_1, pad = dense_output_1531_pad_1, pad_type = dense_output_1531_pad_type_1, strides = dense_output_1531_strides_1, weight = layers_18_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_889_cast_fp16)[name = string("dense_output_1531_cast_fp16")];
            string sparse_output_1531_pad_type_1 = const()[name = string("sparse_output_1531_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1531_strides_1 = const()[name = string("sparse_output_1531_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1531_pad_1 = const()[name = string("sparse_output_1531_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1531_dilations_1 = const()[name = string("sparse_output_1531_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1531_groups_1 = const()[name = string("sparse_output_1531_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_18_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(559556480))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(559472512))))[name = string("layers_18_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1531_cast_fp16 = conv(dilations = sparse_output_1531_dilations_1, groups = sparse_output_1531_groups_1, pad = sparse_output_1531_pad_1, pad_type = sparse_output_1531_pad_type_1, strides = sparse_output_1531_strides_1, weight = layers_18_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_889_cast_fp16)[name = string("sparse_output_1531_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1605_cast_fp16 = add(x = dense_output_1531_cast_fp16, y = sparse_output_1531_cast_fp16)[name = string("x_1605_cast_fp16")];
            tensor<int32, [4]> x_1607_perm_1 = const()[name = string("x_1607_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_24297 = const()[name = string("op_24297"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1607_cast_fp16 = transpose(perm = x_1607_perm_1, x = x_1605_cast_fp16)[name = string("transpose_205")];
            tensor<fp16, [1, 51, 1024]> var_24298_cast_fp16 = reshape(shape = var_24297, x = x_1607_cast_fp16)[name = string("op_24298_cast_fp16")];
            fp16 var_24299_to_fp16 = const()[name = string("op_24299_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_24300_cast_fp16 = mul(x = var_24298_cast_fp16, y = var_24299_to_fp16)[name = string("op_24300_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_891_cast_fp16 = add(x = input_883_cast_fp16, y = var_24300_cast_fp16)[name = string("input_891_cast_fp16")];
            tensor<int32, [1]> input_893_axes_1 = const()[name = string("input_893_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_18_norm_out_weight_to_fp16 = const()[name = string("layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560080832)))];
            tensor<fp16, [1024]> layers_18_norm_out_bias_to_fp16 = const()[name = string("layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560082944)))];
            tensor<fp16, [1, 51, 1024]> input_893_cast_fp16 = layer_norm(axes = input_893_axes_1, beta = layers_18_norm_out_bias_to_fp16, epsilon = var_23054_to_fp16, gamma = layers_18_norm_out_weight_to_fp16, x = input_891_cast_fp16)[name = string("input_893_cast_fp16")];
            int32 var_24308 = const()[name = string("op_24308"), val = int32(-1)];
            tensor<int32, [1]> x_1609_axes_1 = const()[name = string("x_1609_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560085056)))];
            tensor<fp16, [1024]> layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560087168)))];
            fp16 var_24323_to_fp16 = const()[name = string("op_24323_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1609_cast_fp16 = layer_norm(axes = x_1609_axes_1, beta = layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_24323_to_fp16, gamma = layers_19_norm_feed_forward1_weight_to_fp16, x = input_893_cast_fp16)[name = string("x_1609_cast_fp16")];
            tensor<int32, [4]> var_24342 = const()[name = string("op_24342"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1611_cast_fp16 = reshape(shape = var_24342, x = x_1609_cast_fp16)[name = string("x_1611_cast_fp16")];
            tensor<int32, [4]> input_895_perm_1 = const()[name = string("input_895_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1533_pad_type_1 = const()[name = string("dense_output_1533_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1533_strides_1 = const()[name = string("dense_output_1533_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1533_pad_1 = const()[name = string("dense_output_1533_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1533_dilations_1 = const()[name = string("dense_output_1533_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1533_groups_1 = const()[name = string("dense_output_1533_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_19_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560089280))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564283648))))[name = string("layers_19_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_895_cast_fp16 = transpose(perm = input_895_perm_1, x = x_1611_cast_fp16)[name = string("transpose_204")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1533_cast_fp16 = conv(dilations = dense_output_1533_dilations_1, groups = dense_output_1533_groups_1, pad = dense_output_1533_pad_1, pad_type = dense_output_1533_pad_type_1, strides = dense_output_1533_strides_1, weight = layers_19_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_895_cast_fp16)[name = string("dense_output_1533_cast_fp16")];
            string sparse_output_1533_pad_type_1 = const()[name = string("sparse_output_1533_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1533_strides_1 = const()[name = string("sparse_output_1533_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1533_pad_1 = const()[name = string("sparse_output_1533_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1533_dilations_1 = const()[name = string("sparse_output_1533_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1533_groups_1 = const()[name = string("sparse_output_1533_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_19_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564368192))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564284224))))[name = string("layers_19_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1533_cast_fp16 = conv(dilations = sparse_output_1533_dilations_1, groups = sparse_output_1533_groups_1, pad = sparse_output_1533_pad_1, pad_type = sparse_output_1533_pad_type_1, strides = sparse_output_1533_strides_1, weight = layers_19_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_895_cast_fp16)[name = string("sparse_output_1533_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_897_cast_fp16 = add(x = dense_output_1533_cast_fp16, y = sparse_output_1533_cast_fp16)[name = string("input_897_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_899_cast_fp16 = silu(x = input_897_cast_fp16)[name = string("input_899_cast_fp16")];
            string dense_output_1535_pad_type_1 = const()[name = string("dense_output_1535_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1535_strides_1 = const()[name = string("dense_output_1535_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1535_pad_1 = const()[name = string("dense_output_1535_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1535_dilations_1 = const()[name = string("dense_output_1535_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1535_groups_1 = const()[name = string("dense_output_1535_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_19_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564892544))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569086912))))[name = string("layers_19_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1535_cast_fp16 = conv(dilations = dense_output_1535_dilations_1, groups = dense_output_1535_groups_1, pad = dense_output_1535_pad_1, pad_type = dense_output_1535_pad_type_1, strides = dense_output_1535_strides_1, weight = layers_19_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_899_cast_fp16)[name = string("dense_output_1535_cast_fp16")];
            string sparse_output_1535_pad_type_1 = const()[name = string("sparse_output_1535_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1535_strides_1 = const()[name = string("sparse_output_1535_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1535_pad_1 = const()[name = string("sparse_output_1535_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1535_dilations_1 = const()[name = string("sparse_output_1535_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1535_groups_1 = const()[name = string("sparse_output_1535_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_19_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569171456))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569087488))))[name = string("layers_19_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1535_cast_fp16 = conv(dilations = sparse_output_1535_dilations_1, groups = sparse_output_1535_groups_1, pad = sparse_output_1535_pad_1, pad_type = sparse_output_1535_pad_type_1, strides = sparse_output_1535_strides_1, weight = layers_19_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_899_cast_fp16)[name = string("sparse_output_1535_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1613_cast_fp16 = add(x = dense_output_1535_cast_fp16, y = sparse_output_1535_cast_fp16)[name = string("x_1613_cast_fp16")];
            tensor<int32, [4]> x_1615_perm_1 = const()[name = string("x_1615_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_24377 = const()[name = string("op_24377"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1615_cast_fp16 = transpose(perm = x_1615_perm_1, x = x_1613_cast_fp16)[name = string("transpose_203")];
            tensor<fp16, [1, 51, 1024]> var_24378_cast_fp16 = reshape(shape = var_24377, x = x_1615_cast_fp16)[name = string("op_24378_cast_fp16")];
            fp16 var_24379_to_fp16 = const()[name = string("op_24379_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_24380_cast_fp16 = mul(x = var_24378_cast_fp16, y = var_24379_to_fp16)[name = string("op_24380_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_901_cast_fp16 = add(x = input_893_cast_fp16, y = var_24380_cast_fp16)[name = string("input_901_cast_fp16")];
            tensor<int32, [1]> q_39_axes_1 = const()[name = string("q_39_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_19_norm_self_att_weight_to_fp16 = const()[name = string("layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569695808)))];
            tensor<fp16, [1024]> layers_19_norm_self_att_bias_to_fp16 = const()[name = string("layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569697920)))];
            tensor<fp16, [1, 51, 1024]> q_39_cast_fp16 = layer_norm(axes = q_39_axes_1, beta = layers_19_norm_self_att_bias_to_fp16, epsilon = var_24323_to_fp16, gamma = layers_19_norm_self_att_weight_to_fp16, x = input_901_cast_fp16)[name = string("q_39_cast_fp16")];
            tensor<int32, [3]> var_24454 = const()[name = string("op_24454"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_903_axes_1 = const()[name = string("input_903_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_24455_cast_fp16 = transpose(perm = var_24454, x = q_39_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 1024, 51, 1]> input_903_cast_fp16 = expand_dims(axes = input_903_axes_1, x = var_24455_cast_fp16)[name = string("input_903_cast_fp16")];
            string dense_output_1537_pad_type_1 = const()[name = string("dense_output_1537_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1537_strides_1 = const()[name = string("dense_output_1537_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1537_pad_1 = const()[name = string("dense_output_1537_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1537_dilations_1 = const()[name = string("dense_output_1537_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1537_groups_1 = const()[name = string("dense_output_1537_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569700032))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569831168))))[name = string("layers_19_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1537_cast_fp16 = conv(dilations = dense_output_1537_dilations_1, groups = dense_output_1537_groups_1, pad = dense_output_1537_pad_1, pad_type = dense_output_1537_pad_type_1, strides = dense_output_1537_strides_1, weight = layers_19_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1537_cast_fp16")];
            string sparse_output_1537_pad_type_1 = const()[name = string("sparse_output_1537_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1537_strides_1 = const()[name = string("sparse_output_1537_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1537_pad_1 = const()[name = string("sparse_output_1537_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1537_dilations_1 = const()[name = string("sparse_output_1537_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1537_groups_1 = const()[name = string("sparse_output_1537_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569834432))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569831744))))[name = string("layers_19_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1537_cast_fp16 = conv(dilations = sparse_output_1537_dilations_1, groups = sparse_output_1537_groups_1, pad = sparse_output_1537_pad_1, pad_type = sparse_output_1537_pad_type_1, strides = sparse_output_1537_strides_1, weight = layers_19_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1537_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24480_cast_fp16 = add(x = dense_output_1537_cast_fp16, y = sparse_output_1537_cast_fp16)[name = string("op_24480_cast_fp16")];
            tensor<int32, [4]> var_24481 = const()[name = string("op_24481"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24483 = const()[name = string("op_24483"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24482_cast_fp16 = transpose(perm = var_24481, x = var_24480_cast_fp16)[name = string("transpose_201")];
            tensor<fp16, [1, 51, 128]> q_head_305_cast_fp16 = reshape(shape = var_24483, x = var_24482_cast_fp16)[name = string("q_head_305_cast_fp16")];
            string dense_output_1539_pad_type_1 = const()[name = string("dense_output_1539_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1539_strides_1 = const()[name = string("dense_output_1539_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1539_pad_1 = const()[name = string("dense_output_1539_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1539_dilations_1 = const()[name = string("dense_output_1539_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1539_groups_1 = const()[name = string("dense_output_1539_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569850880))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569982016))))[name = string("layers_19_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1539_cast_fp16 = conv(dilations = dense_output_1539_dilations_1, groups = dense_output_1539_groups_1, pad = dense_output_1539_pad_1, pad_type = dense_output_1539_pad_type_1, strides = dense_output_1539_strides_1, weight = layers_19_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1539_cast_fp16")];
            string sparse_output_1539_pad_type_1 = const()[name = string("sparse_output_1539_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1539_strides_1 = const()[name = string("sparse_output_1539_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1539_pad_1 = const()[name = string("sparse_output_1539_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1539_dilations_1 = const()[name = string("sparse_output_1539_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1539_groups_1 = const()[name = string("sparse_output_1539_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569985280))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569982592))))[name = string("layers_19_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1539_cast_fp16 = conv(dilations = sparse_output_1539_dilations_1, groups = sparse_output_1539_groups_1, pad = sparse_output_1539_pad_1, pad_type = sparse_output_1539_pad_type_1, strides = sparse_output_1539_strides_1, weight = layers_19_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1539_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24499_cast_fp16 = add(x = dense_output_1539_cast_fp16, y = sparse_output_1539_cast_fp16)[name = string("op_24499_cast_fp16")];
            tensor<int32, [4]> var_24500 = const()[name = string("op_24500"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24502 = const()[name = string("op_24502"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24501_cast_fp16 = transpose(perm = var_24500, x = var_24499_cast_fp16)[name = string("transpose_200")];
            tensor<fp16, [1, 51, 128]> k_head_609_cast_fp16 = reshape(shape = var_24502, x = var_24501_cast_fp16)[name = string("k_head_609_cast_fp16")];
            string dense_output_1541_pad_type_1 = const()[name = string("dense_output_1541_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1541_strides_1 = const()[name = string("dense_output_1541_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1541_pad_1 = const()[name = string("dense_output_1541_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1541_dilations_1 = const()[name = string("dense_output_1541_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1541_groups_1 = const()[name = string("dense_output_1541_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570001728))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570132864))))[name = string("layers_19_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1541_cast_fp16 = conv(dilations = dense_output_1541_dilations_1, groups = dense_output_1541_groups_1, pad = dense_output_1541_pad_1, pad_type = dense_output_1541_pad_type_1, strides = dense_output_1541_strides_1, weight = layers_19_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1541_cast_fp16")];
            string sparse_output_1541_pad_type_1 = const()[name = string("sparse_output_1541_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1541_strides_1 = const()[name = string("sparse_output_1541_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1541_pad_1 = const()[name = string("sparse_output_1541_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1541_dilations_1 = const()[name = string("sparse_output_1541_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1541_groups_1 = const()[name = string("sparse_output_1541_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570136128))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570133440))))[name = string("layers_19_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1541_cast_fp16 = conv(dilations = sparse_output_1541_dilations_1, groups = sparse_output_1541_groups_1, pad = sparse_output_1541_pad_1, pad_type = sparse_output_1541_pad_type_1, strides = sparse_output_1541_strides_1, weight = layers_19_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1541_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24518_cast_fp16 = add(x = dense_output_1541_cast_fp16, y = sparse_output_1541_cast_fp16)[name = string("op_24518_cast_fp16")];
            tensor<int32, [4]> var_24519 = const()[name = string("op_24519"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24521 = const()[name = string("op_24521"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24520_cast_fp16 = transpose(perm = var_24519, x = var_24518_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 51, 128]> v_head_609_cast_fp16 = reshape(shape = var_24521, x = var_24520_cast_fp16)[name = string("v_head_609_cast_fp16")];
            string dense_output_1543_pad_type_1 = const()[name = string("dense_output_1543_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1543_strides_1 = const()[name = string("dense_output_1543_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1543_pad_1 = const()[name = string("dense_output_1543_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1543_dilations_1 = const()[name = string("dense_output_1543_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1543_groups_1 = const()[name = string("dense_output_1543_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570152576))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570283712))))[name = string("layers_19_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1543_cast_fp16 = conv(dilations = dense_output_1543_dilations_1, groups = dense_output_1543_groups_1, pad = dense_output_1543_pad_1, pad_type = dense_output_1543_pad_type_1, strides = dense_output_1543_strides_1, weight = layers_19_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1543_cast_fp16")];
            string sparse_output_1543_pad_type_1 = const()[name = string("sparse_output_1543_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1543_strides_1 = const()[name = string("sparse_output_1543_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1543_pad_1 = const()[name = string("sparse_output_1543_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1543_dilations_1 = const()[name = string("sparse_output_1543_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1543_groups_1 = const()[name = string("sparse_output_1543_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570286976))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570284288))))[name = string("layers_19_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1543_cast_fp16 = conv(dilations = sparse_output_1543_dilations_1, groups = sparse_output_1543_groups_1, pad = sparse_output_1543_pad_1, pad_type = sparse_output_1543_pad_type_1, strides = sparse_output_1543_strides_1, weight = layers_19_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1543_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_24537_cast_fp16 = add(x = dense_output_1543_cast_fp16, y = sparse_output_1543_cast_fp16)[name = string("op_24537_cast_fp16")];
            tensor<int32, [4]> var_24538 = const()[name = string("op_24538"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24540 = const()[name = string("op_24540"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_24539_cast_fp16 = transpose(perm = var_24538, x = var_24537_cast_fp16)[name = string("transpose_198")];
            tensor<fp16, [1, 101, 128]> p_head_609_cast_fp16 = reshape(shape = var_24540, x = var_24539_cast_fp16)[name = string("p_head_609_cast_fp16")];
            tensor<fp16, [128]> var_24542_to_fp16 = const()[name = string("op_24542_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570303424)))];
            tensor<fp16, [1, 51, 128]> var_24543_cast_fp16 = add(x = q_head_305_cast_fp16, y = var_24542_to_fp16)[name = string("op_24543_cast_fp16")];
            tensor<int32, [1]> q_u_305_axes_1 = const()[name = string("q_u_305_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_305_cast_fp16 = expand_dims(axes = q_u_305_axes_1, x = var_24543_cast_fp16)[name = string("q_u_305_cast_fp16")];
            tensor<fp16, [128]> var_24545_to_fp16 = const()[name = string("op_24545_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570303744)))];
            tensor<fp16, [1, 51, 128]> var_24546_cast_fp16 = add(x = q_head_305_cast_fp16, y = var_24545_to_fp16)[name = string("op_24546_cast_fp16")];
            tensor<int32, [1]> q_v_305_axes_1 = const()[name = string("q_v_305_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_305_cast_fp16 = expand_dims(axes = q_v_305_axes_1, x = var_24546_cast_fp16)[name = string("q_v_305_cast_fp16")];
            tensor<int32, [1]> k_head_611_axes_1 = const()[name = string("k_head_611_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_611_cast_fp16 = expand_dims(axes = k_head_611_axes_1, x = k_head_609_cast_fp16)[name = string("k_head_611_cast_fp16")];
            tensor<int32, [1]> v_head_611_axes_1 = const()[name = string("v_head_611_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_611_cast_fp16 = expand_dims(axes = v_head_611_axes_1, x = v_head_609_cast_fp16)[name = string("v_head_611_cast_fp16")];
            tensor<int32, [1]> p_head_611_axes_1 = const()[name = string("p_head_611_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_611_cast_fp16 = expand_dims(axes = p_head_611_axes_1, x = p_head_609_cast_fp16)[name = string("p_head_611_cast_fp16")];
            bool var_24552_transpose_x_3 = const()[name = string("op_24552_transpose_x_3"), val = bool(false)];
            bool var_24552_transpose_y_3 = const()[name = string("op_24552_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_24552_cast_fp16 = matmul(transpose_x = var_24552_transpose_x_3, transpose_y = var_24552_transpose_y_3, x = q_u_305_cast_fp16, y = k_head_611_cast_fp16)[name = string("op_24552_cast_fp16")];
            fp16 var_24553_to_fp16 = const()[name = string("op_24553_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_305_cast_fp16 = mul(x = var_24552_cast_fp16, y = var_24553_to_fp16)[name = string("scores_content_305_cast_fp16")];
            bool x_1617_transpose_x_3 = const()[name = string("x_1617_transpose_x_3"), val = bool(false)];
            bool x_1617_transpose_y_3 = const()[name = string("x_1617_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1617_cast_fp16 = matmul(transpose_x = x_1617_transpose_x_3, transpose_y = x_1617_transpose_y_3, x = q_v_305_cast_fp16, y = p_head_611_cast_fp16)[name = string("x_1617_cast_fp16")];
            tensor<int32, [8]> x_1619_pad_1 = const()[name = string("x_1619_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1619_mode_1 = const()[name = string("x_1619_mode_1"), val = string("constant")];
            fp16 const_2335_to_fp16 = const()[name = string("const_2335_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1619_cast_fp16 = pad(constant_val = const_2335_to_fp16, mode = x_1619_mode_1, pad = x_1619_pad_1, x = x_1617_cast_fp16)[name = string("x_1619_cast_fp16")];
            tensor<int32, [4]> var_24567 = const()[name = string("op_24567"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1621_cast_fp16 = reshape(shape = var_24567, x = x_1619_cast_fp16)[name = string("x_1621_cast_fp16")];
            tensor<int32, [4]> var_24571_begin_1 = const()[name = string("op_24571_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_24571_end_1 = const()[name = string("op_24571_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_24571_end_mask_1 = const()[name = string("op_24571_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_24571_cast_fp16 = slice_by_index(begin = var_24571_begin_1, end = var_24571_end_1, end_mask = var_24571_end_mask_1, x = x_1621_cast_fp16)[name = string("op_24571_cast_fp16")];
            tensor<int32, [4]> var_24573 = const()[name = string("op_24573"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_24574_cast_fp16 = reshape(shape = var_24573, x = var_24571_cast_fp16)[name = string("op_24574_cast_fp16")];
            tensor<int32, [4]> var_24579_begin_1 = const()[name = string("op_24579_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_24579_end_1 = const()[name = string("op_24579_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_24579_end_mask_1 = const()[name = string("op_24579_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_24579_cast_fp16 = slice_by_index(begin = var_24579_begin_1, end = var_24579_end_1, end_mask = var_24579_end_mask_1, x = var_24574_cast_fp16)[name = string("op_24579_cast_fp16")];
            fp16 var_24580_to_fp16 = const()[name = string("op_24580_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_305_cast_fp16 = mul(x = var_24579_cast_fp16, y = var_24580_to_fp16)[name = string("scores_pos_305_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_305_cast_fp16 = add(x = scores_content_305_cast_fp16, y = scores_pos_305_cast_fp16)[name = string("logits_305_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_24583_cast_fp16 = softmax(axis = var_24308, x = logits_305_cast_fp16)[name = string("op_24583_cast_fp16")];
            bool var_24585_transpose_x_1 = const()[name = string("op_24585_transpose_x_1"), val = bool(false)];
            bool var_24585_transpose_y_1 = const()[name = string("op_24585_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_24585_cast_fp16 = matmul(transpose_x = var_24585_transpose_x_1, transpose_y = var_24585_transpose_y_1, x = var_24583_cast_fp16, y = v_head_611_cast_fp16)[name = string("op_24585_cast_fp16")];
            tensor<int32, [1]> var_24586_axes_1 = const()[name = string("op_24586_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_24586_cast_fp16 = squeeze(axes = var_24586_axes_1, x = var_24585_cast_fp16)[name = string("op_24586_cast_fp16")];
            string dense_output_1545_pad_type_1 = const()[name = string("dense_output_1545_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1545_strides_1 = const()[name = string("dense_output_1545_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1545_pad_1 = const()[name = string("dense_output_1545_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1545_dilations_1 = const()[name = string("dense_output_1545_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1545_groups_1 = const()[name = string("dense_output_1545_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570304064))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570435200))))[name = string("layers_19_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1545_cast_fp16 = conv(dilations = dense_output_1545_dilations_1, groups = dense_output_1545_groups_1, pad = dense_output_1545_pad_1, pad_type = dense_output_1545_pad_type_1, strides = dense_output_1545_strides_1, weight = layers_19_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1545_cast_fp16")];
            string sparse_output_1545_pad_type_1 = const()[name = string("sparse_output_1545_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1545_strides_1 = const()[name = string("sparse_output_1545_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1545_pad_1 = const()[name = string("sparse_output_1545_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1545_dilations_1 = const()[name = string("sparse_output_1545_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1545_groups_1 = const()[name = string("sparse_output_1545_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570438464))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570435776))))[name = string("layers_19_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1545_cast_fp16 = conv(dilations = sparse_output_1545_dilations_1, groups = sparse_output_1545_groups_1, pad = sparse_output_1545_pad_1, pad_type = sparse_output_1545_pad_type_1, strides = sparse_output_1545_strides_1, weight = layers_19_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1545_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24601_cast_fp16 = add(x = dense_output_1545_cast_fp16, y = sparse_output_1545_cast_fp16)[name = string("op_24601_cast_fp16")];
            tensor<int32, [4]> var_24602 = const()[name = string("op_24602"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24604 = const()[name = string("op_24604"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24603_cast_fp16 = transpose(perm = var_24602, x = var_24601_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 51, 128]> q_head_307_cast_fp16 = reshape(shape = var_24604, x = var_24603_cast_fp16)[name = string("q_head_307_cast_fp16")];
            string dense_output_1547_pad_type_1 = const()[name = string("dense_output_1547_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1547_strides_1 = const()[name = string("dense_output_1547_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1547_pad_1 = const()[name = string("dense_output_1547_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1547_dilations_1 = const()[name = string("dense_output_1547_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1547_groups_1 = const()[name = string("dense_output_1547_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570454912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570586048))))[name = string("layers_19_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1547_cast_fp16 = conv(dilations = dense_output_1547_dilations_1, groups = dense_output_1547_groups_1, pad = dense_output_1547_pad_1, pad_type = dense_output_1547_pad_type_1, strides = dense_output_1547_strides_1, weight = layers_19_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1547_cast_fp16")];
            string sparse_output_1547_pad_type_1 = const()[name = string("sparse_output_1547_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1547_strides_1 = const()[name = string("sparse_output_1547_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1547_pad_1 = const()[name = string("sparse_output_1547_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1547_dilations_1 = const()[name = string("sparse_output_1547_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1547_groups_1 = const()[name = string("sparse_output_1547_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570589312))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570586624))))[name = string("layers_19_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1547_cast_fp16 = conv(dilations = sparse_output_1547_dilations_1, groups = sparse_output_1547_groups_1, pad = sparse_output_1547_pad_1, pad_type = sparse_output_1547_pad_type_1, strides = sparse_output_1547_strides_1, weight = layers_19_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1547_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24620_cast_fp16 = add(x = dense_output_1547_cast_fp16, y = sparse_output_1547_cast_fp16)[name = string("op_24620_cast_fp16")];
            tensor<int32, [4]> var_24621 = const()[name = string("op_24621"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24623 = const()[name = string("op_24623"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24622_cast_fp16 = transpose(perm = var_24621, x = var_24620_cast_fp16)[name = string("transpose_196")];
            tensor<fp16, [1, 51, 128]> k_head_613_cast_fp16 = reshape(shape = var_24623, x = var_24622_cast_fp16)[name = string("k_head_613_cast_fp16")];
            string dense_output_1549_pad_type_1 = const()[name = string("dense_output_1549_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1549_strides_1 = const()[name = string("dense_output_1549_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1549_pad_1 = const()[name = string("dense_output_1549_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1549_dilations_1 = const()[name = string("dense_output_1549_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1549_groups_1 = const()[name = string("dense_output_1549_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570605760))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570736896))))[name = string("layers_19_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1549_cast_fp16 = conv(dilations = dense_output_1549_dilations_1, groups = dense_output_1549_groups_1, pad = dense_output_1549_pad_1, pad_type = dense_output_1549_pad_type_1, strides = dense_output_1549_strides_1, weight = layers_19_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1549_cast_fp16")];
            string sparse_output_1549_pad_type_1 = const()[name = string("sparse_output_1549_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1549_strides_1 = const()[name = string("sparse_output_1549_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1549_pad_1 = const()[name = string("sparse_output_1549_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1549_dilations_1 = const()[name = string("sparse_output_1549_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1549_groups_1 = const()[name = string("sparse_output_1549_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570740160))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570737472))))[name = string("layers_19_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1549_cast_fp16 = conv(dilations = sparse_output_1549_dilations_1, groups = sparse_output_1549_groups_1, pad = sparse_output_1549_pad_1, pad_type = sparse_output_1549_pad_type_1, strides = sparse_output_1549_strides_1, weight = layers_19_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1549_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24639_cast_fp16 = add(x = dense_output_1549_cast_fp16, y = sparse_output_1549_cast_fp16)[name = string("op_24639_cast_fp16")];
            tensor<int32, [4]> var_24640 = const()[name = string("op_24640"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24642 = const()[name = string("op_24642"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24641_cast_fp16 = transpose(perm = var_24640, x = var_24639_cast_fp16)[name = string("transpose_195")];
            tensor<fp16, [1, 51, 128]> v_head_613_cast_fp16 = reshape(shape = var_24642, x = var_24641_cast_fp16)[name = string("v_head_613_cast_fp16")];
            string dense_output_1551_pad_type_1 = const()[name = string("dense_output_1551_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1551_strides_1 = const()[name = string("dense_output_1551_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1551_pad_1 = const()[name = string("dense_output_1551_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1551_dilations_1 = const()[name = string("dense_output_1551_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1551_groups_1 = const()[name = string("dense_output_1551_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570756608))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570887744))))[name = string("layers_19_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1551_cast_fp16 = conv(dilations = dense_output_1551_dilations_1, groups = dense_output_1551_groups_1, pad = dense_output_1551_pad_1, pad_type = dense_output_1551_pad_type_1, strides = dense_output_1551_strides_1, weight = layers_19_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1551_cast_fp16")];
            string sparse_output_1551_pad_type_1 = const()[name = string("sparse_output_1551_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1551_strides_1 = const()[name = string("sparse_output_1551_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1551_pad_1 = const()[name = string("sparse_output_1551_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1551_dilations_1 = const()[name = string("sparse_output_1551_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1551_groups_1 = const()[name = string("sparse_output_1551_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570891008))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570888320))))[name = string("layers_19_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1551_cast_fp16 = conv(dilations = sparse_output_1551_dilations_1, groups = sparse_output_1551_groups_1, pad = sparse_output_1551_pad_1, pad_type = sparse_output_1551_pad_type_1, strides = sparse_output_1551_strides_1, weight = layers_19_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1551_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_24658_cast_fp16 = add(x = dense_output_1551_cast_fp16, y = sparse_output_1551_cast_fp16)[name = string("op_24658_cast_fp16")];
            tensor<int32, [4]> var_24659 = const()[name = string("op_24659"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24661 = const()[name = string("op_24661"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_24660_cast_fp16 = transpose(perm = var_24659, x = var_24658_cast_fp16)[name = string("transpose_194")];
            tensor<fp16, [1, 101, 128]> p_head_613_cast_fp16 = reshape(shape = var_24661, x = var_24660_cast_fp16)[name = string("p_head_613_cast_fp16")];
            tensor<fp16, [128]> var_24663_to_fp16 = const()[name = string("op_24663_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570907456)))];
            tensor<fp16, [1, 51, 128]> var_24664_cast_fp16 = add(x = q_head_307_cast_fp16, y = var_24663_to_fp16)[name = string("op_24664_cast_fp16")];
            tensor<int32, [1]> q_u_307_axes_1 = const()[name = string("q_u_307_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_307_cast_fp16 = expand_dims(axes = q_u_307_axes_1, x = var_24664_cast_fp16)[name = string("q_u_307_cast_fp16")];
            tensor<fp16, [128]> var_24666_to_fp16 = const()[name = string("op_24666_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570907776)))];
            tensor<fp16, [1, 51, 128]> var_24667_cast_fp16 = add(x = q_head_307_cast_fp16, y = var_24666_to_fp16)[name = string("op_24667_cast_fp16")];
            tensor<int32, [1]> q_v_307_axes_1 = const()[name = string("q_v_307_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_307_cast_fp16 = expand_dims(axes = q_v_307_axes_1, x = var_24667_cast_fp16)[name = string("q_v_307_cast_fp16")];
            tensor<int32, [1]> k_head_615_axes_1 = const()[name = string("k_head_615_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_615_cast_fp16 = expand_dims(axes = k_head_615_axes_1, x = k_head_613_cast_fp16)[name = string("k_head_615_cast_fp16")];
            tensor<int32, [1]> v_head_615_axes_1 = const()[name = string("v_head_615_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_615_cast_fp16 = expand_dims(axes = v_head_615_axes_1, x = v_head_613_cast_fp16)[name = string("v_head_615_cast_fp16")];
            tensor<int32, [1]> p_head_615_axes_1 = const()[name = string("p_head_615_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_615_cast_fp16 = expand_dims(axes = p_head_615_axes_1, x = p_head_613_cast_fp16)[name = string("p_head_615_cast_fp16")];
            bool var_24673_transpose_x_3 = const()[name = string("op_24673_transpose_x_3"), val = bool(false)];
            bool var_24673_transpose_y_3 = const()[name = string("op_24673_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_24673_cast_fp16 = matmul(transpose_x = var_24673_transpose_x_3, transpose_y = var_24673_transpose_y_3, x = q_u_307_cast_fp16, y = k_head_615_cast_fp16)[name = string("op_24673_cast_fp16")];
            fp16 var_24674_to_fp16 = const()[name = string("op_24674_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_307_cast_fp16 = mul(x = var_24673_cast_fp16, y = var_24674_to_fp16)[name = string("scores_content_307_cast_fp16")];
            bool x_1625_transpose_x_3 = const()[name = string("x_1625_transpose_x_3"), val = bool(false)];
            bool x_1625_transpose_y_3 = const()[name = string("x_1625_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1625_cast_fp16 = matmul(transpose_x = x_1625_transpose_x_3, transpose_y = x_1625_transpose_y_3, x = q_v_307_cast_fp16, y = p_head_615_cast_fp16)[name = string("x_1625_cast_fp16")];
            tensor<int32, [8]> x_1627_pad_1 = const()[name = string("x_1627_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1627_mode_1 = const()[name = string("x_1627_mode_1"), val = string("constant")];
            fp16 const_2341_to_fp16 = const()[name = string("const_2341_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1627_cast_fp16 = pad(constant_val = const_2341_to_fp16, mode = x_1627_mode_1, pad = x_1627_pad_1, x = x_1625_cast_fp16)[name = string("x_1627_cast_fp16")];
            tensor<int32, [4]> var_24688 = const()[name = string("op_24688"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1629_cast_fp16 = reshape(shape = var_24688, x = x_1627_cast_fp16)[name = string("x_1629_cast_fp16")];
            tensor<int32, [4]> var_24692_begin_1 = const()[name = string("op_24692_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_24692_end_1 = const()[name = string("op_24692_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_24692_end_mask_1 = const()[name = string("op_24692_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_24692_cast_fp16 = slice_by_index(begin = var_24692_begin_1, end = var_24692_end_1, end_mask = var_24692_end_mask_1, x = x_1629_cast_fp16)[name = string("op_24692_cast_fp16")];
            tensor<int32, [4]> var_24694 = const()[name = string("op_24694"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_24695_cast_fp16 = reshape(shape = var_24694, x = var_24692_cast_fp16)[name = string("op_24695_cast_fp16")];
            tensor<int32, [4]> var_24700_begin_1 = const()[name = string("op_24700_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_24700_end_1 = const()[name = string("op_24700_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_24700_end_mask_1 = const()[name = string("op_24700_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_24700_cast_fp16 = slice_by_index(begin = var_24700_begin_1, end = var_24700_end_1, end_mask = var_24700_end_mask_1, x = var_24695_cast_fp16)[name = string("op_24700_cast_fp16")];
            fp16 var_24701_to_fp16 = const()[name = string("op_24701_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_307_cast_fp16 = mul(x = var_24700_cast_fp16, y = var_24701_to_fp16)[name = string("scores_pos_307_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_307_cast_fp16 = add(x = scores_content_307_cast_fp16, y = scores_pos_307_cast_fp16)[name = string("logits_307_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_24704_cast_fp16 = softmax(axis = var_24308, x = logits_307_cast_fp16)[name = string("op_24704_cast_fp16")];
            bool var_24706_transpose_x_1 = const()[name = string("op_24706_transpose_x_1"), val = bool(false)];
            bool var_24706_transpose_y_1 = const()[name = string("op_24706_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_24706_cast_fp16 = matmul(transpose_x = var_24706_transpose_x_1, transpose_y = var_24706_transpose_y_1, x = var_24704_cast_fp16, y = v_head_615_cast_fp16)[name = string("op_24706_cast_fp16")];
            tensor<int32, [1]> var_24707_axes_1 = const()[name = string("op_24707_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_24707_cast_fp16 = squeeze(axes = var_24707_axes_1, x = var_24706_cast_fp16)[name = string("op_24707_cast_fp16")];
            string dense_output_1553_pad_type_1 = const()[name = string("dense_output_1553_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1553_strides_1 = const()[name = string("dense_output_1553_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1553_pad_1 = const()[name = string("dense_output_1553_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1553_dilations_1 = const()[name = string("dense_output_1553_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1553_groups_1 = const()[name = string("dense_output_1553_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570908096))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571039232))))[name = string("layers_19_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1553_cast_fp16 = conv(dilations = dense_output_1553_dilations_1, groups = dense_output_1553_groups_1, pad = dense_output_1553_pad_1, pad_type = dense_output_1553_pad_type_1, strides = dense_output_1553_strides_1, weight = layers_19_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1553_cast_fp16")];
            string sparse_output_1553_pad_type_1 = const()[name = string("sparse_output_1553_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1553_strides_1 = const()[name = string("sparse_output_1553_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1553_pad_1 = const()[name = string("sparse_output_1553_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1553_dilations_1 = const()[name = string("sparse_output_1553_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1553_groups_1 = const()[name = string("sparse_output_1553_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571042496))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571039808))))[name = string("layers_19_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1553_cast_fp16 = conv(dilations = sparse_output_1553_dilations_1, groups = sparse_output_1553_groups_1, pad = sparse_output_1553_pad_1, pad_type = sparse_output_1553_pad_type_1, strides = sparse_output_1553_strides_1, weight = layers_19_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1553_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24722_cast_fp16 = add(x = dense_output_1553_cast_fp16, y = sparse_output_1553_cast_fp16)[name = string("op_24722_cast_fp16")];
            tensor<int32, [4]> var_24723 = const()[name = string("op_24723"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24725 = const()[name = string("op_24725"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24724_cast_fp16 = transpose(perm = var_24723, x = var_24722_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [1, 51, 128]> q_head_309_cast_fp16 = reshape(shape = var_24725, x = var_24724_cast_fp16)[name = string("q_head_309_cast_fp16")];
            string dense_output_1555_pad_type_1 = const()[name = string("dense_output_1555_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1555_strides_1 = const()[name = string("dense_output_1555_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1555_pad_1 = const()[name = string("dense_output_1555_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1555_dilations_1 = const()[name = string("dense_output_1555_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1555_groups_1 = const()[name = string("dense_output_1555_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571058944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571190080))))[name = string("layers_19_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1555_cast_fp16 = conv(dilations = dense_output_1555_dilations_1, groups = dense_output_1555_groups_1, pad = dense_output_1555_pad_1, pad_type = dense_output_1555_pad_type_1, strides = dense_output_1555_strides_1, weight = layers_19_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1555_cast_fp16")];
            string sparse_output_1555_pad_type_1 = const()[name = string("sparse_output_1555_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1555_strides_1 = const()[name = string("sparse_output_1555_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1555_pad_1 = const()[name = string("sparse_output_1555_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1555_dilations_1 = const()[name = string("sparse_output_1555_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1555_groups_1 = const()[name = string("sparse_output_1555_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571193344))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571190656))))[name = string("layers_19_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1555_cast_fp16 = conv(dilations = sparse_output_1555_dilations_1, groups = sparse_output_1555_groups_1, pad = sparse_output_1555_pad_1, pad_type = sparse_output_1555_pad_type_1, strides = sparse_output_1555_strides_1, weight = layers_19_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1555_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24741_cast_fp16 = add(x = dense_output_1555_cast_fp16, y = sparse_output_1555_cast_fp16)[name = string("op_24741_cast_fp16")];
            tensor<int32, [4]> var_24742 = const()[name = string("op_24742"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24744 = const()[name = string("op_24744"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24743_cast_fp16 = transpose(perm = var_24742, x = var_24741_cast_fp16)[name = string("transpose_192")];
            tensor<fp16, [1, 51, 128]> k_head_617_cast_fp16 = reshape(shape = var_24744, x = var_24743_cast_fp16)[name = string("k_head_617_cast_fp16")];
            string dense_output_1557_pad_type_1 = const()[name = string("dense_output_1557_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1557_strides_1 = const()[name = string("dense_output_1557_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1557_pad_1 = const()[name = string("dense_output_1557_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1557_dilations_1 = const()[name = string("dense_output_1557_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1557_groups_1 = const()[name = string("dense_output_1557_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571209792))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571340928))))[name = string("layers_19_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1557_cast_fp16 = conv(dilations = dense_output_1557_dilations_1, groups = dense_output_1557_groups_1, pad = dense_output_1557_pad_1, pad_type = dense_output_1557_pad_type_1, strides = dense_output_1557_strides_1, weight = layers_19_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1557_cast_fp16")];
            string sparse_output_1557_pad_type_1 = const()[name = string("sparse_output_1557_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1557_strides_1 = const()[name = string("sparse_output_1557_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1557_pad_1 = const()[name = string("sparse_output_1557_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1557_dilations_1 = const()[name = string("sparse_output_1557_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1557_groups_1 = const()[name = string("sparse_output_1557_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571344192))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571341504))))[name = string("layers_19_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1557_cast_fp16 = conv(dilations = sparse_output_1557_dilations_1, groups = sparse_output_1557_groups_1, pad = sparse_output_1557_pad_1, pad_type = sparse_output_1557_pad_type_1, strides = sparse_output_1557_strides_1, weight = layers_19_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1557_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24760_cast_fp16 = add(x = dense_output_1557_cast_fp16, y = sparse_output_1557_cast_fp16)[name = string("op_24760_cast_fp16")];
            tensor<int32, [4]> var_24761 = const()[name = string("op_24761"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24763 = const()[name = string("op_24763"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24762_cast_fp16 = transpose(perm = var_24761, x = var_24760_cast_fp16)[name = string("transpose_191")];
            tensor<fp16, [1, 51, 128]> v_head_617_cast_fp16 = reshape(shape = var_24763, x = var_24762_cast_fp16)[name = string("v_head_617_cast_fp16")];
            string dense_output_1559_pad_type_1 = const()[name = string("dense_output_1559_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1559_strides_1 = const()[name = string("dense_output_1559_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1559_pad_1 = const()[name = string("dense_output_1559_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1559_dilations_1 = const()[name = string("dense_output_1559_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1559_groups_1 = const()[name = string("dense_output_1559_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571360640))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571491776))))[name = string("layers_19_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1559_cast_fp16 = conv(dilations = dense_output_1559_dilations_1, groups = dense_output_1559_groups_1, pad = dense_output_1559_pad_1, pad_type = dense_output_1559_pad_type_1, strides = dense_output_1559_strides_1, weight = layers_19_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1559_cast_fp16")];
            string sparse_output_1559_pad_type_1 = const()[name = string("sparse_output_1559_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1559_strides_1 = const()[name = string("sparse_output_1559_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1559_pad_1 = const()[name = string("sparse_output_1559_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1559_dilations_1 = const()[name = string("sparse_output_1559_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1559_groups_1 = const()[name = string("sparse_output_1559_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571495040))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571492352))))[name = string("layers_19_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1559_cast_fp16 = conv(dilations = sparse_output_1559_dilations_1, groups = sparse_output_1559_groups_1, pad = sparse_output_1559_pad_1, pad_type = sparse_output_1559_pad_type_1, strides = sparse_output_1559_strides_1, weight = layers_19_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1559_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_24779_cast_fp16 = add(x = dense_output_1559_cast_fp16, y = sparse_output_1559_cast_fp16)[name = string("op_24779_cast_fp16")];
            tensor<int32, [4]> var_24780 = const()[name = string("op_24780"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24782 = const()[name = string("op_24782"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_24781_cast_fp16 = transpose(perm = var_24780, x = var_24779_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 101, 128]> p_head_617_cast_fp16 = reshape(shape = var_24782, x = var_24781_cast_fp16)[name = string("p_head_617_cast_fp16")];
            tensor<fp16, [128]> var_24784_to_fp16 = const()[name = string("op_24784_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571511488)))];
            tensor<fp16, [1, 51, 128]> var_24785_cast_fp16 = add(x = q_head_309_cast_fp16, y = var_24784_to_fp16)[name = string("op_24785_cast_fp16")];
            tensor<int32, [1]> q_u_309_axes_1 = const()[name = string("q_u_309_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_309_cast_fp16 = expand_dims(axes = q_u_309_axes_1, x = var_24785_cast_fp16)[name = string("q_u_309_cast_fp16")];
            tensor<fp16, [128]> var_24787_to_fp16 = const()[name = string("op_24787_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571511808)))];
            tensor<fp16, [1, 51, 128]> var_24788_cast_fp16 = add(x = q_head_309_cast_fp16, y = var_24787_to_fp16)[name = string("op_24788_cast_fp16")];
            tensor<int32, [1]> q_v_309_axes_1 = const()[name = string("q_v_309_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_309_cast_fp16 = expand_dims(axes = q_v_309_axes_1, x = var_24788_cast_fp16)[name = string("q_v_309_cast_fp16")];
            tensor<int32, [1]> k_head_619_axes_1 = const()[name = string("k_head_619_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_619_cast_fp16 = expand_dims(axes = k_head_619_axes_1, x = k_head_617_cast_fp16)[name = string("k_head_619_cast_fp16")];
            tensor<int32, [1]> v_head_619_axes_1 = const()[name = string("v_head_619_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_619_cast_fp16 = expand_dims(axes = v_head_619_axes_1, x = v_head_617_cast_fp16)[name = string("v_head_619_cast_fp16")];
            tensor<int32, [1]> p_head_619_axes_1 = const()[name = string("p_head_619_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_619_cast_fp16 = expand_dims(axes = p_head_619_axes_1, x = p_head_617_cast_fp16)[name = string("p_head_619_cast_fp16")];
            bool var_24794_transpose_x_3 = const()[name = string("op_24794_transpose_x_3"), val = bool(false)];
            bool var_24794_transpose_y_3 = const()[name = string("op_24794_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_24794_cast_fp16 = matmul(transpose_x = var_24794_transpose_x_3, transpose_y = var_24794_transpose_y_3, x = q_u_309_cast_fp16, y = k_head_619_cast_fp16)[name = string("op_24794_cast_fp16")];
            fp16 var_24795_to_fp16 = const()[name = string("op_24795_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_309_cast_fp16 = mul(x = var_24794_cast_fp16, y = var_24795_to_fp16)[name = string("scores_content_309_cast_fp16")];
            bool x_1633_transpose_x_3 = const()[name = string("x_1633_transpose_x_3"), val = bool(false)];
            bool x_1633_transpose_y_3 = const()[name = string("x_1633_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1633_cast_fp16 = matmul(transpose_x = x_1633_transpose_x_3, transpose_y = x_1633_transpose_y_3, x = q_v_309_cast_fp16, y = p_head_619_cast_fp16)[name = string("x_1633_cast_fp16")];
            tensor<int32, [8]> x_1635_pad_1 = const()[name = string("x_1635_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1635_mode_1 = const()[name = string("x_1635_mode_1"), val = string("constant")];
            fp16 const_2347_to_fp16 = const()[name = string("const_2347_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1635_cast_fp16 = pad(constant_val = const_2347_to_fp16, mode = x_1635_mode_1, pad = x_1635_pad_1, x = x_1633_cast_fp16)[name = string("x_1635_cast_fp16")];
            tensor<int32, [4]> var_24809 = const()[name = string("op_24809"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1637_cast_fp16 = reshape(shape = var_24809, x = x_1635_cast_fp16)[name = string("x_1637_cast_fp16")];
            tensor<int32, [4]> var_24813_begin_1 = const()[name = string("op_24813_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_24813_end_1 = const()[name = string("op_24813_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_24813_end_mask_1 = const()[name = string("op_24813_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_24813_cast_fp16 = slice_by_index(begin = var_24813_begin_1, end = var_24813_end_1, end_mask = var_24813_end_mask_1, x = x_1637_cast_fp16)[name = string("op_24813_cast_fp16")];
            tensor<int32, [4]> var_24815 = const()[name = string("op_24815"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_24816_cast_fp16 = reshape(shape = var_24815, x = var_24813_cast_fp16)[name = string("op_24816_cast_fp16")];
            tensor<int32, [4]> var_24821_begin_1 = const()[name = string("op_24821_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_24821_end_1 = const()[name = string("op_24821_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_24821_end_mask_1 = const()[name = string("op_24821_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_24821_cast_fp16 = slice_by_index(begin = var_24821_begin_1, end = var_24821_end_1, end_mask = var_24821_end_mask_1, x = var_24816_cast_fp16)[name = string("op_24821_cast_fp16")];
            fp16 var_24822_to_fp16 = const()[name = string("op_24822_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_309_cast_fp16 = mul(x = var_24821_cast_fp16, y = var_24822_to_fp16)[name = string("scores_pos_309_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_309_cast_fp16 = add(x = scores_content_309_cast_fp16, y = scores_pos_309_cast_fp16)[name = string("logits_309_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_24825_cast_fp16 = softmax(axis = var_24308, x = logits_309_cast_fp16)[name = string("op_24825_cast_fp16")];
            bool var_24827_transpose_x_1 = const()[name = string("op_24827_transpose_x_1"), val = bool(false)];
            bool var_24827_transpose_y_1 = const()[name = string("op_24827_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_24827_cast_fp16 = matmul(transpose_x = var_24827_transpose_x_1, transpose_y = var_24827_transpose_y_1, x = var_24825_cast_fp16, y = v_head_619_cast_fp16)[name = string("op_24827_cast_fp16")];
            tensor<int32, [1]> var_24828_axes_1 = const()[name = string("op_24828_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_24828_cast_fp16 = squeeze(axes = var_24828_axes_1, x = var_24827_cast_fp16)[name = string("op_24828_cast_fp16")];
            string dense_output_1561_pad_type_1 = const()[name = string("dense_output_1561_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1561_strides_1 = const()[name = string("dense_output_1561_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1561_pad_1 = const()[name = string("dense_output_1561_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1561_dilations_1 = const()[name = string("dense_output_1561_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1561_groups_1 = const()[name = string("dense_output_1561_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571512128))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571643264))))[name = string("layers_19_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1561_cast_fp16 = conv(dilations = dense_output_1561_dilations_1, groups = dense_output_1561_groups_1, pad = dense_output_1561_pad_1, pad_type = dense_output_1561_pad_type_1, strides = dense_output_1561_strides_1, weight = layers_19_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1561_cast_fp16")];
            string sparse_output_1561_pad_type_1 = const()[name = string("sparse_output_1561_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1561_strides_1 = const()[name = string("sparse_output_1561_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1561_pad_1 = const()[name = string("sparse_output_1561_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1561_dilations_1 = const()[name = string("sparse_output_1561_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1561_groups_1 = const()[name = string("sparse_output_1561_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571646528))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571643840))))[name = string("layers_19_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1561_cast_fp16 = conv(dilations = sparse_output_1561_dilations_1, groups = sparse_output_1561_groups_1, pad = sparse_output_1561_pad_1, pad_type = sparse_output_1561_pad_type_1, strides = sparse_output_1561_strides_1, weight = layers_19_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1561_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24843_cast_fp16 = add(x = dense_output_1561_cast_fp16, y = sparse_output_1561_cast_fp16)[name = string("op_24843_cast_fp16")];
            tensor<int32, [4]> var_24844 = const()[name = string("op_24844"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24846 = const()[name = string("op_24846"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24845_cast_fp16 = transpose(perm = var_24844, x = var_24843_cast_fp16)[name = string("transpose_189")];
            tensor<fp16, [1, 51, 128]> q_head_311_cast_fp16 = reshape(shape = var_24846, x = var_24845_cast_fp16)[name = string("q_head_311_cast_fp16")];
            string dense_output_1563_pad_type_1 = const()[name = string("dense_output_1563_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1563_strides_1 = const()[name = string("dense_output_1563_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1563_pad_1 = const()[name = string("dense_output_1563_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1563_dilations_1 = const()[name = string("dense_output_1563_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1563_groups_1 = const()[name = string("dense_output_1563_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571662976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571794112))))[name = string("layers_19_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1563_cast_fp16 = conv(dilations = dense_output_1563_dilations_1, groups = dense_output_1563_groups_1, pad = dense_output_1563_pad_1, pad_type = dense_output_1563_pad_type_1, strides = dense_output_1563_strides_1, weight = layers_19_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1563_cast_fp16")];
            string sparse_output_1563_pad_type_1 = const()[name = string("sparse_output_1563_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1563_strides_1 = const()[name = string("sparse_output_1563_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1563_pad_1 = const()[name = string("sparse_output_1563_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1563_dilations_1 = const()[name = string("sparse_output_1563_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1563_groups_1 = const()[name = string("sparse_output_1563_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571797376))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571794688))))[name = string("layers_19_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1563_cast_fp16 = conv(dilations = sparse_output_1563_dilations_1, groups = sparse_output_1563_groups_1, pad = sparse_output_1563_pad_1, pad_type = sparse_output_1563_pad_type_1, strides = sparse_output_1563_strides_1, weight = layers_19_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1563_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24862_cast_fp16 = add(x = dense_output_1563_cast_fp16, y = sparse_output_1563_cast_fp16)[name = string("op_24862_cast_fp16")];
            tensor<int32, [4]> var_24863 = const()[name = string("op_24863"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24865 = const()[name = string("op_24865"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24864_cast_fp16 = transpose(perm = var_24863, x = var_24862_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 51, 128]> k_head_621_cast_fp16 = reshape(shape = var_24865, x = var_24864_cast_fp16)[name = string("k_head_621_cast_fp16")];
            string dense_output_1565_pad_type_1 = const()[name = string("dense_output_1565_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1565_strides_1 = const()[name = string("dense_output_1565_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1565_pad_1 = const()[name = string("dense_output_1565_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1565_dilations_1 = const()[name = string("dense_output_1565_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1565_groups_1 = const()[name = string("dense_output_1565_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571813824))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571944960))))[name = string("layers_19_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1565_cast_fp16 = conv(dilations = dense_output_1565_dilations_1, groups = dense_output_1565_groups_1, pad = dense_output_1565_pad_1, pad_type = dense_output_1565_pad_type_1, strides = dense_output_1565_strides_1, weight = layers_19_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1565_cast_fp16")];
            string sparse_output_1565_pad_type_1 = const()[name = string("sparse_output_1565_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1565_strides_1 = const()[name = string("sparse_output_1565_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1565_pad_1 = const()[name = string("sparse_output_1565_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1565_dilations_1 = const()[name = string("sparse_output_1565_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1565_groups_1 = const()[name = string("sparse_output_1565_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571948224))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571945536))))[name = string("layers_19_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1565_cast_fp16 = conv(dilations = sparse_output_1565_dilations_1, groups = sparse_output_1565_groups_1, pad = sparse_output_1565_pad_1, pad_type = sparse_output_1565_pad_type_1, strides = sparse_output_1565_strides_1, weight = layers_19_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1565_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24881_cast_fp16 = add(x = dense_output_1565_cast_fp16, y = sparse_output_1565_cast_fp16)[name = string("op_24881_cast_fp16")];
            tensor<int32, [4]> var_24882 = const()[name = string("op_24882"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24884 = const()[name = string("op_24884"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24883_cast_fp16 = transpose(perm = var_24882, x = var_24881_cast_fp16)[name = string("transpose_187")];
            tensor<fp16, [1, 51, 128]> v_head_621_cast_fp16 = reshape(shape = var_24884, x = var_24883_cast_fp16)[name = string("v_head_621_cast_fp16")];
            string dense_output_1567_pad_type_1 = const()[name = string("dense_output_1567_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1567_strides_1 = const()[name = string("dense_output_1567_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1567_pad_1 = const()[name = string("dense_output_1567_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1567_dilations_1 = const()[name = string("dense_output_1567_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1567_groups_1 = const()[name = string("dense_output_1567_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571964672))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572095808))))[name = string("layers_19_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1567_cast_fp16 = conv(dilations = dense_output_1567_dilations_1, groups = dense_output_1567_groups_1, pad = dense_output_1567_pad_1, pad_type = dense_output_1567_pad_type_1, strides = dense_output_1567_strides_1, weight = layers_19_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1567_cast_fp16")];
            string sparse_output_1567_pad_type_1 = const()[name = string("sparse_output_1567_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1567_strides_1 = const()[name = string("sparse_output_1567_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1567_pad_1 = const()[name = string("sparse_output_1567_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1567_dilations_1 = const()[name = string("sparse_output_1567_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1567_groups_1 = const()[name = string("sparse_output_1567_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572099072))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572096384))))[name = string("layers_19_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1567_cast_fp16 = conv(dilations = sparse_output_1567_dilations_1, groups = sparse_output_1567_groups_1, pad = sparse_output_1567_pad_1, pad_type = sparse_output_1567_pad_type_1, strides = sparse_output_1567_strides_1, weight = layers_19_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1567_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_24900_cast_fp16 = add(x = dense_output_1567_cast_fp16, y = sparse_output_1567_cast_fp16)[name = string("op_24900_cast_fp16")];
            tensor<int32, [4]> var_24901 = const()[name = string("op_24901"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24903 = const()[name = string("op_24903"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_24902_cast_fp16 = transpose(perm = var_24901, x = var_24900_cast_fp16)[name = string("transpose_186")];
            tensor<fp16, [1, 101, 128]> p_head_621_cast_fp16 = reshape(shape = var_24903, x = var_24902_cast_fp16)[name = string("p_head_621_cast_fp16")];
            tensor<fp16, [128]> var_24905_to_fp16 = const()[name = string("op_24905_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572115520)))];
            tensor<fp16, [1, 51, 128]> var_24906_cast_fp16 = add(x = q_head_311_cast_fp16, y = var_24905_to_fp16)[name = string("op_24906_cast_fp16")];
            tensor<int32, [1]> q_u_311_axes_1 = const()[name = string("q_u_311_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_311_cast_fp16 = expand_dims(axes = q_u_311_axes_1, x = var_24906_cast_fp16)[name = string("q_u_311_cast_fp16")];
            tensor<fp16, [128]> var_24908_to_fp16 = const()[name = string("op_24908_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572115840)))];
            tensor<fp16, [1, 51, 128]> var_24909_cast_fp16 = add(x = q_head_311_cast_fp16, y = var_24908_to_fp16)[name = string("op_24909_cast_fp16")];
            tensor<int32, [1]> q_v_311_axes_1 = const()[name = string("q_v_311_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_311_cast_fp16 = expand_dims(axes = q_v_311_axes_1, x = var_24909_cast_fp16)[name = string("q_v_311_cast_fp16")];
            tensor<int32, [1]> k_head_623_axes_1 = const()[name = string("k_head_623_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_623_cast_fp16 = expand_dims(axes = k_head_623_axes_1, x = k_head_621_cast_fp16)[name = string("k_head_623_cast_fp16")];
            tensor<int32, [1]> v_head_623_axes_1 = const()[name = string("v_head_623_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_623_cast_fp16 = expand_dims(axes = v_head_623_axes_1, x = v_head_621_cast_fp16)[name = string("v_head_623_cast_fp16")];
            tensor<int32, [1]> p_head_623_axes_1 = const()[name = string("p_head_623_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_623_cast_fp16 = expand_dims(axes = p_head_623_axes_1, x = p_head_621_cast_fp16)[name = string("p_head_623_cast_fp16")];
            bool var_24915_transpose_x_3 = const()[name = string("op_24915_transpose_x_3"), val = bool(false)];
            bool var_24915_transpose_y_3 = const()[name = string("op_24915_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_24915_cast_fp16 = matmul(transpose_x = var_24915_transpose_x_3, transpose_y = var_24915_transpose_y_3, x = q_u_311_cast_fp16, y = k_head_623_cast_fp16)[name = string("op_24915_cast_fp16")];
            fp16 var_24916_to_fp16 = const()[name = string("op_24916_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_311_cast_fp16 = mul(x = var_24915_cast_fp16, y = var_24916_to_fp16)[name = string("scores_content_311_cast_fp16")];
            bool x_1641_transpose_x_3 = const()[name = string("x_1641_transpose_x_3"), val = bool(false)];
            bool x_1641_transpose_y_3 = const()[name = string("x_1641_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1641_cast_fp16 = matmul(transpose_x = x_1641_transpose_x_3, transpose_y = x_1641_transpose_y_3, x = q_v_311_cast_fp16, y = p_head_623_cast_fp16)[name = string("x_1641_cast_fp16")];
            tensor<int32, [8]> x_1643_pad_1 = const()[name = string("x_1643_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1643_mode_1 = const()[name = string("x_1643_mode_1"), val = string("constant")];
            fp16 const_2353_to_fp16 = const()[name = string("const_2353_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1643_cast_fp16 = pad(constant_val = const_2353_to_fp16, mode = x_1643_mode_1, pad = x_1643_pad_1, x = x_1641_cast_fp16)[name = string("x_1643_cast_fp16")];
            tensor<int32, [4]> var_24930 = const()[name = string("op_24930"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1645_cast_fp16 = reshape(shape = var_24930, x = x_1643_cast_fp16)[name = string("x_1645_cast_fp16")];
            tensor<int32, [4]> var_24934_begin_1 = const()[name = string("op_24934_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_24934_end_1 = const()[name = string("op_24934_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_24934_end_mask_1 = const()[name = string("op_24934_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_24934_cast_fp16 = slice_by_index(begin = var_24934_begin_1, end = var_24934_end_1, end_mask = var_24934_end_mask_1, x = x_1645_cast_fp16)[name = string("op_24934_cast_fp16")];
            tensor<int32, [4]> var_24936 = const()[name = string("op_24936"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_24937_cast_fp16 = reshape(shape = var_24936, x = var_24934_cast_fp16)[name = string("op_24937_cast_fp16")];
            tensor<int32, [4]> var_24942_begin_1 = const()[name = string("op_24942_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_24942_end_1 = const()[name = string("op_24942_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_24942_end_mask_1 = const()[name = string("op_24942_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_24942_cast_fp16 = slice_by_index(begin = var_24942_begin_1, end = var_24942_end_1, end_mask = var_24942_end_mask_1, x = var_24937_cast_fp16)[name = string("op_24942_cast_fp16")];
            fp16 var_24943_to_fp16 = const()[name = string("op_24943_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_311_cast_fp16 = mul(x = var_24942_cast_fp16, y = var_24943_to_fp16)[name = string("scores_pos_311_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_311_cast_fp16 = add(x = scores_content_311_cast_fp16, y = scores_pos_311_cast_fp16)[name = string("logits_311_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_24946_cast_fp16 = softmax(axis = var_24308, x = logits_311_cast_fp16)[name = string("op_24946_cast_fp16")];
            bool var_24948_transpose_x_1 = const()[name = string("op_24948_transpose_x_1"), val = bool(false)];
            bool var_24948_transpose_y_1 = const()[name = string("op_24948_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_24948_cast_fp16 = matmul(transpose_x = var_24948_transpose_x_1, transpose_y = var_24948_transpose_y_1, x = var_24946_cast_fp16, y = v_head_623_cast_fp16)[name = string("op_24948_cast_fp16")];
            tensor<int32, [1]> var_24949_axes_1 = const()[name = string("op_24949_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_24949_cast_fp16 = squeeze(axes = var_24949_axes_1, x = var_24948_cast_fp16)[name = string("op_24949_cast_fp16")];
            string dense_output_1569_pad_type_1 = const()[name = string("dense_output_1569_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1569_strides_1 = const()[name = string("dense_output_1569_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1569_pad_1 = const()[name = string("dense_output_1569_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1569_dilations_1 = const()[name = string("dense_output_1569_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1569_groups_1 = const()[name = string("dense_output_1569_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572116160))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572247296))))[name = string("layers_19_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1569_cast_fp16 = conv(dilations = dense_output_1569_dilations_1, groups = dense_output_1569_groups_1, pad = dense_output_1569_pad_1, pad_type = dense_output_1569_pad_type_1, strides = dense_output_1569_strides_1, weight = layers_19_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1569_cast_fp16")];
            string sparse_output_1569_pad_type_1 = const()[name = string("sparse_output_1569_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1569_strides_1 = const()[name = string("sparse_output_1569_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1569_pad_1 = const()[name = string("sparse_output_1569_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1569_dilations_1 = const()[name = string("sparse_output_1569_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1569_groups_1 = const()[name = string("sparse_output_1569_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572250560))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572247872))))[name = string("layers_19_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1569_cast_fp16 = conv(dilations = sparse_output_1569_dilations_1, groups = sparse_output_1569_groups_1, pad = sparse_output_1569_pad_1, pad_type = sparse_output_1569_pad_type_1, strides = sparse_output_1569_strides_1, weight = layers_19_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1569_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24964_cast_fp16 = add(x = dense_output_1569_cast_fp16, y = sparse_output_1569_cast_fp16)[name = string("op_24964_cast_fp16")];
            tensor<int32, [4]> var_24965 = const()[name = string("op_24965"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24967 = const()[name = string("op_24967"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24966_cast_fp16 = transpose(perm = var_24965, x = var_24964_cast_fp16)[name = string("transpose_185")];
            tensor<fp16, [1, 51, 128]> q_head_313_cast_fp16 = reshape(shape = var_24967, x = var_24966_cast_fp16)[name = string("q_head_313_cast_fp16")];
            string dense_output_1571_pad_type_1 = const()[name = string("dense_output_1571_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1571_strides_1 = const()[name = string("dense_output_1571_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1571_pad_1 = const()[name = string("dense_output_1571_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1571_dilations_1 = const()[name = string("dense_output_1571_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1571_groups_1 = const()[name = string("dense_output_1571_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572267008))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572398144))))[name = string("layers_19_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1571_cast_fp16 = conv(dilations = dense_output_1571_dilations_1, groups = dense_output_1571_groups_1, pad = dense_output_1571_pad_1, pad_type = dense_output_1571_pad_type_1, strides = dense_output_1571_strides_1, weight = layers_19_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1571_cast_fp16")];
            string sparse_output_1571_pad_type_1 = const()[name = string("sparse_output_1571_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1571_strides_1 = const()[name = string("sparse_output_1571_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1571_pad_1 = const()[name = string("sparse_output_1571_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1571_dilations_1 = const()[name = string("sparse_output_1571_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1571_groups_1 = const()[name = string("sparse_output_1571_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572401408))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572398720))))[name = string("layers_19_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1571_cast_fp16 = conv(dilations = sparse_output_1571_dilations_1, groups = sparse_output_1571_groups_1, pad = sparse_output_1571_pad_1, pad_type = sparse_output_1571_pad_type_1, strides = sparse_output_1571_strides_1, weight = layers_19_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1571_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_24983_cast_fp16 = add(x = dense_output_1571_cast_fp16, y = sparse_output_1571_cast_fp16)[name = string("op_24983_cast_fp16")];
            tensor<int32, [4]> var_24984 = const()[name = string("op_24984"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_24986 = const()[name = string("op_24986"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_24985_cast_fp16 = transpose(perm = var_24984, x = var_24983_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [1, 51, 128]> k_head_625_cast_fp16 = reshape(shape = var_24986, x = var_24985_cast_fp16)[name = string("k_head_625_cast_fp16")];
            string dense_output_1573_pad_type_1 = const()[name = string("dense_output_1573_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1573_strides_1 = const()[name = string("dense_output_1573_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1573_pad_1 = const()[name = string("dense_output_1573_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1573_dilations_1 = const()[name = string("dense_output_1573_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1573_groups_1 = const()[name = string("dense_output_1573_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572417856))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572548992))))[name = string("layers_19_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1573_cast_fp16 = conv(dilations = dense_output_1573_dilations_1, groups = dense_output_1573_groups_1, pad = dense_output_1573_pad_1, pad_type = dense_output_1573_pad_type_1, strides = dense_output_1573_strides_1, weight = layers_19_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1573_cast_fp16")];
            string sparse_output_1573_pad_type_1 = const()[name = string("sparse_output_1573_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1573_strides_1 = const()[name = string("sparse_output_1573_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1573_pad_1 = const()[name = string("sparse_output_1573_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1573_dilations_1 = const()[name = string("sparse_output_1573_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1573_groups_1 = const()[name = string("sparse_output_1573_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572552256))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572549568))))[name = string("layers_19_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1573_cast_fp16 = conv(dilations = sparse_output_1573_dilations_1, groups = sparse_output_1573_groups_1, pad = sparse_output_1573_pad_1, pad_type = sparse_output_1573_pad_type_1, strides = sparse_output_1573_strides_1, weight = layers_19_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1573_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25002_cast_fp16 = add(x = dense_output_1573_cast_fp16, y = sparse_output_1573_cast_fp16)[name = string("op_25002_cast_fp16")];
            tensor<int32, [4]> var_25003 = const()[name = string("op_25003"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25005 = const()[name = string("op_25005"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25004_cast_fp16 = transpose(perm = var_25003, x = var_25002_cast_fp16)[name = string("transpose_183")];
            tensor<fp16, [1, 51, 128]> v_head_625_cast_fp16 = reshape(shape = var_25005, x = var_25004_cast_fp16)[name = string("v_head_625_cast_fp16")];
            string dense_output_1575_pad_type_1 = const()[name = string("dense_output_1575_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1575_strides_1 = const()[name = string("dense_output_1575_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1575_pad_1 = const()[name = string("dense_output_1575_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1575_dilations_1 = const()[name = string("dense_output_1575_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1575_groups_1 = const()[name = string("dense_output_1575_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572568704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572699840))))[name = string("layers_19_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1575_cast_fp16 = conv(dilations = dense_output_1575_dilations_1, groups = dense_output_1575_groups_1, pad = dense_output_1575_pad_1, pad_type = dense_output_1575_pad_type_1, strides = dense_output_1575_strides_1, weight = layers_19_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1575_cast_fp16")];
            string sparse_output_1575_pad_type_1 = const()[name = string("sparse_output_1575_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1575_strides_1 = const()[name = string("sparse_output_1575_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1575_pad_1 = const()[name = string("sparse_output_1575_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1575_dilations_1 = const()[name = string("sparse_output_1575_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1575_groups_1 = const()[name = string("sparse_output_1575_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572703104))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572700416))))[name = string("layers_19_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1575_cast_fp16 = conv(dilations = sparse_output_1575_dilations_1, groups = sparse_output_1575_groups_1, pad = sparse_output_1575_pad_1, pad_type = sparse_output_1575_pad_type_1, strides = sparse_output_1575_strides_1, weight = layers_19_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1575_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_25021_cast_fp16 = add(x = dense_output_1575_cast_fp16, y = sparse_output_1575_cast_fp16)[name = string("op_25021_cast_fp16")];
            tensor<int32, [4]> var_25022 = const()[name = string("op_25022"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25024 = const()[name = string("op_25024"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_25023_cast_fp16 = transpose(perm = var_25022, x = var_25021_cast_fp16)[name = string("transpose_182")];
            tensor<fp16, [1, 101, 128]> p_head_625_cast_fp16 = reshape(shape = var_25024, x = var_25023_cast_fp16)[name = string("p_head_625_cast_fp16")];
            tensor<fp16, [128]> var_25026_to_fp16 = const()[name = string("op_25026_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572719552)))];
            tensor<fp16, [1, 51, 128]> var_25027_cast_fp16 = add(x = q_head_313_cast_fp16, y = var_25026_to_fp16)[name = string("op_25027_cast_fp16")];
            tensor<int32, [1]> q_u_313_axes_1 = const()[name = string("q_u_313_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_313_cast_fp16 = expand_dims(axes = q_u_313_axes_1, x = var_25027_cast_fp16)[name = string("q_u_313_cast_fp16")];
            tensor<fp16, [128]> var_25029_to_fp16 = const()[name = string("op_25029_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572719872)))];
            tensor<fp16, [1, 51, 128]> var_25030_cast_fp16 = add(x = q_head_313_cast_fp16, y = var_25029_to_fp16)[name = string("op_25030_cast_fp16")];
            tensor<int32, [1]> q_v_313_axes_1 = const()[name = string("q_v_313_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_313_cast_fp16 = expand_dims(axes = q_v_313_axes_1, x = var_25030_cast_fp16)[name = string("q_v_313_cast_fp16")];
            tensor<int32, [1]> k_head_627_axes_1 = const()[name = string("k_head_627_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_627_cast_fp16 = expand_dims(axes = k_head_627_axes_1, x = k_head_625_cast_fp16)[name = string("k_head_627_cast_fp16")];
            tensor<int32, [1]> v_head_627_axes_1 = const()[name = string("v_head_627_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_627_cast_fp16 = expand_dims(axes = v_head_627_axes_1, x = v_head_625_cast_fp16)[name = string("v_head_627_cast_fp16")];
            tensor<int32, [1]> p_head_627_axes_1 = const()[name = string("p_head_627_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_627_cast_fp16 = expand_dims(axes = p_head_627_axes_1, x = p_head_625_cast_fp16)[name = string("p_head_627_cast_fp16")];
            bool var_25036_transpose_x_3 = const()[name = string("op_25036_transpose_x_3"), val = bool(false)];
            bool var_25036_transpose_y_3 = const()[name = string("op_25036_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_25036_cast_fp16 = matmul(transpose_x = var_25036_transpose_x_3, transpose_y = var_25036_transpose_y_3, x = q_u_313_cast_fp16, y = k_head_627_cast_fp16)[name = string("op_25036_cast_fp16")];
            fp16 var_25037_to_fp16 = const()[name = string("op_25037_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_313_cast_fp16 = mul(x = var_25036_cast_fp16, y = var_25037_to_fp16)[name = string("scores_content_313_cast_fp16")];
            bool x_1649_transpose_x_3 = const()[name = string("x_1649_transpose_x_3"), val = bool(false)];
            bool x_1649_transpose_y_3 = const()[name = string("x_1649_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1649_cast_fp16 = matmul(transpose_x = x_1649_transpose_x_3, transpose_y = x_1649_transpose_y_3, x = q_v_313_cast_fp16, y = p_head_627_cast_fp16)[name = string("x_1649_cast_fp16")];
            tensor<int32, [8]> x_1651_pad_1 = const()[name = string("x_1651_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1651_mode_1 = const()[name = string("x_1651_mode_1"), val = string("constant")];
            fp16 const_2359_to_fp16 = const()[name = string("const_2359_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1651_cast_fp16 = pad(constant_val = const_2359_to_fp16, mode = x_1651_mode_1, pad = x_1651_pad_1, x = x_1649_cast_fp16)[name = string("x_1651_cast_fp16")];
            tensor<int32, [4]> var_25051 = const()[name = string("op_25051"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1653_cast_fp16 = reshape(shape = var_25051, x = x_1651_cast_fp16)[name = string("x_1653_cast_fp16")];
            tensor<int32, [4]> var_25055_begin_1 = const()[name = string("op_25055_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_25055_end_1 = const()[name = string("op_25055_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_25055_end_mask_1 = const()[name = string("op_25055_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_25055_cast_fp16 = slice_by_index(begin = var_25055_begin_1, end = var_25055_end_1, end_mask = var_25055_end_mask_1, x = x_1653_cast_fp16)[name = string("op_25055_cast_fp16")];
            tensor<int32, [4]> var_25057 = const()[name = string("op_25057"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_25058_cast_fp16 = reshape(shape = var_25057, x = var_25055_cast_fp16)[name = string("op_25058_cast_fp16")];
            tensor<int32, [4]> var_25063_begin_1 = const()[name = string("op_25063_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_25063_end_1 = const()[name = string("op_25063_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_25063_end_mask_1 = const()[name = string("op_25063_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_25063_cast_fp16 = slice_by_index(begin = var_25063_begin_1, end = var_25063_end_1, end_mask = var_25063_end_mask_1, x = var_25058_cast_fp16)[name = string("op_25063_cast_fp16")];
            fp16 var_25064_to_fp16 = const()[name = string("op_25064_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_313_cast_fp16 = mul(x = var_25063_cast_fp16, y = var_25064_to_fp16)[name = string("scores_pos_313_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_313_cast_fp16 = add(x = scores_content_313_cast_fp16, y = scores_pos_313_cast_fp16)[name = string("logits_313_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_25067_cast_fp16 = softmax(axis = var_24308, x = logits_313_cast_fp16)[name = string("op_25067_cast_fp16")];
            bool var_25069_transpose_x_1 = const()[name = string("op_25069_transpose_x_1"), val = bool(false)];
            bool var_25069_transpose_y_1 = const()[name = string("op_25069_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_25069_cast_fp16 = matmul(transpose_x = var_25069_transpose_x_1, transpose_y = var_25069_transpose_y_1, x = var_25067_cast_fp16, y = v_head_627_cast_fp16)[name = string("op_25069_cast_fp16")];
            tensor<int32, [1]> var_25070_axes_1 = const()[name = string("op_25070_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_25070_cast_fp16 = squeeze(axes = var_25070_axes_1, x = var_25069_cast_fp16)[name = string("op_25070_cast_fp16")];
            string dense_output_1577_pad_type_1 = const()[name = string("dense_output_1577_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1577_strides_1 = const()[name = string("dense_output_1577_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1577_pad_1 = const()[name = string("dense_output_1577_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1577_dilations_1 = const()[name = string("dense_output_1577_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1577_groups_1 = const()[name = string("dense_output_1577_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572720192))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572851328))))[name = string("layers_19_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1577_cast_fp16 = conv(dilations = dense_output_1577_dilations_1, groups = dense_output_1577_groups_1, pad = dense_output_1577_pad_1, pad_type = dense_output_1577_pad_type_1, strides = dense_output_1577_strides_1, weight = layers_19_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1577_cast_fp16")];
            string sparse_output_1577_pad_type_1 = const()[name = string("sparse_output_1577_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1577_strides_1 = const()[name = string("sparse_output_1577_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1577_pad_1 = const()[name = string("sparse_output_1577_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1577_dilations_1 = const()[name = string("sparse_output_1577_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1577_groups_1 = const()[name = string("sparse_output_1577_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572854592))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572851904))))[name = string("layers_19_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1577_cast_fp16 = conv(dilations = sparse_output_1577_dilations_1, groups = sparse_output_1577_groups_1, pad = sparse_output_1577_pad_1, pad_type = sparse_output_1577_pad_type_1, strides = sparse_output_1577_strides_1, weight = layers_19_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1577_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25085_cast_fp16 = add(x = dense_output_1577_cast_fp16, y = sparse_output_1577_cast_fp16)[name = string("op_25085_cast_fp16")];
            tensor<int32, [4]> var_25086 = const()[name = string("op_25086"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25088 = const()[name = string("op_25088"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25087_cast_fp16 = transpose(perm = var_25086, x = var_25085_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 51, 128]> q_head_315_cast_fp16 = reshape(shape = var_25088, x = var_25087_cast_fp16)[name = string("q_head_315_cast_fp16")];
            string dense_output_1579_pad_type_1 = const()[name = string("dense_output_1579_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1579_strides_1 = const()[name = string("dense_output_1579_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1579_pad_1 = const()[name = string("dense_output_1579_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1579_dilations_1 = const()[name = string("dense_output_1579_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1579_groups_1 = const()[name = string("dense_output_1579_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572871040))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573002176))))[name = string("layers_19_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1579_cast_fp16 = conv(dilations = dense_output_1579_dilations_1, groups = dense_output_1579_groups_1, pad = dense_output_1579_pad_1, pad_type = dense_output_1579_pad_type_1, strides = dense_output_1579_strides_1, weight = layers_19_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1579_cast_fp16")];
            string sparse_output_1579_pad_type_1 = const()[name = string("sparse_output_1579_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1579_strides_1 = const()[name = string("sparse_output_1579_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1579_pad_1 = const()[name = string("sparse_output_1579_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1579_dilations_1 = const()[name = string("sparse_output_1579_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1579_groups_1 = const()[name = string("sparse_output_1579_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573005440))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573002752))))[name = string("layers_19_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1579_cast_fp16 = conv(dilations = sparse_output_1579_dilations_1, groups = sparse_output_1579_groups_1, pad = sparse_output_1579_pad_1, pad_type = sparse_output_1579_pad_type_1, strides = sparse_output_1579_strides_1, weight = layers_19_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1579_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25104_cast_fp16 = add(x = dense_output_1579_cast_fp16, y = sparse_output_1579_cast_fp16)[name = string("op_25104_cast_fp16")];
            tensor<int32, [4]> var_25105 = const()[name = string("op_25105"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25107 = const()[name = string("op_25107"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25106_cast_fp16 = transpose(perm = var_25105, x = var_25104_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 51, 128]> k_head_629_cast_fp16 = reshape(shape = var_25107, x = var_25106_cast_fp16)[name = string("k_head_629_cast_fp16")];
            string dense_output_1581_pad_type_1 = const()[name = string("dense_output_1581_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1581_strides_1 = const()[name = string("dense_output_1581_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1581_pad_1 = const()[name = string("dense_output_1581_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1581_dilations_1 = const()[name = string("dense_output_1581_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1581_groups_1 = const()[name = string("dense_output_1581_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573021888))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573153024))))[name = string("layers_19_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1581_cast_fp16 = conv(dilations = dense_output_1581_dilations_1, groups = dense_output_1581_groups_1, pad = dense_output_1581_pad_1, pad_type = dense_output_1581_pad_type_1, strides = dense_output_1581_strides_1, weight = layers_19_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1581_cast_fp16")];
            string sparse_output_1581_pad_type_1 = const()[name = string("sparse_output_1581_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1581_strides_1 = const()[name = string("sparse_output_1581_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1581_pad_1 = const()[name = string("sparse_output_1581_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1581_dilations_1 = const()[name = string("sparse_output_1581_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1581_groups_1 = const()[name = string("sparse_output_1581_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573156288))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573153600))))[name = string("layers_19_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1581_cast_fp16 = conv(dilations = sparse_output_1581_dilations_1, groups = sparse_output_1581_groups_1, pad = sparse_output_1581_pad_1, pad_type = sparse_output_1581_pad_type_1, strides = sparse_output_1581_strides_1, weight = layers_19_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1581_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25123_cast_fp16 = add(x = dense_output_1581_cast_fp16, y = sparse_output_1581_cast_fp16)[name = string("op_25123_cast_fp16")];
            tensor<int32, [4]> var_25124 = const()[name = string("op_25124"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25126 = const()[name = string("op_25126"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25125_cast_fp16 = transpose(perm = var_25124, x = var_25123_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 51, 128]> v_head_629_cast_fp16 = reshape(shape = var_25126, x = var_25125_cast_fp16)[name = string("v_head_629_cast_fp16")];
            string dense_output_1583_pad_type_1 = const()[name = string("dense_output_1583_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1583_strides_1 = const()[name = string("dense_output_1583_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1583_pad_1 = const()[name = string("dense_output_1583_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1583_dilations_1 = const()[name = string("dense_output_1583_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1583_groups_1 = const()[name = string("dense_output_1583_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573172736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573303872))))[name = string("layers_19_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1583_cast_fp16 = conv(dilations = dense_output_1583_dilations_1, groups = dense_output_1583_groups_1, pad = dense_output_1583_pad_1, pad_type = dense_output_1583_pad_type_1, strides = dense_output_1583_strides_1, weight = layers_19_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1583_cast_fp16")];
            string sparse_output_1583_pad_type_1 = const()[name = string("sparse_output_1583_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1583_strides_1 = const()[name = string("sparse_output_1583_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1583_pad_1 = const()[name = string("sparse_output_1583_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1583_dilations_1 = const()[name = string("sparse_output_1583_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1583_groups_1 = const()[name = string("sparse_output_1583_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573307136))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573304448))))[name = string("layers_19_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1583_cast_fp16 = conv(dilations = sparse_output_1583_dilations_1, groups = sparse_output_1583_groups_1, pad = sparse_output_1583_pad_1, pad_type = sparse_output_1583_pad_type_1, strides = sparse_output_1583_strides_1, weight = layers_19_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1583_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_25142_cast_fp16 = add(x = dense_output_1583_cast_fp16, y = sparse_output_1583_cast_fp16)[name = string("op_25142_cast_fp16")];
            tensor<int32, [4]> var_25143 = const()[name = string("op_25143"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25145 = const()[name = string("op_25145"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_25144_cast_fp16 = transpose(perm = var_25143, x = var_25142_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 101, 128]> p_head_629_cast_fp16 = reshape(shape = var_25145, x = var_25144_cast_fp16)[name = string("p_head_629_cast_fp16")];
            tensor<fp16, [128]> var_25147_to_fp16 = const()[name = string("op_25147_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573323584)))];
            tensor<fp16, [1, 51, 128]> var_25148_cast_fp16 = add(x = q_head_315_cast_fp16, y = var_25147_to_fp16)[name = string("op_25148_cast_fp16")];
            tensor<int32, [1]> q_u_315_axes_1 = const()[name = string("q_u_315_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_315_cast_fp16 = expand_dims(axes = q_u_315_axes_1, x = var_25148_cast_fp16)[name = string("q_u_315_cast_fp16")];
            tensor<fp16, [128]> var_25150_to_fp16 = const()[name = string("op_25150_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573323904)))];
            tensor<fp16, [1, 51, 128]> var_25151_cast_fp16 = add(x = q_head_315_cast_fp16, y = var_25150_to_fp16)[name = string("op_25151_cast_fp16")];
            tensor<int32, [1]> q_v_315_axes_1 = const()[name = string("q_v_315_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_315_cast_fp16 = expand_dims(axes = q_v_315_axes_1, x = var_25151_cast_fp16)[name = string("q_v_315_cast_fp16")];
            tensor<int32, [1]> k_head_631_axes_1 = const()[name = string("k_head_631_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_631_cast_fp16 = expand_dims(axes = k_head_631_axes_1, x = k_head_629_cast_fp16)[name = string("k_head_631_cast_fp16")];
            tensor<int32, [1]> v_head_631_axes_1 = const()[name = string("v_head_631_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_631_cast_fp16 = expand_dims(axes = v_head_631_axes_1, x = v_head_629_cast_fp16)[name = string("v_head_631_cast_fp16")];
            tensor<int32, [1]> p_head_631_axes_1 = const()[name = string("p_head_631_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_631_cast_fp16 = expand_dims(axes = p_head_631_axes_1, x = p_head_629_cast_fp16)[name = string("p_head_631_cast_fp16")];
            bool var_25157_transpose_x_3 = const()[name = string("op_25157_transpose_x_3"), val = bool(false)];
            bool var_25157_transpose_y_3 = const()[name = string("op_25157_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_25157_cast_fp16 = matmul(transpose_x = var_25157_transpose_x_3, transpose_y = var_25157_transpose_y_3, x = q_u_315_cast_fp16, y = k_head_631_cast_fp16)[name = string("op_25157_cast_fp16")];
            fp16 var_25158_to_fp16 = const()[name = string("op_25158_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_315_cast_fp16 = mul(x = var_25157_cast_fp16, y = var_25158_to_fp16)[name = string("scores_content_315_cast_fp16")];
            bool x_1657_transpose_x_3 = const()[name = string("x_1657_transpose_x_3"), val = bool(false)];
            bool x_1657_transpose_y_3 = const()[name = string("x_1657_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1657_cast_fp16 = matmul(transpose_x = x_1657_transpose_x_3, transpose_y = x_1657_transpose_y_3, x = q_v_315_cast_fp16, y = p_head_631_cast_fp16)[name = string("x_1657_cast_fp16")];
            tensor<int32, [8]> x_1659_pad_1 = const()[name = string("x_1659_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1659_mode_1 = const()[name = string("x_1659_mode_1"), val = string("constant")];
            fp16 const_2365_to_fp16 = const()[name = string("const_2365_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1659_cast_fp16 = pad(constant_val = const_2365_to_fp16, mode = x_1659_mode_1, pad = x_1659_pad_1, x = x_1657_cast_fp16)[name = string("x_1659_cast_fp16")];
            tensor<int32, [4]> var_25172 = const()[name = string("op_25172"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1661_cast_fp16 = reshape(shape = var_25172, x = x_1659_cast_fp16)[name = string("x_1661_cast_fp16")];
            tensor<int32, [4]> var_25176_begin_1 = const()[name = string("op_25176_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_25176_end_1 = const()[name = string("op_25176_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_25176_end_mask_1 = const()[name = string("op_25176_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_25176_cast_fp16 = slice_by_index(begin = var_25176_begin_1, end = var_25176_end_1, end_mask = var_25176_end_mask_1, x = x_1661_cast_fp16)[name = string("op_25176_cast_fp16")];
            tensor<int32, [4]> var_25178 = const()[name = string("op_25178"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_25179_cast_fp16 = reshape(shape = var_25178, x = var_25176_cast_fp16)[name = string("op_25179_cast_fp16")];
            tensor<int32, [4]> var_25184_begin_1 = const()[name = string("op_25184_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_25184_end_1 = const()[name = string("op_25184_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_25184_end_mask_1 = const()[name = string("op_25184_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_25184_cast_fp16 = slice_by_index(begin = var_25184_begin_1, end = var_25184_end_1, end_mask = var_25184_end_mask_1, x = var_25179_cast_fp16)[name = string("op_25184_cast_fp16")];
            fp16 var_25185_to_fp16 = const()[name = string("op_25185_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_315_cast_fp16 = mul(x = var_25184_cast_fp16, y = var_25185_to_fp16)[name = string("scores_pos_315_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_315_cast_fp16 = add(x = scores_content_315_cast_fp16, y = scores_pos_315_cast_fp16)[name = string("logits_315_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_25188_cast_fp16 = softmax(axis = var_24308, x = logits_315_cast_fp16)[name = string("op_25188_cast_fp16")];
            bool var_25190_transpose_x_1 = const()[name = string("op_25190_transpose_x_1"), val = bool(false)];
            bool var_25190_transpose_y_1 = const()[name = string("op_25190_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_25190_cast_fp16 = matmul(transpose_x = var_25190_transpose_x_1, transpose_y = var_25190_transpose_y_1, x = var_25188_cast_fp16, y = v_head_631_cast_fp16)[name = string("op_25190_cast_fp16")];
            tensor<int32, [1]> var_25191_axes_1 = const()[name = string("op_25191_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_25191_cast_fp16 = squeeze(axes = var_25191_axes_1, x = var_25190_cast_fp16)[name = string("op_25191_cast_fp16")];
            string dense_output_1585_pad_type_1 = const()[name = string("dense_output_1585_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1585_strides_1 = const()[name = string("dense_output_1585_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1585_pad_1 = const()[name = string("dense_output_1585_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1585_dilations_1 = const()[name = string("dense_output_1585_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1585_groups_1 = const()[name = string("dense_output_1585_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573324224))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573455360))))[name = string("layers_19_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1585_cast_fp16 = conv(dilations = dense_output_1585_dilations_1, groups = dense_output_1585_groups_1, pad = dense_output_1585_pad_1, pad_type = dense_output_1585_pad_type_1, strides = dense_output_1585_strides_1, weight = layers_19_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1585_cast_fp16")];
            string sparse_output_1585_pad_type_1 = const()[name = string("sparse_output_1585_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1585_strides_1 = const()[name = string("sparse_output_1585_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1585_pad_1 = const()[name = string("sparse_output_1585_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1585_dilations_1 = const()[name = string("sparse_output_1585_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1585_groups_1 = const()[name = string("sparse_output_1585_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573458624))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573455936))))[name = string("layers_19_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1585_cast_fp16 = conv(dilations = sparse_output_1585_dilations_1, groups = sparse_output_1585_groups_1, pad = sparse_output_1585_pad_1, pad_type = sparse_output_1585_pad_type_1, strides = sparse_output_1585_strides_1, weight = layers_19_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1585_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25206_cast_fp16 = add(x = dense_output_1585_cast_fp16, y = sparse_output_1585_cast_fp16)[name = string("op_25206_cast_fp16")];
            tensor<int32, [4]> var_25207 = const()[name = string("op_25207"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25209 = const()[name = string("op_25209"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25208_cast_fp16 = transpose(perm = var_25207, x = var_25206_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 51, 128]> q_head_317_cast_fp16 = reshape(shape = var_25209, x = var_25208_cast_fp16)[name = string("q_head_317_cast_fp16")];
            string dense_output_1587_pad_type_1 = const()[name = string("dense_output_1587_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1587_strides_1 = const()[name = string("dense_output_1587_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1587_pad_1 = const()[name = string("dense_output_1587_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1587_dilations_1 = const()[name = string("dense_output_1587_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1587_groups_1 = const()[name = string("dense_output_1587_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573475072))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573606208))))[name = string("layers_19_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1587_cast_fp16 = conv(dilations = dense_output_1587_dilations_1, groups = dense_output_1587_groups_1, pad = dense_output_1587_pad_1, pad_type = dense_output_1587_pad_type_1, strides = dense_output_1587_strides_1, weight = layers_19_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1587_cast_fp16")];
            string sparse_output_1587_pad_type_1 = const()[name = string("sparse_output_1587_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1587_strides_1 = const()[name = string("sparse_output_1587_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1587_pad_1 = const()[name = string("sparse_output_1587_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1587_dilations_1 = const()[name = string("sparse_output_1587_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1587_groups_1 = const()[name = string("sparse_output_1587_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573609472))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573606784))))[name = string("layers_19_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1587_cast_fp16 = conv(dilations = sparse_output_1587_dilations_1, groups = sparse_output_1587_groups_1, pad = sparse_output_1587_pad_1, pad_type = sparse_output_1587_pad_type_1, strides = sparse_output_1587_strides_1, weight = layers_19_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1587_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25225_cast_fp16 = add(x = dense_output_1587_cast_fp16, y = sparse_output_1587_cast_fp16)[name = string("op_25225_cast_fp16")];
            tensor<int32, [4]> var_25226 = const()[name = string("op_25226"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25228 = const()[name = string("op_25228"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25227_cast_fp16 = transpose(perm = var_25226, x = var_25225_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [1, 51, 128]> k_head_633_cast_fp16 = reshape(shape = var_25228, x = var_25227_cast_fp16)[name = string("k_head_633_cast_fp16")];
            string dense_output_1589_pad_type_1 = const()[name = string("dense_output_1589_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1589_strides_1 = const()[name = string("dense_output_1589_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1589_pad_1 = const()[name = string("dense_output_1589_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1589_dilations_1 = const()[name = string("dense_output_1589_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1589_groups_1 = const()[name = string("dense_output_1589_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573625920))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573757056))))[name = string("layers_19_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1589_cast_fp16 = conv(dilations = dense_output_1589_dilations_1, groups = dense_output_1589_groups_1, pad = dense_output_1589_pad_1, pad_type = dense_output_1589_pad_type_1, strides = dense_output_1589_strides_1, weight = layers_19_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1589_cast_fp16")];
            string sparse_output_1589_pad_type_1 = const()[name = string("sparse_output_1589_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1589_strides_1 = const()[name = string("sparse_output_1589_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1589_pad_1 = const()[name = string("sparse_output_1589_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1589_dilations_1 = const()[name = string("sparse_output_1589_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1589_groups_1 = const()[name = string("sparse_output_1589_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573760320))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573757632))))[name = string("layers_19_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1589_cast_fp16 = conv(dilations = sparse_output_1589_dilations_1, groups = sparse_output_1589_groups_1, pad = sparse_output_1589_pad_1, pad_type = sparse_output_1589_pad_type_1, strides = sparse_output_1589_strides_1, weight = layers_19_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1589_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25244_cast_fp16 = add(x = dense_output_1589_cast_fp16, y = sparse_output_1589_cast_fp16)[name = string("op_25244_cast_fp16")];
            tensor<int32, [4]> var_25245 = const()[name = string("op_25245"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25247 = const()[name = string("op_25247"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25246_cast_fp16 = transpose(perm = var_25245, x = var_25244_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [1, 51, 128]> v_head_633_cast_fp16 = reshape(shape = var_25247, x = var_25246_cast_fp16)[name = string("v_head_633_cast_fp16")];
            string dense_output_1591_pad_type_1 = const()[name = string("dense_output_1591_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1591_strides_1 = const()[name = string("dense_output_1591_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1591_pad_1 = const()[name = string("dense_output_1591_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1591_dilations_1 = const()[name = string("dense_output_1591_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1591_groups_1 = const()[name = string("dense_output_1591_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573776768))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573907904))))[name = string("layers_19_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1591_cast_fp16 = conv(dilations = dense_output_1591_dilations_1, groups = dense_output_1591_groups_1, pad = dense_output_1591_pad_1, pad_type = dense_output_1591_pad_type_1, strides = dense_output_1591_strides_1, weight = layers_19_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1591_cast_fp16")];
            string sparse_output_1591_pad_type_1 = const()[name = string("sparse_output_1591_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1591_strides_1 = const()[name = string("sparse_output_1591_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1591_pad_1 = const()[name = string("sparse_output_1591_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1591_dilations_1 = const()[name = string("sparse_output_1591_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1591_groups_1 = const()[name = string("sparse_output_1591_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573911168))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573908480))))[name = string("layers_19_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1591_cast_fp16 = conv(dilations = sparse_output_1591_dilations_1, groups = sparse_output_1591_groups_1, pad = sparse_output_1591_pad_1, pad_type = sparse_output_1591_pad_type_1, strides = sparse_output_1591_strides_1, weight = layers_19_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1591_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_25263_cast_fp16 = add(x = dense_output_1591_cast_fp16, y = sparse_output_1591_cast_fp16)[name = string("op_25263_cast_fp16")];
            tensor<int32, [4]> var_25264 = const()[name = string("op_25264"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25266 = const()[name = string("op_25266"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_25265_cast_fp16 = transpose(perm = var_25264, x = var_25263_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [1, 101, 128]> p_head_633_cast_fp16 = reshape(shape = var_25266, x = var_25265_cast_fp16)[name = string("p_head_633_cast_fp16")];
            tensor<fp16, [128]> var_25268_to_fp16 = const()[name = string("op_25268_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573927616)))];
            tensor<fp16, [1, 51, 128]> var_25269_cast_fp16 = add(x = q_head_317_cast_fp16, y = var_25268_to_fp16)[name = string("op_25269_cast_fp16")];
            tensor<int32, [1]> q_u_317_axes_1 = const()[name = string("q_u_317_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_317_cast_fp16 = expand_dims(axes = q_u_317_axes_1, x = var_25269_cast_fp16)[name = string("q_u_317_cast_fp16")];
            tensor<fp16, [128]> var_25271_to_fp16 = const()[name = string("op_25271_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573927936)))];
            tensor<fp16, [1, 51, 128]> var_25272_cast_fp16 = add(x = q_head_317_cast_fp16, y = var_25271_to_fp16)[name = string("op_25272_cast_fp16")];
            tensor<int32, [1]> q_v_317_axes_1 = const()[name = string("q_v_317_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_317_cast_fp16 = expand_dims(axes = q_v_317_axes_1, x = var_25272_cast_fp16)[name = string("q_v_317_cast_fp16")];
            tensor<int32, [1]> k_head_635_axes_1 = const()[name = string("k_head_635_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_635_cast_fp16 = expand_dims(axes = k_head_635_axes_1, x = k_head_633_cast_fp16)[name = string("k_head_635_cast_fp16")];
            tensor<int32, [1]> v_head_635_axes_1 = const()[name = string("v_head_635_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_635_cast_fp16 = expand_dims(axes = v_head_635_axes_1, x = v_head_633_cast_fp16)[name = string("v_head_635_cast_fp16")];
            tensor<int32, [1]> p_head_635_axes_1 = const()[name = string("p_head_635_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_635_cast_fp16 = expand_dims(axes = p_head_635_axes_1, x = p_head_633_cast_fp16)[name = string("p_head_635_cast_fp16")];
            bool var_25278_transpose_x_3 = const()[name = string("op_25278_transpose_x_3"), val = bool(false)];
            bool var_25278_transpose_y_3 = const()[name = string("op_25278_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_25278_cast_fp16 = matmul(transpose_x = var_25278_transpose_x_3, transpose_y = var_25278_transpose_y_3, x = q_u_317_cast_fp16, y = k_head_635_cast_fp16)[name = string("op_25278_cast_fp16")];
            fp16 var_25279_to_fp16 = const()[name = string("op_25279_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_317_cast_fp16 = mul(x = var_25278_cast_fp16, y = var_25279_to_fp16)[name = string("scores_content_317_cast_fp16")];
            bool x_1665_transpose_x_3 = const()[name = string("x_1665_transpose_x_3"), val = bool(false)];
            bool x_1665_transpose_y_3 = const()[name = string("x_1665_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1665_cast_fp16 = matmul(transpose_x = x_1665_transpose_x_3, transpose_y = x_1665_transpose_y_3, x = q_v_317_cast_fp16, y = p_head_635_cast_fp16)[name = string("x_1665_cast_fp16")];
            tensor<int32, [8]> x_1667_pad_1 = const()[name = string("x_1667_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1667_mode_1 = const()[name = string("x_1667_mode_1"), val = string("constant")];
            fp16 const_2371_to_fp16 = const()[name = string("const_2371_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1667_cast_fp16 = pad(constant_val = const_2371_to_fp16, mode = x_1667_mode_1, pad = x_1667_pad_1, x = x_1665_cast_fp16)[name = string("x_1667_cast_fp16")];
            tensor<int32, [4]> var_25293 = const()[name = string("op_25293"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1669_cast_fp16 = reshape(shape = var_25293, x = x_1667_cast_fp16)[name = string("x_1669_cast_fp16")];
            tensor<int32, [4]> var_25297_begin_1 = const()[name = string("op_25297_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_25297_end_1 = const()[name = string("op_25297_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_25297_end_mask_1 = const()[name = string("op_25297_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_25297_cast_fp16 = slice_by_index(begin = var_25297_begin_1, end = var_25297_end_1, end_mask = var_25297_end_mask_1, x = x_1669_cast_fp16)[name = string("op_25297_cast_fp16")];
            tensor<int32, [4]> var_25299 = const()[name = string("op_25299"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_25300_cast_fp16 = reshape(shape = var_25299, x = var_25297_cast_fp16)[name = string("op_25300_cast_fp16")];
            tensor<int32, [4]> var_25305_begin_1 = const()[name = string("op_25305_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_25305_end_1 = const()[name = string("op_25305_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_25305_end_mask_1 = const()[name = string("op_25305_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_25305_cast_fp16 = slice_by_index(begin = var_25305_begin_1, end = var_25305_end_1, end_mask = var_25305_end_mask_1, x = var_25300_cast_fp16)[name = string("op_25305_cast_fp16")];
            fp16 var_25306_to_fp16 = const()[name = string("op_25306_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_317_cast_fp16 = mul(x = var_25305_cast_fp16, y = var_25306_to_fp16)[name = string("scores_pos_317_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_317_cast_fp16 = add(x = scores_content_317_cast_fp16, y = scores_pos_317_cast_fp16)[name = string("logits_317_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_25309_cast_fp16 = softmax(axis = var_24308, x = logits_317_cast_fp16)[name = string("op_25309_cast_fp16")];
            bool var_25311_transpose_x_1 = const()[name = string("op_25311_transpose_x_1"), val = bool(false)];
            bool var_25311_transpose_y_1 = const()[name = string("op_25311_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_25311_cast_fp16 = matmul(transpose_x = var_25311_transpose_x_1, transpose_y = var_25311_transpose_y_1, x = var_25309_cast_fp16, y = v_head_635_cast_fp16)[name = string("op_25311_cast_fp16")];
            tensor<int32, [1]> var_25312_axes_1 = const()[name = string("op_25312_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_25312_cast_fp16 = squeeze(axes = var_25312_axes_1, x = var_25311_cast_fp16)[name = string("op_25312_cast_fp16")];
            string dense_output_1593_pad_type_1 = const()[name = string("dense_output_1593_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1593_strides_1 = const()[name = string("dense_output_1593_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1593_pad_1 = const()[name = string("dense_output_1593_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1593_dilations_1 = const()[name = string("dense_output_1593_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1593_groups_1 = const()[name = string("dense_output_1593_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573928256))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574059392))))[name = string("layers_19_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1593_cast_fp16 = conv(dilations = dense_output_1593_dilations_1, groups = dense_output_1593_groups_1, pad = dense_output_1593_pad_1, pad_type = dense_output_1593_pad_type_1, strides = dense_output_1593_strides_1, weight = layers_19_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1593_cast_fp16")];
            string sparse_output_1593_pad_type_1 = const()[name = string("sparse_output_1593_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1593_strides_1 = const()[name = string("sparse_output_1593_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1593_pad_1 = const()[name = string("sparse_output_1593_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1593_dilations_1 = const()[name = string("sparse_output_1593_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1593_groups_1 = const()[name = string("sparse_output_1593_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574062656))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574059968))))[name = string("layers_19_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1593_cast_fp16 = conv(dilations = sparse_output_1593_dilations_1, groups = sparse_output_1593_groups_1, pad = sparse_output_1593_pad_1, pad_type = sparse_output_1593_pad_type_1, strides = sparse_output_1593_strides_1, weight = layers_19_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1593_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25327_cast_fp16 = add(x = dense_output_1593_cast_fp16, y = sparse_output_1593_cast_fp16)[name = string("op_25327_cast_fp16")];
            tensor<int32, [4]> var_25328 = const()[name = string("op_25328"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25330 = const()[name = string("op_25330"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25329_cast_fp16 = transpose(perm = var_25328, x = var_25327_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [1, 51, 128]> q_head_319_cast_fp16 = reshape(shape = var_25330, x = var_25329_cast_fp16)[name = string("q_head_319_cast_fp16")];
            string dense_output_1595_pad_type_1 = const()[name = string("dense_output_1595_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1595_strides_1 = const()[name = string("dense_output_1595_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1595_pad_1 = const()[name = string("dense_output_1595_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1595_dilations_1 = const()[name = string("dense_output_1595_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1595_groups_1 = const()[name = string("dense_output_1595_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574079104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574210240))))[name = string("layers_19_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1595_cast_fp16 = conv(dilations = dense_output_1595_dilations_1, groups = dense_output_1595_groups_1, pad = dense_output_1595_pad_1, pad_type = dense_output_1595_pad_type_1, strides = dense_output_1595_strides_1, weight = layers_19_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1595_cast_fp16")];
            string sparse_output_1595_pad_type_1 = const()[name = string("sparse_output_1595_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1595_strides_1 = const()[name = string("sparse_output_1595_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1595_pad_1 = const()[name = string("sparse_output_1595_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1595_dilations_1 = const()[name = string("sparse_output_1595_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1595_groups_1 = const()[name = string("sparse_output_1595_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574213504))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574210816))))[name = string("layers_19_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1595_cast_fp16 = conv(dilations = sparse_output_1595_dilations_1, groups = sparse_output_1595_groups_1, pad = sparse_output_1595_pad_1, pad_type = sparse_output_1595_pad_type_1, strides = sparse_output_1595_strides_1, weight = layers_19_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1595_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25346_cast_fp16 = add(x = dense_output_1595_cast_fp16, y = sparse_output_1595_cast_fp16)[name = string("op_25346_cast_fp16")];
            tensor<int32, [4]> var_25347 = const()[name = string("op_25347"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25349 = const()[name = string("op_25349"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25348_cast_fp16 = transpose(perm = var_25347, x = var_25346_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 51, 128]> k_head_637_cast_fp16 = reshape(shape = var_25349, x = var_25348_cast_fp16)[name = string("k_head_637_cast_fp16")];
            string dense_output_1597_pad_type_1 = const()[name = string("dense_output_1597_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1597_strides_1 = const()[name = string("dense_output_1597_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1597_pad_1 = const()[name = string("dense_output_1597_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1597_dilations_1 = const()[name = string("dense_output_1597_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1597_groups_1 = const()[name = string("dense_output_1597_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574229952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574361088))))[name = string("layers_19_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1597_cast_fp16 = conv(dilations = dense_output_1597_dilations_1, groups = dense_output_1597_groups_1, pad = dense_output_1597_pad_1, pad_type = dense_output_1597_pad_type_1, strides = dense_output_1597_strides_1, weight = layers_19_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_903_cast_fp16)[name = string("dense_output_1597_cast_fp16")];
            string sparse_output_1597_pad_type_1 = const()[name = string("sparse_output_1597_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1597_strides_1 = const()[name = string("sparse_output_1597_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1597_pad_1 = const()[name = string("sparse_output_1597_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1597_dilations_1 = const()[name = string("sparse_output_1597_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1597_groups_1 = const()[name = string("sparse_output_1597_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574364352))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574361664))))[name = string("layers_19_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1597_cast_fp16 = conv(dilations = sparse_output_1597_dilations_1, groups = sparse_output_1597_groups_1, pad = sparse_output_1597_pad_1, pad_type = sparse_output_1597_pad_type_1, strides = sparse_output_1597_strides_1, weight = layers_19_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_903_cast_fp16)[name = string("sparse_output_1597_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25365_cast_fp16 = add(x = dense_output_1597_cast_fp16, y = sparse_output_1597_cast_fp16)[name = string("op_25365_cast_fp16")];
            tensor<int32, [4]> var_25366 = const()[name = string("op_25366"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25368 = const()[name = string("op_25368"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25367_cast_fp16 = transpose(perm = var_25366, x = var_25365_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 51, 128]> v_head_637_cast_fp16 = reshape(shape = var_25368, x = var_25367_cast_fp16)[name = string("v_head_637_cast_fp16")];
            string dense_output_1599_pad_type_1 = const()[name = string("dense_output_1599_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1599_strides_1 = const()[name = string("dense_output_1599_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1599_pad_1 = const()[name = string("dense_output_1599_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1599_dilations_1 = const()[name = string("dense_output_1599_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1599_groups_1 = const()[name = string("dense_output_1599_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574380800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574511936))))[name = string("layers_19_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1599_cast_fp16 = conv(dilations = dense_output_1599_dilations_1, groups = dense_output_1599_groups_1, pad = dense_output_1599_pad_1, pad_type = dense_output_1599_pad_type_1, strides = dense_output_1599_strides_1, weight = layers_19_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1599_cast_fp16")];
            string sparse_output_1599_pad_type_1 = const()[name = string("sparse_output_1599_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1599_strides_1 = const()[name = string("sparse_output_1599_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1599_pad_1 = const()[name = string("sparse_output_1599_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1599_dilations_1 = const()[name = string("sparse_output_1599_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1599_groups_1 = const()[name = string("sparse_output_1599_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_19_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574515200))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574512512))))[name = string("layers_19_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1599_cast_fp16 = conv(dilations = sparse_output_1599_dilations_1, groups = sparse_output_1599_groups_1, pad = sparse_output_1599_pad_1, pad_type = sparse_output_1599_pad_type_1, strides = sparse_output_1599_strides_1, weight = layers_19_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1599_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_25384_cast_fp16 = add(x = dense_output_1599_cast_fp16, y = sparse_output_1599_cast_fp16)[name = string("op_25384_cast_fp16")];
            tensor<int32, [4]> var_25385 = const()[name = string("op_25385"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25387 = const()[name = string("op_25387"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_25386_cast_fp16 = transpose(perm = var_25385, x = var_25384_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 101, 128]> p_head_637_cast_fp16 = reshape(shape = var_25387, x = var_25386_cast_fp16)[name = string("p_head_637_cast_fp16")];
            tensor<fp16, [128]> var_25389_to_fp16 = const()[name = string("op_25389_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574531648)))];
            tensor<fp16, [1, 51, 128]> var_25390_cast_fp16 = add(x = q_head_319_cast_fp16, y = var_25389_to_fp16)[name = string("op_25390_cast_fp16")];
            tensor<int32, [1]> q_u_319_axes_1 = const()[name = string("q_u_319_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_319_cast_fp16 = expand_dims(axes = q_u_319_axes_1, x = var_25390_cast_fp16)[name = string("q_u_319_cast_fp16")];
            tensor<fp16, [128]> var_25392_to_fp16 = const()[name = string("op_25392_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574531968)))];
            tensor<fp16, [1, 51, 128]> var_25393_cast_fp16 = add(x = q_head_319_cast_fp16, y = var_25392_to_fp16)[name = string("op_25393_cast_fp16")];
            tensor<int32, [1]> q_v_319_axes_1 = const()[name = string("q_v_319_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_319_cast_fp16 = expand_dims(axes = q_v_319_axes_1, x = var_25393_cast_fp16)[name = string("q_v_319_cast_fp16")];
            tensor<int32, [1]> k_head_639_axes_1 = const()[name = string("k_head_639_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_639_cast_fp16 = expand_dims(axes = k_head_639_axes_1, x = k_head_637_cast_fp16)[name = string("k_head_639_cast_fp16")];
            tensor<int32, [1]> v_head_639_axes_1 = const()[name = string("v_head_639_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_639_cast_fp16 = expand_dims(axes = v_head_639_axes_1, x = v_head_637_cast_fp16)[name = string("v_head_639_cast_fp16")];
            tensor<int32, [1]> p_head_639_axes_1 = const()[name = string("p_head_639_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_639_cast_fp16 = expand_dims(axes = p_head_639_axes_1, x = p_head_637_cast_fp16)[name = string("p_head_639_cast_fp16")];
            bool var_25399_transpose_x_3 = const()[name = string("op_25399_transpose_x_3"), val = bool(false)];
            bool var_25399_transpose_y_3 = const()[name = string("op_25399_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_25399_cast_fp16 = matmul(transpose_x = var_25399_transpose_x_3, transpose_y = var_25399_transpose_y_3, x = q_u_319_cast_fp16, y = k_head_639_cast_fp16)[name = string("op_25399_cast_fp16")];
            fp16 var_25400_to_fp16 = const()[name = string("op_25400_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_319_cast_fp16 = mul(x = var_25399_cast_fp16, y = var_25400_to_fp16)[name = string("scores_content_319_cast_fp16")];
            bool x_1673_transpose_x_3 = const()[name = string("x_1673_transpose_x_3"), val = bool(false)];
            bool x_1673_transpose_y_3 = const()[name = string("x_1673_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1673_cast_fp16 = matmul(transpose_x = x_1673_transpose_x_3, transpose_y = x_1673_transpose_y_3, x = q_v_319_cast_fp16, y = p_head_639_cast_fp16)[name = string("x_1673_cast_fp16")];
            tensor<int32, [8]> x_1675_pad_1 = const()[name = string("x_1675_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1675_mode_1 = const()[name = string("x_1675_mode_1"), val = string("constant")];
            fp16 const_2377_to_fp16 = const()[name = string("const_2377_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1675_cast_fp16 = pad(constant_val = const_2377_to_fp16, mode = x_1675_mode_1, pad = x_1675_pad_1, x = x_1673_cast_fp16)[name = string("x_1675_cast_fp16")];
            tensor<int32, [4]> var_25414 = const()[name = string("op_25414"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1677_cast_fp16 = reshape(shape = var_25414, x = x_1675_cast_fp16)[name = string("x_1677_cast_fp16")];
            tensor<int32, [4]> var_25418_begin_1 = const()[name = string("op_25418_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_25418_end_1 = const()[name = string("op_25418_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_25418_end_mask_1 = const()[name = string("op_25418_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_25418_cast_fp16 = slice_by_index(begin = var_25418_begin_1, end = var_25418_end_1, end_mask = var_25418_end_mask_1, x = x_1677_cast_fp16)[name = string("op_25418_cast_fp16")];
            tensor<int32, [4]> var_25420 = const()[name = string("op_25420"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_25421_cast_fp16 = reshape(shape = var_25420, x = var_25418_cast_fp16)[name = string("op_25421_cast_fp16")];
            tensor<int32, [4]> var_25426_begin_1 = const()[name = string("op_25426_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_25426_end_1 = const()[name = string("op_25426_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_25426_end_mask_1 = const()[name = string("op_25426_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_25426_cast_fp16 = slice_by_index(begin = var_25426_begin_1, end = var_25426_end_1, end_mask = var_25426_end_mask_1, x = var_25421_cast_fp16)[name = string("op_25426_cast_fp16")];
            fp16 var_25427_to_fp16 = const()[name = string("op_25427_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_319_cast_fp16 = mul(x = var_25426_cast_fp16, y = var_25427_to_fp16)[name = string("scores_pos_319_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_319_cast_fp16 = add(x = scores_content_319_cast_fp16, y = scores_pos_319_cast_fp16)[name = string("logits_319_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_25430_cast_fp16 = softmax(axis = var_24308, x = logits_319_cast_fp16)[name = string("op_25430_cast_fp16")];
            bool var_25432_transpose_x_1 = const()[name = string("op_25432_transpose_x_1"), val = bool(false)];
            bool var_25432_transpose_y_1 = const()[name = string("op_25432_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_25432_cast_fp16 = matmul(transpose_x = var_25432_transpose_x_1, transpose_y = var_25432_transpose_y_1, x = var_25430_cast_fp16, y = v_head_639_cast_fp16)[name = string("op_25432_cast_fp16")];
            tensor<int32, [1]> o_head_39_axes_1 = const()[name = string("o_head_39_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_39_cast_fp16 = squeeze(axes = o_head_39_axes_1, x = var_25432_cast_fp16)[name = string("o_head_39_cast_fp16")];
            bool out_39_interleave_1 = const()[name = string("out_39_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_39_cast_fp16 = concat(axis = var_24308, interleave = out_39_interleave_1, values = (var_24586_cast_fp16, var_24707_cast_fp16, var_24828_cast_fp16, var_24949_cast_fp16, var_25070_cast_fp16, var_25191_cast_fp16, var_25312_cast_fp16, o_head_39_cast_fp16))[name = string("out_39_cast_fp16")];
            tensor<int32, [3]> var_25436_perm_1 = const()[name = string("op_25436_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_911_axes_1 = const()[name = string("input_911_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_25436_cast_fp16 = transpose(perm = var_25436_perm_1, x = out_39_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 1024, 51, 1]> input_911_cast_fp16 = expand_dims(axes = input_911_axes_1, x = var_25436_cast_fp16)[name = string("input_911_cast_fp16")];
            string dense_output_1601_pad_type_1 = const()[name = string("dense_output_1601_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1601_strides_1 = const()[name = string("dense_output_1601_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1601_pad_1 = const()[name = string("dense_output_1601_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1601_dilations_1 = const()[name = string("dense_output_1601_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1601_groups_1 = const()[name = string("dense_output_1601_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574532288))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575580928))))[name = string("layers_19_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1601_cast_fp16 = conv(dilations = dense_output_1601_dilations_1, groups = dense_output_1601_groups_1, pad = dense_output_1601_pad_1, pad_type = dense_output_1601_pad_type_1, strides = dense_output_1601_strides_1, weight = layers_19_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("dense_output_1601_cast_fp16")];
            string sparse_output_1601_pad_type_1 = const()[name = string("sparse_output_1601_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1601_strides_1 = const()[name = string("sparse_output_1601_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1601_pad_1 = const()[name = string("sparse_output_1601_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1601_dilations_1 = const()[name = string("sparse_output_1601_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1601_groups_1 = const()[name = string("sparse_output_1601_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575602560))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575581504))))[name = string("layers_19_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1601_cast_fp16 = conv(dilations = sparse_output_1601_dilations_1, groups = sparse_output_1601_groups_1, pad = sparse_output_1601_pad_1, pad_type = sparse_output_1601_pad_type_1, strides = sparse_output_1601_strides_1, weight = layers_19_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_911_cast_fp16)[name = string("sparse_output_1601_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_39_cast_fp16 = add(x = dense_output_1601_cast_fp16, y = sparse_output_1601_cast_fp16)[name = string("out_conv_39_cast_fp16")];
            tensor<int32, [1]> var_25453_axes_1 = const()[name = string("op_25453_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_25453_cast_fp16 = squeeze(axes = var_25453_axes_1, x = out_conv_39_cast_fp16)[name = string("op_25453_cast_fp16")];
            tensor<int32, [3]> var_25454_perm_1 = const()[name = string("op_25454_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_25454_cast_fp16 = transpose(perm = var_25454_perm_1, x = var_25453_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 51, 1024]> input_913_cast_fp16 = add(x = input_901_cast_fp16, y = var_25454_cast_fp16)[name = string("input_913_cast_fp16")];
            tensor<int32, [1]> x_1681_axes_1 = const()[name = string("x_1681_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_19_norm_conv_weight_to_fp16 = const()[name = string("layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575733696)))];
            tensor<fp16, [1024]> layers_19_norm_conv_bias_to_fp16 = const()[name = string("layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575735808)))];
            tensor<fp16, [1, 51, 1024]> x_1681_cast_fp16 = layer_norm(axes = x_1681_axes_1, beta = layers_19_norm_conv_bias_to_fp16, epsilon = var_24323_to_fp16, gamma = layers_19_norm_conv_weight_to_fp16, x = input_913_cast_fp16)[name = string("x_1681_cast_fp16")];
            tensor<int32, [3]> var_25464_perm_1 = const()[name = string("op_25464_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_915_axes_1 = const()[name = string("input_915_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_25464_cast_fp16 = transpose(perm = var_25464_perm_1, x = x_1681_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 1024, 51, 1]> input_915_cast_fp16 = expand_dims(axes = input_915_axes_1, x = var_25464_cast_fp16)[name = string("input_915_cast_fp16")];
            string dense_output_1603_pad_type_1 = const()[name = string("dense_output_1603_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1603_strides_1 = const()[name = string("dense_output_1603_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1603_pad_1 = const()[name = string("dense_output_1603_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1603_dilations_1 = const()[name = string("dense_output_1603_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1603_groups_1 = const()[name = string("dense_output_1603_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_19_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575737920))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577835136))))[name = string("layers_19_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1603_cast_fp16 = conv(dilations = dense_output_1603_dilations_1, groups = dense_output_1603_groups_1, pad = dense_output_1603_pad_1, pad_type = dense_output_1603_pad_type_1, strides = dense_output_1603_strides_1, weight = layers_19_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("dense_output_1603_cast_fp16")];
            string sparse_output_1603_pad_type_1 = const()[name = string("sparse_output_1603_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1603_strides_1 = const()[name = string("sparse_output_1603_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1603_pad_1 = const()[name = string("sparse_output_1603_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1603_dilations_1 = const()[name = string("sparse_output_1603_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1603_groups_1 = const()[name = string("sparse_output_1603_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_19_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577877760))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(577835712))))[name = string("layers_19_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1603_cast_fp16 = conv(dilations = sparse_output_1603_dilations_1, groups = sparse_output_1603_groups_1, pad = sparse_output_1603_pad_1, pad_type = sparse_output_1603_pad_type_1, strides = sparse_output_1603_strides_1, weight = layers_19_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_915_cast_fp16)[name = string("sparse_output_1603_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_917_cast_fp16 = add(x = dense_output_1603_cast_fp16, y = sparse_output_1603_cast_fp16)[name = string("input_917_cast_fp16")];
            int32 input_919_split_num_splits_1 = const()[name = string("input_919_split_num_splits_1"), val = int32(2)];
            int32 input_919_split_axis_1 = const()[name = string("input_919_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_919_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_919_split_cast_fp16_1 = split(axis = input_919_split_axis_1, num_splits = input_919_split_num_splits_1, x = input_917_cast_fp16)[name = string("input_919_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_919_split_1_sigmoid_cast_fp16 = sigmoid(x = input_919_split_cast_fp16_1)[name = string("input_919_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_919_cast_fp16 = mul(x = input_919_split_cast_fp16_0, y = input_919_split_1_sigmoid_cast_fp16)[name = string("input_919_cast_fp16")];
            tensor<int32, [8]> input_921_pad_1 = const()[name = string("input_921_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_921_mode_1 = const()[name = string("input_921_mode_1"), val = string("constant")];
            fp16 const_2379_to_fp16 = const()[name = string("const_2379_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_921_cast_fp16 = pad(constant_val = const_2379_to_fp16, mode = input_921_mode_1, pad = input_921_pad_1, x = input_919_cast_fp16)[name = string("input_921_cast_fp16")];
            string dense_output_1605_pad_type_1 = const()[name = string("dense_output_1605_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1605_strides_1 = const()[name = string("dense_output_1605_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1605_pad_1 = const()[name = string("dense_output_1605_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1605_dilations_1 = const()[name = string("dense_output_1605_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1605_groups_1 = const()[name = string("dense_output_1605_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1605_cast_fp16 = conv(dilations = dense_output_1605_dilations_1, groups = dense_output_1605_groups_1, pad = dense_output_1605_pad_1, pad_type = dense_output_1605_pad_type_1, strides = dense_output_1605_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_921_cast_fp16)[name = string("dense_output_1605_cast_fp16")];
            string sparse_output_1605_pad_type_1 = const()[name = string("sparse_output_1605_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1605_strides_1 = const()[name = string("sparse_output_1605_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1605_pad_1 = const()[name = string("sparse_output_1605_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1605_dilations_1 = const()[name = string("sparse_output_1605_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1605_groups_1 = const()[name = string("sparse_output_1605_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_19_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578139968))))[name = string("layers_19_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1605_cast_fp16 = conv(dilations = sparse_output_1605_dilations_1, groups = sparse_output_1605_groups_1, pad = sparse_output_1605_pad_1, pad_type = sparse_output_1605_pad_type_1, strides = sparse_output_1605_strides_1, weight = layers_19_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_921_cast_fp16)[name = string("sparse_output_1605_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_923_cast_fp16 = add(x = dense_output_1605_cast_fp16, y = sparse_output_1605_cast_fp16)[name = string("input_923_cast_fp16")];
            tensor<fp16, [1024]> layers_19_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_19_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578158464)))];
            tensor<fp16, [1024]> layers_19_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_19_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578160576)))];
            tensor<fp16, [1024]> layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_19_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578162688)))];
            tensor<fp16, [1024]> layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_19_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578164800)))];
            tensor<fp16, [1, 1024, 51, 1]> input_925_cast_fp16 = batch_norm(beta = layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_24323_to_fp16, gamma = layers_19_conv_batch_norm_weight_to_fp16, mean = layers_19_conv_batch_norm_running_mean_to_fp16, variance = layers_19_conv_batch_norm_running_var_to_fp16, x = input_923_cast_fp16)[name = string("input_925_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_927_cast_fp16 = silu(x = input_925_cast_fp16)[name = string("input_927_cast_fp16")];
            string dense_output_1607_pad_type_1 = const()[name = string("dense_output_1607_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1607_strides_1 = const()[name = string("dense_output_1607_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1607_pad_1 = const()[name = string("dense_output_1607_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1607_dilations_1 = const()[name = string("dense_output_1607_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1607_groups_1 = const()[name = string("dense_output_1607_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578166912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579215552))))[name = string("layers_19_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1607_cast_fp16 = conv(dilations = dense_output_1607_dilations_1, groups = dense_output_1607_groups_1, pad = dense_output_1607_pad_1, pad_type = dense_output_1607_pad_type_1, strides = dense_output_1607_strides_1, weight = layers_19_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("dense_output_1607_cast_fp16")];
            string sparse_output_1607_pad_type_1 = const()[name = string("sparse_output_1607_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1607_strides_1 = const()[name = string("sparse_output_1607_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1607_pad_1 = const()[name = string("sparse_output_1607_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1607_dilations_1 = const()[name = string("sparse_output_1607_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1607_groups_1 = const()[name = string("sparse_output_1607_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_19_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579237184))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579216128))))[name = string("layers_19_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1607_cast_fp16 = conv(dilations = sparse_output_1607_dilations_1, groups = sparse_output_1607_groups_1, pad = sparse_output_1607_pad_1, pad_type = sparse_output_1607_pad_type_1, strides = sparse_output_1607_strides_1, weight = layers_19_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_927_cast_fp16)[name = string("sparse_output_1607_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1683_cast_fp16 = add(x = dense_output_1607_cast_fp16, y = sparse_output_1607_cast_fp16)[name = string("x_1683_cast_fp16")];
            tensor<int32, [1]> var_25520_axes_1 = const()[name = string("op_25520_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_25520_cast_fp16 = squeeze(axes = var_25520_axes_1, x = x_1683_cast_fp16)[name = string("op_25520_cast_fp16")];
            tensor<int32, [3]> var_25521_perm_1 = const()[name = string("op_25521_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_25521_cast_fp16 = transpose(perm = var_25521_perm_1, x = var_25520_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [1, 51, 1024]> input_929_cast_fp16 = add(x = input_913_cast_fp16, y = var_25521_cast_fp16)[name = string("input_929_cast_fp16")];
            tensor<int32, [1]> x_1685_axes_1 = const()[name = string("x_1685_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579368320)))];
            tensor<fp16, [1024]> layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579370432)))];
            tensor<fp16, [1, 51, 1024]> x_1685_cast_fp16 = layer_norm(axes = x_1685_axes_1, beta = layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_24323_to_fp16, gamma = layers_19_norm_feed_forward2_weight_to_fp16, x = input_929_cast_fp16)[name = string("x_1685_cast_fp16")];
            tensor<int32, [4]> var_25531 = const()[name = string("op_25531"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1687_cast_fp16 = reshape(shape = var_25531, x = x_1685_cast_fp16)[name = string("x_1687_cast_fp16")];
            tensor<int32, [4]> input_931_perm_1 = const()[name = string("input_931_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1609_pad_type_1 = const()[name = string("dense_output_1609_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1609_strides_1 = const()[name = string("dense_output_1609_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1609_pad_1 = const()[name = string("dense_output_1609_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1609_dilations_1 = const()[name = string("dense_output_1609_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1609_groups_1 = const()[name = string("dense_output_1609_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_19_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579372544))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583566912))))[name = string("layers_19_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_931_cast_fp16 = transpose(perm = input_931_perm_1, x = x_1687_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1609_cast_fp16 = conv(dilations = dense_output_1609_dilations_1, groups = dense_output_1609_groups_1, pad = dense_output_1609_pad_1, pad_type = dense_output_1609_pad_type_1, strides = dense_output_1609_strides_1, weight = layers_19_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_931_cast_fp16)[name = string("dense_output_1609_cast_fp16")];
            string sparse_output_1609_pad_type_1 = const()[name = string("sparse_output_1609_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1609_strides_1 = const()[name = string("sparse_output_1609_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1609_pad_1 = const()[name = string("sparse_output_1609_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1609_dilations_1 = const()[name = string("sparse_output_1609_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1609_groups_1 = const()[name = string("sparse_output_1609_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_19_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583651456))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583567488))))[name = string("layers_19_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1609_cast_fp16 = conv(dilations = sparse_output_1609_dilations_1, groups = sparse_output_1609_groups_1, pad = sparse_output_1609_pad_1, pad_type = sparse_output_1609_pad_type_1, strides = sparse_output_1609_strides_1, weight = layers_19_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_931_cast_fp16)[name = string("sparse_output_1609_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_933_cast_fp16 = add(x = dense_output_1609_cast_fp16, y = sparse_output_1609_cast_fp16)[name = string("input_933_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_935_cast_fp16 = silu(x = input_933_cast_fp16)[name = string("input_935_cast_fp16")];
            string dense_output_1611_pad_type_1 = const()[name = string("dense_output_1611_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1611_strides_1 = const()[name = string("dense_output_1611_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1611_pad_1 = const()[name = string("dense_output_1611_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1611_dilations_1 = const()[name = string("dense_output_1611_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1611_groups_1 = const()[name = string("dense_output_1611_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_19_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(584175808))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588370176))))[name = string("layers_19_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1611_cast_fp16 = conv(dilations = dense_output_1611_dilations_1, groups = dense_output_1611_groups_1, pad = dense_output_1611_pad_1, pad_type = dense_output_1611_pad_type_1, strides = dense_output_1611_strides_1, weight = layers_19_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_935_cast_fp16)[name = string("dense_output_1611_cast_fp16")];
            string sparse_output_1611_pad_type_1 = const()[name = string("sparse_output_1611_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1611_strides_1 = const()[name = string("sparse_output_1611_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1611_pad_1 = const()[name = string("sparse_output_1611_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1611_dilations_1 = const()[name = string("sparse_output_1611_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1611_groups_1 = const()[name = string("sparse_output_1611_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_19_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588454720))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588370752))))[name = string("layers_19_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1611_cast_fp16 = conv(dilations = sparse_output_1611_dilations_1, groups = sparse_output_1611_groups_1, pad = sparse_output_1611_pad_1, pad_type = sparse_output_1611_pad_type_1, strides = sparse_output_1611_strides_1, weight = layers_19_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_935_cast_fp16)[name = string("sparse_output_1611_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1689_cast_fp16 = add(x = dense_output_1611_cast_fp16, y = sparse_output_1611_cast_fp16)[name = string("x_1689_cast_fp16")];
            tensor<int32, [4]> x_1691_perm_1 = const()[name = string("x_1691_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_25566 = const()[name = string("op_25566"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1691_cast_fp16 = transpose(perm = x_1691_perm_1, x = x_1689_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [1, 51, 1024]> var_25567_cast_fp16 = reshape(shape = var_25566, x = x_1691_cast_fp16)[name = string("op_25567_cast_fp16")];
            fp16 var_25568_to_fp16 = const()[name = string("op_25568_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_25569_cast_fp16 = mul(x = var_25567_cast_fp16, y = var_25568_to_fp16)[name = string("op_25569_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_937_cast_fp16 = add(x = input_929_cast_fp16, y = var_25569_cast_fp16)[name = string("input_937_cast_fp16")];
            tensor<int32, [1]> input_939_axes_1 = const()[name = string("input_939_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_19_norm_out_weight_to_fp16 = const()[name = string("layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588979072)))];
            tensor<fp16, [1024]> layers_19_norm_out_bias_to_fp16 = const()[name = string("layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588981184)))];
            tensor<fp16, [1, 51, 1024]> input_939_cast_fp16 = layer_norm(axes = input_939_axes_1, beta = layers_19_norm_out_bias_to_fp16, epsilon = var_24323_to_fp16, gamma = layers_19_norm_out_weight_to_fp16, x = input_937_cast_fp16)[name = string("input_939_cast_fp16")];
            int32 var_25577 = const()[name = string("op_25577"), val = int32(-1)];
            tensor<int32, [1]> x_1693_axes_1 = const()[name = string("x_1693_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588983296)))];
            tensor<fp16, [1024]> layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588985408)))];
            fp16 var_25592_to_fp16 = const()[name = string("op_25592_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1693_cast_fp16 = layer_norm(axes = x_1693_axes_1, beta = layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_25592_to_fp16, gamma = layers_20_norm_feed_forward1_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_1693_cast_fp16")];
            tensor<int32, [4]> var_25611 = const()[name = string("op_25611"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1695_cast_fp16 = reshape(shape = var_25611, x = x_1693_cast_fp16)[name = string("x_1695_cast_fp16")];
            tensor<int32, [4]> input_941_perm_1 = const()[name = string("input_941_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1613_pad_type_1 = const()[name = string("dense_output_1613_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1613_strides_1 = const()[name = string("dense_output_1613_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1613_pad_1 = const()[name = string("dense_output_1613_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1613_dilations_1 = const()[name = string("dense_output_1613_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1613_groups_1 = const()[name = string("dense_output_1613_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_20_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588987520))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593181888))))[name = string("layers_20_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_941_cast_fp16 = transpose(perm = input_941_perm_1, x = x_1695_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1613_cast_fp16 = conv(dilations = dense_output_1613_dilations_1, groups = dense_output_1613_groups_1, pad = dense_output_1613_pad_1, pad_type = dense_output_1613_pad_type_1, strides = dense_output_1613_strides_1, weight = layers_20_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_941_cast_fp16)[name = string("dense_output_1613_cast_fp16")];
            string sparse_output_1613_pad_type_1 = const()[name = string("sparse_output_1613_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1613_strides_1 = const()[name = string("sparse_output_1613_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1613_pad_1 = const()[name = string("sparse_output_1613_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1613_dilations_1 = const()[name = string("sparse_output_1613_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1613_groups_1 = const()[name = string("sparse_output_1613_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_20_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593266432))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593182464))))[name = string("layers_20_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1613_cast_fp16 = conv(dilations = sparse_output_1613_dilations_1, groups = sparse_output_1613_groups_1, pad = sparse_output_1613_pad_1, pad_type = sparse_output_1613_pad_type_1, strides = sparse_output_1613_strides_1, weight = layers_20_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_941_cast_fp16)[name = string("sparse_output_1613_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_943_cast_fp16 = add(x = dense_output_1613_cast_fp16, y = sparse_output_1613_cast_fp16)[name = string("input_943_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_945_cast_fp16 = silu(x = input_943_cast_fp16)[name = string("input_945_cast_fp16")];
            string dense_output_1615_pad_type_1 = const()[name = string("dense_output_1615_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1615_strides_1 = const()[name = string("dense_output_1615_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1615_pad_1 = const()[name = string("dense_output_1615_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1615_dilations_1 = const()[name = string("dense_output_1615_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1615_groups_1 = const()[name = string("dense_output_1615_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_20_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593790784))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597985152))))[name = string("layers_20_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1615_cast_fp16 = conv(dilations = dense_output_1615_dilations_1, groups = dense_output_1615_groups_1, pad = dense_output_1615_pad_1, pad_type = dense_output_1615_pad_type_1, strides = dense_output_1615_strides_1, weight = layers_20_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_945_cast_fp16)[name = string("dense_output_1615_cast_fp16")];
            string sparse_output_1615_pad_type_1 = const()[name = string("sparse_output_1615_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1615_strides_1 = const()[name = string("sparse_output_1615_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1615_pad_1 = const()[name = string("sparse_output_1615_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1615_dilations_1 = const()[name = string("sparse_output_1615_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1615_groups_1 = const()[name = string("sparse_output_1615_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_20_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598069696))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597985728))))[name = string("layers_20_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1615_cast_fp16 = conv(dilations = sparse_output_1615_dilations_1, groups = sparse_output_1615_groups_1, pad = sparse_output_1615_pad_1, pad_type = sparse_output_1615_pad_type_1, strides = sparse_output_1615_strides_1, weight = layers_20_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_945_cast_fp16)[name = string("sparse_output_1615_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1697_cast_fp16 = add(x = dense_output_1615_cast_fp16, y = sparse_output_1615_cast_fp16)[name = string("x_1697_cast_fp16")];
            tensor<int32, [4]> x_1699_perm_1 = const()[name = string("x_1699_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_25646 = const()[name = string("op_25646"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1699_cast_fp16 = transpose(perm = x_1699_perm_1, x = x_1697_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 51, 1024]> var_25647_cast_fp16 = reshape(shape = var_25646, x = x_1699_cast_fp16)[name = string("op_25647_cast_fp16")];
            fp16 var_25648_to_fp16 = const()[name = string("op_25648_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_25649_cast_fp16 = mul(x = var_25647_cast_fp16, y = var_25648_to_fp16)[name = string("op_25649_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_947_cast_fp16 = add(x = input_939_cast_fp16, y = var_25649_cast_fp16)[name = string("input_947_cast_fp16")];
            tensor<int32, [1]> q_41_axes_1 = const()[name = string("q_41_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_20_norm_self_att_weight_to_fp16 = const()[name = string("layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598594048)))];
            tensor<fp16, [1024]> layers_20_norm_self_att_bias_to_fp16 = const()[name = string("layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598596160)))];
            tensor<fp16, [1, 51, 1024]> q_41_cast_fp16 = layer_norm(axes = q_41_axes_1, beta = layers_20_norm_self_att_bias_to_fp16, epsilon = var_25592_to_fp16, gamma = layers_20_norm_self_att_weight_to_fp16, x = input_947_cast_fp16)[name = string("q_41_cast_fp16")];
            tensor<int32, [3]> var_25723 = const()[name = string("op_25723"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_949_axes_1 = const()[name = string("input_949_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_25724_cast_fp16 = transpose(perm = var_25723, x = q_41_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 1024, 51, 1]> input_949_cast_fp16 = expand_dims(axes = input_949_axes_1, x = var_25724_cast_fp16)[name = string("input_949_cast_fp16")];
            string dense_output_1617_pad_type_1 = const()[name = string("dense_output_1617_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1617_strides_1 = const()[name = string("dense_output_1617_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1617_pad_1 = const()[name = string("dense_output_1617_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1617_dilations_1 = const()[name = string("dense_output_1617_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1617_groups_1 = const()[name = string("dense_output_1617_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598598272))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598729408))))[name = string("layers_20_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1617_cast_fp16 = conv(dilations = dense_output_1617_dilations_1, groups = dense_output_1617_groups_1, pad = dense_output_1617_pad_1, pad_type = dense_output_1617_pad_type_1, strides = dense_output_1617_strides_1, weight = layers_20_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1617_cast_fp16")];
            string sparse_output_1617_pad_type_1 = const()[name = string("sparse_output_1617_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1617_strides_1 = const()[name = string("sparse_output_1617_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1617_pad_1 = const()[name = string("sparse_output_1617_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1617_dilations_1 = const()[name = string("sparse_output_1617_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1617_groups_1 = const()[name = string("sparse_output_1617_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598732672))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598729984))))[name = string("layers_20_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1617_cast_fp16 = conv(dilations = sparse_output_1617_dilations_1, groups = sparse_output_1617_groups_1, pad = sparse_output_1617_pad_1, pad_type = sparse_output_1617_pad_type_1, strides = sparse_output_1617_strides_1, weight = layers_20_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1617_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25749_cast_fp16 = add(x = dense_output_1617_cast_fp16, y = sparse_output_1617_cast_fp16)[name = string("op_25749_cast_fp16")];
            tensor<int32, [4]> var_25750 = const()[name = string("op_25750"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25752 = const()[name = string("op_25752"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25751_cast_fp16 = transpose(perm = var_25750, x = var_25749_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 51, 128]> q_head_321_cast_fp16 = reshape(shape = var_25752, x = var_25751_cast_fp16)[name = string("q_head_321_cast_fp16")];
            string dense_output_1619_pad_type_1 = const()[name = string("dense_output_1619_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1619_strides_1 = const()[name = string("dense_output_1619_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1619_pad_1 = const()[name = string("dense_output_1619_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1619_dilations_1 = const()[name = string("dense_output_1619_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1619_groups_1 = const()[name = string("dense_output_1619_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598749120))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598880256))))[name = string("layers_20_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1619_cast_fp16 = conv(dilations = dense_output_1619_dilations_1, groups = dense_output_1619_groups_1, pad = dense_output_1619_pad_1, pad_type = dense_output_1619_pad_type_1, strides = dense_output_1619_strides_1, weight = layers_20_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1619_cast_fp16")];
            string sparse_output_1619_pad_type_1 = const()[name = string("sparse_output_1619_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1619_strides_1 = const()[name = string("sparse_output_1619_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1619_pad_1 = const()[name = string("sparse_output_1619_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1619_dilations_1 = const()[name = string("sparse_output_1619_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1619_groups_1 = const()[name = string("sparse_output_1619_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598883520))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598880832))))[name = string("layers_20_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1619_cast_fp16 = conv(dilations = sparse_output_1619_dilations_1, groups = sparse_output_1619_groups_1, pad = sparse_output_1619_pad_1, pad_type = sparse_output_1619_pad_type_1, strides = sparse_output_1619_strides_1, weight = layers_20_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1619_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25768_cast_fp16 = add(x = dense_output_1619_cast_fp16, y = sparse_output_1619_cast_fp16)[name = string("op_25768_cast_fp16")];
            tensor<int32, [4]> var_25769 = const()[name = string("op_25769"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25771 = const()[name = string("op_25771"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25770_cast_fp16 = transpose(perm = var_25769, x = var_25768_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 51, 128]> k_head_641_cast_fp16 = reshape(shape = var_25771, x = var_25770_cast_fp16)[name = string("k_head_641_cast_fp16")];
            string dense_output_1621_pad_type_1 = const()[name = string("dense_output_1621_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1621_strides_1 = const()[name = string("dense_output_1621_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1621_pad_1 = const()[name = string("dense_output_1621_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1621_dilations_1 = const()[name = string("dense_output_1621_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1621_groups_1 = const()[name = string("dense_output_1621_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598899968))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599031104))))[name = string("layers_20_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1621_cast_fp16 = conv(dilations = dense_output_1621_dilations_1, groups = dense_output_1621_groups_1, pad = dense_output_1621_pad_1, pad_type = dense_output_1621_pad_type_1, strides = dense_output_1621_strides_1, weight = layers_20_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1621_cast_fp16")];
            string sparse_output_1621_pad_type_1 = const()[name = string("sparse_output_1621_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1621_strides_1 = const()[name = string("sparse_output_1621_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1621_pad_1 = const()[name = string("sparse_output_1621_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1621_dilations_1 = const()[name = string("sparse_output_1621_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1621_groups_1 = const()[name = string("sparse_output_1621_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599034368))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599031680))))[name = string("layers_20_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1621_cast_fp16 = conv(dilations = sparse_output_1621_dilations_1, groups = sparse_output_1621_groups_1, pad = sparse_output_1621_pad_1, pad_type = sparse_output_1621_pad_type_1, strides = sparse_output_1621_strides_1, weight = layers_20_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1621_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25787_cast_fp16 = add(x = dense_output_1621_cast_fp16, y = sparse_output_1621_cast_fp16)[name = string("op_25787_cast_fp16")];
            tensor<int32, [4]> var_25788 = const()[name = string("op_25788"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25790 = const()[name = string("op_25790"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25789_cast_fp16 = transpose(perm = var_25788, x = var_25787_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 51, 128]> v_head_641_cast_fp16 = reshape(shape = var_25790, x = var_25789_cast_fp16)[name = string("v_head_641_cast_fp16")];
            string dense_output_1623_pad_type_1 = const()[name = string("dense_output_1623_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1623_strides_1 = const()[name = string("dense_output_1623_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1623_pad_1 = const()[name = string("dense_output_1623_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1623_dilations_1 = const()[name = string("dense_output_1623_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1623_groups_1 = const()[name = string("dense_output_1623_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599050816))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599181952))))[name = string("layers_20_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1623_cast_fp16 = conv(dilations = dense_output_1623_dilations_1, groups = dense_output_1623_groups_1, pad = dense_output_1623_pad_1, pad_type = dense_output_1623_pad_type_1, strides = dense_output_1623_strides_1, weight = layers_20_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1623_cast_fp16")];
            string sparse_output_1623_pad_type_1 = const()[name = string("sparse_output_1623_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1623_strides_1 = const()[name = string("sparse_output_1623_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1623_pad_1 = const()[name = string("sparse_output_1623_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1623_dilations_1 = const()[name = string("sparse_output_1623_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1623_groups_1 = const()[name = string("sparse_output_1623_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599185216))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599182528))))[name = string("layers_20_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1623_cast_fp16 = conv(dilations = sparse_output_1623_dilations_1, groups = sparse_output_1623_groups_1, pad = sparse_output_1623_pad_1, pad_type = sparse_output_1623_pad_type_1, strides = sparse_output_1623_strides_1, weight = layers_20_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1623_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_25806_cast_fp16 = add(x = dense_output_1623_cast_fp16, y = sparse_output_1623_cast_fp16)[name = string("op_25806_cast_fp16")];
            tensor<int32, [4]> var_25807 = const()[name = string("op_25807"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25809 = const()[name = string("op_25809"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_25808_cast_fp16 = transpose(perm = var_25807, x = var_25806_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 101, 128]> p_head_641_cast_fp16 = reshape(shape = var_25809, x = var_25808_cast_fp16)[name = string("p_head_641_cast_fp16")];
            tensor<fp16, [128]> var_25811_to_fp16 = const()[name = string("op_25811_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599201664)))];
            tensor<fp16, [1, 51, 128]> var_25812_cast_fp16 = add(x = q_head_321_cast_fp16, y = var_25811_to_fp16)[name = string("op_25812_cast_fp16")];
            tensor<int32, [1]> q_u_321_axes_1 = const()[name = string("q_u_321_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_321_cast_fp16 = expand_dims(axes = q_u_321_axes_1, x = var_25812_cast_fp16)[name = string("q_u_321_cast_fp16")];
            tensor<fp16, [128]> var_25814_to_fp16 = const()[name = string("op_25814_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599201984)))];
            tensor<fp16, [1, 51, 128]> var_25815_cast_fp16 = add(x = q_head_321_cast_fp16, y = var_25814_to_fp16)[name = string("op_25815_cast_fp16")];
            tensor<int32, [1]> q_v_321_axes_1 = const()[name = string("q_v_321_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_321_cast_fp16 = expand_dims(axes = q_v_321_axes_1, x = var_25815_cast_fp16)[name = string("q_v_321_cast_fp16")];
            tensor<int32, [1]> k_head_643_axes_1 = const()[name = string("k_head_643_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_643_cast_fp16 = expand_dims(axes = k_head_643_axes_1, x = k_head_641_cast_fp16)[name = string("k_head_643_cast_fp16")];
            tensor<int32, [1]> v_head_643_axes_1 = const()[name = string("v_head_643_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_643_cast_fp16 = expand_dims(axes = v_head_643_axes_1, x = v_head_641_cast_fp16)[name = string("v_head_643_cast_fp16")];
            tensor<int32, [1]> p_head_643_axes_1 = const()[name = string("p_head_643_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_643_cast_fp16 = expand_dims(axes = p_head_643_axes_1, x = p_head_641_cast_fp16)[name = string("p_head_643_cast_fp16")];
            bool var_25821_transpose_x_3 = const()[name = string("op_25821_transpose_x_3"), val = bool(false)];
            bool var_25821_transpose_y_3 = const()[name = string("op_25821_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_25821_cast_fp16 = matmul(transpose_x = var_25821_transpose_x_3, transpose_y = var_25821_transpose_y_3, x = q_u_321_cast_fp16, y = k_head_643_cast_fp16)[name = string("op_25821_cast_fp16")];
            fp16 var_25822_to_fp16 = const()[name = string("op_25822_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_321_cast_fp16 = mul(x = var_25821_cast_fp16, y = var_25822_to_fp16)[name = string("scores_content_321_cast_fp16")];
            bool x_1701_transpose_x_3 = const()[name = string("x_1701_transpose_x_3"), val = bool(false)];
            bool x_1701_transpose_y_3 = const()[name = string("x_1701_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1701_cast_fp16 = matmul(transpose_x = x_1701_transpose_x_3, transpose_y = x_1701_transpose_y_3, x = q_v_321_cast_fp16, y = p_head_643_cast_fp16)[name = string("x_1701_cast_fp16")];
            tensor<int32, [8]> x_1703_pad_1 = const()[name = string("x_1703_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1703_mode_1 = const()[name = string("x_1703_mode_1"), val = string("constant")];
            fp16 const_2389_to_fp16 = const()[name = string("const_2389_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1703_cast_fp16 = pad(constant_val = const_2389_to_fp16, mode = x_1703_mode_1, pad = x_1703_pad_1, x = x_1701_cast_fp16)[name = string("x_1703_cast_fp16")];
            tensor<int32, [4]> var_25836 = const()[name = string("op_25836"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1705_cast_fp16 = reshape(shape = var_25836, x = x_1703_cast_fp16)[name = string("x_1705_cast_fp16")];
            tensor<int32, [4]> var_25840_begin_1 = const()[name = string("op_25840_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_25840_end_1 = const()[name = string("op_25840_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_25840_end_mask_1 = const()[name = string("op_25840_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_25840_cast_fp16 = slice_by_index(begin = var_25840_begin_1, end = var_25840_end_1, end_mask = var_25840_end_mask_1, x = x_1705_cast_fp16)[name = string("op_25840_cast_fp16")];
            tensor<int32, [4]> var_25842 = const()[name = string("op_25842"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_25843_cast_fp16 = reshape(shape = var_25842, x = var_25840_cast_fp16)[name = string("op_25843_cast_fp16")];
            tensor<int32, [4]> var_25848_begin_1 = const()[name = string("op_25848_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_25848_end_1 = const()[name = string("op_25848_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_25848_end_mask_1 = const()[name = string("op_25848_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_25848_cast_fp16 = slice_by_index(begin = var_25848_begin_1, end = var_25848_end_1, end_mask = var_25848_end_mask_1, x = var_25843_cast_fp16)[name = string("op_25848_cast_fp16")];
            fp16 var_25849_to_fp16 = const()[name = string("op_25849_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_321_cast_fp16 = mul(x = var_25848_cast_fp16, y = var_25849_to_fp16)[name = string("scores_pos_321_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_321_cast_fp16 = add(x = scores_content_321_cast_fp16, y = scores_pos_321_cast_fp16)[name = string("logits_321_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_25852_cast_fp16 = softmax(axis = var_25577, x = logits_321_cast_fp16)[name = string("op_25852_cast_fp16")];
            bool var_25854_transpose_x_1 = const()[name = string("op_25854_transpose_x_1"), val = bool(false)];
            bool var_25854_transpose_y_1 = const()[name = string("op_25854_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_25854_cast_fp16 = matmul(transpose_x = var_25854_transpose_x_1, transpose_y = var_25854_transpose_y_1, x = var_25852_cast_fp16, y = v_head_643_cast_fp16)[name = string("op_25854_cast_fp16")];
            tensor<int32, [1]> var_25855_axes_1 = const()[name = string("op_25855_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_25855_cast_fp16 = squeeze(axes = var_25855_axes_1, x = var_25854_cast_fp16)[name = string("op_25855_cast_fp16")];
            string dense_output_1625_pad_type_1 = const()[name = string("dense_output_1625_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1625_strides_1 = const()[name = string("dense_output_1625_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1625_pad_1 = const()[name = string("dense_output_1625_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1625_dilations_1 = const()[name = string("dense_output_1625_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1625_groups_1 = const()[name = string("dense_output_1625_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599202304))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599333440))))[name = string("layers_20_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1625_cast_fp16 = conv(dilations = dense_output_1625_dilations_1, groups = dense_output_1625_groups_1, pad = dense_output_1625_pad_1, pad_type = dense_output_1625_pad_type_1, strides = dense_output_1625_strides_1, weight = layers_20_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1625_cast_fp16")];
            string sparse_output_1625_pad_type_1 = const()[name = string("sparse_output_1625_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1625_strides_1 = const()[name = string("sparse_output_1625_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1625_pad_1 = const()[name = string("sparse_output_1625_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1625_dilations_1 = const()[name = string("sparse_output_1625_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1625_groups_1 = const()[name = string("sparse_output_1625_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599336704))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599334016))))[name = string("layers_20_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1625_cast_fp16 = conv(dilations = sparse_output_1625_dilations_1, groups = sparse_output_1625_groups_1, pad = sparse_output_1625_pad_1, pad_type = sparse_output_1625_pad_type_1, strides = sparse_output_1625_strides_1, weight = layers_20_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1625_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25870_cast_fp16 = add(x = dense_output_1625_cast_fp16, y = sparse_output_1625_cast_fp16)[name = string("op_25870_cast_fp16")];
            tensor<int32, [4]> var_25871 = const()[name = string("op_25871"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25873 = const()[name = string("op_25873"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25872_cast_fp16 = transpose(perm = var_25871, x = var_25870_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 51, 128]> q_head_323_cast_fp16 = reshape(shape = var_25873, x = var_25872_cast_fp16)[name = string("q_head_323_cast_fp16")];
            string dense_output_1627_pad_type_1 = const()[name = string("dense_output_1627_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1627_strides_1 = const()[name = string("dense_output_1627_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1627_pad_1 = const()[name = string("dense_output_1627_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1627_dilations_1 = const()[name = string("dense_output_1627_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1627_groups_1 = const()[name = string("dense_output_1627_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599353152))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599484288))))[name = string("layers_20_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1627_cast_fp16 = conv(dilations = dense_output_1627_dilations_1, groups = dense_output_1627_groups_1, pad = dense_output_1627_pad_1, pad_type = dense_output_1627_pad_type_1, strides = dense_output_1627_strides_1, weight = layers_20_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1627_cast_fp16")];
            string sparse_output_1627_pad_type_1 = const()[name = string("sparse_output_1627_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1627_strides_1 = const()[name = string("sparse_output_1627_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1627_pad_1 = const()[name = string("sparse_output_1627_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1627_dilations_1 = const()[name = string("sparse_output_1627_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1627_groups_1 = const()[name = string("sparse_output_1627_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599487552))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599484864))))[name = string("layers_20_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1627_cast_fp16 = conv(dilations = sparse_output_1627_dilations_1, groups = sparse_output_1627_groups_1, pad = sparse_output_1627_pad_1, pad_type = sparse_output_1627_pad_type_1, strides = sparse_output_1627_strides_1, weight = layers_20_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1627_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25889_cast_fp16 = add(x = dense_output_1627_cast_fp16, y = sparse_output_1627_cast_fp16)[name = string("op_25889_cast_fp16")];
            tensor<int32, [4]> var_25890 = const()[name = string("op_25890"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25892 = const()[name = string("op_25892"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25891_cast_fp16 = transpose(perm = var_25890, x = var_25889_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [1, 51, 128]> k_head_645_cast_fp16 = reshape(shape = var_25892, x = var_25891_cast_fp16)[name = string("k_head_645_cast_fp16")];
            string dense_output_1629_pad_type_1 = const()[name = string("dense_output_1629_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1629_strides_1 = const()[name = string("dense_output_1629_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1629_pad_1 = const()[name = string("dense_output_1629_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1629_dilations_1 = const()[name = string("dense_output_1629_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1629_groups_1 = const()[name = string("dense_output_1629_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599504000))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599635136))))[name = string("layers_20_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1629_cast_fp16 = conv(dilations = dense_output_1629_dilations_1, groups = dense_output_1629_groups_1, pad = dense_output_1629_pad_1, pad_type = dense_output_1629_pad_type_1, strides = dense_output_1629_strides_1, weight = layers_20_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1629_cast_fp16")];
            string sparse_output_1629_pad_type_1 = const()[name = string("sparse_output_1629_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1629_strides_1 = const()[name = string("sparse_output_1629_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1629_pad_1 = const()[name = string("sparse_output_1629_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1629_dilations_1 = const()[name = string("sparse_output_1629_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1629_groups_1 = const()[name = string("sparse_output_1629_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599638400))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599635712))))[name = string("layers_20_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1629_cast_fp16 = conv(dilations = sparse_output_1629_dilations_1, groups = sparse_output_1629_groups_1, pad = sparse_output_1629_pad_1, pad_type = sparse_output_1629_pad_type_1, strides = sparse_output_1629_strides_1, weight = layers_20_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1629_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25908_cast_fp16 = add(x = dense_output_1629_cast_fp16, y = sparse_output_1629_cast_fp16)[name = string("op_25908_cast_fp16")];
            tensor<int32, [4]> var_25909 = const()[name = string("op_25909"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25911 = const()[name = string("op_25911"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25910_cast_fp16 = transpose(perm = var_25909, x = var_25908_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [1, 51, 128]> v_head_645_cast_fp16 = reshape(shape = var_25911, x = var_25910_cast_fp16)[name = string("v_head_645_cast_fp16")];
            string dense_output_1631_pad_type_1 = const()[name = string("dense_output_1631_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1631_strides_1 = const()[name = string("dense_output_1631_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1631_pad_1 = const()[name = string("dense_output_1631_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1631_dilations_1 = const()[name = string("dense_output_1631_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1631_groups_1 = const()[name = string("dense_output_1631_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599654848))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599785984))))[name = string("layers_20_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1631_cast_fp16 = conv(dilations = dense_output_1631_dilations_1, groups = dense_output_1631_groups_1, pad = dense_output_1631_pad_1, pad_type = dense_output_1631_pad_type_1, strides = dense_output_1631_strides_1, weight = layers_20_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1631_cast_fp16")];
            string sparse_output_1631_pad_type_1 = const()[name = string("sparse_output_1631_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1631_strides_1 = const()[name = string("sparse_output_1631_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1631_pad_1 = const()[name = string("sparse_output_1631_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1631_dilations_1 = const()[name = string("sparse_output_1631_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1631_groups_1 = const()[name = string("sparse_output_1631_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599789248))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599786560))))[name = string("layers_20_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1631_cast_fp16 = conv(dilations = sparse_output_1631_dilations_1, groups = sparse_output_1631_groups_1, pad = sparse_output_1631_pad_1, pad_type = sparse_output_1631_pad_type_1, strides = sparse_output_1631_strides_1, weight = layers_20_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1631_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_25927_cast_fp16 = add(x = dense_output_1631_cast_fp16, y = sparse_output_1631_cast_fp16)[name = string("op_25927_cast_fp16")];
            tensor<int32, [4]> var_25928 = const()[name = string("op_25928"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25930 = const()[name = string("op_25930"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_25929_cast_fp16 = transpose(perm = var_25928, x = var_25927_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 101, 128]> p_head_645_cast_fp16 = reshape(shape = var_25930, x = var_25929_cast_fp16)[name = string("p_head_645_cast_fp16")];
            tensor<fp16, [128]> var_25932_to_fp16 = const()[name = string("op_25932_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599805696)))];
            tensor<fp16, [1, 51, 128]> var_25933_cast_fp16 = add(x = q_head_323_cast_fp16, y = var_25932_to_fp16)[name = string("op_25933_cast_fp16")];
            tensor<int32, [1]> q_u_323_axes_1 = const()[name = string("q_u_323_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_323_cast_fp16 = expand_dims(axes = q_u_323_axes_1, x = var_25933_cast_fp16)[name = string("q_u_323_cast_fp16")];
            tensor<fp16, [128]> var_25935_to_fp16 = const()[name = string("op_25935_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599806016)))];
            tensor<fp16, [1, 51, 128]> var_25936_cast_fp16 = add(x = q_head_323_cast_fp16, y = var_25935_to_fp16)[name = string("op_25936_cast_fp16")];
            tensor<int32, [1]> q_v_323_axes_1 = const()[name = string("q_v_323_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_323_cast_fp16 = expand_dims(axes = q_v_323_axes_1, x = var_25936_cast_fp16)[name = string("q_v_323_cast_fp16")];
            tensor<int32, [1]> k_head_647_axes_1 = const()[name = string("k_head_647_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_647_cast_fp16 = expand_dims(axes = k_head_647_axes_1, x = k_head_645_cast_fp16)[name = string("k_head_647_cast_fp16")];
            tensor<int32, [1]> v_head_647_axes_1 = const()[name = string("v_head_647_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_647_cast_fp16 = expand_dims(axes = v_head_647_axes_1, x = v_head_645_cast_fp16)[name = string("v_head_647_cast_fp16")];
            tensor<int32, [1]> p_head_647_axes_1 = const()[name = string("p_head_647_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_647_cast_fp16 = expand_dims(axes = p_head_647_axes_1, x = p_head_645_cast_fp16)[name = string("p_head_647_cast_fp16")];
            bool var_25942_transpose_x_3 = const()[name = string("op_25942_transpose_x_3"), val = bool(false)];
            bool var_25942_transpose_y_3 = const()[name = string("op_25942_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_25942_cast_fp16 = matmul(transpose_x = var_25942_transpose_x_3, transpose_y = var_25942_transpose_y_3, x = q_u_323_cast_fp16, y = k_head_647_cast_fp16)[name = string("op_25942_cast_fp16")];
            fp16 var_25943_to_fp16 = const()[name = string("op_25943_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_323_cast_fp16 = mul(x = var_25942_cast_fp16, y = var_25943_to_fp16)[name = string("scores_content_323_cast_fp16")];
            bool x_1709_transpose_x_3 = const()[name = string("x_1709_transpose_x_3"), val = bool(false)];
            bool x_1709_transpose_y_3 = const()[name = string("x_1709_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1709_cast_fp16 = matmul(transpose_x = x_1709_transpose_x_3, transpose_y = x_1709_transpose_y_3, x = q_v_323_cast_fp16, y = p_head_647_cast_fp16)[name = string("x_1709_cast_fp16")];
            tensor<int32, [8]> x_1711_pad_1 = const()[name = string("x_1711_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1711_mode_1 = const()[name = string("x_1711_mode_1"), val = string("constant")];
            fp16 const_2395_to_fp16 = const()[name = string("const_2395_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1711_cast_fp16 = pad(constant_val = const_2395_to_fp16, mode = x_1711_mode_1, pad = x_1711_pad_1, x = x_1709_cast_fp16)[name = string("x_1711_cast_fp16")];
            tensor<int32, [4]> var_25957 = const()[name = string("op_25957"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1713_cast_fp16 = reshape(shape = var_25957, x = x_1711_cast_fp16)[name = string("x_1713_cast_fp16")];
            tensor<int32, [4]> var_25961_begin_1 = const()[name = string("op_25961_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_25961_end_1 = const()[name = string("op_25961_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_25961_end_mask_1 = const()[name = string("op_25961_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_25961_cast_fp16 = slice_by_index(begin = var_25961_begin_1, end = var_25961_end_1, end_mask = var_25961_end_mask_1, x = x_1713_cast_fp16)[name = string("op_25961_cast_fp16")];
            tensor<int32, [4]> var_25963 = const()[name = string("op_25963"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_25964_cast_fp16 = reshape(shape = var_25963, x = var_25961_cast_fp16)[name = string("op_25964_cast_fp16")];
            tensor<int32, [4]> var_25969_begin_1 = const()[name = string("op_25969_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_25969_end_1 = const()[name = string("op_25969_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_25969_end_mask_1 = const()[name = string("op_25969_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_25969_cast_fp16 = slice_by_index(begin = var_25969_begin_1, end = var_25969_end_1, end_mask = var_25969_end_mask_1, x = var_25964_cast_fp16)[name = string("op_25969_cast_fp16")];
            fp16 var_25970_to_fp16 = const()[name = string("op_25970_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_323_cast_fp16 = mul(x = var_25969_cast_fp16, y = var_25970_to_fp16)[name = string("scores_pos_323_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_323_cast_fp16 = add(x = scores_content_323_cast_fp16, y = scores_pos_323_cast_fp16)[name = string("logits_323_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_25973_cast_fp16 = softmax(axis = var_25577, x = logits_323_cast_fp16)[name = string("op_25973_cast_fp16")];
            bool var_25975_transpose_x_1 = const()[name = string("op_25975_transpose_x_1"), val = bool(false)];
            bool var_25975_transpose_y_1 = const()[name = string("op_25975_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_25975_cast_fp16 = matmul(transpose_x = var_25975_transpose_x_1, transpose_y = var_25975_transpose_y_1, x = var_25973_cast_fp16, y = v_head_647_cast_fp16)[name = string("op_25975_cast_fp16")];
            tensor<int32, [1]> var_25976_axes_1 = const()[name = string("op_25976_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_25976_cast_fp16 = squeeze(axes = var_25976_axes_1, x = var_25975_cast_fp16)[name = string("op_25976_cast_fp16")];
            string dense_output_1633_pad_type_1 = const()[name = string("dense_output_1633_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1633_strides_1 = const()[name = string("dense_output_1633_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1633_pad_1 = const()[name = string("dense_output_1633_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1633_dilations_1 = const()[name = string("dense_output_1633_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1633_groups_1 = const()[name = string("dense_output_1633_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599806336))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599937472))))[name = string("layers_20_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1633_cast_fp16 = conv(dilations = dense_output_1633_dilations_1, groups = dense_output_1633_groups_1, pad = dense_output_1633_pad_1, pad_type = dense_output_1633_pad_type_1, strides = dense_output_1633_strides_1, weight = layers_20_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1633_cast_fp16")];
            string sparse_output_1633_pad_type_1 = const()[name = string("sparse_output_1633_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1633_strides_1 = const()[name = string("sparse_output_1633_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1633_pad_1 = const()[name = string("sparse_output_1633_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1633_dilations_1 = const()[name = string("sparse_output_1633_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1633_groups_1 = const()[name = string("sparse_output_1633_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599940736))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599938048))))[name = string("layers_20_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1633_cast_fp16 = conv(dilations = sparse_output_1633_dilations_1, groups = sparse_output_1633_groups_1, pad = sparse_output_1633_pad_1, pad_type = sparse_output_1633_pad_type_1, strides = sparse_output_1633_strides_1, weight = layers_20_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1633_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_25991_cast_fp16 = add(x = dense_output_1633_cast_fp16, y = sparse_output_1633_cast_fp16)[name = string("op_25991_cast_fp16")];
            tensor<int32, [4]> var_25992 = const()[name = string("op_25992"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_25994 = const()[name = string("op_25994"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_25993_cast_fp16 = transpose(perm = var_25992, x = var_25991_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 51, 128]> q_head_325_cast_fp16 = reshape(shape = var_25994, x = var_25993_cast_fp16)[name = string("q_head_325_cast_fp16")];
            string dense_output_1635_pad_type_1 = const()[name = string("dense_output_1635_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1635_strides_1 = const()[name = string("dense_output_1635_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1635_pad_1 = const()[name = string("dense_output_1635_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1635_dilations_1 = const()[name = string("dense_output_1635_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1635_groups_1 = const()[name = string("dense_output_1635_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599957184))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600088320))))[name = string("layers_20_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1635_cast_fp16 = conv(dilations = dense_output_1635_dilations_1, groups = dense_output_1635_groups_1, pad = dense_output_1635_pad_1, pad_type = dense_output_1635_pad_type_1, strides = dense_output_1635_strides_1, weight = layers_20_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1635_cast_fp16")];
            string sparse_output_1635_pad_type_1 = const()[name = string("sparse_output_1635_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1635_strides_1 = const()[name = string("sparse_output_1635_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1635_pad_1 = const()[name = string("sparse_output_1635_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1635_dilations_1 = const()[name = string("sparse_output_1635_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1635_groups_1 = const()[name = string("sparse_output_1635_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600091584))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600088896))))[name = string("layers_20_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1635_cast_fp16 = conv(dilations = sparse_output_1635_dilations_1, groups = sparse_output_1635_groups_1, pad = sparse_output_1635_pad_1, pad_type = sparse_output_1635_pad_type_1, strides = sparse_output_1635_strides_1, weight = layers_20_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1635_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26010_cast_fp16 = add(x = dense_output_1635_cast_fp16, y = sparse_output_1635_cast_fp16)[name = string("op_26010_cast_fp16")];
            tensor<int32, [4]> var_26011 = const()[name = string("op_26011"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26013 = const()[name = string("op_26013"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26012_cast_fp16 = transpose(perm = var_26011, x = var_26010_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 51, 128]> k_head_649_cast_fp16 = reshape(shape = var_26013, x = var_26012_cast_fp16)[name = string("k_head_649_cast_fp16")];
            string dense_output_1637_pad_type_1 = const()[name = string("dense_output_1637_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1637_strides_1 = const()[name = string("dense_output_1637_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1637_pad_1 = const()[name = string("dense_output_1637_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1637_dilations_1 = const()[name = string("dense_output_1637_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1637_groups_1 = const()[name = string("dense_output_1637_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600108032))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600239168))))[name = string("layers_20_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1637_cast_fp16 = conv(dilations = dense_output_1637_dilations_1, groups = dense_output_1637_groups_1, pad = dense_output_1637_pad_1, pad_type = dense_output_1637_pad_type_1, strides = dense_output_1637_strides_1, weight = layers_20_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1637_cast_fp16")];
            string sparse_output_1637_pad_type_1 = const()[name = string("sparse_output_1637_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1637_strides_1 = const()[name = string("sparse_output_1637_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1637_pad_1 = const()[name = string("sparse_output_1637_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1637_dilations_1 = const()[name = string("sparse_output_1637_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1637_groups_1 = const()[name = string("sparse_output_1637_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600242432))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600239744))))[name = string("layers_20_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1637_cast_fp16 = conv(dilations = sparse_output_1637_dilations_1, groups = sparse_output_1637_groups_1, pad = sparse_output_1637_pad_1, pad_type = sparse_output_1637_pad_type_1, strides = sparse_output_1637_strides_1, weight = layers_20_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1637_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26029_cast_fp16 = add(x = dense_output_1637_cast_fp16, y = sparse_output_1637_cast_fp16)[name = string("op_26029_cast_fp16")];
            tensor<int32, [4]> var_26030 = const()[name = string("op_26030"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26032 = const()[name = string("op_26032"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26031_cast_fp16 = transpose(perm = var_26030, x = var_26029_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 51, 128]> v_head_649_cast_fp16 = reshape(shape = var_26032, x = var_26031_cast_fp16)[name = string("v_head_649_cast_fp16")];
            string dense_output_1639_pad_type_1 = const()[name = string("dense_output_1639_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1639_strides_1 = const()[name = string("dense_output_1639_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1639_pad_1 = const()[name = string("dense_output_1639_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1639_dilations_1 = const()[name = string("dense_output_1639_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1639_groups_1 = const()[name = string("dense_output_1639_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600258880))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600390016))))[name = string("layers_20_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1639_cast_fp16 = conv(dilations = dense_output_1639_dilations_1, groups = dense_output_1639_groups_1, pad = dense_output_1639_pad_1, pad_type = dense_output_1639_pad_type_1, strides = dense_output_1639_strides_1, weight = layers_20_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1639_cast_fp16")];
            string sparse_output_1639_pad_type_1 = const()[name = string("sparse_output_1639_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1639_strides_1 = const()[name = string("sparse_output_1639_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1639_pad_1 = const()[name = string("sparse_output_1639_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1639_dilations_1 = const()[name = string("sparse_output_1639_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1639_groups_1 = const()[name = string("sparse_output_1639_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600393280))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600390592))))[name = string("layers_20_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1639_cast_fp16 = conv(dilations = sparse_output_1639_dilations_1, groups = sparse_output_1639_groups_1, pad = sparse_output_1639_pad_1, pad_type = sparse_output_1639_pad_type_1, strides = sparse_output_1639_strides_1, weight = layers_20_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1639_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_26048_cast_fp16 = add(x = dense_output_1639_cast_fp16, y = sparse_output_1639_cast_fp16)[name = string("op_26048_cast_fp16")];
            tensor<int32, [4]> var_26049 = const()[name = string("op_26049"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26051 = const()[name = string("op_26051"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_26050_cast_fp16 = transpose(perm = var_26049, x = var_26048_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 101, 128]> p_head_649_cast_fp16 = reshape(shape = var_26051, x = var_26050_cast_fp16)[name = string("p_head_649_cast_fp16")];
            tensor<fp16, [128]> var_26053_to_fp16 = const()[name = string("op_26053_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600409728)))];
            tensor<fp16, [1, 51, 128]> var_26054_cast_fp16 = add(x = q_head_325_cast_fp16, y = var_26053_to_fp16)[name = string("op_26054_cast_fp16")];
            tensor<int32, [1]> q_u_325_axes_1 = const()[name = string("q_u_325_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_325_cast_fp16 = expand_dims(axes = q_u_325_axes_1, x = var_26054_cast_fp16)[name = string("q_u_325_cast_fp16")];
            tensor<fp16, [128]> var_26056_to_fp16 = const()[name = string("op_26056_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600410048)))];
            tensor<fp16, [1, 51, 128]> var_26057_cast_fp16 = add(x = q_head_325_cast_fp16, y = var_26056_to_fp16)[name = string("op_26057_cast_fp16")];
            tensor<int32, [1]> q_v_325_axes_1 = const()[name = string("q_v_325_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_325_cast_fp16 = expand_dims(axes = q_v_325_axes_1, x = var_26057_cast_fp16)[name = string("q_v_325_cast_fp16")];
            tensor<int32, [1]> k_head_651_axes_1 = const()[name = string("k_head_651_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_651_cast_fp16 = expand_dims(axes = k_head_651_axes_1, x = k_head_649_cast_fp16)[name = string("k_head_651_cast_fp16")];
            tensor<int32, [1]> v_head_651_axes_1 = const()[name = string("v_head_651_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_651_cast_fp16 = expand_dims(axes = v_head_651_axes_1, x = v_head_649_cast_fp16)[name = string("v_head_651_cast_fp16")];
            tensor<int32, [1]> p_head_651_axes_1 = const()[name = string("p_head_651_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_651_cast_fp16 = expand_dims(axes = p_head_651_axes_1, x = p_head_649_cast_fp16)[name = string("p_head_651_cast_fp16")];
            bool var_26063_transpose_x_3 = const()[name = string("op_26063_transpose_x_3"), val = bool(false)];
            bool var_26063_transpose_y_3 = const()[name = string("op_26063_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_26063_cast_fp16 = matmul(transpose_x = var_26063_transpose_x_3, transpose_y = var_26063_transpose_y_3, x = q_u_325_cast_fp16, y = k_head_651_cast_fp16)[name = string("op_26063_cast_fp16")];
            fp16 var_26064_to_fp16 = const()[name = string("op_26064_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_325_cast_fp16 = mul(x = var_26063_cast_fp16, y = var_26064_to_fp16)[name = string("scores_content_325_cast_fp16")];
            bool x_1717_transpose_x_3 = const()[name = string("x_1717_transpose_x_3"), val = bool(false)];
            bool x_1717_transpose_y_3 = const()[name = string("x_1717_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1717_cast_fp16 = matmul(transpose_x = x_1717_transpose_x_3, transpose_y = x_1717_transpose_y_3, x = q_v_325_cast_fp16, y = p_head_651_cast_fp16)[name = string("x_1717_cast_fp16")];
            tensor<int32, [8]> x_1719_pad_1 = const()[name = string("x_1719_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1719_mode_1 = const()[name = string("x_1719_mode_1"), val = string("constant")];
            fp16 const_2401_to_fp16 = const()[name = string("const_2401_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1719_cast_fp16 = pad(constant_val = const_2401_to_fp16, mode = x_1719_mode_1, pad = x_1719_pad_1, x = x_1717_cast_fp16)[name = string("x_1719_cast_fp16")];
            tensor<int32, [4]> var_26078 = const()[name = string("op_26078"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1721_cast_fp16 = reshape(shape = var_26078, x = x_1719_cast_fp16)[name = string("x_1721_cast_fp16")];
            tensor<int32, [4]> var_26082_begin_1 = const()[name = string("op_26082_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_26082_end_1 = const()[name = string("op_26082_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_26082_end_mask_1 = const()[name = string("op_26082_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_26082_cast_fp16 = slice_by_index(begin = var_26082_begin_1, end = var_26082_end_1, end_mask = var_26082_end_mask_1, x = x_1721_cast_fp16)[name = string("op_26082_cast_fp16")];
            tensor<int32, [4]> var_26084 = const()[name = string("op_26084"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_26085_cast_fp16 = reshape(shape = var_26084, x = var_26082_cast_fp16)[name = string("op_26085_cast_fp16")];
            tensor<int32, [4]> var_26090_begin_1 = const()[name = string("op_26090_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_26090_end_1 = const()[name = string("op_26090_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_26090_end_mask_1 = const()[name = string("op_26090_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_26090_cast_fp16 = slice_by_index(begin = var_26090_begin_1, end = var_26090_end_1, end_mask = var_26090_end_mask_1, x = var_26085_cast_fp16)[name = string("op_26090_cast_fp16")];
            fp16 var_26091_to_fp16 = const()[name = string("op_26091_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_325_cast_fp16 = mul(x = var_26090_cast_fp16, y = var_26091_to_fp16)[name = string("scores_pos_325_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_325_cast_fp16 = add(x = scores_content_325_cast_fp16, y = scores_pos_325_cast_fp16)[name = string("logits_325_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_26094_cast_fp16 = softmax(axis = var_25577, x = logits_325_cast_fp16)[name = string("op_26094_cast_fp16")];
            bool var_26096_transpose_x_1 = const()[name = string("op_26096_transpose_x_1"), val = bool(false)];
            bool var_26096_transpose_y_1 = const()[name = string("op_26096_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_26096_cast_fp16 = matmul(transpose_x = var_26096_transpose_x_1, transpose_y = var_26096_transpose_y_1, x = var_26094_cast_fp16, y = v_head_651_cast_fp16)[name = string("op_26096_cast_fp16")];
            tensor<int32, [1]> var_26097_axes_1 = const()[name = string("op_26097_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_26097_cast_fp16 = squeeze(axes = var_26097_axes_1, x = var_26096_cast_fp16)[name = string("op_26097_cast_fp16")];
            string dense_output_1641_pad_type_1 = const()[name = string("dense_output_1641_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1641_strides_1 = const()[name = string("dense_output_1641_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1641_pad_1 = const()[name = string("dense_output_1641_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1641_dilations_1 = const()[name = string("dense_output_1641_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1641_groups_1 = const()[name = string("dense_output_1641_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600410368))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600541504))))[name = string("layers_20_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1641_cast_fp16 = conv(dilations = dense_output_1641_dilations_1, groups = dense_output_1641_groups_1, pad = dense_output_1641_pad_1, pad_type = dense_output_1641_pad_type_1, strides = dense_output_1641_strides_1, weight = layers_20_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1641_cast_fp16")];
            string sparse_output_1641_pad_type_1 = const()[name = string("sparse_output_1641_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1641_strides_1 = const()[name = string("sparse_output_1641_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1641_pad_1 = const()[name = string("sparse_output_1641_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1641_dilations_1 = const()[name = string("sparse_output_1641_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1641_groups_1 = const()[name = string("sparse_output_1641_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600544768))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600542080))))[name = string("layers_20_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1641_cast_fp16 = conv(dilations = sparse_output_1641_dilations_1, groups = sparse_output_1641_groups_1, pad = sparse_output_1641_pad_1, pad_type = sparse_output_1641_pad_type_1, strides = sparse_output_1641_strides_1, weight = layers_20_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1641_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26112_cast_fp16 = add(x = dense_output_1641_cast_fp16, y = sparse_output_1641_cast_fp16)[name = string("op_26112_cast_fp16")];
            tensor<int32, [4]> var_26113 = const()[name = string("op_26113"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26115 = const()[name = string("op_26115"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26114_cast_fp16 = transpose(perm = var_26113, x = var_26112_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 51, 128]> q_head_327_cast_fp16 = reshape(shape = var_26115, x = var_26114_cast_fp16)[name = string("q_head_327_cast_fp16")];
            string dense_output_1643_pad_type_1 = const()[name = string("dense_output_1643_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1643_strides_1 = const()[name = string("dense_output_1643_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1643_pad_1 = const()[name = string("dense_output_1643_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1643_dilations_1 = const()[name = string("dense_output_1643_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1643_groups_1 = const()[name = string("dense_output_1643_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600561216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600692352))))[name = string("layers_20_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1643_cast_fp16 = conv(dilations = dense_output_1643_dilations_1, groups = dense_output_1643_groups_1, pad = dense_output_1643_pad_1, pad_type = dense_output_1643_pad_type_1, strides = dense_output_1643_strides_1, weight = layers_20_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1643_cast_fp16")];
            string sparse_output_1643_pad_type_1 = const()[name = string("sparse_output_1643_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1643_strides_1 = const()[name = string("sparse_output_1643_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1643_pad_1 = const()[name = string("sparse_output_1643_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1643_dilations_1 = const()[name = string("sparse_output_1643_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1643_groups_1 = const()[name = string("sparse_output_1643_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600695616))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600692928))))[name = string("layers_20_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1643_cast_fp16 = conv(dilations = sparse_output_1643_dilations_1, groups = sparse_output_1643_groups_1, pad = sparse_output_1643_pad_1, pad_type = sparse_output_1643_pad_type_1, strides = sparse_output_1643_strides_1, weight = layers_20_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1643_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26131_cast_fp16 = add(x = dense_output_1643_cast_fp16, y = sparse_output_1643_cast_fp16)[name = string("op_26131_cast_fp16")];
            tensor<int32, [4]> var_26132 = const()[name = string("op_26132"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26134 = const()[name = string("op_26134"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26133_cast_fp16 = transpose(perm = var_26132, x = var_26131_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 51, 128]> k_head_653_cast_fp16 = reshape(shape = var_26134, x = var_26133_cast_fp16)[name = string("k_head_653_cast_fp16")];
            string dense_output_1645_pad_type_1 = const()[name = string("dense_output_1645_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1645_strides_1 = const()[name = string("dense_output_1645_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1645_pad_1 = const()[name = string("dense_output_1645_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1645_dilations_1 = const()[name = string("dense_output_1645_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1645_groups_1 = const()[name = string("dense_output_1645_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600712064))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600843200))))[name = string("layers_20_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1645_cast_fp16 = conv(dilations = dense_output_1645_dilations_1, groups = dense_output_1645_groups_1, pad = dense_output_1645_pad_1, pad_type = dense_output_1645_pad_type_1, strides = dense_output_1645_strides_1, weight = layers_20_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1645_cast_fp16")];
            string sparse_output_1645_pad_type_1 = const()[name = string("sparse_output_1645_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1645_strides_1 = const()[name = string("sparse_output_1645_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1645_pad_1 = const()[name = string("sparse_output_1645_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1645_dilations_1 = const()[name = string("sparse_output_1645_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1645_groups_1 = const()[name = string("sparse_output_1645_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600846464))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600843776))))[name = string("layers_20_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1645_cast_fp16 = conv(dilations = sparse_output_1645_dilations_1, groups = sparse_output_1645_groups_1, pad = sparse_output_1645_pad_1, pad_type = sparse_output_1645_pad_type_1, strides = sparse_output_1645_strides_1, weight = layers_20_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1645_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26150_cast_fp16 = add(x = dense_output_1645_cast_fp16, y = sparse_output_1645_cast_fp16)[name = string("op_26150_cast_fp16")];
            tensor<int32, [4]> var_26151 = const()[name = string("op_26151"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26153 = const()[name = string("op_26153"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26152_cast_fp16 = transpose(perm = var_26151, x = var_26150_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [1, 51, 128]> v_head_653_cast_fp16 = reshape(shape = var_26153, x = var_26152_cast_fp16)[name = string("v_head_653_cast_fp16")];
            string dense_output_1647_pad_type_1 = const()[name = string("dense_output_1647_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1647_strides_1 = const()[name = string("dense_output_1647_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1647_pad_1 = const()[name = string("dense_output_1647_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1647_dilations_1 = const()[name = string("dense_output_1647_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1647_groups_1 = const()[name = string("dense_output_1647_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600862912))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600994048))))[name = string("layers_20_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1647_cast_fp16 = conv(dilations = dense_output_1647_dilations_1, groups = dense_output_1647_groups_1, pad = dense_output_1647_pad_1, pad_type = dense_output_1647_pad_type_1, strides = dense_output_1647_strides_1, weight = layers_20_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1647_cast_fp16")];
            string sparse_output_1647_pad_type_1 = const()[name = string("sparse_output_1647_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1647_strides_1 = const()[name = string("sparse_output_1647_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1647_pad_1 = const()[name = string("sparse_output_1647_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1647_dilations_1 = const()[name = string("sparse_output_1647_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1647_groups_1 = const()[name = string("sparse_output_1647_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600997312))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600994624))))[name = string("layers_20_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1647_cast_fp16 = conv(dilations = sparse_output_1647_dilations_1, groups = sparse_output_1647_groups_1, pad = sparse_output_1647_pad_1, pad_type = sparse_output_1647_pad_type_1, strides = sparse_output_1647_strides_1, weight = layers_20_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1647_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_26169_cast_fp16 = add(x = dense_output_1647_cast_fp16, y = sparse_output_1647_cast_fp16)[name = string("op_26169_cast_fp16")];
            tensor<int32, [4]> var_26170 = const()[name = string("op_26170"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26172 = const()[name = string("op_26172"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_26171_cast_fp16 = transpose(perm = var_26170, x = var_26169_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [1, 101, 128]> p_head_653_cast_fp16 = reshape(shape = var_26172, x = var_26171_cast_fp16)[name = string("p_head_653_cast_fp16")];
            tensor<fp16, [128]> var_26174_to_fp16 = const()[name = string("op_26174_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601013760)))];
            tensor<fp16, [1, 51, 128]> var_26175_cast_fp16 = add(x = q_head_327_cast_fp16, y = var_26174_to_fp16)[name = string("op_26175_cast_fp16")];
            tensor<int32, [1]> q_u_327_axes_1 = const()[name = string("q_u_327_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_327_cast_fp16 = expand_dims(axes = q_u_327_axes_1, x = var_26175_cast_fp16)[name = string("q_u_327_cast_fp16")];
            tensor<fp16, [128]> var_26177_to_fp16 = const()[name = string("op_26177_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601014080)))];
            tensor<fp16, [1, 51, 128]> var_26178_cast_fp16 = add(x = q_head_327_cast_fp16, y = var_26177_to_fp16)[name = string("op_26178_cast_fp16")];
            tensor<int32, [1]> q_v_327_axes_1 = const()[name = string("q_v_327_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_327_cast_fp16 = expand_dims(axes = q_v_327_axes_1, x = var_26178_cast_fp16)[name = string("q_v_327_cast_fp16")];
            tensor<int32, [1]> k_head_655_axes_1 = const()[name = string("k_head_655_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_655_cast_fp16 = expand_dims(axes = k_head_655_axes_1, x = k_head_653_cast_fp16)[name = string("k_head_655_cast_fp16")];
            tensor<int32, [1]> v_head_655_axes_1 = const()[name = string("v_head_655_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_655_cast_fp16 = expand_dims(axes = v_head_655_axes_1, x = v_head_653_cast_fp16)[name = string("v_head_655_cast_fp16")];
            tensor<int32, [1]> p_head_655_axes_1 = const()[name = string("p_head_655_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_655_cast_fp16 = expand_dims(axes = p_head_655_axes_1, x = p_head_653_cast_fp16)[name = string("p_head_655_cast_fp16")];
            bool var_26184_transpose_x_3 = const()[name = string("op_26184_transpose_x_3"), val = bool(false)];
            bool var_26184_transpose_y_3 = const()[name = string("op_26184_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_26184_cast_fp16 = matmul(transpose_x = var_26184_transpose_x_3, transpose_y = var_26184_transpose_y_3, x = q_u_327_cast_fp16, y = k_head_655_cast_fp16)[name = string("op_26184_cast_fp16")];
            fp16 var_26185_to_fp16 = const()[name = string("op_26185_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_327_cast_fp16 = mul(x = var_26184_cast_fp16, y = var_26185_to_fp16)[name = string("scores_content_327_cast_fp16")];
            bool x_1725_transpose_x_3 = const()[name = string("x_1725_transpose_x_3"), val = bool(false)];
            bool x_1725_transpose_y_3 = const()[name = string("x_1725_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1725_cast_fp16 = matmul(transpose_x = x_1725_transpose_x_3, transpose_y = x_1725_transpose_y_3, x = q_v_327_cast_fp16, y = p_head_655_cast_fp16)[name = string("x_1725_cast_fp16")];
            tensor<int32, [8]> x_1727_pad_1 = const()[name = string("x_1727_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1727_mode_1 = const()[name = string("x_1727_mode_1"), val = string("constant")];
            fp16 const_2407_to_fp16 = const()[name = string("const_2407_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1727_cast_fp16 = pad(constant_val = const_2407_to_fp16, mode = x_1727_mode_1, pad = x_1727_pad_1, x = x_1725_cast_fp16)[name = string("x_1727_cast_fp16")];
            tensor<int32, [4]> var_26199 = const()[name = string("op_26199"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1729_cast_fp16 = reshape(shape = var_26199, x = x_1727_cast_fp16)[name = string("x_1729_cast_fp16")];
            tensor<int32, [4]> var_26203_begin_1 = const()[name = string("op_26203_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_26203_end_1 = const()[name = string("op_26203_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_26203_end_mask_1 = const()[name = string("op_26203_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_26203_cast_fp16 = slice_by_index(begin = var_26203_begin_1, end = var_26203_end_1, end_mask = var_26203_end_mask_1, x = x_1729_cast_fp16)[name = string("op_26203_cast_fp16")];
            tensor<int32, [4]> var_26205 = const()[name = string("op_26205"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_26206_cast_fp16 = reshape(shape = var_26205, x = var_26203_cast_fp16)[name = string("op_26206_cast_fp16")];
            tensor<int32, [4]> var_26211_begin_1 = const()[name = string("op_26211_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_26211_end_1 = const()[name = string("op_26211_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_26211_end_mask_1 = const()[name = string("op_26211_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_26211_cast_fp16 = slice_by_index(begin = var_26211_begin_1, end = var_26211_end_1, end_mask = var_26211_end_mask_1, x = var_26206_cast_fp16)[name = string("op_26211_cast_fp16")];
            fp16 var_26212_to_fp16 = const()[name = string("op_26212_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_327_cast_fp16 = mul(x = var_26211_cast_fp16, y = var_26212_to_fp16)[name = string("scores_pos_327_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_327_cast_fp16 = add(x = scores_content_327_cast_fp16, y = scores_pos_327_cast_fp16)[name = string("logits_327_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_26215_cast_fp16 = softmax(axis = var_25577, x = logits_327_cast_fp16)[name = string("op_26215_cast_fp16")];
            bool var_26217_transpose_x_1 = const()[name = string("op_26217_transpose_x_1"), val = bool(false)];
            bool var_26217_transpose_y_1 = const()[name = string("op_26217_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_26217_cast_fp16 = matmul(transpose_x = var_26217_transpose_x_1, transpose_y = var_26217_transpose_y_1, x = var_26215_cast_fp16, y = v_head_655_cast_fp16)[name = string("op_26217_cast_fp16")];
            tensor<int32, [1]> var_26218_axes_1 = const()[name = string("op_26218_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_26218_cast_fp16 = squeeze(axes = var_26218_axes_1, x = var_26217_cast_fp16)[name = string("op_26218_cast_fp16")];
            string dense_output_1649_pad_type_1 = const()[name = string("dense_output_1649_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1649_strides_1 = const()[name = string("dense_output_1649_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1649_pad_1 = const()[name = string("dense_output_1649_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1649_dilations_1 = const()[name = string("dense_output_1649_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1649_groups_1 = const()[name = string("dense_output_1649_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601014400))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601145536))))[name = string("layers_20_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1649_cast_fp16 = conv(dilations = dense_output_1649_dilations_1, groups = dense_output_1649_groups_1, pad = dense_output_1649_pad_1, pad_type = dense_output_1649_pad_type_1, strides = dense_output_1649_strides_1, weight = layers_20_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1649_cast_fp16")];
            string sparse_output_1649_pad_type_1 = const()[name = string("sparse_output_1649_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1649_strides_1 = const()[name = string("sparse_output_1649_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1649_pad_1 = const()[name = string("sparse_output_1649_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1649_dilations_1 = const()[name = string("sparse_output_1649_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1649_groups_1 = const()[name = string("sparse_output_1649_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601148800))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601146112))))[name = string("layers_20_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1649_cast_fp16 = conv(dilations = sparse_output_1649_dilations_1, groups = sparse_output_1649_groups_1, pad = sparse_output_1649_pad_1, pad_type = sparse_output_1649_pad_type_1, strides = sparse_output_1649_strides_1, weight = layers_20_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1649_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26233_cast_fp16 = add(x = dense_output_1649_cast_fp16, y = sparse_output_1649_cast_fp16)[name = string("op_26233_cast_fp16")];
            tensor<int32, [4]> var_26234 = const()[name = string("op_26234"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26236 = const()[name = string("op_26236"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26235_cast_fp16 = transpose(perm = var_26234, x = var_26233_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 51, 128]> q_head_329_cast_fp16 = reshape(shape = var_26236, x = var_26235_cast_fp16)[name = string("q_head_329_cast_fp16")];
            string dense_output_1651_pad_type_1 = const()[name = string("dense_output_1651_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1651_strides_1 = const()[name = string("dense_output_1651_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1651_pad_1 = const()[name = string("dense_output_1651_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1651_dilations_1 = const()[name = string("dense_output_1651_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1651_groups_1 = const()[name = string("dense_output_1651_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601165248))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601296384))))[name = string("layers_20_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1651_cast_fp16 = conv(dilations = dense_output_1651_dilations_1, groups = dense_output_1651_groups_1, pad = dense_output_1651_pad_1, pad_type = dense_output_1651_pad_type_1, strides = dense_output_1651_strides_1, weight = layers_20_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1651_cast_fp16")];
            string sparse_output_1651_pad_type_1 = const()[name = string("sparse_output_1651_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1651_strides_1 = const()[name = string("sparse_output_1651_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1651_pad_1 = const()[name = string("sparse_output_1651_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1651_dilations_1 = const()[name = string("sparse_output_1651_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1651_groups_1 = const()[name = string("sparse_output_1651_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601299648))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601296960))))[name = string("layers_20_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1651_cast_fp16 = conv(dilations = sparse_output_1651_dilations_1, groups = sparse_output_1651_groups_1, pad = sparse_output_1651_pad_1, pad_type = sparse_output_1651_pad_type_1, strides = sparse_output_1651_strides_1, weight = layers_20_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1651_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26252_cast_fp16 = add(x = dense_output_1651_cast_fp16, y = sparse_output_1651_cast_fp16)[name = string("op_26252_cast_fp16")];
            tensor<int32, [4]> var_26253 = const()[name = string("op_26253"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26255 = const()[name = string("op_26255"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26254_cast_fp16 = transpose(perm = var_26253, x = var_26252_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [1, 51, 128]> k_head_657_cast_fp16 = reshape(shape = var_26255, x = var_26254_cast_fp16)[name = string("k_head_657_cast_fp16")];
            string dense_output_1653_pad_type_1 = const()[name = string("dense_output_1653_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1653_strides_1 = const()[name = string("dense_output_1653_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1653_pad_1 = const()[name = string("dense_output_1653_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1653_dilations_1 = const()[name = string("dense_output_1653_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1653_groups_1 = const()[name = string("dense_output_1653_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601316096))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601447232))))[name = string("layers_20_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1653_cast_fp16 = conv(dilations = dense_output_1653_dilations_1, groups = dense_output_1653_groups_1, pad = dense_output_1653_pad_1, pad_type = dense_output_1653_pad_type_1, strides = dense_output_1653_strides_1, weight = layers_20_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1653_cast_fp16")];
            string sparse_output_1653_pad_type_1 = const()[name = string("sparse_output_1653_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1653_strides_1 = const()[name = string("sparse_output_1653_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1653_pad_1 = const()[name = string("sparse_output_1653_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1653_dilations_1 = const()[name = string("sparse_output_1653_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1653_groups_1 = const()[name = string("sparse_output_1653_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601450496))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601447808))))[name = string("layers_20_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1653_cast_fp16 = conv(dilations = sparse_output_1653_dilations_1, groups = sparse_output_1653_groups_1, pad = sparse_output_1653_pad_1, pad_type = sparse_output_1653_pad_type_1, strides = sparse_output_1653_strides_1, weight = layers_20_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1653_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26271_cast_fp16 = add(x = dense_output_1653_cast_fp16, y = sparse_output_1653_cast_fp16)[name = string("op_26271_cast_fp16")];
            tensor<int32, [4]> var_26272 = const()[name = string("op_26272"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26274 = const()[name = string("op_26274"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26273_cast_fp16 = transpose(perm = var_26272, x = var_26271_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 51, 128]> v_head_657_cast_fp16 = reshape(shape = var_26274, x = var_26273_cast_fp16)[name = string("v_head_657_cast_fp16")];
            string dense_output_1655_pad_type_1 = const()[name = string("dense_output_1655_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1655_strides_1 = const()[name = string("dense_output_1655_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1655_pad_1 = const()[name = string("dense_output_1655_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1655_dilations_1 = const()[name = string("dense_output_1655_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1655_groups_1 = const()[name = string("dense_output_1655_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601466944))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601598080))))[name = string("layers_20_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1655_cast_fp16 = conv(dilations = dense_output_1655_dilations_1, groups = dense_output_1655_groups_1, pad = dense_output_1655_pad_1, pad_type = dense_output_1655_pad_type_1, strides = dense_output_1655_strides_1, weight = layers_20_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1655_cast_fp16")];
            string sparse_output_1655_pad_type_1 = const()[name = string("sparse_output_1655_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1655_strides_1 = const()[name = string("sparse_output_1655_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1655_pad_1 = const()[name = string("sparse_output_1655_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1655_dilations_1 = const()[name = string("sparse_output_1655_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1655_groups_1 = const()[name = string("sparse_output_1655_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601601344))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601598656))))[name = string("layers_20_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1655_cast_fp16 = conv(dilations = sparse_output_1655_dilations_1, groups = sparse_output_1655_groups_1, pad = sparse_output_1655_pad_1, pad_type = sparse_output_1655_pad_type_1, strides = sparse_output_1655_strides_1, weight = layers_20_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1655_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_26290_cast_fp16 = add(x = dense_output_1655_cast_fp16, y = sparse_output_1655_cast_fp16)[name = string("op_26290_cast_fp16")];
            tensor<int32, [4]> var_26291 = const()[name = string("op_26291"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26293 = const()[name = string("op_26293"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_26292_cast_fp16 = transpose(perm = var_26291, x = var_26290_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 101, 128]> p_head_657_cast_fp16 = reshape(shape = var_26293, x = var_26292_cast_fp16)[name = string("p_head_657_cast_fp16")];
            tensor<fp16, [128]> var_26295_to_fp16 = const()[name = string("op_26295_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601617792)))];
            tensor<fp16, [1, 51, 128]> var_26296_cast_fp16 = add(x = q_head_329_cast_fp16, y = var_26295_to_fp16)[name = string("op_26296_cast_fp16")];
            tensor<int32, [1]> q_u_329_axes_1 = const()[name = string("q_u_329_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_329_cast_fp16 = expand_dims(axes = q_u_329_axes_1, x = var_26296_cast_fp16)[name = string("q_u_329_cast_fp16")];
            tensor<fp16, [128]> var_26298_to_fp16 = const()[name = string("op_26298_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601618112)))];
            tensor<fp16, [1, 51, 128]> var_26299_cast_fp16 = add(x = q_head_329_cast_fp16, y = var_26298_to_fp16)[name = string("op_26299_cast_fp16")];
            tensor<int32, [1]> q_v_329_axes_1 = const()[name = string("q_v_329_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_329_cast_fp16 = expand_dims(axes = q_v_329_axes_1, x = var_26299_cast_fp16)[name = string("q_v_329_cast_fp16")];
            tensor<int32, [1]> k_head_659_axes_1 = const()[name = string("k_head_659_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_659_cast_fp16 = expand_dims(axes = k_head_659_axes_1, x = k_head_657_cast_fp16)[name = string("k_head_659_cast_fp16")];
            tensor<int32, [1]> v_head_659_axes_1 = const()[name = string("v_head_659_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_659_cast_fp16 = expand_dims(axes = v_head_659_axes_1, x = v_head_657_cast_fp16)[name = string("v_head_659_cast_fp16")];
            tensor<int32, [1]> p_head_659_axes_1 = const()[name = string("p_head_659_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_659_cast_fp16 = expand_dims(axes = p_head_659_axes_1, x = p_head_657_cast_fp16)[name = string("p_head_659_cast_fp16")];
            bool var_26305_transpose_x_3 = const()[name = string("op_26305_transpose_x_3"), val = bool(false)];
            bool var_26305_transpose_y_3 = const()[name = string("op_26305_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_26305_cast_fp16 = matmul(transpose_x = var_26305_transpose_x_3, transpose_y = var_26305_transpose_y_3, x = q_u_329_cast_fp16, y = k_head_659_cast_fp16)[name = string("op_26305_cast_fp16")];
            fp16 var_26306_to_fp16 = const()[name = string("op_26306_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_329_cast_fp16 = mul(x = var_26305_cast_fp16, y = var_26306_to_fp16)[name = string("scores_content_329_cast_fp16")];
            bool x_1733_transpose_x_3 = const()[name = string("x_1733_transpose_x_3"), val = bool(false)];
            bool x_1733_transpose_y_3 = const()[name = string("x_1733_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1733_cast_fp16 = matmul(transpose_x = x_1733_transpose_x_3, transpose_y = x_1733_transpose_y_3, x = q_v_329_cast_fp16, y = p_head_659_cast_fp16)[name = string("x_1733_cast_fp16")];
            tensor<int32, [8]> x_1735_pad_1 = const()[name = string("x_1735_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1735_mode_1 = const()[name = string("x_1735_mode_1"), val = string("constant")];
            fp16 const_2413_to_fp16 = const()[name = string("const_2413_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1735_cast_fp16 = pad(constant_val = const_2413_to_fp16, mode = x_1735_mode_1, pad = x_1735_pad_1, x = x_1733_cast_fp16)[name = string("x_1735_cast_fp16")];
            tensor<int32, [4]> var_26320 = const()[name = string("op_26320"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1737_cast_fp16 = reshape(shape = var_26320, x = x_1735_cast_fp16)[name = string("x_1737_cast_fp16")];
            tensor<int32, [4]> var_26324_begin_1 = const()[name = string("op_26324_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_26324_end_1 = const()[name = string("op_26324_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_26324_end_mask_1 = const()[name = string("op_26324_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_26324_cast_fp16 = slice_by_index(begin = var_26324_begin_1, end = var_26324_end_1, end_mask = var_26324_end_mask_1, x = x_1737_cast_fp16)[name = string("op_26324_cast_fp16")];
            tensor<int32, [4]> var_26326 = const()[name = string("op_26326"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_26327_cast_fp16 = reshape(shape = var_26326, x = var_26324_cast_fp16)[name = string("op_26327_cast_fp16")];
            tensor<int32, [4]> var_26332_begin_1 = const()[name = string("op_26332_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_26332_end_1 = const()[name = string("op_26332_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_26332_end_mask_1 = const()[name = string("op_26332_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_26332_cast_fp16 = slice_by_index(begin = var_26332_begin_1, end = var_26332_end_1, end_mask = var_26332_end_mask_1, x = var_26327_cast_fp16)[name = string("op_26332_cast_fp16")];
            fp16 var_26333_to_fp16 = const()[name = string("op_26333_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_329_cast_fp16 = mul(x = var_26332_cast_fp16, y = var_26333_to_fp16)[name = string("scores_pos_329_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_329_cast_fp16 = add(x = scores_content_329_cast_fp16, y = scores_pos_329_cast_fp16)[name = string("logits_329_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_26336_cast_fp16 = softmax(axis = var_25577, x = logits_329_cast_fp16)[name = string("op_26336_cast_fp16")];
            bool var_26338_transpose_x_1 = const()[name = string("op_26338_transpose_x_1"), val = bool(false)];
            bool var_26338_transpose_y_1 = const()[name = string("op_26338_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_26338_cast_fp16 = matmul(transpose_x = var_26338_transpose_x_1, transpose_y = var_26338_transpose_y_1, x = var_26336_cast_fp16, y = v_head_659_cast_fp16)[name = string("op_26338_cast_fp16")];
            tensor<int32, [1]> var_26339_axes_1 = const()[name = string("op_26339_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_26339_cast_fp16 = squeeze(axes = var_26339_axes_1, x = var_26338_cast_fp16)[name = string("op_26339_cast_fp16")];
            string dense_output_1657_pad_type_1 = const()[name = string("dense_output_1657_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1657_strides_1 = const()[name = string("dense_output_1657_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1657_pad_1 = const()[name = string("dense_output_1657_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1657_dilations_1 = const()[name = string("dense_output_1657_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1657_groups_1 = const()[name = string("dense_output_1657_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601618432))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601749568))))[name = string("layers_20_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1657_cast_fp16 = conv(dilations = dense_output_1657_dilations_1, groups = dense_output_1657_groups_1, pad = dense_output_1657_pad_1, pad_type = dense_output_1657_pad_type_1, strides = dense_output_1657_strides_1, weight = layers_20_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1657_cast_fp16")];
            string sparse_output_1657_pad_type_1 = const()[name = string("sparse_output_1657_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1657_strides_1 = const()[name = string("sparse_output_1657_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1657_pad_1 = const()[name = string("sparse_output_1657_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1657_dilations_1 = const()[name = string("sparse_output_1657_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1657_groups_1 = const()[name = string("sparse_output_1657_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601752832))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601750144))))[name = string("layers_20_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1657_cast_fp16 = conv(dilations = sparse_output_1657_dilations_1, groups = sparse_output_1657_groups_1, pad = sparse_output_1657_pad_1, pad_type = sparse_output_1657_pad_type_1, strides = sparse_output_1657_strides_1, weight = layers_20_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1657_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26354_cast_fp16 = add(x = dense_output_1657_cast_fp16, y = sparse_output_1657_cast_fp16)[name = string("op_26354_cast_fp16")];
            tensor<int32, [4]> var_26355 = const()[name = string("op_26355"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26357 = const()[name = string("op_26357"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26356_cast_fp16 = transpose(perm = var_26355, x = var_26354_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 51, 128]> q_head_331_cast_fp16 = reshape(shape = var_26357, x = var_26356_cast_fp16)[name = string("q_head_331_cast_fp16")];
            string dense_output_1659_pad_type_1 = const()[name = string("dense_output_1659_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1659_strides_1 = const()[name = string("dense_output_1659_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1659_pad_1 = const()[name = string("dense_output_1659_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1659_dilations_1 = const()[name = string("dense_output_1659_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1659_groups_1 = const()[name = string("dense_output_1659_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601769280))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601900416))))[name = string("layers_20_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1659_cast_fp16 = conv(dilations = dense_output_1659_dilations_1, groups = dense_output_1659_groups_1, pad = dense_output_1659_pad_1, pad_type = dense_output_1659_pad_type_1, strides = dense_output_1659_strides_1, weight = layers_20_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1659_cast_fp16")];
            string sparse_output_1659_pad_type_1 = const()[name = string("sparse_output_1659_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1659_strides_1 = const()[name = string("sparse_output_1659_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1659_pad_1 = const()[name = string("sparse_output_1659_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1659_dilations_1 = const()[name = string("sparse_output_1659_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1659_groups_1 = const()[name = string("sparse_output_1659_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601903680))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601900992))))[name = string("layers_20_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1659_cast_fp16 = conv(dilations = sparse_output_1659_dilations_1, groups = sparse_output_1659_groups_1, pad = sparse_output_1659_pad_1, pad_type = sparse_output_1659_pad_type_1, strides = sparse_output_1659_strides_1, weight = layers_20_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1659_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26373_cast_fp16 = add(x = dense_output_1659_cast_fp16, y = sparse_output_1659_cast_fp16)[name = string("op_26373_cast_fp16")];
            tensor<int32, [4]> var_26374 = const()[name = string("op_26374"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26376 = const()[name = string("op_26376"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26375_cast_fp16 = transpose(perm = var_26374, x = var_26373_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 51, 128]> k_head_661_cast_fp16 = reshape(shape = var_26376, x = var_26375_cast_fp16)[name = string("k_head_661_cast_fp16")];
            string dense_output_1661_pad_type_1 = const()[name = string("dense_output_1661_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1661_strides_1 = const()[name = string("dense_output_1661_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1661_pad_1 = const()[name = string("dense_output_1661_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1661_dilations_1 = const()[name = string("dense_output_1661_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1661_groups_1 = const()[name = string("dense_output_1661_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601920128))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602051264))))[name = string("layers_20_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1661_cast_fp16 = conv(dilations = dense_output_1661_dilations_1, groups = dense_output_1661_groups_1, pad = dense_output_1661_pad_1, pad_type = dense_output_1661_pad_type_1, strides = dense_output_1661_strides_1, weight = layers_20_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1661_cast_fp16")];
            string sparse_output_1661_pad_type_1 = const()[name = string("sparse_output_1661_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1661_strides_1 = const()[name = string("sparse_output_1661_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1661_pad_1 = const()[name = string("sparse_output_1661_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1661_dilations_1 = const()[name = string("sparse_output_1661_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1661_groups_1 = const()[name = string("sparse_output_1661_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602054528))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602051840))))[name = string("layers_20_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1661_cast_fp16 = conv(dilations = sparse_output_1661_dilations_1, groups = sparse_output_1661_groups_1, pad = sparse_output_1661_pad_1, pad_type = sparse_output_1661_pad_type_1, strides = sparse_output_1661_strides_1, weight = layers_20_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1661_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26392_cast_fp16 = add(x = dense_output_1661_cast_fp16, y = sparse_output_1661_cast_fp16)[name = string("op_26392_cast_fp16")];
            tensor<int32, [4]> var_26393 = const()[name = string("op_26393"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26395 = const()[name = string("op_26395"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26394_cast_fp16 = transpose(perm = var_26393, x = var_26392_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 51, 128]> v_head_661_cast_fp16 = reshape(shape = var_26395, x = var_26394_cast_fp16)[name = string("v_head_661_cast_fp16")];
            string dense_output_1663_pad_type_1 = const()[name = string("dense_output_1663_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1663_strides_1 = const()[name = string("dense_output_1663_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1663_pad_1 = const()[name = string("dense_output_1663_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1663_dilations_1 = const()[name = string("dense_output_1663_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1663_groups_1 = const()[name = string("dense_output_1663_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602070976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602202112))))[name = string("layers_20_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1663_cast_fp16 = conv(dilations = dense_output_1663_dilations_1, groups = dense_output_1663_groups_1, pad = dense_output_1663_pad_1, pad_type = dense_output_1663_pad_type_1, strides = dense_output_1663_strides_1, weight = layers_20_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1663_cast_fp16")];
            string sparse_output_1663_pad_type_1 = const()[name = string("sparse_output_1663_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1663_strides_1 = const()[name = string("sparse_output_1663_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1663_pad_1 = const()[name = string("sparse_output_1663_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1663_dilations_1 = const()[name = string("sparse_output_1663_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1663_groups_1 = const()[name = string("sparse_output_1663_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602205376))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602202688))))[name = string("layers_20_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1663_cast_fp16 = conv(dilations = sparse_output_1663_dilations_1, groups = sparse_output_1663_groups_1, pad = sparse_output_1663_pad_1, pad_type = sparse_output_1663_pad_type_1, strides = sparse_output_1663_strides_1, weight = layers_20_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1663_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_26411_cast_fp16 = add(x = dense_output_1663_cast_fp16, y = sparse_output_1663_cast_fp16)[name = string("op_26411_cast_fp16")];
            tensor<int32, [4]> var_26412 = const()[name = string("op_26412"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26414 = const()[name = string("op_26414"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_26413_cast_fp16 = transpose(perm = var_26412, x = var_26411_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 101, 128]> p_head_661_cast_fp16 = reshape(shape = var_26414, x = var_26413_cast_fp16)[name = string("p_head_661_cast_fp16")];
            tensor<fp16, [128]> var_26416_to_fp16 = const()[name = string("op_26416_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602221824)))];
            tensor<fp16, [1, 51, 128]> var_26417_cast_fp16 = add(x = q_head_331_cast_fp16, y = var_26416_to_fp16)[name = string("op_26417_cast_fp16")];
            tensor<int32, [1]> q_u_331_axes_1 = const()[name = string("q_u_331_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_331_cast_fp16 = expand_dims(axes = q_u_331_axes_1, x = var_26417_cast_fp16)[name = string("q_u_331_cast_fp16")];
            tensor<fp16, [128]> var_26419_to_fp16 = const()[name = string("op_26419_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602222144)))];
            tensor<fp16, [1, 51, 128]> var_26420_cast_fp16 = add(x = q_head_331_cast_fp16, y = var_26419_to_fp16)[name = string("op_26420_cast_fp16")];
            tensor<int32, [1]> q_v_331_axes_1 = const()[name = string("q_v_331_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_331_cast_fp16 = expand_dims(axes = q_v_331_axes_1, x = var_26420_cast_fp16)[name = string("q_v_331_cast_fp16")];
            tensor<int32, [1]> k_head_663_axes_1 = const()[name = string("k_head_663_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_663_cast_fp16 = expand_dims(axes = k_head_663_axes_1, x = k_head_661_cast_fp16)[name = string("k_head_663_cast_fp16")];
            tensor<int32, [1]> v_head_663_axes_1 = const()[name = string("v_head_663_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_663_cast_fp16 = expand_dims(axes = v_head_663_axes_1, x = v_head_661_cast_fp16)[name = string("v_head_663_cast_fp16")];
            tensor<int32, [1]> p_head_663_axes_1 = const()[name = string("p_head_663_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_663_cast_fp16 = expand_dims(axes = p_head_663_axes_1, x = p_head_661_cast_fp16)[name = string("p_head_663_cast_fp16")];
            bool var_26426_transpose_x_3 = const()[name = string("op_26426_transpose_x_3"), val = bool(false)];
            bool var_26426_transpose_y_3 = const()[name = string("op_26426_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_26426_cast_fp16 = matmul(transpose_x = var_26426_transpose_x_3, transpose_y = var_26426_transpose_y_3, x = q_u_331_cast_fp16, y = k_head_663_cast_fp16)[name = string("op_26426_cast_fp16")];
            fp16 var_26427_to_fp16 = const()[name = string("op_26427_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_331_cast_fp16 = mul(x = var_26426_cast_fp16, y = var_26427_to_fp16)[name = string("scores_content_331_cast_fp16")];
            bool x_1741_transpose_x_3 = const()[name = string("x_1741_transpose_x_3"), val = bool(false)];
            bool x_1741_transpose_y_3 = const()[name = string("x_1741_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1741_cast_fp16 = matmul(transpose_x = x_1741_transpose_x_3, transpose_y = x_1741_transpose_y_3, x = q_v_331_cast_fp16, y = p_head_663_cast_fp16)[name = string("x_1741_cast_fp16")];
            tensor<int32, [8]> x_1743_pad_1 = const()[name = string("x_1743_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1743_mode_1 = const()[name = string("x_1743_mode_1"), val = string("constant")];
            fp16 const_2419_to_fp16 = const()[name = string("const_2419_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1743_cast_fp16 = pad(constant_val = const_2419_to_fp16, mode = x_1743_mode_1, pad = x_1743_pad_1, x = x_1741_cast_fp16)[name = string("x_1743_cast_fp16")];
            tensor<int32, [4]> var_26441 = const()[name = string("op_26441"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1745_cast_fp16 = reshape(shape = var_26441, x = x_1743_cast_fp16)[name = string("x_1745_cast_fp16")];
            tensor<int32, [4]> var_26445_begin_1 = const()[name = string("op_26445_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_26445_end_1 = const()[name = string("op_26445_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_26445_end_mask_1 = const()[name = string("op_26445_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_26445_cast_fp16 = slice_by_index(begin = var_26445_begin_1, end = var_26445_end_1, end_mask = var_26445_end_mask_1, x = x_1745_cast_fp16)[name = string("op_26445_cast_fp16")];
            tensor<int32, [4]> var_26447 = const()[name = string("op_26447"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_26448_cast_fp16 = reshape(shape = var_26447, x = var_26445_cast_fp16)[name = string("op_26448_cast_fp16")];
            tensor<int32, [4]> var_26453_begin_1 = const()[name = string("op_26453_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_26453_end_1 = const()[name = string("op_26453_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_26453_end_mask_1 = const()[name = string("op_26453_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_26453_cast_fp16 = slice_by_index(begin = var_26453_begin_1, end = var_26453_end_1, end_mask = var_26453_end_mask_1, x = var_26448_cast_fp16)[name = string("op_26453_cast_fp16")];
            fp16 var_26454_to_fp16 = const()[name = string("op_26454_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_331_cast_fp16 = mul(x = var_26453_cast_fp16, y = var_26454_to_fp16)[name = string("scores_pos_331_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_331_cast_fp16 = add(x = scores_content_331_cast_fp16, y = scores_pos_331_cast_fp16)[name = string("logits_331_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_26457_cast_fp16 = softmax(axis = var_25577, x = logits_331_cast_fp16)[name = string("op_26457_cast_fp16")];
            bool var_26459_transpose_x_1 = const()[name = string("op_26459_transpose_x_1"), val = bool(false)];
            bool var_26459_transpose_y_1 = const()[name = string("op_26459_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_26459_cast_fp16 = matmul(transpose_x = var_26459_transpose_x_1, transpose_y = var_26459_transpose_y_1, x = var_26457_cast_fp16, y = v_head_663_cast_fp16)[name = string("op_26459_cast_fp16")];
            tensor<int32, [1]> var_26460_axes_1 = const()[name = string("op_26460_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_26460_cast_fp16 = squeeze(axes = var_26460_axes_1, x = var_26459_cast_fp16)[name = string("op_26460_cast_fp16")];
            string dense_output_1665_pad_type_1 = const()[name = string("dense_output_1665_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1665_strides_1 = const()[name = string("dense_output_1665_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1665_pad_1 = const()[name = string("dense_output_1665_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1665_dilations_1 = const()[name = string("dense_output_1665_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1665_groups_1 = const()[name = string("dense_output_1665_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602222464))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602353600))))[name = string("layers_20_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1665_cast_fp16 = conv(dilations = dense_output_1665_dilations_1, groups = dense_output_1665_groups_1, pad = dense_output_1665_pad_1, pad_type = dense_output_1665_pad_type_1, strides = dense_output_1665_strides_1, weight = layers_20_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1665_cast_fp16")];
            string sparse_output_1665_pad_type_1 = const()[name = string("sparse_output_1665_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1665_strides_1 = const()[name = string("sparse_output_1665_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1665_pad_1 = const()[name = string("sparse_output_1665_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1665_dilations_1 = const()[name = string("sparse_output_1665_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1665_groups_1 = const()[name = string("sparse_output_1665_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602356864))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602354176))))[name = string("layers_20_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1665_cast_fp16 = conv(dilations = sparse_output_1665_dilations_1, groups = sparse_output_1665_groups_1, pad = sparse_output_1665_pad_1, pad_type = sparse_output_1665_pad_type_1, strides = sparse_output_1665_strides_1, weight = layers_20_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1665_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26475_cast_fp16 = add(x = dense_output_1665_cast_fp16, y = sparse_output_1665_cast_fp16)[name = string("op_26475_cast_fp16")];
            tensor<int32, [4]> var_26476 = const()[name = string("op_26476"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26478 = const()[name = string("op_26478"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26477_cast_fp16 = transpose(perm = var_26476, x = var_26475_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 51, 128]> q_head_333_cast_fp16 = reshape(shape = var_26478, x = var_26477_cast_fp16)[name = string("q_head_333_cast_fp16")];
            string dense_output_1667_pad_type_1 = const()[name = string("dense_output_1667_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1667_strides_1 = const()[name = string("dense_output_1667_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1667_pad_1 = const()[name = string("dense_output_1667_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1667_dilations_1 = const()[name = string("dense_output_1667_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1667_groups_1 = const()[name = string("dense_output_1667_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602373312))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602504448))))[name = string("layers_20_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1667_cast_fp16 = conv(dilations = dense_output_1667_dilations_1, groups = dense_output_1667_groups_1, pad = dense_output_1667_pad_1, pad_type = dense_output_1667_pad_type_1, strides = dense_output_1667_strides_1, weight = layers_20_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1667_cast_fp16")];
            string sparse_output_1667_pad_type_1 = const()[name = string("sparse_output_1667_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1667_strides_1 = const()[name = string("sparse_output_1667_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1667_pad_1 = const()[name = string("sparse_output_1667_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1667_dilations_1 = const()[name = string("sparse_output_1667_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1667_groups_1 = const()[name = string("sparse_output_1667_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602507712))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602505024))))[name = string("layers_20_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1667_cast_fp16 = conv(dilations = sparse_output_1667_dilations_1, groups = sparse_output_1667_groups_1, pad = sparse_output_1667_pad_1, pad_type = sparse_output_1667_pad_type_1, strides = sparse_output_1667_strides_1, weight = layers_20_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1667_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26494_cast_fp16 = add(x = dense_output_1667_cast_fp16, y = sparse_output_1667_cast_fp16)[name = string("op_26494_cast_fp16")];
            tensor<int32, [4]> var_26495 = const()[name = string("op_26495"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26497 = const()[name = string("op_26497"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26496_cast_fp16 = transpose(perm = var_26495, x = var_26494_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 51, 128]> k_head_665_cast_fp16 = reshape(shape = var_26497, x = var_26496_cast_fp16)[name = string("k_head_665_cast_fp16")];
            string dense_output_1669_pad_type_1 = const()[name = string("dense_output_1669_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1669_strides_1 = const()[name = string("dense_output_1669_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1669_pad_1 = const()[name = string("dense_output_1669_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1669_dilations_1 = const()[name = string("dense_output_1669_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1669_groups_1 = const()[name = string("dense_output_1669_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602524160))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602655296))))[name = string("layers_20_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1669_cast_fp16 = conv(dilations = dense_output_1669_dilations_1, groups = dense_output_1669_groups_1, pad = dense_output_1669_pad_1, pad_type = dense_output_1669_pad_type_1, strides = dense_output_1669_strides_1, weight = layers_20_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1669_cast_fp16")];
            string sparse_output_1669_pad_type_1 = const()[name = string("sparse_output_1669_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1669_strides_1 = const()[name = string("sparse_output_1669_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1669_pad_1 = const()[name = string("sparse_output_1669_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1669_dilations_1 = const()[name = string("sparse_output_1669_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1669_groups_1 = const()[name = string("sparse_output_1669_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602658560))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602655872))))[name = string("layers_20_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1669_cast_fp16 = conv(dilations = sparse_output_1669_dilations_1, groups = sparse_output_1669_groups_1, pad = sparse_output_1669_pad_1, pad_type = sparse_output_1669_pad_type_1, strides = sparse_output_1669_strides_1, weight = layers_20_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1669_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26513_cast_fp16 = add(x = dense_output_1669_cast_fp16, y = sparse_output_1669_cast_fp16)[name = string("op_26513_cast_fp16")];
            tensor<int32, [4]> var_26514 = const()[name = string("op_26514"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26516 = const()[name = string("op_26516"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26515_cast_fp16 = transpose(perm = var_26514, x = var_26513_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 51, 128]> v_head_665_cast_fp16 = reshape(shape = var_26516, x = var_26515_cast_fp16)[name = string("v_head_665_cast_fp16")];
            string dense_output_1671_pad_type_1 = const()[name = string("dense_output_1671_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1671_strides_1 = const()[name = string("dense_output_1671_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1671_pad_1 = const()[name = string("dense_output_1671_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1671_dilations_1 = const()[name = string("dense_output_1671_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1671_groups_1 = const()[name = string("dense_output_1671_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602675008))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602806144))))[name = string("layers_20_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1671_cast_fp16 = conv(dilations = dense_output_1671_dilations_1, groups = dense_output_1671_groups_1, pad = dense_output_1671_pad_1, pad_type = dense_output_1671_pad_type_1, strides = dense_output_1671_strides_1, weight = layers_20_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1671_cast_fp16")];
            string sparse_output_1671_pad_type_1 = const()[name = string("sparse_output_1671_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1671_strides_1 = const()[name = string("sparse_output_1671_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1671_pad_1 = const()[name = string("sparse_output_1671_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1671_dilations_1 = const()[name = string("sparse_output_1671_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1671_groups_1 = const()[name = string("sparse_output_1671_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602809408))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602806720))))[name = string("layers_20_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1671_cast_fp16 = conv(dilations = sparse_output_1671_dilations_1, groups = sparse_output_1671_groups_1, pad = sparse_output_1671_pad_1, pad_type = sparse_output_1671_pad_type_1, strides = sparse_output_1671_strides_1, weight = layers_20_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1671_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_26532_cast_fp16 = add(x = dense_output_1671_cast_fp16, y = sparse_output_1671_cast_fp16)[name = string("op_26532_cast_fp16")];
            tensor<int32, [4]> var_26533 = const()[name = string("op_26533"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26535 = const()[name = string("op_26535"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_26534_cast_fp16 = transpose(perm = var_26533, x = var_26532_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 101, 128]> p_head_665_cast_fp16 = reshape(shape = var_26535, x = var_26534_cast_fp16)[name = string("p_head_665_cast_fp16")];
            tensor<fp16, [128]> var_26537_to_fp16 = const()[name = string("op_26537_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602825856)))];
            tensor<fp16, [1, 51, 128]> var_26538_cast_fp16 = add(x = q_head_333_cast_fp16, y = var_26537_to_fp16)[name = string("op_26538_cast_fp16")];
            tensor<int32, [1]> q_u_333_axes_1 = const()[name = string("q_u_333_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_333_cast_fp16 = expand_dims(axes = q_u_333_axes_1, x = var_26538_cast_fp16)[name = string("q_u_333_cast_fp16")];
            tensor<fp16, [128]> var_26540_to_fp16 = const()[name = string("op_26540_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602826176)))];
            tensor<fp16, [1, 51, 128]> var_26541_cast_fp16 = add(x = q_head_333_cast_fp16, y = var_26540_to_fp16)[name = string("op_26541_cast_fp16")];
            tensor<int32, [1]> q_v_333_axes_1 = const()[name = string("q_v_333_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_333_cast_fp16 = expand_dims(axes = q_v_333_axes_1, x = var_26541_cast_fp16)[name = string("q_v_333_cast_fp16")];
            tensor<int32, [1]> k_head_667_axes_1 = const()[name = string("k_head_667_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_667_cast_fp16 = expand_dims(axes = k_head_667_axes_1, x = k_head_665_cast_fp16)[name = string("k_head_667_cast_fp16")];
            tensor<int32, [1]> v_head_667_axes_1 = const()[name = string("v_head_667_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_667_cast_fp16 = expand_dims(axes = v_head_667_axes_1, x = v_head_665_cast_fp16)[name = string("v_head_667_cast_fp16")];
            tensor<int32, [1]> p_head_667_axes_1 = const()[name = string("p_head_667_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_667_cast_fp16 = expand_dims(axes = p_head_667_axes_1, x = p_head_665_cast_fp16)[name = string("p_head_667_cast_fp16")];
            bool var_26547_transpose_x_3 = const()[name = string("op_26547_transpose_x_3"), val = bool(false)];
            bool var_26547_transpose_y_3 = const()[name = string("op_26547_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_26547_cast_fp16 = matmul(transpose_x = var_26547_transpose_x_3, transpose_y = var_26547_transpose_y_3, x = q_u_333_cast_fp16, y = k_head_667_cast_fp16)[name = string("op_26547_cast_fp16")];
            fp16 var_26548_to_fp16 = const()[name = string("op_26548_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_333_cast_fp16 = mul(x = var_26547_cast_fp16, y = var_26548_to_fp16)[name = string("scores_content_333_cast_fp16")];
            bool x_1749_transpose_x_3 = const()[name = string("x_1749_transpose_x_3"), val = bool(false)];
            bool x_1749_transpose_y_3 = const()[name = string("x_1749_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1749_cast_fp16 = matmul(transpose_x = x_1749_transpose_x_3, transpose_y = x_1749_transpose_y_3, x = q_v_333_cast_fp16, y = p_head_667_cast_fp16)[name = string("x_1749_cast_fp16")];
            tensor<int32, [8]> x_1751_pad_1 = const()[name = string("x_1751_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1751_mode_1 = const()[name = string("x_1751_mode_1"), val = string("constant")];
            fp16 const_2425_to_fp16 = const()[name = string("const_2425_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1751_cast_fp16 = pad(constant_val = const_2425_to_fp16, mode = x_1751_mode_1, pad = x_1751_pad_1, x = x_1749_cast_fp16)[name = string("x_1751_cast_fp16")];
            tensor<int32, [4]> var_26562 = const()[name = string("op_26562"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1753_cast_fp16 = reshape(shape = var_26562, x = x_1751_cast_fp16)[name = string("x_1753_cast_fp16")];
            tensor<int32, [4]> var_26566_begin_1 = const()[name = string("op_26566_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_26566_end_1 = const()[name = string("op_26566_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_26566_end_mask_1 = const()[name = string("op_26566_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_26566_cast_fp16 = slice_by_index(begin = var_26566_begin_1, end = var_26566_end_1, end_mask = var_26566_end_mask_1, x = x_1753_cast_fp16)[name = string("op_26566_cast_fp16")];
            tensor<int32, [4]> var_26568 = const()[name = string("op_26568"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_26569_cast_fp16 = reshape(shape = var_26568, x = var_26566_cast_fp16)[name = string("op_26569_cast_fp16")];
            tensor<int32, [4]> var_26574_begin_1 = const()[name = string("op_26574_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_26574_end_1 = const()[name = string("op_26574_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_26574_end_mask_1 = const()[name = string("op_26574_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_26574_cast_fp16 = slice_by_index(begin = var_26574_begin_1, end = var_26574_end_1, end_mask = var_26574_end_mask_1, x = var_26569_cast_fp16)[name = string("op_26574_cast_fp16")];
            fp16 var_26575_to_fp16 = const()[name = string("op_26575_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_333_cast_fp16 = mul(x = var_26574_cast_fp16, y = var_26575_to_fp16)[name = string("scores_pos_333_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_333_cast_fp16 = add(x = scores_content_333_cast_fp16, y = scores_pos_333_cast_fp16)[name = string("logits_333_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_26578_cast_fp16 = softmax(axis = var_25577, x = logits_333_cast_fp16)[name = string("op_26578_cast_fp16")];
            bool var_26580_transpose_x_1 = const()[name = string("op_26580_transpose_x_1"), val = bool(false)];
            bool var_26580_transpose_y_1 = const()[name = string("op_26580_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_26580_cast_fp16 = matmul(transpose_x = var_26580_transpose_x_1, transpose_y = var_26580_transpose_y_1, x = var_26578_cast_fp16, y = v_head_667_cast_fp16)[name = string("op_26580_cast_fp16")];
            tensor<int32, [1]> var_26581_axes_1 = const()[name = string("op_26581_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_26581_cast_fp16 = squeeze(axes = var_26581_axes_1, x = var_26580_cast_fp16)[name = string("op_26581_cast_fp16")];
            string dense_output_1673_pad_type_1 = const()[name = string("dense_output_1673_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1673_strides_1 = const()[name = string("dense_output_1673_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1673_pad_1 = const()[name = string("dense_output_1673_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1673_dilations_1 = const()[name = string("dense_output_1673_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1673_groups_1 = const()[name = string("dense_output_1673_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602826496))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602957632))))[name = string("layers_20_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1673_cast_fp16 = conv(dilations = dense_output_1673_dilations_1, groups = dense_output_1673_groups_1, pad = dense_output_1673_pad_1, pad_type = dense_output_1673_pad_type_1, strides = dense_output_1673_strides_1, weight = layers_20_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1673_cast_fp16")];
            string sparse_output_1673_pad_type_1 = const()[name = string("sparse_output_1673_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1673_strides_1 = const()[name = string("sparse_output_1673_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1673_pad_1 = const()[name = string("sparse_output_1673_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1673_dilations_1 = const()[name = string("sparse_output_1673_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1673_groups_1 = const()[name = string("sparse_output_1673_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602960896))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602958208))))[name = string("layers_20_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1673_cast_fp16 = conv(dilations = sparse_output_1673_dilations_1, groups = sparse_output_1673_groups_1, pad = sparse_output_1673_pad_1, pad_type = sparse_output_1673_pad_type_1, strides = sparse_output_1673_strides_1, weight = layers_20_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1673_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26596_cast_fp16 = add(x = dense_output_1673_cast_fp16, y = sparse_output_1673_cast_fp16)[name = string("op_26596_cast_fp16")];
            tensor<int32, [4]> var_26597 = const()[name = string("op_26597"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26599 = const()[name = string("op_26599"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26598_cast_fp16 = transpose(perm = var_26597, x = var_26596_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 51, 128]> q_head_335_cast_fp16 = reshape(shape = var_26599, x = var_26598_cast_fp16)[name = string("q_head_335_cast_fp16")];
            string dense_output_1675_pad_type_1 = const()[name = string("dense_output_1675_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1675_strides_1 = const()[name = string("dense_output_1675_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1675_pad_1 = const()[name = string("dense_output_1675_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1675_dilations_1 = const()[name = string("dense_output_1675_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1675_groups_1 = const()[name = string("dense_output_1675_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(602977344))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603108480))))[name = string("layers_20_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1675_cast_fp16 = conv(dilations = dense_output_1675_dilations_1, groups = dense_output_1675_groups_1, pad = dense_output_1675_pad_1, pad_type = dense_output_1675_pad_type_1, strides = dense_output_1675_strides_1, weight = layers_20_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1675_cast_fp16")];
            string sparse_output_1675_pad_type_1 = const()[name = string("sparse_output_1675_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1675_strides_1 = const()[name = string("sparse_output_1675_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1675_pad_1 = const()[name = string("sparse_output_1675_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1675_dilations_1 = const()[name = string("sparse_output_1675_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1675_groups_1 = const()[name = string("sparse_output_1675_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603111744))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603109056))))[name = string("layers_20_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1675_cast_fp16 = conv(dilations = sparse_output_1675_dilations_1, groups = sparse_output_1675_groups_1, pad = sparse_output_1675_pad_1, pad_type = sparse_output_1675_pad_type_1, strides = sparse_output_1675_strides_1, weight = layers_20_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1675_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26615_cast_fp16 = add(x = dense_output_1675_cast_fp16, y = sparse_output_1675_cast_fp16)[name = string("op_26615_cast_fp16")];
            tensor<int32, [4]> var_26616 = const()[name = string("op_26616"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26618 = const()[name = string("op_26618"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26617_cast_fp16 = transpose(perm = var_26616, x = var_26615_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 51, 128]> k_head_669_cast_fp16 = reshape(shape = var_26618, x = var_26617_cast_fp16)[name = string("k_head_669_cast_fp16")];
            string dense_output_1677_pad_type_1 = const()[name = string("dense_output_1677_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1677_strides_1 = const()[name = string("dense_output_1677_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1677_pad_1 = const()[name = string("dense_output_1677_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1677_dilations_1 = const()[name = string("dense_output_1677_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1677_groups_1 = const()[name = string("dense_output_1677_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603128192))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603259328))))[name = string("layers_20_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1677_cast_fp16 = conv(dilations = dense_output_1677_dilations_1, groups = dense_output_1677_groups_1, pad = dense_output_1677_pad_1, pad_type = dense_output_1677_pad_type_1, strides = dense_output_1677_strides_1, weight = layers_20_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("dense_output_1677_cast_fp16")];
            string sparse_output_1677_pad_type_1 = const()[name = string("sparse_output_1677_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1677_strides_1 = const()[name = string("sparse_output_1677_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1677_pad_1 = const()[name = string("sparse_output_1677_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1677_dilations_1 = const()[name = string("sparse_output_1677_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1677_groups_1 = const()[name = string("sparse_output_1677_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603262592))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603259904))))[name = string("layers_20_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1677_cast_fp16 = conv(dilations = sparse_output_1677_dilations_1, groups = sparse_output_1677_groups_1, pad = sparse_output_1677_pad_1, pad_type = sparse_output_1677_pad_type_1, strides = sparse_output_1677_strides_1, weight = layers_20_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_949_cast_fp16)[name = string("sparse_output_1677_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_26634_cast_fp16 = add(x = dense_output_1677_cast_fp16, y = sparse_output_1677_cast_fp16)[name = string("op_26634_cast_fp16")];
            tensor<int32, [4]> var_26635 = const()[name = string("op_26635"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26637 = const()[name = string("op_26637"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_26636_cast_fp16 = transpose(perm = var_26635, x = var_26634_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 51, 128]> v_head_669_cast_fp16 = reshape(shape = var_26637, x = var_26636_cast_fp16)[name = string("v_head_669_cast_fp16")];
            string dense_output_1679_pad_type_1 = const()[name = string("dense_output_1679_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1679_strides_1 = const()[name = string("dense_output_1679_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1679_pad_1 = const()[name = string("dense_output_1679_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1679_dilations_1 = const()[name = string("dense_output_1679_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1679_groups_1 = const()[name = string("dense_output_1679_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603279040))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603410176))))[name = string("layers_20_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1679_cast_fp16 = conv(dilations = dense_output_1679_dilations_1, groups = dense_output_1679_groups_1, pad = dense_output_1679_pad_1, pad_type = dense_output_1679_pad_type_1, strides = dense_output_1679_strides_1, weight = layers_20_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1679_cast_fp16")];
            string sparse_output_1679_pad_type_1 = const()[name = string("sparse_output_1679_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1679_strides_1 = const()[name = string("sparse_output_1679_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1679_pad_1 = const()[name = string("sparse_output_1679_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1679_dilations_1 = const()[name = string("sparse_output_1679_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1679_groups_1 = const()[name = string("sparse_output_1679_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_20_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603413440))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603410752))))[name = string("layers_20_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1679_cast_fp16 = conv(dilations = sparse_output_1679_dilations_1, groups = sparse_output_1679_groups_1, pad = sparse_output_1679_pad_1, pad_type = sparse_output_1679_pad_type_1, strides = sparse_output_1679_strides_1, weight = layers_20_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1679_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_26653_cast_fp16 = add(x = dense_output_1679_cast_fp16, y = sparse_output_1679_cast_fp16)[name = string("op_26653_cast_fp16")];
            tensor<int32, [4]> var_26654 = const()[name = string("op_26654"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_26656 = const()[name = string("op_26656"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_26655_cast_fp16 = transpose(perm = var_26654, x = var_26653_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 101, 128]> p_head_669_cast_fp16 = reshape(shape = var_26656, x = var_26655_cast_fp16)[name = string("p_head_669_cast_fp16")];
            tensor<fp16, [128]> var_26658_to_fp16 = const()[name = string("op_26658_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603429888)))];
            tensor<fp16, [1, 51, 128]> var_26659_cast_fp16 = add(x = q_head_335_cast_fp16, y = var_26658_to_fp16)[name = string("op_26659_cast_fp16")];
            tensor<int32, [1]> q_u_335_axes_1 = const()[name = string("q_u_335_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_335_cast_fp16 = expand_dims(axes = q_u_335_axes_1, x = var_26659_cast_fp16)[name = string("q_u_335_cast_fp16")];
            tensor<fp16, [128]> var_26661_to_fp16 = const()[name = string("op_26661_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603430208)))];
            tensor<fp16, [1, 51, 128]> var_26662_cast_fp16 = add(x = q_head_335_cast_fp16, y = var_26661_to_fp16)[name = string("op_26662_cast_fp16")];
            tensor<int32, [1]> q_v_335_axes_1 = const()[name = string("q_v_335_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_335_cast_fp16 = expand_dims(axes = q_v_335_axes_1, x = var_26662_cast_fp16)[name = string("q_v_335_cast_fp16")];
            tensor<int32, [1]> k_head_671_axes_1 = const()[name = string("k_head_671_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_671_cast_fp16 = expand_dims(axes = k_head_671_axes_1, x = k_head_669_cast_fp16)[name = string("k_head_671_cast_fp16")];
            tensor<int32, [1]> v_head_671_axes_1 = const()[name = string("v_head_671_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_671_cast_fp16 = expand_dims(axes = v_head_671_axes_1, x = v_head_669_cast_fp16)[name = string("v_head_671_cast_fp16")];
            tensor<int32, [1]> p_head_671_axes_1 = const()[name = string("p_head_671_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_671_cast_fp16 = expand_dims(axes = p_head_671_axes_1, x = p_head_669_cast_fp16)[name = string("p_head_671_cast_fp16")];
            bool var_26668_transpose_x_3 = const()[name = string("op_26668_transpose_x_3"), val = bool(false)];
            bool var_26668_transpose_y_3 = const()[name = string("op_26668_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_26668_cast_fp16 = matmul(transpose_x = var_26668_transpose_x_3, transpose_y = var_26668_transpose_y_3, x = q_u_335_cast_fp16, y = k_head_671_cast_fp16)[name = string("op_26668_cast_fp16")];
            fp16 var_26669_to_fp16 = const()[name = string("op_26669_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_335_cast_fp16 = mul(x = var_26668_cast_fp16, y = var_26669_to_fp16)[name = string("scores_content_335_cast_fp16")];
            bool x_1757_transpose_x_3 = const()[name = string("x_1757_transpose_x_3"), val = bool(false)];
            bool x_1757_transpose_y_3 = const()[name = string("x_1757_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1757_cast_fp16 = matmul(transpose_x = x_1757_transpose_x_3, transpose_y = x_1757_transpose_y_3, x = q_v_335_cast_fp16, y = p_head_671_cast_fp16)[name = string("x_1757_cast_fp16")];
            tensor<int32, [8]> x_1759_pad_1 = const()[name = string("x_1759_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1759_mode_1 = const()[name = string("x_1759_mode_1"), val = string("constant")];
            fp16 const_2431_to_fp16 = const()[name = string("const_2431_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1759_cast_fp16 = pad(constant_val = const_2431_to_fp16, mode = x_1759_mode_1, pad = x_1759_pad_1, x = x_1757_cast_fp16)[name = string("x_1759_cast_fp16")];
            tensor<int32, [4]> var_26683 = const()[name = string("op_26683"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1761_cast_fp16 = reshape(shape = var_26683, x = x_1759_cast_fp16)[name = string("x_1761_cast_fp16")];
            tensor<int32, [4]> var_26687_begin_1 = const()[name = string("op_26687_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_26687_end_1 = const()[name = string("op_26687_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_26687_end_mask_1 = const()[name = string("op_26687_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_26687_cast_fp16 = slice_by_index(begin = var_26687_begin_1, end = var_26687_end_1, end_mask = var_26687_end_mask_1, x = x_1761_cast_fp16)[name = string("op_26687_cast_fp16")];
            tensor<int32, [4]> var_26689 = const()[name = string("op_26689"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_26690_cast_fp16 = reshape(shape = var_26689, x = var_26687_cast_fp16)[name = string("op_26690_cast_fp16")];
            tensor<int32, [4]> var_26695_begin_1 = const()[name = string("op_26695_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_26695_end_1 = const()[name = string("op_26695_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_26695_end_mask_1 = const()[name = string("op_26695_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_26695_cast_fp16 = slice_by_index(begin = var_26695_begin_1, end = var_26695_end_1, end_mask = var_26695_end_mask_1, x = var_26690_cast_fp16)[name = string("op_26695_cast_fp16")];
            fp16 var_26696_to_fp16 = const()[name = string("op_26696_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_335_cast_fp16 = mul(x = var_26695_cast_fp16, y = var_26696_to_fp16)[name = string("scores_pos_335_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_335_cast_fp16 = add(x = scores_content_335_cast_fp16, y = scores_pos_335_cast_fp16)[name = string("logits_335_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_26699_cast_fp16 = softmax(axis = var_25577, x = logits_335_cast_fp16)[name = string("op_26699_cast_fp16")];
            bool var_26701_transpose_x_1 = const()[name = string("op_26701_transpose_x_1"), val = bool(false)];
            bool var_26701_transpose_y_1 = const()[name = string("op_26701_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_26701_cast_fp16 = matmul(transpose_x = var_26701_transpose_x_1, transpose_y = var_26701_transpose_y_1, x = var_26699_cast_fp16, y = v_head_671_cast_fp16)[name = string("op_26701_cast_fp16")];
            tensor<int32, [1]> o_head_41_axes_1 = const()[name = string("o_head_41_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_41_cast_fp16 = squeeze(axes = o_head_41_axes_1, x = var_26701_cast_fp16)[name = string("o_head_41_cast_fp16")];
            bool out_41_interleave_1 = const()[name = string("out_41_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_41_cast_fp16 = concat(axis = var_25577, interleave = out_41_interleave_1, values = (var_25855_cast_fp16, var_25976_cast_fp16, var_26097_cast_fp16, var_26218_cast_fp16, var_26339_cast_fp16, var_26460_cast_fp16, var_26581_cast_fp16, o_head_41_cast_fp16))[name = string("out_41_cast_fp16")];
            tensor<int32, [3]> var_26705_perm_1 = const()[name = string("op_26705_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_957_axes_1 = const()[name = string("input_957_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_26705_cast_fp16 = transpose(perm = var_26705_perm_1, x = out_41_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 1024, 51, 1]> input_957_cast_fp16 = expand_dims(axes = input_957_axes_1, x = var_26705_cast_fp16)[name = string("input_957_cast_fp16")];
            string dense_output_1681_pad_type_1 = const()[name = string("dense_output_1681_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1681_strides_1 = const()[name = string("dense_output_1681_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1681_pad_1 = const()[name = string("dense_output_1681_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1681_dilations_1 = const()[name = string("dense_output_1681_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1681_groups_1 = const()[name = string("dense_output_1681_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603430528))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604479168))))[name = string("layers_20_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1681_cast_fp16 = conv(dilations = dense_output_1681_dilations_1, groups = dense_output_1681_groups_1, pad = dense_output_1681_pad_1, pad_type = dense_output_1681_pad_type_1, strides = dense_output_1681_strides_1, weight = layers_20_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_957_cast_fp16)[name = string("dense_output_1681_cast_fp16")];
            string sparse_output_1681_pad_type_1 = const()[name = string("sparse_output_1681_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1681_strides_1 = const()[name = string("sparse_output_1681_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1681_pad_1 = const()[name = string("sparse_output_1681_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1681_dilations_1 = const()[name = string("sparse_output_1681_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1681_groups_1 = const()[name = string("sparse_output_1681_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604500800))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604479744))))[name = string("layers_20_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1681_cast_fp16 = conv(dilations = sparse_output_1681_dilations_1, groups = sparse_output_1681_groups_1, pad = sparse_output_1681_pad_1, pad_type = sparse_output_1681_pad_type_1, strides = sparse_output_1681_strides_1, weight = layers_20_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_957_cast_fp16)[name = string("sparse_output_1681_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_41_cast_fp16 = add(x = dense_output_1681_cast_fp16, y = sparse_output_1681_cast_fp16)[name = string("out_conv_41_cast_fp16")];
            tensor<int32, [1]> var_26722_axes_1 = const()[name = string("op_26722_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_26722_cast_fp16 = squeeze(axes = var_26722_axes_1, x = out_conv_41_cast_fp16)[name = string("op_26722_cast_fp16")];
            tensor<int32, [3]> var_26723_perm_1 = const()[name = string("op_26723_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_26723_cast_fp16 = transpose(perm = var_26723_perm_1, x = var_26722_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 51, 1024]> input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_26723_cast_fp16)[name = string("input_959_cast_fp16")];
            tensor<int32, [1]> x_1765_axes_1 = const()[name = string("x_1765_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_20_norm_conv_weight_to_fp16 = const()[name = string("layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604631936)))];
            tensor<fp16, [1024]> layers_20_norm_conv_bias_to_fp16 = const()[name = string("layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604634048)))];
            tensor<fp16, [1, 51, 1024]> x_1765_cast_fp16 = layer_norm(axes = x_1765_axes_1, beta = layers_20_norm_conv_bias_to_fp16, epsilon = var_25592_to_fp16, gamma = layers_20_norm_conv_weight_to_fp16, x = input_959_cast_fp16)[name = string("x_1765_cast_fp16")];
            tensor<int32, [3]> var_26733_perm_1 = const()[name = string("op_26733_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_961_axes_1 = const()[name = string("input_961_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_26733_cast_fp16 = transpose(perm = var_26733_perm_1, x = x_1765_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 1024, 51, 1]> input_961_cast_fp16 = expand_dims(axes = input_961_axes_1, x = var_26733_cast_fp16)[name = string("input_961_cast_fp16")];
            string dense_output_1683_pad_type_1 = const()[name = string("dense_output_1683_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1683_strides_1 = const()[name = string("dense_output_1683_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1683_pad_1 = const()[name = string("dense_output_1683_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1683_dilations_1 = const()[name = string("dense_output_1683_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1683_groups_1 = const()[name = string("dense_output_1683_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_20_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604636160))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606733376))))[name = string("layers_20_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1683_cast_fp16 = conv(dilations = dense_output_1683_dilations_1, groups = dense_output_1683_groups_1, pad = dense_output_1683_pad_1, pad_type = dense_output_1683_pad_type_1, strides = dense_output_1683_strides_1, weight = layers_20_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_961_cast_fp16)[name = string("dense_output_1683_cast_fp16")];
            string sparse_output_1683_pad_type_1 = const()[name = string("sparse_output_1683_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1683_strides_1 = const()[name = string("sparse_output_1683_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1683_pad_1 = const()[name = string("sparse_output_1683_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1683_dilations_1 = const()[name = string("sparse_output_1683_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1683_groups_1 = const()[name = string("sparse_output_1683_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_20_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606776000))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606733952))))[name = string("layers_20_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1683_cast_fp16 = conv(dilations = sparse_output_1683_dilations_1, groups = sparse_output_1683_groups_1, pad = sparse_output_1683_pad_1, pad_type = sparse_output_1683_pad_type_1, strides = sparse_output_1683_strides_1, weight = layers_20_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_961_cast_fp16)[name = string("sparse_output_1683_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_963_cast_fp16 = add(x = dense_output_1683_cast_fp16, y = sparse_output_1683_cast_fp16)[name = string("input_963_cast_fp16")];
            int32 input_965_split_num_splits_1 = const()[name = string("input_965_split_num_splits_1"), val = int32(2)];
            int32 input_965_split_axis_1 = const()[name = string("input_965_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_965_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_965_split_cast_fp16_1 = split(axis = input_965_split_axis_1, num_splits = input_965_split_num_splits_1, x = input_963_cast_fp16)[name = string("input_965_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_965_split_1_sigmoid_cast_fp16 = sigmoid(x = input_965_split_cast_fp16_1)[name = string("input_965_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_965_cast_fp16 = mul(x = input_965_split_cast_fp16_0, y = input_965_split_1_sigmoid_cast_fp16)[name = string("input_965_cast_fp16")];
            tensor<int32, [8]> input_967_pad_1 = const()[name = string("input_967_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_967_mode_1 = const()[name = string("input_967_mode_1"), val = string("constant")];
            fp16 const_2433_to_fp16 = const()[name = string("const_2433_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_967_cast_fp16 = pad(constant_val = const_2433_to_fp16, mode = input_967_mode_1, pad = input_967_pad_1, x = input_965_cast_fp16)[name = string("input_967_cast_fp16")];
            string dense_output_1685_pad_type_1 = const()[name = string("dense_output_1685_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1685_strides_1 = const()[name = string("dense_output_1685_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1685_pad_1 = const()[name = string("dense_output_1685_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1685_dilations_1 = const()[name = string("dense_output_1685_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1685_groups_1 = const()[name = string("dense_output_1685_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1685_cast_fp16 = conv(dilations = dense_output_1685_dilations_1, groups = dense_output_1685_groups_1, pad = dense_output_1685_pad_1, pad_type = dense_output_1685_pad_type_1, strides = dense_output_1685_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_967_cast_fp16)[name = string("dense_output_1685_cast_fp16")];
            string sparse_output_1685_pad_type_1 = const()[name = string("sparse_output_1685_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1685_strides_1 = const()[name = string("sparse_output_1685_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1685_pad_1 = const()[name = string("sparse_output_1685_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1685_dilations_1 = const()[name = string("sparse_output_1685_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1685_groups_1 = const()[name = string("sparse_output_1685_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_20_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607056704))), nonzero_data = tensor<fp16, [9215]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607038208))))[name = string("layers_20_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1685_cast_fp16 = conv(dilations = sparse_output_1685_dilations_1, groups = sparse_output_1685_groups_1, pad = sparse_output_1685_pad_1, pad_type = sparse_output_1685_pad_type_1, strides = sparse_output_1685_strides_1, weight = layers_20_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_967_cast_fp16)[name = string("sparse_output_1685_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_969_cast_fp16 = add(x = dense_output_1685_cast_fp16, y = sparse_output_1685_cast_fp16)[name = string("input_969_cast_fp16")];
            tensor<fp16, [1024]> layers_20_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_20_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608236416)))];
            tensor<fp16, [1024]> layers_20_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_20_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608238528)))];
            tensor<fp16, [1024]> layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_20_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608240640)))];
            tensor<fp16, [1024]> layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_20_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608242752)))];
            tensor<fp16, [1, 1024, 51, 1]> input_971_cast_fp16 = batch_norm(beta = layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_25592_to_fp16, gamma = layers_20_conv_batch_norm_weight_to_fp16, mean = layers_20_conv_batch_norm_running_mean_to_fp16, variance = layers_20_conv_batch_norm_running_var_to_fp16, x = input_969_cast_fp16)[name = string("input_971_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_973_cast_fp16 = silu(x = input_971_cast_fp16)[name = string("input_973_cast_fp16")];
            string dense_output_1687_pad_type_1 = const()[name = string("dense_output_1687_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1687_strides_1 = const()[name = string("dense_output_1687_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1687_pad_1 = const()[name = string("dense_output_1687_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1687_dilations_1 = const()[name = string("dense_output_1687_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1687_groups_1 = const()[name = string("dense_output_1687_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608244864))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609293504))))[name = string("layers_20_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1687_cast_fp16 = conv(dilations = dense_output_1687_dilations_1, groups = dense_output_1687_groups_1, pad = dense_output_1687_pad_1, pad_type = dense_output_1687_pad_type_1, strides = dense_output_1687_strides_1, weight = layers_20_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_973_cast_fp16)[name = string("dense_output_1687_cast_fp16")];
            string sparse_output_1687_pad_type_1 = const()[name = string("sparse_output_1687_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1687_strides_1 = const()[name = string("sparse_output_1687_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1687_pad_1 = const()[name = string("sparse_output_1687_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1687_dilations_1 = const()[name = string("sparse_output_1687_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1687_groups_1 = const()[name = string("sparse_output_1687_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_20_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609315136))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609294080))))[name = string("layers_20_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1687_cast_fp16 = conv(dilations = sparse_output_1687_dilations_1, groups = sparse_output_1687_groups_1, pad = sparse_output_1687_pad_1, pad_type = sparse_output_1687_pad_type_1, strides = sparse_output_1687_strides_1, weight = layers_20_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_973_cast_fp16)[name = string("sparse_output_1687_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1767_cast_fp16 = add(x = dense_output_1687_cast_fp16, y = sparse_output_1687_cast_fp16)[name = string("x_1767_cast_fp16")];
            tensor<int32, [1]> var_26789_axes_1 = const()[name = string("op_26789_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_26789_cast_fp16 = squeeze(axes = var_26789_axes_1, x = x_1767_cast_fp16)[name = string("op_26789_cast_fp16")];
            tensor<int32, [3]> var_26790_perm_1 = const()[name = string("op_26790_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_26790_cast_fp16 = transpose(perm = var_26790_perm_1, x = var_26789_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 51, 1024]> input_975_cast_fp16 = add(x = input_959_cast_fp16, y = var_26790_cast_fp16)[name = string("input_975_cast_fp16")];
            tensor<int32, [1]> x_1769_axes_1 = const()[name = string("x_1769_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609446272)))];
            tensor<fp16, [1024]> layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609448384)))];
            tensor<fp16, [1, 51, 1024]> x_1769_cast_fp16 = layer_norm(axes = x_1769_axes_1, beta = layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_25592_to_fp16, gamma = layers_20_norm_feed_forward2_weight_to_fp16, x = input_975_cast_fp16)[name = string("x_1769_cast_fp16")];
            tensor<int32, [4]> var_26800 = const()[name = string("op_26800"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1771_cast_fp16 = reshape(shape = var_26800, x = x_1769_cast_fp16)[name = string("x_1771_cast_fp16")];
            tensor<int32, [4]> input_977_perm_1 = const()[name = string("input_977_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1689_pad_type_1 = const()[name = string("dense_output_1689_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1689_strides_1 = const()[name = string("dense_output_1689_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1689_pad_1 = const()[name = string("dense_output_1689_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1689_dilations_1 = const()[name = string("dense_output_1689_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1689_groups_1 = const()[name = string("dense_output_1689_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_20_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609450496))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613644864))))[name = string("layers_20_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_977_cast_fp16 = transpose(perm = input_977_perm_1, x = x_1771_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1689_cast_fp16 = conv(dilations = dense_output_1689_dilations_1, groups = dense_output_1689_groups_1, pad = dense_output_1689_pad_1, pad_type = dense_output_1689_pad_type_1, strides = dense_output_1689_strides_1, weight = layers_20_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_977_cast_fp16)[name = string("dense_output_1689_cast_fp16")];
            string sparse_output_1689_pad_type_1 = const()[name = string("sparse_output_1689_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1689_strides_1 = const()[name = string("sparse_output_1689_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1689_pad_1 = const()[name = string("sparse_output_1689_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1689_dilations_1 = const()[name = string("sparse_output_1689_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1689_groups_1 = const()[name = string("sparse_output_1689_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_20_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613729408))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613645440))))[name = string("layers_20_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1689_cast_fp16 = conv(dilations = sparse_output_1689_dilations_1, groups = sparse_output_1689_groups_1, pad = sparse_output_1689_pad_1, pad_type = sparse_output_1689_pad_type_1, strides = sparse_output_1689_strides_1, weight = layers_20_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_977_cast_fp16)[name = string("sparse_output_1689_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_979_cast_fp16 = add(x = dense_output_1689_cast_fp16, y = sparse_output_1689_cast_fp16)[name = string("input_979_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_981_cast_fp16 = silu(x = input_979_cast_fp16)[name = string("input_981_cast_fp16")];
            string dense_output_1691_pad_type_1 = const()[name = string("dense_output_1691_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1691_strides_1 = const()[name = string("dense_output_1691_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1691_pad_1 = const()[name = string("dense_output_1691_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1691_dilations_1 = const()[name = string("dense_output_1691_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1691_groups_1 = const()[name = string("dense_output_1691_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_20_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614253760))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618448128))))[name = string("layers_20_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1691_cast_fp16 = conv(dilations = dense_output_1691_dilations_1, groups = dense_output_1691_groups_1, pad = dense_output_1691_pad_1, pad_type = dense_output_1691_pad_type_1, strides = dense_output_1691_strides_1, weight = layers_20_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_981_cast_fp16)[name = string("dense_output_1691_cast_fp16")];
            string sparse_output_1691_pad_type_1 = const()[name = string("sparse_output_1691_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1691_strides_1 = const()[name = string("sparse_output_1691_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1691_pad_1 = const()[name = string("sparse_output_1691_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1691_dilations_1 = const()[name = string("sparse_output_1691_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1691_groups_1 = const()[name = string("sparse_output_1691_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_20_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618532672))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(618448704))))[name = string("layers_20_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1691_cast_fp16 = conv(dilations = sparse_output_1691_dilations_1, groups = sparse_output_1691_groups_1, pad = sparse_output_1691_pad_1, pad_type = sparse_output_1691_pad_type_1, strides = sparse_output_1691_strides_1, weight = layers_20_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_981_cast_fp16)[name = string("sparse_output_1691_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1773_cast_fp16 = add(x = dense_output_1691_cast_fp16, y = sparse_output_1691_cast_fp16)[name = string("x_1773_cast_fp16")];
            tensor<int32, [4]> x_1775_perm_1 = const()[name = string("x_1775_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_26835 = const()[name = string("op_26835"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1775_cast_fp16 = transpose(perm = x_1775_perm_1, x = x_1773_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 51, 1024]> var_26836_cast_fp16 = reshape(shape = var_26835, x = x_1775_cast_fp16)[name = string("op_26836_cast_fp16")];
            fp16 var_26837_to_fp16 = const()[name = string("op_26837_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_26838_cast_fp16 = mul(x = var_26836_cast_fp16, y = var_26837_to_fp16)[name = string("op_26838_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_983_cast_fp16 = add(x = input_975_cast_fp16, y = var_26838_cast_fp16)[name = string("input_983_cast_fp16")];
            tensor<int32, [1]> input_985_axes_1 = const()[name = string("input_985_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_20_norm_out_weight_to_fp16 = const()[name = string("layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(619057024)))];
            tensor<fp16, [1024]> layers_20_norm_out_bias_to_fp16 = const()[name = string("layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(619059136)))];
            tensor<fp16, [1, 51, 1024]> input_985_cast_fp16 = layer_norm(axes = input_985_axes_1, beta = layers_20_norm_out_bias_to_fp16, epsilon = var_25592_to_fp16, gamma = layers_20_norm_out_weight_to_fp16, x = input_983_cast_fp16)[name = string("input_985_cast_fp16")];
            int32 var_26846 = const()[name = string("op_26846"), val = int32(-1)];
            tensor<int32, [1]> x_1777_axes_1 = const()[name = string("x_1777_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(619061248)))];
            tensor<fp16, [1024]> layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(619063360)))];
            fp16 var_26861_to_fp16 = const()[name = string("op_26861_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1777_cast_fp16 = layer_norm(axes = x_1777_axes_1, beta = layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_26861_to_fp16, gamma = layers_21_norm_feed_forward1_weight_to_fp16, x = input_985_cast_fp16)[name = string("x_1777_cast_fp16")];
            tensor<int32, [4]> var_26880 = const()[name = string("op_26880"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1779_cast_fp16 = reshape(shape = var_26880, x = x_1777_cast_fp16)[name = string("x_1779_cast_fp16")];
            tensor<int32, [4]> input_987_perm_1 = const()[name = string("input_987_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1693_pad_type_1 = const()[name = string("dense_output_1693_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1693_strides_1 = const()[name = string("dense_output_1693_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1693_pad_1 = const()[name = string("dense_output_1693_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1693_dilations_1 = const()[name = string("dense_output_1693_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1693_groups_1 = const()[name = string("dense_output_1693_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_21_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(619065472))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623259840))))[name = string("layers_21_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_987_cast_fp16 = transpose(perm = input_987_perm_1, x = x_1779_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1693_cast_fp16 = conv(dilations = dense_output_1693_dilations_1, groups = dense_output_1693_groups_1, pad = dense_output_1693_pad_1, pad_type = dense_output_1693_pad_type_1, strides = dense_output_1693_strides_1, weight = layers_21_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_987_cast_fp16)[name = string("dense_output_1693_cast_fp16")];
            string sparse_output_1693_pad_type_1 = const()[name = string("sparse_output_1693_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1693_strides_1 = const()[name = string("sparse_output_1693_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1693_pad_1 = const()[name = string("sparse_output_1693_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1693_dilations_1 = const()[name = string("sparse_output_1693_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1693_groups_1 = const()[name = string("sparse_output_1693_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_21_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623344384))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623260416))))[name = string("layers_21_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1693_cast_fp16 = conv(dilations = sparse_output_1693_dilations_1, groups = sparse_output_1693_groups_1, pad = sparse_output_1693_pad_1, pad_type = sparse_output_1693_pad_type_1, strides = sparse_output_1693_strides_1, weight = layers_21_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_987_cast_fp16)[name = string("sparse_output_1693_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_989_cast_fp16 = add(x = dense_output_1693_cast_fp16, y = sparse_output_1693_cast_fp16)[name = string("input_989_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_991_cast_fp16 = silu(x = input_989_cast_fp16)[name = string("input_991_cast_fp16")];
            string dense_output_1695_pad_type_1 = const()[name = string("dense_output_1695_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1695_strides_1 = const()[name = string("dense_output_1695_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1695_pad_1 = const()[name = string("dense_output_1695_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1695_dilations_1 = const()[name = string("dense_output_1695_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1695_groups_1 = const()[name = string("dense_output_1695_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_21_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623868736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628063104))))[name = string("layers_21_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1695_cast_fp16 = conv(dilations = dense_output_1695_dilations_1, groups = dense_output_1695_groups_1, pad = dense_output_1695_pad_1, pad_type = dense_output_1695_pad_type_1, strides = dense_output_1695_strides_1, weight = layers_21_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_991_cast_fp16)[name = string("dense_output_1695_cast_fp16")];
            string sparse_output_1695_pad_type_1 = const()[name = string("sparse_output_1695_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1695_strides_1 = const()[name = string("sparse_output_1695_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1695_pad_1 = const()[name = string("sparse_output_1695_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1695_dilations_1 = const()[name = string("sparse_output_1695_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1695_groups_1 = const()[name = string("sparse_output_1695_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_21_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628147648))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628063680))))[name = string("layers_21_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1695_cast_fp16 = conv(dilations = sparse_output_1695_dilations_1, groups = sparse_output_1695_groups_1, pad = sparse_output_1695_pad_1, pad_type = sparse_output_1695_pad_type_1, strides = sparse_output_1695_strides_1, weight = layers_21_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_991_cast_fp16)[name = string("sparse_output_1695_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1781_cast_fp16 = add(x = dense_output_1695_cast_fp16, y = sparse_output_1695_cast_fp16)[name = string("x_1781_cast_fp16")];
            tensor<int32, [4]> x_1783_perm_1 = const()[name = string("x_1783_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_26915 = const()[name = string("op_26915"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1783_cast_fp16 = transpose(perm = x_1783_perm_1, x = x_1781_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 51, 1024]> var_26916_cast_fp16 = reshape(shape = var_26915, x = x_1783_cast_fp16)[name = string("op_26916_cast_fp16")];
            fp16 var_26917_to_fp16 = const()[name = string("op_26917_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_26918_cast_fp16 = mul(x = var_26916_cast_fp16, y = var_26917_to_fp16)[name = string("op_26918_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_993_cast_fp16 = add(x = input_985_cast_fp16, y = var_26918_cast_fp16)[name = string("input_993_cast_fp16")];
            tensor<int32, [1]> q_43_axes_1 = const()[name = string("q_43_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_21_norm_self_att_weight_to_fp16 = const()[name = string("layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628672000)))];
            tensor<fp16, [1024]> layers_21_norm_self_att_bias_to_fp16 = const()[name = string("layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628674112)))];
            tensor<fp16, [1, 51, 1024]> q_43_cast_fp16 = layer_norm(axes = q_43_axes_1, beta = layers_21_norm_self_att_bias_to_fp16, epsilon = var_26861_to_fp16, gamma = layers_21_norm_self_att_weight_to_fp16, x = input_993_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<int32, [3]> var_26992 = const()[name = string("op_26992"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_995_axes_1 = const()[name = string("input_995_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_26993_cast_fp16 = transpose(perm = var_26992, x = q_43_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 1024, 51, 1]> input_995_cast_fp16 = expand_dims(axes = input_995_axes_1, x = var_26993_cast_fp16)[name = string("input_995_cast_fp16")];
            string dense_output_1697_pad_type_1 = const()[name = string("dense_output_1697_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1697_strides_1 = const()[name = string("dense_output_1697_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1697_pad_1 = const()[name = string("dense_output_1697_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1697_dilations_1 = const()[name = string("dense_output_1697_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1697_groups_1 = const()[name = string("dense_output_1697_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628676224))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628807360))))[name = string("layers_21_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1697_cast_fp16 = conv(dilations = dense_output_1697_dilations_1, groups = dense_output_1697_groups_1, pad = dense_output_1697_pad_1, pad_type = dense_output_1697_pad_type_1, strides = dense_output_1697_strides_1, weight = layers_21_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1697_cast_fp16")];
            string sparse_output_1697_pad_type_1 = const()[name = string("sparse_output_1697_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1697_strides_1 = const()[name = string("sparse_output_1697_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1697_pad_1 = const()[name = string("sparse_output_1697_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1697_dilations_1 = const()[name = string("sparse_output_1697_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1697_groups_1 = const()[name = string("sparse_output_1697_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628810624))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628807936))))[name = string("layers_21_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1697_cast_fp16 = conv(dilations = sparse_output_1697_dilations_1, groups = sparse_output_1697_groups_1, pad = sparse_output_1697_pad_1, pad_type = sparse_output_1697_pad_type_1, strides = sparse_output_1697_strides_1, weight = layers_21_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1697_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27018_cast_fp16 = add(x = dense_output_1697_cast_fp16, y = sparse_output_1697_cast_fp16)[name = string("op_27018_cast_fp16")];
            tensor<int32, [4]> var_27019 = const()[name = string("op_27019"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27021 = const()[name = string("op_27021"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27020_cast_fp16 = transpose(perm = var_27019, x = var_27018_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 51, 128]> q_head_337_cast_fp16 = reshape(shape = var_27021, x = var_27020_cast_fp16)[name = string("q_head_337_cast_fp16")];
            string dense_output_1699_pad_type_1 = const()[name = string("dense_output_1699_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1699_strides_1 = const()[name = string("dense_output_1699_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1699_pad_1 = const()[name = string("dense_output_1699_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1699_dilations_1 = const()[name = string("dense_output_1699_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1699_groups_1 = const()[name = string("dense_output_1699_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628827072))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628958208))))[name = string("layers_21_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1699_cast_fp16 = conv(dilations = dense_output_1699_dilations_1, groups = dense_output_1699_groups_1, pad = dense_output_1699_pad_1, pad_type = dense_output_1699_pad_type_1, strides = dense_output_1699_strides_1, weight = layers_21_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1699_cast_fp16")];
            string sparse_output_1699_pad_type_1 = const()[name = string("sparse_output_1699_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1699_strides_1 = const()[name = string("sparse_output_1699_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1699_pad_1 = const()[name = string("sparse_output_1699_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1699_dilations_1 = const()[name = string("sparse_output_1699_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1699_groups_1 = const()[name = string("sparse_output_1699_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628961472))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628958784))))[name = string("layers_21_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1699_cast_fp16 = conv(dilations = sparse_output_1699_dilations_1, groups = sparse_output_1699_groups_1, pad = sparse_output_1699_pad_1, pad_type = sparse_output_1699_pad_type_1, strides = sparse_output_1699_strides_1, weight = layers_21_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1699_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27037_cast_fp16 = add(x = dense_output_1699_cast_fp16, y = sparse_output_1699_cast_fp16)[name = string("op_27037_cast_fp16")];
            tensor<int32, [4]> var_27038 = const()[name = string("op_27038"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27040 = const()[name = string("op_27040"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27039_cast_fp16 = transpose(perm = var_27038, x = var_27037_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 51, 128]> k_head_673_cast_fp16 = reshape(shape = var_27040, x = var_27039_cast_fp16)[name = string("k_head_673_cast_fp16")];
            string dense_output_1701_pad_type_1 = const()[name = string("dense_output_1701_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1701_strides_1 = const()[name = string("dense_output_1701_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1701_pad_1 = const()[name = string("dense_output_1701_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1701_dilations_1 = const()[name = string("dense_output_1701_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1701_groups_1 = const()[name = string("dense_output_1701_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628977920))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629109056))))[name = string("layers_21_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1701_cast_fp16 = conv(dilations = dense_output_1701_dilations_1, groups = dense_output_1701_groups_1, pad = dense_output_1701_pad_1, pad_type = dense_output_1701_pad_type_1, strides = dense_output_1701_strides_1, weight = layers_21_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1701_cast_fp16")];
            string sparse_output_1701_pad_type_1 = const()[name = string("sparse_output_1701_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1701_strides_1 = const()[name = string("sparse_output_1701_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1701_pad_1 = const()[name = string("sparse_output_1701_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1701_dilations_1 = const()[name = string("sparse_output_1701_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1701_groups_1 = const()[name = string("sparse_output_1701_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629112320))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629109632))))[name = string("layers_21_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1701_cast_fp16 = conv(dilations = sparse_output_1701_dilations_1, groups = sparse_output_1701_groups_1, pad = sparse_output_1701_pad_1, pad_type = sparse_output_1701_pad_type_1, strides = sparse_output_1701_strides_1, weight = layers_21_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1701_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27056_cast_fp16 = add(x = dense_output_1701_cast_fp16, y = sparse_output_1701_cast_fp16)[name = string("op_27056_cast_fp16")];
            tensor<int32, [4]> var_27057 = const()[name = string("op_27057"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27059 = const()[name = string("op_27059"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27058_cast_fp16 = transpose(perm = var_27057, x = var_27056_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 51, 128]> v_head_673_cast_fp16 = reshape(shape = var_27059, x = var_27058_cast_fp16)[name = string("v_head_673_cast_fp16")];
            string dense_output_1703_pad_type_1 = const()[name = string("dense_output_1703_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1703_strides_1 = const()[name = string("dense_output_1703_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1703_pad_1 = const()[name = string("dense_output_1703_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1703_dilations_1 = const()[name = string("dense_output_1703_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1703_groups_1 = const()[name = string("dense_output_1703_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629128768))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629259904))))[name = string("layers_21_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1703_cast_fp16 = conv(dilations = dense_output_1703_dilations_1, groups = dense_output_1703_groups_1, pad = dense_output_1703_pad_1, pad_type = dense_output_1703_pad_type_1, strides = dense_output_1703_strides_1, weight = layers_21_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1703_cast_fp16")];
            string sparse_output_1703_pad_type_1 = const()[name = string("sparse_output_1703_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1703_strides_1 = const()[name = string("sparse_output_1703_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1703_pad_1 = const()[name = string("sparse_output_1703_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1703_dilations_1 = const()[name = string("sparse_output_1703_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1703_groups_1 = const()[name = string("sparse_output_1703_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629263168))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629260480))))[name = string("layers_21_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1703_cast_fp16 = conv(dilations = sparse_output_1703_dilations_1, groups = sparse_output_1703_groups_1, pad = sparse_output_1703_pad_1, pad_type = sparse_output_1703_pad_type_1, strides = sparse_output_1703_strides_1, weight = layers_21_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1703_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_27075_cast_fp16 = add(x = dense_output_1703_cast_fp16, y = sparse_output_1703_cast_fp16)[name = string("op_27075_cast_fp16")];
            tensor<int32, [4]> var_27076 = const()[name = string("op_27076"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27078 = const()[name = string("op_27078"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_27077_cast_fp16 = transpose(perm = var_27076, x = var_27075_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 101, 128]> p_head_673_cast_fp16 = reshape(shape = var_27078, x = var_27077_cast_fp16)[name = string("p_head_673_cast_fp16")];
            tensor<fp16, [128]> var_27080_to_fp16 = const()[name = string("op_27080_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629279616)))];
            tensor<fp16, [1, 51, 128]> var_27081_cast_fp16 = add(x = q_head_337_cast_fp16, y = var_27080_to_fp16)[name = string("op_27081_cast_fp16")];
            tensor<int32, [1]> q_u_337_axes_1 = const()[name = string("q_u_337_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_337_cast_fp16 = expand_dims(axes = q_u_337_axes_1, x = var_27081_cast_fp16)[name = string("q_u_337_cast_fp16")];
            tensor<fp16, [128]> var_27083_to_fp16 = const()[name = string("op_27083_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629279936)))];
            tensor<fp16, [1, 51, 128]> var_27084_cast_fp16 = add(x = q_head_337_cast_fp16, y = var_27083_to_fp16)[name = string("op_27084_cast_fp16")];
            tensor<int32, [1]> q_v_337_axes_1 = const()[name = string("q_v_337_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_337_cast_fp16 = expand_dims(axes = q_v_337_axes_1, x = var_27084_cast_fp16)[name = string("q_v_337_cast_fp16")];
            tensor<int32, [1]> k_head_675_axes_1 = const()[name = string("k_head_675_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_675_cast_fp16 = expand_dims(axes = k_head_675_axes_1, x = k_head_673_cast_fp16)[name = string("k_head_675_cast_fp16")];
            tensor<int32, [1]> v_head_675_axes_1 = const()[name = string("v_head_675_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_675_cast_fp16 = expand_dims(axes = v_head_675_axes_1, x = v_head_673_cast_fp16)[name = string("v_head_675_cast_fp16")];
            tensor<int32, [1]> p_head_675_axes_1 = const()[name = string("p_head_675_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_675_cast_fp16 = expand_dims(axes = p_head_675_axes_1, x = p_head_673_cast_fp16)[name = string("p_head_675_cast_fp16")];
            bool var_27090_transpose_x_3 = const()[name = string("op_27090_transpose_x_3"), val = bool(false)];
            bool var_27090_transpose_y_3 = const()[name = string("op_27090_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_27090_cast_fp16 = matmul(transpose_x = var_27090_transpose_x_3, transpose_y = var_27090_transpose_y_3, x = q_u_337_cast_fp16, y = k_head_675_cast_fp16)[name = string("op_27090_cast_fp16")];
            fp16 var_27091_to_fp16 = const()[name = string("op_27091_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_337_cast_fp16 = mul(x = var_27090_cast_fp16, y = var_27091_to_fp16)[name = string("scores_content_337_cast_fp16")];
            bool x_1785_transpose_x_3 = const()[name = string("x_1785_transpose_x_3"), val = bool(false)];
            bool x_1785_transpose_y_3 = const()[name = string("x_1785_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1785_cast_fp16 = matmul(transpose_x = x_1785_transpose_x_3, transpose_y = x_1785_transpose_y_3, x = q_v_337_cast_fp16, y = p_head_675_cast_fp16)[name = string("x_1785_cast_fp16")];
            tensor<int32, [8]> x_1787_pad_1 = const()[name = string("x_1787_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1787_mode_1 = const()[name = string("x_1787_mode_1"), val = string("constant")];
            fp16 const_2443_to_fp16 = const()[name = string("const_2443_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1787_cast_fp16 = pad(constant_val = const_2443_to_fp16, mode = x_1787_mode_1, pad = x_1787_pad_1, x = x_1785_cast_fp16)[name = string("x_1787_cast_fp16")];
            tensor<int32, [4]> var_27105 = const()[name = string("op_27105"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1789_cast_fp16 = reshape(shape = var_27105, x = x_1787_cast_fp16)[name = string("x_1789_cast_fp16")];
            tensor<int32, [4]> var_27109_begin_1 = const()[name = string("op_27109_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_27109_end_1 = const()[name = string("op_27109_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_27109_end_mask_1 = const()[name = string("op_27109_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_27109_cast_fp16 = slice_by_index(begin = var_27109_begin_1, end = var_27109_end_1, end_mask = var_27109_end_mask_1, x = x_1789_cast_fp16)[name = string("op_27109_cast_fp16")];
            tensor<int32, [4]> var_27111 = const()[name = string("op_27111"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_27112_cast_fp16 = reshape(shape = var_27111, x = var_27109_cast_fp16)[name = string("op_27112_cast_fp16")];
            tensor<int32, [4]> var_27117_begin_1 = const()[name = string("op_27117_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_27117_end_1 = const()[name = string("op_27117_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_27117_end_mask_1 = const()[name = string("op_27117_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_27117_cast_fp16 = slice_by_index(begin = var_27117_begin_1, end = var_27117_end_1, end_mask = var_27117_end_mask_1, x = var_27112_cast_fp16)[name = string("op_27117_cast_fp16")];
            fp16 var_27118_to_fp16 = const()[name = string("op_27118_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_337_cast_fp16 = mul(x = var_27117_cast_fp16, y = var_27118_to_fp16)[name = string("scores_pos_337_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_337_cast_fp16 = add(x = scores_content_337_cast_fp16, y = scores_pos_337_cast_fp16)[name = string("logits_337_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_27121_cast_fp16 = softmax(axis = var_26846, x = logits_337_cast_fp16)[name = string("op_27121_cast_fp16")];
            bool var_27123_transpose_x_1 = const()[name = string("op_27123_transpose_x_1"), val = bool(false)];
            bool var_27123_transpose_y_1 = const()[name = string("op_27123_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_27123_cast_fp16 = matmul(transpose_x = var_27123_transpose_x_1, transpose_y = var_27123_transpose_y_1, x = var_27121_cast_fp16, y = v_head_675_cast_fp16)[name = string("op_27123_cast_fp16")];
            tensor<int32, [1]> var_27124_axes_1 = const()[name = string("op_27124_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_27124_cast_fp16 = squeeze(axes = var_27124_axes_1, x = var_27123_cast_fp16)[name = string("op_27124_cast_fp16")];
            string dense_output_1705_pad_type_1 = const()[name = string("dense_output_1705_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1705_strides_1 = const()[name = string("dense_output_1705_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1705_pad_1 = const()[name = string("dense_output_1705_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1705_dilations_1 = const()[name = string("dense_output_1705_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1705_groups_1 = const()[name = string("dense_output_1705_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629280256))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629411392))))[name = string("layers_21_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1705_cast_fp16 = conv(dilations = dense_output_1705_dilations_1, groups = dense_output_1705_groups_1, pad = dense_output_1705_pad_1, pad_type = dense_output_1705_pad_type_1, strides = dense_output_1705_strides_1, weight = layers_21_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1705_cast_fp16")];
            string sparse_output_1705_pad_type_1 = const()[name = string("sparse_output_1705_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1705_strides_1 = const()[name = string("sparse_output_1705_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1705_pad_1 = const()[name = string("sparse_output_1705_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1705_dilations_1 = const()[name = string("sparse_output_1705_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1705_groups_1 = const()[name = string("sparse_output_1705_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629414656))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629411968))))[name = string("layers_21_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1705_cast_fp16 = conv(dilations = sparse_output_1705_dilations_1, groups = sparse_output_1705_groups_1, pad = sparse_output_1705_pad_1, pad_type = sparse_output_1705_pad_type_1, strides = sparse_output_1705_strides_1, weight = layers_21_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1705_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27139_cast_fp16 = add(x = dense_output_1705_cast_fp16, y = sparse_output_1705_cast_fp16)[name = string("op_27139_cast_fp16")];
            tensor<int32, [4]> var_27140 = const()[name = string("op_27140"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27142 = const()[name = string("op_27142"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27141_cast_fp16 = transpose(perm = var_27140, x = var_27139_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 51, 128]> q_head_339_cast_fp16 = reshape(shape = var_27142, x = var_27141_cast_fp16)[name = string("q_head_339_cast_fp16")];
            string dense_output_1707_pad_type_1 = const()[name = string("dense_output_1707_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1707_strides_1 = const()[name = string("dense_output_1707_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1707_pad_1 = const()[name = string("dense_output_1707_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1707_dilations_1 = const()[name = string("dense_output_1707_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1707_groups_1 = const()[name = string("dense_output_1707_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629431104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629562240))))[name = string("layers_21_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1707_cast_fp16 = conv(dilations = dense_output_1707_dilations_1, groups = dense_output_1707_groups_1, pad = dense_output_1707_pad_1, pad_type = dense_output_1707_pad_type_1, strides = dense_output_1707_strides_1, weight = layers_21_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1707_cast_fp16")];
            string sparse_output_1707_pad_type_1 = const()[name = string("sparse_output_1707_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1707_strides_1 = const()[name = string("sparse_output_1707_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1707_pad_1 = const()[name = string("sparse_output_1707_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1707_dilations_1 = const()[name = string("sparse_output_1707_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1707_groups_1 = const()[name = string("sparse_output_1707_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629565504))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629562816))))[name = string("layers_21_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1707_cast_fp16 = conv(dilations = sparse_output_1707_dilations_1, groups = sparse_output_1707_groups_1, pad = sparse_output_1707_pad_1, pad_type = sparse_output_1707_pad_type_1, strides = sparse_output_1707_strides_1, weight = layers_21_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1707_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27158_cast_fp16 = add(x = dense_output_1707_cast_fp16, y = sparse_output_1707_cast_fp16)[name = string("op_27158_cast_fp16")];
            tensor<int32, [4]> var_27159 = const()[name = string("op_27159"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27161 = const()[name = string("op_27161"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27160_cast_fp16 = transpose(perm = var_27159, x = var_27158_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 51, 128]> k_head_677_cast_fp16 = reshape(shape = var_27161, x = var_27160_cast_fp16)[name = string("k_head_677_cast_fp16")];
            string dense_output_1709_pad_type_1 = const()[name = string("dense_output_1709_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1709_strides_1 = const()[name = string("dense_output_1709_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1709_pad_1 = const()[name = string("dense_output_1709_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1709_dilations_1 = const()[name = string("dense_output_1709_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1709_groups_1 = const()[name = string("dense_output_1709_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629581952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629713088))))[name = string("layers_21_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1709_cast_fp16 = conv(dilations = dense_output_1709_dilations_1, groups = dense_output_1709_groups_1, pad = dense_output_1709_pad_1, pad_type = dense_output_1709_pad_type_1, strides = dense_output_1709_strides_1, weight = layers_21_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1709_cast_fp16")];
            string sparse_output_1709_pad_type_1 = const()[name = string("sparse_output_1709_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1709_strides_1 = const()[name = string("sparse_output_1709_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1709_pad_1 = const()[name = string("sparse_output_1709_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1709_dilations_1 = const()[name = string("sparse_output_1709_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1709_groups_1 = const()[name = string("sparse_output_1709_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629716352))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629713664))))[name = string("layers_21_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1709_cast_fp16 = conv(dilations = sparse_output_1709_dilations_1, groups = sparse_output_1709_groups_1, pad = sparse_output_1709_pad_1, pad_type = sparse_output_1709_pad_type_1, strides = sparse_output_1709_strides_1, weight = layers_21_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1709_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27177_cast_fp16 = add(x = dense_output_1709_cast_fp16, y = sparse_output_1709_cast_fp16)[name = string("op_27177_cast_fp16")];
            tensor<int32, [4]> var_27178 = const()[name = string("op_27178"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27180 = const()[name = string("op_27180"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27179_cast_fp16 = transpose(perm = var_27178, x = var_27177_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 51, 128]> v_head_677_cast_fp16 = reshape(shape = var_27180, x = var_27179_cast_fp16)[name = string("v_head_677_cast_fp16")];
            string dense_output_1711_pad_type_1 = const()[name = string("dense_output_1711_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1711_strides_1 = const()[name = string("dense_output_1711_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1711_pad_1 = const()[name = string("dense_output_1711_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1711_dilations_1 = const()[name = string("dense_output_1711_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1711_groups_1 = const()[name = string("dense_output_1711_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629732800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629863936))))[name = string("layers_21_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1711_cast_fp16 = conv(dilations = dense_output_1711_dilations_1, groups = dense_output_1711_groups_1, pad = dense_output_1711_pad_1, pad_type = dense_output_1711_pad_type_1, strides = dense_output_1711_strides_1, weight = layers_21_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1711_cast_fp16")];
            string sparse_output_1711_pad_type_1 = const()[name = string("sparse_output_1711_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1711_strides_1 = const()[name = string("sparse_output_1711_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1711_pad_1 = const()[name = string("sparse_output_1711_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1711_dilations_1 = const()[name = string("sparse_output_1711_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1711_groups_1 = const()[name = string("sparse_output_1711_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629867200))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629864512))))[name = string("layers_21_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1711_cast_fp16 = conv(dilations = sparse_output_1711_dilations_1, groups = sparse_output_1711_groups_1, pad = sparse_output_1711_pad_1, pad_type = sparse_output_1711_pad_type_1, strides = sparse_output_1711_strides_1, weight = layers_21_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1711_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_27196_cast_fp16 = add(x = dense_output_1711_cast_fp16, y = sparse_output_1711_cast_fp16)[name = string("op_27196_cast_fp16")];
            tensor<int32, [4]> var_27197 = const()[name = string("op_27197"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27199 = const()[name = string("op_27199"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_27198_cast_fp16 = transpose(perm = var_27197, x = var_27196_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 101, 128]> p_head_677_cast_fp16 = reshape(shape = var_27199, x = var_27198_cast_fp16)[name = string("p_head_677_cast_fp16")];
            tensor<fp16, [128]> var_27201_to_fp16 = const()[name = string("op_27201_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629883648)))];
            tensor<fp16, [1, 51, 128]> var_27202_cast_fp16 = add(x = q_head_339_cast_fp16, y = var_27201_to_fp16)[name = string("op_27202_cast_fp16")];
            tensor<int32, [1]> q_u_339_axes_1 = const()[name = string("q_u_339_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_339_cast_fp16 = expand_dims(axes = q_u_339_axes_1, x = var_27202_cast_fp16)[name = string("q_u_339_cast_fp16")];
            tensor<fp16, [128]> var_27204_to_fp16 = const()[name = string("op_27204_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629883968)))];
            tensor<fp16, [1, 51, 128]> var_27205_cast_fp16 = add(x = q_head_339_cast_fp16, y = var_27204_to_fp16)[name = string("op_27205_cast_fp16")];
            tensor<int32, [1]> q_v_339_axes_1 = const()[name = string("q_v_339_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_339_cast_fp16 = expand_dims(axes = q_v_339_axes_1, x = var_27205_cast_fp16)[name = string("q_v_339_cast_fp16")];
            tensor<int32, [1]> k_head_679_axes_1 = const()[name = string("k_head_679_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_679_cast_fp16 = expand_dims(axes = k_head_679_axes_1, x = k_head_677_cast_fp16)[name = string("k_head_679_cast_fp16")];
            tensor<int32, [1]> v_head_679_axes_1 = const()[name = string("v_head_679_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_679_cast_fp16 = expand_dims(axes = v_head_679_axes_1, x = v_head_677_cast_fp16)[name = string("v_head_679_cast_fp16")];
            tensor<int32, [1]> p_head_679_axes_1 = const()[name = string("p_head_679_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_679_cast_fp16 = expand_dims(axes = p_head_679_axes_1, x = p_head_677_cast_fp16)[name = string("p_head_679_cast_fp16")];
            bool var_27211_transpose_x_3 = const()[name = string("op_27211_transpose_x_3"), val = bool(false)];
            bool var_27211_transpose_y_3 = const()[name = string("op_27211_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_27211_cast_fp16 = matmul(transpose_x = var_27211_transpose_x_3, transpose_y = var_27211_transpose_y_3, x = q_u_339_cast_fp16, y = k_head_679_cast_fp16)[name = string("op_27211_cast_fp16")];
            fp16 var_27212_to_fp16 = const()[name = string("op_27212_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_339_cast_fp16 = mul(x = var_27211_cast_fp16, y = var_27212_to_fp16)[name = string("scores_content_339_cast_fp16")];
            bool x_1793_transpose_x_3 = const()[name = string("x_1793_transpose_x_3"), val = bool(false)];
            bool x_1793_transpose_y_3 = const()[name = string("x_1793_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1793_cast_fp16 = matmul(transpose_x = x_1793_transpose_x_3, transpose_y = x_1793_transpose_y_3, x = q_v_339_cast_fp16, y = p_head_679_cast_fp16)[name = string("x_1793_cast_fp16")];
            tensor<int32, [8]> x_1795_pad_1 = const()[name = string("x_1795_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1795_mode_1 = const()[name = string("x_1795_mode_1"), val = string("constant")];
            fp16 const_2449_to_fp16 = const()[name = string("const_2449_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1795_cast_fp16 = pad(constant_val = const_2449_to_fp16, mode = x_1795_mode_1, pad = x_1795_pad_1, x = x_1793_cast_fp16)[name = string("x_1795_cast_fp16")];
            tensor<int32, [4]> var_27226 = const()[name = string("op_27226"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1797_cast_fp16 = reshape(shape = var_27226, x = x_1795_cast_fp16)[name = string("x_1797_cast_fp16")];
            tensor<int32, [4]> var_27230_begin_1 = const()[name = string("op_27230_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_27230_end_1 = const()[name = string("op_27230_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_27230_end_mask_1 = const()[name = string("op_27230_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_27230_cast_fp16 = slice_by_index(begin = var_27230_begin_1, end = var_27230_end_1, end_mask = var_27230_end_mask_1, x = x_1797_cast_fp16)[name = string("op_27230_cast_fp16")];
            tensor<int32, [4]> var_27232 = const()[name = string("op_27232"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_27233_cast_fp16 = reshape(shape = var_27232, x = var_27230_cast_fp16)[name = string("op_27233_cast_fp16")];
            tensor<int32, [4]> var_27238_begin_1 = const()[name = string("op_27238_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_27238_end_1 = const()[name = string("op_27238_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_27238_end_mask_1 = const()[name = string("op_27238_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_27238_cast_fp16 = slice_by_index(begin = var_27238_begin_1, end = var_27238_end_1, end_mask = var_27238_end_mask_1, x = var_27233_cast_fp16)[name = string("op_27238_cast_fp16")];
            fp16 var_27239_to_fp16 = const()[name = string("op_27239_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_339_cast_fp16 = mul(x = var_27238_cast_fp16, y = var_27239_to_fp16)[name = string("scores_pos_339_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_339_cast_fp16 = add(x = scores_content_339_cast_fp16, y = scores_pos_339_cast_fp16)[name = string("logits_339_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_27242_cast_fp16 = softmax(axis = var_26846, x = logits_339_cast_fp16)[name = string("op_27242_cast_fp16")];
            bool var_27244_transpose_x_1 = const()[name = string("op_27244_transpose_x_1"), val = bool(false)];
            bool var_27244_transpose_y_1 = const()[name = string("op_27244_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_27244_cast_fp16 = matmul(transpose_x = var_27244_transpose_x_1, transpose_y = var_27244_transpose_y_1, x = var_27242_cast_fp16, y = v_head_679_cast_fp16)[name = string("op_27244_cast_fp16")];
            tensor<int32, [1]> var_27245_axes_1 = const()[name = string("op_27245_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_27245_cast_fp16 = squeeze(axes = var_27245_axes_1, x = var_27244_cast_fp16)[name = string("op_27245_cast_fp16")];
            string dense_output_1713_pad_type_1 = const()[name = string("dense_output_1713_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1713_strides_1 = const()[name = string("dense_output_1713_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1713_pad_1 = const()[name = string("dense_output_1713_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1713_dilations_1 = const()[name = string("dense_output_1713_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1713_groups_1 = const()[name = string("dense_output_1713_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629884288))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630015424))))[name = string("layers_21_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1713_cast_fp16 = conv(dilations = dense_output_1713_dilations_1, groups = dense_output_1713_groups_1, pad = dense_output_1713_pad_1, pad_type = dense_output_1713_pad_type_1, strides = dense_output_1713_strides_1, weight = layers_21_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1713_cast_fp16")];
            string sparse_output_1713_pad_type_1 = const()[name = string("sparse_output_1713_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1713_strides_1 = const()[name = string("sparse_output_1713_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1713_pad_1 = const()[name = string("sparse_output_1713_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1713_dilations_1 = const()[name = string("sparse_output_1713_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1713_groups_1 = const()[name = string("sparse_output_1713_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630018688))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630016000))))[name = string("layers_21_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1713_cast_fp16 = conv(dilations = sparse_output_1713_dilations_1, groups = sparse_output_1713_groups_1, pad = sparse_output_1713_pad_1, pad_type = sparse_output_1713_pad_type_1, strides = sparse_output_1713_strides_1, weight = layers_21_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1713_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27260_cast_fp16 = add(x = dense_output_1713_cast_fp16, y = sparse_output_1713_cast_fp16)[name = string("op_27260_cast_fp16")];
            tensor<int32, [4]> var_27261 = const()[name = string("op_27261"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27263 = const()[name = string("op_27263"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27262_cast_fp16 = transpose(perm = var_27261, x = var_27260_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 51, 128]> q_head_341_cast_fp16 = reshape(shape = var_27263, x = var_27262_cast_fp16)[name = string("q_head_341_cast_fp16")];
            string dense_output_1715_pad_type_1 = const()[name = string("dense_output_1715_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1715_strides_1 = const()[name = string("dense_output_1715_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1715_pad_1 = const()[name = string("dense_output_1715_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1715_dilations_1 = const()[name = string("dense_output_1715_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1715_groups_1 = const()[name = string("dense_output_1715_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630035136))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630166272))))[name = string("layers_21_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1715_cast_fp16 = conv(dilations = dense_output_1715_dilations_1, groups = dense_output_1715_groups_1, pad = dense_output_1715_pad_1, pad_type = dense_output_1715_pad_type_1, strides = dense_output_1715_strides_1, weight = layers_21_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1715_cast_fp16")];
            string sparse_output_1715_pad_type_1 = const()[name = string("sparse_output_1715_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1715_strides_1 = const()[name = string("sparse_output_1715_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1715_pad_1 = const()[name = string("sparse_output_1715_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1715_dilations_1 = const()[name = string("sparse_output_1715_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1715_groups_1 = const()[name = string("sparse_output_1715_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630169536))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630166848))))[name = string("layers_21_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1715_cast_fp16 = conv(dilations = sparse_output_1715_dilations_1, groups = sparse_output_1715_groups_1, pad = sparse_output_1715_pad_1, pad_type = sparse_output_1715_pad_type_1, strides = sparse_output_1715_strides_1, weight = layers_21_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1715_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27279_cast_fp16 = add(x = dense_output_1715_cast_fp16, y = sparse_output_1715_cast_fp16)[name = string("op_27279_cast_fp16")];
            tensor<int32, [4]> var_27280 = const()[name = string("op_27280"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27282 = const()[name = string("op_27282"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27281_cast_fp16 = transpose(perm = var_27280, x = var_27279_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 51, 128]> k_head_681_cast_fp16 = reshape(shape = var_27282, x = var_27281_cast_fp16)[name = string("k_head_681_cast_fp16")];
            string dense_output_1717_pad_type_1 = const()[name = string("dense_output_1717_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1717_strides_1 = const()[name = string("dense_output_1717_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1717_pad_1 = const()[name = string("dense_output_1717_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1717_dilations_1 = const()[name = string("dense_output_1717_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1717_groups_1 = const()[name = string("dense_output_1717_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630185984))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630317120))))[name = string("layers_21_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1717_cast_fp16 = conv(dilations = dense_output_1717_dilations_1, groups = dense_output_1717_groups_1, pad = dense_output_1717_pad_1, pad_type = dense_output_1717_pad_type_1, strides = dense_output_1717_strides_1, weight = layers_21_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1717_cast_fp16")];
            string sparse_output_1717_pad_type_1 = const()[name = string("sparse_output_1717_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1717_strides_1 = const()[name = string("sparse_output_1717_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1717_pad_1 = const()[name = string("sparse_output_1717_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1717_dilations_1 = const()[name = string("sparse_output_1717_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1717_groups_1 = const()[name = string("sparse_output_1717_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630320384))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630317696))))[name = string("layers_21_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1717_cast_fp16 = conv(dilations = sparse_output_1717_dilations_1, groups = sparse_output_1717_groups_1, pad = sparse_output_1717_pad_1, pad_type = sparse_output_1717_pad_type_1, strides = sparse_output_1717_strides_1, weight = layers_21_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1717_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27298_cast_fp16 = add(x = dense_output_1717_cast_fp16, y = sparse_output_1717_cast_fp16)[name = string("op_27298_cast_fp16")];
            tensor<int32, [4]> var_27299 = const()[name = string("op_27299"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27301 = const()[name = string("op_27301"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27300_cast_fp16 = transpose(perm = var_27299, x = var_27298_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 51, 128]> v_head_681_cast_fp16 = reshape(shape = var_27301, x = var_27300_cast_fp16)[name = string("v_head_681_cast_fp16")];
            string dense_output_1719_pad_type_1 = const()[name = string("dense_output_1719_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1719_strides_1 = const()[name = string("dense_output_1719_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1719_pad_1 = const()[name = string("dense_output_1719_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1719_dilations_1 = const()[name = string("dense_output_1719_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1719_groups_1 = const()[name = string("dense_output_1719_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630336832))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630467968))))[name = string("layers_21_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1719_cast_fp16 = conv(dilations = dense_output_1719_dilations_1, groups = dense_output_1719_groups_1, pad = dense_output_1719_pad_1, pad_type = dense_output_1719_pad_type_1, strides = dense_output_1719_strides_1, weight = layers_21_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1719_cast_fp16")];
            string sparse_output_1719_pad_type_1 = const()[name = string("sparse_output_1719_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1719_strides_1 = const()[name = string("sparse_output_1719_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1719_pad_1 = const()[name = string("sparse_output_1719_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1719_dilations_1 = const()[name = string("sparse_output_1719_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1719_groups_1 = const()[name = string("sparse_output_1719_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630471232))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630468544))))[name = string("layers_21_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1719_cast_fp16 = conv(dilations = sparse_output_1719_dilations_1, groups = sparse_output_1719_groups_1, pad = sparse_output_1719_pad_1, pad_type = sparse_output_1719_pad_type_1, strides = sparse_output_1719_strides_1, weight = layers_21_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1719_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_27317_cast_fp16 = add(x = dense_output_1719_cast_fp16, y = sparse_output_1719_cast_fp16)[name = string("op_27317_cast_fp16")];
            tensor<int32, [4]> var_27318 = const()[name = string("op_27318"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27320 = const()[name = string("op_27320"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_27319_cast_fp16 = transpose(perm = var_27318, x = var_27317_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 101, 128]> p_head_681_cast_fp16 = reshape(shape = var_27320, x = var_27319_cast_fp16)[name = string("p_head_681_cast_fp16")];
            tensor<fp16, [128]> var_27322_to_fp16 = const()[name = string("op_27322_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630487680)))];
            tensor<fp16, [1, 51, 128]> var_27323_cast_fp16 = add(x = q_head_341_cast_fp16, y = var_27322_to_fp16)[name = string("op_27323_cast_fp16")];
            tensor<int32, [1]> q_u_341_axes_1 = const()[name = string("q_u_341_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_341_cast_fp16 = expand_dims(axes = q_u_341_axes_1, x = var_27323_cast_fp16)[name = string("q_u_341_cast_fp16")];
            tensor<fp16, [128]> var_27325_to_fp16 = const()[name = string("op_27325_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630488000)))];
            tensor<fp16, [1, 51, 128]> var_27326_cast_fp16 = add(x = q_head_341_cast_fp16, y = var_27325_to_fp16)[name = string("op_27326_cast_fp16")];
            tensor<int32, [1]> q_v_341_axes_1 = const()[name = string("q_v_341_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_341_cast_fp16 = expand_dims(axes = q_v_341_axes_1, x = var_27326_cast_fp16)[name = string("q_v_341_cast_fp16")];
            tensor<int32, [1]> k_head_683_axes_1 = const()[name = string("k_head_683_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_683_cast_fp16 = expand_dims(axes = k_head_683_axes_1, x = k_head_681_cast_fp16)[name = string("k_head_683_cast_fp16")];
            tensor<int32, [1]> v_head_683_axes_1 = const()[name = string("v_head_683_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_683_cast_fp16 = expand_dims(axes = v_head_683_axes_1, x = v_head_681_cast_fp16)[name = string("v_head_683_cast_fp16")];
            tensor<int32, [1]> p_head_683_axes_1 = const()[name = string("p_head_683_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_683_cast_fp16 = expand_dims(axes = p_head_683_axes_1, x = p_head_681_cast_fp16)[name = string("p_head_683_cast_fp16")];
            bool var_27332_transpose_x_3 = const()[name = string("op_27332_transpose_x_3"), val = bool(false)];
            bool var_27332_transpose_y_3 = const()[name = string("op_27332_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_27332_cast_fp16 = matmul(transpose_x = var_27332_transpose_x_3, transpose_y = var_27332_transpose_y_3, x = q_u_341_cast_fp16, y = k_head_683_cast_fp16)[name = string("op_27332_cast_fp16")];
            fp16 var_27333_to_fp16 = const()[name = string("op_27333_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_341_cast_fp16 = mul(x = var_27332_cast_fp16, y = var_27333_to_fp16)[name = string("scores_content_341_cast_fp16")];
            bool x_1801_transpose_x_3 = const()[name = string("x_1801_transpose_x_3"), val = bool(false)];
            bool x_1801_transpose_y_3 = const()[name = string("x_1801_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1801_cast_fp16 = matmul(transpose_x = x_1801_transpose_x_3, transpose_y = x_1801_transpose_y_3, x = q_v_341_cast_fp16, y = p_head_683_cast_fp16)[name = string("x_1801_cast_fp16")];
            tensor<int32, [8]> x_1803_pad_1 = const()[name = string("x_1803_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1803_mode_1 = const()[name = string("x_1803_mode_1"), val = string("constant")];
            fp16 const_2455_to_fp16 = const()[name = string("const_2455_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1803_cast_fp16 = pad(constant_val = const_2455_to_fp16, mode = x_1803_mode_1, pad = x_1803_pad_1, x = x_1801_cast_fp16)[name = string("x_1803_cast_fp16")];
            tensor<int32, [4]> var_27347 = const()[name = string("op_27347"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1805_cast_fp16 = reshape(shape = var_27347, x = x_1803_cast_fp16)[name = string("x_1805_cast_fp16")];
            tensor<int32, [4]> var_27351_begin_1 = const()[name = string("op_27351_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_27351_end_1 = const()[name = string("op_27351_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_27351_end_mask_1 = const()[name = string("op_27351_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_27351_cast_fp16 = slice_by_index(begin = var_27351_begin_1, end = var_27351_end_1, end_mask = var_27351_end_mask_1, x = x_1805_cast_fp16)[name = string("op_27351_cast_fp16")];
            tensor<int32, [4]> var_27353 = const()[name = string("op_27353"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_27354_cast_fp16 = reshape(shape = var_27353, x = var_27351_cast_fp16)[name = string("op_27354_cast_fp16")];
            tensor<int32, [4]> var_27359_begin_1 = const()[name = string("op_27359_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_27359_end_1 = const()[name = string("op_27359_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_27359_end_mask_1 = const()[name = string("op_27359_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_27359_cast_fp16 = slice_by_index(begin = var_27359_begin_1, end = var_27359_end_1, end_mask = var_27359_end_mask_1, x = var_27354_cast_fp16)[name = string("op_27359_cast_fp16")];
            fp16 var_27360_to_fp16 = const()[name = string("op_27360_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_341_cast_fp16 = mul(x = var_27359_cast_fp16, y = var_27360_to_fp16)[name = string("scores_pos_341_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_341_cast_fp16 = add(x = scores_content_341_cast_fp16, y = scores_pos_341_cast_fp16)[name = string("logits_341_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_27363_cast_fp16 = softmax(axis = var_26846, x = logits_341_cast_fp16)[name = string("op_27363_cast_fp16")];
            bool var_27365_transpose_x_1 = const()[name = string("op_27365_transpose_x_1"), val = bool(false)];
            bool var_27365_transpose_y_1 = const()[name = string("op_27365_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_27365_cast_fp16 = matmul(transpose_x = var_27365_transpose_x_1, transpose_y = var_27365_transpose_y_1, x = var_27363_cast_fp16, y = v_head_683_cast_fp16)[name = string("op_27365_cast_fp16")];
            tensor<int32, [1]> var_27366_axes_1 = const()[name = string("op_27366_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_27366_cast_fp16 = squeeze(axes = var_27366_axes_1, x = var_27365_cast_fp16)[name = string("op_27366_cast_fp16")];
            string dense_output_1721_pad_type_1 = const()[name = string("dense_output_1721_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1721_strides_1 = const()[name = string("dense_output_1721_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1721_pad_1 = const()[name = string("dense_output_1721_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1721_dilations_1 = const()[name = string("dense_output_1721_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1721_groups_1 = const()[name = string("dense_output_1721_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630488320))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630619456))))[name = string("layers_21_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1721_cast_fp16 = conv(dilations = dense_output_1721_dilations_1, groups = dense_output_1721_groups_1, pad = dense_output_1721_pad_1, pad_type = dense_output_1721_pad_type_1, strides = dense_output_1721_strides_1, weight = layers_21_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1721_cast_fp16")];
            string sparse_output_1721_pad_type_1 = const()[name = string("sparse_output_1721_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1721_strides_1 = const()[name = string("sparse_output_1721_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1721_pad_1 = const()[name = string("sparse_output_1721_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1721_dilations_1 = const()[name = string("sparse_output_1721_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1721_groups_1 = const()[name = string("sparse_output_1721_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630622720))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630620032))))[name = string("layers_21_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1721_cast_fp16 = conv(dilations = sparse_output_1721_dilations_1, groups = sparse_output_1721_groups_1, pad = sparse_output_1721_pad_1, pad_type = sparse_output_1721_pad_type_1, strides = sparse_output_1721_strides_1, weight = layers_21_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1721_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27381_cast_fp16 = add(x = dense_output_1721_cast_fp16, y = sparse_output_1721_cast_fp16)[name = string("op_27381_cast_fp16")];
            tensor<int32, [4]> var_27382 = const()[name = string("op_27382"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27384 = const()[name = string("op_27384"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27383_cast_fp16 = transpose(perm = var_27382, x = var_27381_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 51, 128]> q_head_343_cast_fp16 = reshape(shape = var_27384, x = var_27383_cast_fp16)[name = string("q_head_343_cast_fp16")];
            string dense_output_1723_pad_type_1 = const()[name = string("dense_output_1723_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1723_strides_1 = const()[name = string("dense_output_1723_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1723_pad_1 = const()[name = string("dense_output_1723_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1723_dilations_1 = const()[name = string("dense_output_1723_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1723_groups_1 = const()[name = string("dense_output_1723_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630639168))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630770304))))[name = string("layers_21_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1723_cast_fp16 = conv(dilations = dense_output_1723_dilations_1, groups = dense_output_1723_groups_1, pad = dense_output_1723_pad_1, pad_type = dense_output_1723_pad_type_1, strides = dense_output_1723_strides_1, weight = layers_21_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1723_cast_fp16")];
            string sparse_output_1723_pad_type_1 = const()[name = string("sparse_output_1723_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1723_strides_1 = const()[name = string("sparse_output_1723_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1723_pad_1 = const()[name = string("sparse_output_1723_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1723_dilations_1 = const()[name = string("sparse_output_1723_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1723_groups_1 = const()[name = string("sparse_output_1723_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630773568))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630770880))))[name = string("layers_21_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1723_cast_fp16 = conv(dilations = sparse_output_1723_dilations_1, groups = sparse_output_1723_groups_1, pad = sparse_output_1723_pad_1, pad_type = sparse_output_1723_pad_type_1, strides = sparse_output_1723_strides_1, weight = layers_21_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1723_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27400_cast_fp16 = add(x = dense_output_1723_cast_fp16, y = sparse_output_1723_cast_fp16)[name = string("op_27400_cast_fp16")];
            tensor<int32, [4]> var_27401 = const()[name = string("op_27401"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27403 = const()[name = string("op_27403"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27402_cast_fp16 = transpose(perm = var_27401, x = var_27400_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 51, 128]> k_head_685_cast_fp16 = reshape(shape = var_27403, x = var_27402_cast_fp16)[name = string("k_head_685_cast_fp16")];
            string dense_output_1725_pad_type_1 = const()[name = string("dense_output_1725_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1725_strides_1 = const()[name = string("dense_output_1725_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1725_pad_1 = const()[name = string("dense_output_1725_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1725_dilations_1 = const()[name = string("dense_output_1725_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1725_groups_1 = const()[name = string("dense_output_1725_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630790016))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630921152))))[name = string("layers_21_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1725_cast_fp16 = conv(dilations = dense_output_1725_dilations_1, groups = dense_output_1725_groups_1, pad = dense_output_1725_pad_1, pad_type = dense_output_1725_pad_type_1, strides = dense_output_1725_strides_1, weight = layers_21_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1725_cast_fp16")];
            string sparse_output_1725_pad_type_1 = const()[name = string("sparse_output_1725_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1725_strides_1 = const()[name = string("sparse_output_1725_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1725_pad_1 = const()[name = string("sparse_output_1725_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1725_dilations_1 = const()[name = string("sparse_output_1725_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1725_groups_1 = const()[name = string("sparse_output_1725_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630924416))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630921728))))[name = string("layers_21_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1725_cast_fp16 = conv(dilations = sparse_output_1725_dilations_1, groups = sparse_output_1725_groups_1, pad = sparse_output_1725_pad_1, pad_type = sparse_output_1725_pad_type_1, strides = sparse_output_1725_strides_1, weight = layers_21_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1725_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27419_cast_fp16 = add(x = dense_output_1725_cast_fp16, y = sparse_output_1725_cast_fp16)[name = string("op_27419_cast_fp16")];
            tensor<int32, [4]> var_27420 = const()[name = string("op_27420"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27422 = const()[name = string("op_27422"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27421_cast_fp16 = transpose(perm = var_27420, x = var_27419_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 51, 128]> v_head_685_cast_fp16 = reshape(shape = var_27422, x = var_27421_cast_fp16)[name = string("v_head_685_cast_fp16")];
            string dense_output_1727_pad_type_1 = const()[name = string("dense_output_1727_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1727_strides_1 = const()[name = string("dense_output_1727_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1727_pad_1 = const()[name = string("dense_output_1727_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1727_dilations_1 = const()[name = string("dense_output_1727_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1727_groups_1 = const()[name = string("dense_output_1727_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630940864))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631072000))))[name = string("layers_21_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1727_cast_fp16 = conv(dilations = dense_output_1727_dilations_1, groups = dense_output_1727_groups_1, pad = dense_output_1727_pad_1, pad_type = dense_output_1727_pad_type_1, strides = dense_output_1727_strides_1, weight = layers_21_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1727_cast_fp16")];
            string sparse_output_1727_pad_type_1 = const()[name = string("sparse_output_1727_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1727_strides_1 = const()[name = string("sparse_output_1727_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1727_pad_1 = const()[name = string("sparse_output_1727_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1727_dilations_1 = const()[name = string("sparse_output_1727_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1727_groups_1 = const()[name = string("sparse_output_1727_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631075264))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631072576))))[name = string("layers_21_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1727_cast_fp16 = conv(dilations = sparse_output_1727_dilations_1, groups = sparse_output_1727_groups_1, pad = sparse_output_1727_pad_1, pad_type = sparse_output_1727_pad_type_1, strides = sparse_output_1727_strides_1, weight = layers_21_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1727_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_27438_cast_fp16 = add(x = dense_output_1727_cast_fp16, y = sparse_output_1727_cast_fp16)[name = string("op_27438_cast_fp16")];
            tensor<int32, [4]> var_27439 = const()[name = string("op_27439"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27441 = const()[name = string("op_27441"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_27440_cast_fp16 = transpose(perm = var_27439, x = var_27438_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 101, 128]> p_head_685_cast_fp16 = reshape(shape = var_27441, x = var_27440_cast_fp16)[name = string("p_head_685_cast_fp16")];
            tensor<fp16, [128]> var_27443_to_fp16 = const()[name = string("op_27443_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631091712)))];
            tensor<fp16, [1, 51, 128]> var_27444_cast_fp16 = add(x = q_head_343_cast_fp16, y = var_27443_to_fp16)[name = string("op_27444_cast_fp16")];
            tensor<int32, [1]> q_u_343_axes_1 = const()[name = string("q_u_343_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_343_cast_fp16 = expand_dims(axes = q_u_343_axes_1, x = var_27444_cast_fp16)[name = string("q_u_343_cast_fp16")];
            tensor<fp16, [128]> var_27446_to_fp16 = const()[name = string("op_27446_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631092032)))];
            tensor<fp16, [1, 51, 128]> var_27447_cast_fp16 = add(x = q_head_343_cast_fp16, y = var_27446_to_fp16)[name = string("op_27447_cast_fp16")];
            tensor<int32, [1]> q_v_343_axes_1 = const()[name = string("q_v_343_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_343_cast_fp16 = expand_dims(axes = q_v_343_axes_1, x = var_27447_cast_fp16)[name = string("q_v_343_cast_fp16")];
            tensor<int32, [1]> k_head_687_axes_1 = const()[name = string("k_head_687_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_687_cast_fp16 = expand_dims(axes = k_head_687_axes_1, x = k_head_685_cast_fp16)[name = string("k_head_687_cast_fp16")];
            tensor<int32, [1]> v_head_687_axes_1 = const()[name = string("v_head_687_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_687_cast_fp16 = expand_dims(axes = v_head_687_axes_1, x = v_head_685_cast_fp16)[name = string("v_head_687_cast_fp16")];
            tensor<int32, [1]> p_head_687_axes_1 = const()[name = string("p_head_687_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_687_cast_fp16 = expand_dims(axes = p_head_687_axes_1, x = p_head_685_cast_fp16)[name = string("p_head_687_cast_fp16")];
            bool var_27453_transpose_x_3 = const()[name = string("op_27453_transpose_x_3"), val = bool(false)];
            bool var_27453_transpose_y_3 = const()[name = string("op_27453_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_27453_cast_fp16 = matmul(transpose_x = var_27453_transpose_x_3, transpose_y = var_27453_transpose_y_3, x = q_u_343_cast_fp16, y = k_head_687_cast_fp16)[name = string("op_27453_cast_fp16")];
            fp16 var_27454_to_fp16 = const()[name = string("op_27454_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_343_cast_fp16 = mul(x = var_27453_cast_fp16, y = var_27454_to_fp16)[name = string("scores_content_343_cast_fp16")];
            bool x_1809_transpose_x_3 = const()[name = string("x_1809_transpose_x_3"), val = bool(false)];
            bool x_1809_transpose_y_3 = const()[name = string("x_1809_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1809_cast_fp16 = matmul(transpose_x = x_1809_transpose_x_3, transpose_y = x_1809_transpose_y_3, x = q_v_343_cast_fp16, y = p_head_687_cast_fp16)[name = string("x_1809_cast_fp16")];
            tensor<int32, [8]> x_1811_pad_1 = const()[name = string("x_1811_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1811_mode_1 = const()[name = string("x_1811_mode_1"), val = string("constant")];
            fp16 const_2461_to_fp16 = const()[name = string("const_2461_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1811_cast_fp16 = pad(constant_val = const_2461_to_fp16, mode = x_1811_mode_1, pad = x_1811_pad_1, x = x_1809_cast_fp16)[name = string("x_1811_cast_fp16")];
            tensor<int32, [4]> var_27468 = const()[name = string("op_27468"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1813_cast_fp16 = reshape(shape = var_27468, x = x_1811_cast_fp16)[name = string("x_1813_cast_fp16")];
            tensor<int32, [4]> var_27472_begin_1 = const()[name = string("op_27472_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_27472_end_1 = const()[name = string("op_27472_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_27472_end_mask_1 = const()[name = string("op_27472_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_27472_cast_fp16 = slice_by_index(begin = var_27472_begin_1, end = var_27472_end_1, end_mask = var_27472_end_mask_1, x = x_1813_cast_fp16)[name = string("op_27472_cast_fp16")];
            tensor<int32, [4]> var_27474 = const()[name = string("op_27474"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_27475_cast_fp16 = reshape(shape = var_27474, x = var_27472_cast_fp16)[name = string("op_27475_cast_fp16")];
            tensor<int32, [4]> var_27480_begin_1 = const()[name = string("op_27480_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_27480_end_1 = const()[name = string("op_27480_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_27480_end_mask_1 = const()[name = string("op_27480_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_27480_cast_fp16 = slice_by_index(begin = var_27480_begin_1, end = var_27480_end_1, end_mask = var_27480_end_mask_1, x = var_27475_cast_fp16)[name = string("op_27480_cast_fp16")];
            fp16 var_27481_to_fp16 = const()[name = string("op_27481_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_343_cast_fp16 = mul(x = var_27480_cast_fp16, y = var_27481_to_fp16)[name = string("scores_pos_343_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_343_cast_fp16 = add(x = scores_content_343_cast_fp16, y = scores_pos_343_cast_fp16)[name = string("logits_343_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_27484_cast_fp16 = softmax(axis = var_26846, x = logits_343_cast_fp16)[name = string("op_27484_cast_fp16")];
            bool var_27486_transpose_x_1 = const()[name = string("op_27486_transpose_x_1"), val = bool(false)];
            bool var_27486_transpose_y_1 = const()[name = string("op_27486_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_27486_cast_fp16 = matmul(transpose_x = var_27486_transpose_x_1, transpose_y = var_27486_transpose_y_1, x = var_27484_cast_fp16, y = v_head_687_cast_fp16)[name = string("op_27486_cast_fp16")];
            tensor<int32, [1]> var_27487_axes_1 = const()[name = string("op_27487_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_27487_cast_fp16 = squeeze(axes = var_27487_axes_1, x = var_27486_cast_fp16)[name = string("op_27487_cast_fp16")];
            string dense_output_1729_pad_type_1 = const()[name = string("dense_output_1729_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1729_strides_1 = const()[name = string("dense_output_1729_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1729_pad_1 = const()[name = string("dense_output_1729_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1729_dilations_1 = const()[name = string("dense_output_1729_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1729_groups_1 = const()[name = string("dense_output_1729_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631092352))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631223488))))[name = string("layers_21_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1729_cast_fp16 = conv(dilations = dense_output_1729_dilations_1, groups = dense_output_1729_groups_1, pad = dense_output_1729_pad_1, pad_type = dense_output_1729_pad_type_1, strides = dense_output_1729_strides_1, weight = layers_21_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1729_cast_fp16")];
            string sparse_output_1729_pad_type_1 = const()[name = string("sparse_output_1729_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1729_strides_1 = const()[name = string("sparse_output_1729_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1729_pad_1 = const()[name = string("sparse_output_1729_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1729_dilations_1 = const()[name = string("sparse_output_1729_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1729_groups_1 = const()[name = string("sparse_output_1729_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631226752))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631224064))))[name = string("layers_21_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1729_cast_fp16 = conv(dilations = sparse_output_1729_dilations_1, groups = sparse_output_1729_groups_1, pad = sparse_output_1729_pad_1, pad_type = sparse_output_1729_pad_type_1, strides = sparse_output_1729_strides_1, weight = layers_21_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1729_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27502_cast_fp16 = add(x = dense_output_1729_cast_fp16, y = sparse_output_1729_cast_fp16)[name = string("op_27502_cast_fp16")];
            tensor<int32, [4]> var_27503 = const()[name = string("op_27503"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27505 = const()[name = string("op_27505"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27504_cast_fp16 = transpose(perm = var_27503, x = var_27502_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 51, 128]> q_head_345_cast_fp16 = reshape(shape = var_27505, x = var_27504_cast_fp16)[name = string("q_head_345_cast_fp16")];
            string dense_output_1731_pad_type_1 = const()[name = string("dense_output_1731_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1731_strides_1 = const()[name = string("dense_output_1731_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1731_pad_1 = const()[name = string("dense_output_1731_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1731_dilations_1 = const()[name = string("dense_output_1731_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1731_groups_1 = const()[name = string("dense_output_1731_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631243200))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631374336))))[name = string("layers_21_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1731_cast_fp16 = conv(dilations = dense_output_1731_dilations_1, groups = dense_output_1731_groups_1, pad = dense_output_1731_pad_1, pad_type = dense_output_1731_pad_type_1, strides = dense_output_1731_strides_1, weight = layers_21_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1731_cast_fp16")];
            string sparse_output_1731_pad_type_1 = const()[name = string("sparse_output_1731_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1731_strides_1 = const()[name = string("sparse_output_1731_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1731_pad_1 = const()[name = string("sparse_output_1731_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1731_dilations_1 = const()[name = string("sparse_output_1731_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1731_groups_1 = const()[name = string("sparse_output_1731_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631377600))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631374912))))[name = string("layers_21_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1731_cast_fp16 = conv(dilations = sparse_output_1731_dilations_1, groups = sparse_output_1731_groups_1, pad = sparse_output_1731_pad_1, pad_type = sparse_output_1731_pad_type_1, strides = sparse_output_1731_strides_1, weight = layers_21_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1731_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27521_cast_fp16 = add(x = dense_output_1731_cast_fp16, y = sparse_output_1731_cast_fp16)[name = string("op_27521_cast_fp16")];
            tensor<int32, [4]> var_27522 = const()[name = string("op_27522"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27524 = const()[name = string("op_27524"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27523_cast_fp16 = transpose(perm = var_27522, x = var_27521_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 51, 128]> k_head_689_cast_fp16 = reshape(shape = var_27524, x = var_27523_cast_fp16)[name = string("k_head_689_cast_fp16")];
            string dense_output_1733_pad_type_1 = const()[name = string("dense_output_1733_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1733_strides_1 = const()[name = string("dense_output_1733_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1733_pad_1 = const()[name = string("dense_output_1733_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1733_dilations_1 = const()[name = string("dense_output_1733_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1733_groups_1 = const()[name = string("dense_output_1733_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631394048))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631525184))))[name = string("layers_21_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1733_cast_fp16 = conv(dilations = dense_output_1733_dilations_1, groups = dense_output_1733_groups_1, pad = dense_output_1733_pad_1, pad_type = dense_output_1733_pad_type_1, strides = dense_output_1733_strides_1, weight = layers_21_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1733_cast_fp16")];
            string sparse_output_1733_pad_type_1 = const()[name = string("sparse_output_1733_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1733_strides_1 = const()[name = string("sparse_output_1733_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1733_pad_1 = const()[name = string("sparse_output_1733_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1733_dilations_1 = const()[name = string("sparse_output_1733_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1733_groups_1 = const()[name = string("sparse_output_1733_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631528448))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631525760))))[name = string("layers_21_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1733_cast_fp16 = conv(dilations = sparse_output_1733_dilations_1, groups = sparse_output_1733_groups_1, pad = sparse_output_1733_pad_1, pad_type = sparse_output_1733_pad_type_1, strides = sparse_output_1733_strides_1, weight = layers_21_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1733_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27540_cast_fp16 = add(x = dense_output_1733_cast_fp16, y = sparse_output_1733_cast_fp16)[name = string("op_27540_cast_fp16")];
            tensor<int32, [4]> var_27541 = const()[name = string("op_27541"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27543 = const()[name = string("op_27543"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27542_cast_fp16 = transpose(perm = var_27541, x = var_27540_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 51, 128]> v_head_689_cast_fp16 = reshape(shape = var_27543, x = var_27542_cast_fp16)[name = string("v_head_689_cast_fp16")];
            string dense_output_1735_pad_type_1 = const()[name = string("dense_output_1735_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1735_strides_1 = const()[name = string("dense_output_1735_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1735_pad_1 = const()[name = string("dense_output_1735_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1735_dilations_1 = const()[name = string("dense_output_1735_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1735_groups_1 = const()[name = string("dense_output_1735_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631544896))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631676032))))[name = string("layers_21_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1735_cast_fp16 = conv(dilations = dense_output_1735_dilations_1, groups = dense_output_1735_groups_1, pad = dense_output_1735_pad_1, pad_type = dense_output_1735_pad_type_1, strides = dense_output_1735_strides_1, weight = layers_21_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1735_cast_fp16")];
            string sparse_output_1735_pad_type_1 = const()[name = string("sparse_output_1735_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1735_strides_1 = const()[name = string("sparse_output_1735_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1735_pad_1 = const()[name = string("sparse_output_1735_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1735_dilations_1 = const()[name = string("sparse_output_1735_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1735_groups_1 = const()[name = string("sparse_output_1735_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631679296))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631676608))))[name = string("layers_21_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1735_cast_fp16 = conv(dilations = sparse_output_1735_dilations_1, groups = sparse_output_1735_groups_1, pad = sparse_output_1735_pad_1, pad_type = sparse_output_1735_pad_type_1, strides = sparse_output_1735_strides_1, weight = layers_21_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1735_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_27559_cast_fp16 = add(x = dense_output_1735_cast_fp16, y = sparse_output_1735_cast_fp16)[name = string("op_27559_cast_fp16")];
            tensor<int32, [4]> var_27560 = const()[name = string("op_27560"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27562 = const()[name = string("op_27562"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_27561_cast_fp16 = transpose(perm = var_27560, x = var_27559_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 101, 128]> p_head_689_cast_fp16 = reshape(shape = var_27562, x = var_27561_cast_fp16)[name = string("p_head_689_cast_fp16")];
            tensor<fp16, [128]> var_27564_to_fp16 = const()[name = string("op_27564_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631695744)))];
            tensor<fp16, [1, 51, 128]> var_27565_cast_fp16 = add(x = q_head_345_cast_fp16, y = var_27564_to_fp16)[name = string("op_27565_cast_fp16")];
            tensor<int32, [1]> q_u_345_axes_1 = const()[name = string("q_u_345_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_345_cast_fp16 = expand_dims(axes = q_u_345_axes_1, x = var_27565_cast_fp16)[name = string("q_u_345_cast_fp16")];
            tensor<fp16, [128]> var_27567_to_fp16 = const()[name = string("op_27567_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631696064)))];
            tensor<fp16, [1, 51, 128]> var_27568_cast_fp16 = add(x = q_head_345_cast_fp16, y = var_27567_to_fp16)[name = string("op_27568_cast_fp16")];
            tensor<int32, [1]> q_v_345_axes_1 = const()[name = string("q_v_345_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_345_cast_fp16 = expand_dims(axes = q_v_345_axes_1, x = var_27568_cast_fp16)[name = string("q_v_345_cast_fp16")];
            tensor<int32, [1]> k_head_691_axes_1 = const()[name = string("k_head_691_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_691_cast_fp16 = expand_dims(axes = k_head_691_axes_1, x = k_head_689_cast_fp16)[name = string("k_head_691_cast_fp16")];
            tensor<int32, [1]> v_head_691_axes_1 = const()[name = string("v_head_691_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_691_cast_fp16 = expand_dims(axes = v_head_691_axes_1, x = v_head_689_cast_fp16)[name = string("v_head_691_cast_fp16")];
            tensor<int32, [1]> p_head_691_axes_1 = const()[name = string("p_head_691_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_691_cast_fp16 = expand_dims(axes = p_head_691_axes_1, x = p_head_689_cast_fp16)[name = string("p_head_691_cast_fp16")];
            bool var_27574_transpose_x_3 = const()[name = string("op_27574_transpose_x_3"), val = bool(false)];
            bool var_27574_transpose_y_3 = const()[name = string("op_27574_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_27574_cast_fp16 = matmul(transpose_x = var_27574_transpose_x_3, transpose_y = var_27574_transpose_y_3, x = q_u_345_cast_fp16, y = k_head_691_cast_fp16)[name = string("op_27574_cast_fp16")];
            fp16 var_27575_to_fp16 = const()[name = string("op_27575_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_345_cast_fp16 = mul(x = var_27574_cast_fp16, y = var_27575_to_fp16)[name = string("scores_content_345_cast_fp16")];
            bool x_1817_transpose_x_3 = const()[name = string("x_1817_transpose_x_3"), val = bool(false)];
            bool x_1817_transpose_y_3 = const()[name = string("x_1817_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1817_cast_fp16 = matmul(transpose_x = x_1817_transpose_x_3, transpose_y = x_1817_transpose_y_3, x = q_v_345_cast_fp16, y = p_head_691_cast_fp16)[name = string("x_1817_cast_fp16")];
            tensor<int32, [8]> x_1819_pad_1 = const()[name = string("x_1819_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1819_mode_1 = const()[name = string("x_1819_mode_1"), val = string("constant")];
            fp16 const_2467_to_fp16 = const()[name = string("const_2467_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1819_cast_fp16 = pad(constant_val = const_2467_to_fp16, mode = x_1819_mode_1, pad = x_1819_pad_1, x = x_1817_cast_fp16)[name = string("x_1819_cast_fp16")];
            tensor<int32, [4]> var_27589 = const()[name = string("op_27589"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1821_cast_fp16 = reshape(shape = var_27589, x = x_1819_cast_fp16)[name = string("x_1821_cast_fp16")];
            tensor<int32, [4]> var_27593_begin_1 = const()[name = string("op_27593_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_27593_end_1 = const()[name = string("op_27593_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_27593_end_mask_1 = const()[name = string("op_27593_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_27593_cast_fp16 = slice_by_index(begin = var_27593_begin_1, end = var_27593_end_1, end_mask = var_27593_end_mask_1, x = x_1821_cast_fp16)[name = string("op_27593_cast_fp16")];
            tensor<int32, [4]> var_27595 = const()[name = string("op_27595"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_27596_cast_fp16 = reshape(shape = var_27595, x = var_27593_cast_fp16)[name = string("op_27596_cast_fp16")];
            tensor<int32, [4]> var_27601_begin_1 = const()[name = string("op_27601_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_27601_end_1 = const()[name = string("op_27601_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_27601_end_mask_1 = const()[name = string("op_27601_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_27601_cast_fp16 = slice_by_index(begin = var_27601_begin_1, end = var_27601_end_1, end_mask = var_27601_end_mask_1, x = var_27596_cast_fp16)[name = string("op_27601_cast_fp16")];
            fp16 var_27602_to_fp16 = const()[name = string("op_27602_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_345_cast_fp16 = mul(x = var_27601_cast_fp16, y = var_27602_to_fp16)[name = string("scores_pos_345_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_345_cast_fp16 = add(x = scores_content_345_cast_fp16, y = scores_pos_345_cast_fp16)[name = string("logits_345_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_27605_cast_fp16 = softmax(axis = var_26846, x = logits_345_cast_fp16)[name = string("op_27605_cast_fp16")];
            bool var_27607_transpose_x_1 = const()[name = string("op_27607_transpose_x_1"), val = bool(false)];
            bool var_27607_transpose_y_1 = const()[name = string("op_27607_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_27607_cast_fp16 = matmul(transpose_x = var_27607_transpose_x_1, transpose_y = var_27607_transpose_y_1, x = var_27605_cast_fp16, y = v_head_691_cast_fp16)[name = string("op_27607_cast_fp16")];
            tensor<int32, [1]> var_27608_axes_1 = const()[name = string("op_27608_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_27608_cast_fp16 = squeeze(axes = var_27608_axes_1, x = var_27607_cast_fp16)[name = string("op_27608_cast_fp16")];
            string dense_output_1737_pad_type_1 = const()[name = string("dense_output_1737_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1737_strides_1 = const()[name = string("dense_output_1737_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1737_pad_1 = const()[name = string("dense_output_1737_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1737_dilations_1 = const()[name = string("dense_output_1737_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1737_groups_1 = const()[name = string("dense_output_1737_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631696384))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631827520))))[name = string("layers_21_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1737_cast_fp16 = conv(dilations = dense_output_1737_dilations_1, groups = dense_output_1737_groups_1, pad = dense_output_1737_pad_1, pad_type = dense_output_1737_pad_type_1, strides = dense_output_1737_strides_1, weight = layers_21_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1737_cast_fp16")];
            string sparse_output_1737_pad_type_1 = const()[name = string("sparse_output_1737_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1737_strides_1 = const()[name = string("sparse_output_1737_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1737_pad_1 = const()[name = string("sparse_output_1737_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1737_dilations_1 = const()[name = string("sparse_output_1737_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1737_groups_1 = const()[name = string("sparse_output_1737_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631830784))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631828096))))[name = string("layers_21_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1737_cast_fp16 = conv(dilations = sparse_output_1737_dilations_1, groups = sparse_output_1737_groups_1, pad = sparse_output_1737_pad_1, pad_type = sparse_output_1737_pad_type_1, strides = sparse_output_1737_strides_1, weight = layers_21_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1737_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27623_cast_fp16 = add(x = dense_output_1737_cast_fp16, y = sparse_output_1737_cast_fp16)[name = string("op_27623_cast_fp16")];
            tensor<int32, [4]> var_27624 = const()[name = string("op_27624"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27626 = const()[name = string("op_27626"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27625_cast_fp16 = transpose(perm = var_27624, x = var_27623_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 51, 128]> q_head_347_cast_fp16 = reshape(shape = var_27626, x = var_27625_cast_fp16)[name = string("q_head_347_cast_fp16")];
            string dense_output_1739_pad_type_1 = const()[name = string("dense_output_1739_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1739_strides_1 = const()[name = string("dense_output_1739_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1739_pad_1 = const()[name = string("dense_output_1739_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1739_dilations_1 = const()[name = string("dense_output_1739_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1739_groups_1 = const()[name = string("dense_output_1739_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631847232))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631978368))))[name = string("layers_21_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1739_cast_fp16 = conv(dilations = dense_output_1739_dilations_1, groups = dense_output_1739_groups_1, pad = dense_output_1739_pad_1, pad_type = dense_output_1739_pad_type_1, strides = dense_output_1739_strides_1, weight = layers_21_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1739_cast_fp16")];
            string sparse_output_1739_pad_type_1 = const()[name = string("sparse_output_1739_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1739_strides_1 = const()[name = string("sparse_output_1739_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1739_pad_1 = const()[name = string("sparse_output_1739_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1739_dilations_1 = const()[name = string("sparse_output_1739_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1739_groups_1 = const()[name = string("sparse_output_1739_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631981632))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631978944))))[name = string("layers_21_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1739_cast_fp16 = conv(dilations = sparse_output_1739_dilations_1, groups = sparse_output_1739_groups_1, pad = sparse_output_1739_pad_1, pad_type = sparse_output_1739_pad_type_1, strides = sparse_output_1739_strides_1, weight = layers_21_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1739_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27642_cast_fp16 = add(x = dense_output_1739_cast_fp16, y = sparse_output_1739_cast_fp16)[name = string("op_27642_cast_fp16")];
            tensor<int32, [4]> var_27643 = const()[name = string("op_27643"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27645 = const()[name = string("op_27645"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27644_cast_fp16 = transpose(perm = var_27643, x = var_27642_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 51, 128]> k_head_693_cast_fp16 = reshape(shape = var_27645, x = var_27644_cast_fp16)[name = string("k_head_693_cast_fp16")];
            string dense_output_1741_pad_type_1 = const()[name = string("dense_output_1741_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1741_strides_1 = const()[name = string("dense_output_1741_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1741_pad_1 = const()[name = string("dense_output_1741_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1741_dilations_1 = const()[name = string("dense_output_1741_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1741_groups_1 = const()[name = string("dense_output_1741_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631998080))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632129216))))[name = string("layers_21_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1741_cast_fp16 = conv(dilations = dense_output_1741_dilations_1, groups = dense_output_1741_groups_1, pad = dense_output_1741_pad_1, pad_type = dense_output_1741_pad_type_1, strides = dense_output_1741_strides_1, weight = layers_21_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1741_cast_fp16")];
            string sparse_output_1741_pad_type_1 = const()[name = string("sparse_output_1741_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1741_strides_1 = const()[name = string("sparse_output_1741_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1741_pad_1 = const()[name = string("sparse_output_1741_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1741_dilations_1 = const()[name = string("sparse_output_1741_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1741_groups_1 = const()[name = string("sparse_output_1741_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632132480))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632129792))))[name = string("layers_21_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1741_cast_fp16 = conv(dilations = sparse_output_1741_dilations_1, groups = sparse_output_1741_groups_1, pad = sparse_output_1741_pad_1, pad_type = sparse_output_1741_pad_type_1, strides = sparse_output_1741_strides_1, weight = layers_21_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1741_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27661_cast_fp16 = add(x = dense_output_1741_cast_fp16, y = sparse_output_1741_cast_fp16)[name = string("op_27661_cast_fp16")];
            tensor<int32, [4]> var_27662 = const()[name = string("op_27662"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27664 = const()[name = string("op_27664"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27663_cast_fp16 = transpose(perm = var_27662, x = var_27661_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 51, 128]> v_head_693_cast_fp16 = reshape(shape = var_27664, x = var_27663_cast_fp16)[name = string("v_head_693_cast_fp16")];
            string dense_output_1743_pad_type_1 = const()[name = string("dense_output_1743_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1743_strides_1 = const()[name = string("dense_output_1743_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1743_pad_1 = const()[name = string("dense_output_1743_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1743_dilations_1 = const()[name = string("dense_output_1743_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1743_groups_1 = const()[name = string("dense_output_1743_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632148928))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632280064))))[name = string("layers_21_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1743_cast_fp16 = conv(dilations = dense_output_1743_dilations_1, groups = dense_output_1743_groups_1, pad = dense_output_1743_pad_1, pad_type = dense_output_1743_pad_type_1, strides = dense_output_1743_strides_1, weight = layers_21_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1743_cast_fp16")];
            string sparse_output_1743_pad_type_1 = const()[name = string("sparse_output_1743_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1743_strides_1 = const()[name = string("sparse_output_1743_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1743_pad_1 = const()[name = string("sparse_output_1743_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1743_dilations_1 = const()[name = string("sparse_output_1743_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1743_groups_1 = const()[name = string("sparse_output_1743_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632283328))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632280640))))[name = string("layers_21_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1743_cast_fp16 = conv(dilations = sparse_output_1743_dilations_1, groups = sparse_output_1743_groups_1, pad = sparse_output_1743_pad_1, pad_type = sparse_output_1743_pad_type_1, strides = sparse_output_1743_strides_1, weight = layers_21_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1743_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_27680_cast_fp16 = add(x = dense_output_1743_cast_fp16, y = sparse_output_1743_cast_fp16)[name = string("op_27680_cast_fp16")];
            tensor<int32, [4]> var_27681 = const()[name = string("op_27681"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27683 = const()[name = string("op_27683"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_27682_cast_fp16 = transpose(perm = var_27681, x = var_27680_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 101, 128]> p_head_693_cast_fp16 = reshape(shape = var_27683, x = var_27682_cast_fp16)[name = string("p_head_693_cast_fp16")];
            tensor<fp16, [128]> var_27685_to_fp16 = const()[name = string("op_27685_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632299776)))];
            tensor<fp16, [1, 51, 128]> var_27686_cast_fp16 = add(x = q_head_347_cast_fp16, y = var_27685_to_fp16)[name = string("op_27686_cast_fp16")];
            tensor<int32, [1]> q_u_347_axes_1 = const()[name = string("q_u_347_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_347_cast_fp16 = expand_dims(axes = q_u_347_axes_1, x = var_27686_cast_fp16)[name = string("q_u_347_cast_fp16")];
            tensor<fp16, [128]> var_27688_to_fp16 = const()[name = string("op_27688_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632300096)))];
            tensor<fp16, [1, 51, 128]> var_27689_cast_fp16 = add(x = q_head_347_cast_fp16, y = var_27688_to_fp16)[name = string("op_27689_cast_fp16")];
            tensor<int32, [1]> q_v_347_axes_1 = const()[name = string("q_v_347_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_347_cast_fp16 = expand_dims(axes = q_v_347_axes_1, x = var_27689_cast_fp16)[name = string("q_v_347_cast_fp16")];
            tensor<int32, [1]> k_head_695_axes_1 = const()[name = string("k_head_695_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_695_cast_fp16 = expand_dims(axes = k_head_695_axes_1, x = k_head_693_cast_fp16)[name = string("k_head_695_cast_fp16")];
            tensor<int32, [1]> v_head_695_axes_1 = const()[name = string("v_head_695_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_695_cast_fp16 = expand_dims(axes = v_head_695_axes_1, x = v_head_693_cast_fp16)[name = string("v_head_695_cast_fp16")];
            tensor<int32, [1]> p_head_695_axes_1 = const()[name = string("p_head_695_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_695_cast_fp16 = expand_dims(axes = p_head_695_axes_1, x = p_head_693_cast_fp16)[name = string("p_head_695_cast_fp16")];
            bool var_27695_transpose_x_3 = const()[name = string("op_27695_transpose_x_3"), val = bool(false)];
            bool var_27695_transpose_y_3 = const()[name = string("op_27695_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_27695_cast_fp16 = matmul(transpose_x = var_27695_transpose_x_3, transpose_y = var_27695_transpose_y_3, x = q_u_347_cast_fp16, y = k_head_695_cast_fp16)[name = string("op_27695_cast_fp16")];
            fp16 var_27696_to_fp16 = const()[name = string("op_27696_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_347_cast_fp16 = mul(x = var_27695_cast_fp16, y = var_27696_to_fp16)[name = string("scores_content_347_cast_fp16")];
            bool x_1825_transpose_x_3 = const()[name = string("x_1825_transpose_x_3"), val = bool(false)];
            bool x_1825_transpose_y_3 = const()[name = string("x_1825_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1825_cast_fp16 = matmul(transpose_x = x_1825_transpose_x_3, transpose_y = x_1825_transpose_y_3, x = q_v_347_cast_fp16, y = p_head_695_cast_fp16)[name = string("x_1825_cast_fp16")];
            tensor<int32, [8]> x_1827_pad_1 = const()[name = string("x_1827_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1827_mode_1 = const()[name = string("x_1827_mode_1"), val = string("constant")];
            fp16 const_2473_to_fp16 = const()[name = string("const_2473_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1827_cast_fp16 = pad(constant_val = const_2473_to_fp16, mode = x_1827_mode_1, pad = x_1827_pad_1, x = x_1825_cast_fp16)[name = string("x_1827_cast_fp16")];
            tensor<int32, [4]> var_27710 = const()[name = string("op_27710"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1829_cast_fp16 = reshape(shape = var_27710, x = x_1827_cast_fp16)[name = string("x_1829_cast_fp16")];
            tensor<int32, [4]> var_27714_begin_1 = const()[name = string("op_27714_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_27714_end_1 = const()[name = string("op_27714_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_27714_end_mask_1 = const()[name = string("op_27714_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_27714_cast_fp16 = slice_by_index(begin = var_27714_begin_1, end = var_27714_end_1, end_mask = var_27714_end_mask_1, x = x_1829_cast_fp16)[name = string("op_27714_cast_fp16")];
            tensor<int32, [4]> var_27716 = const()[name = string("op_27716"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_27717_cast_fp16 = reshape(shape = var_27716, x = var_27714_cast_fp16)[name = string("op_27717_cast_fp16")];
            tensor<int32, [4]> var_27722_begin_1 = const()[name = string("op_27722_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_27722_end_1 = const()[name = string("op_27722_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_27722_end_mask_1 = const()[name = string("op_27722_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_27722_cast_fp16 = slice_by_index(begin = var_27722_begin_1, end = var_27722_end_1, end_mask = var_27722_end_mask_1, x = var_27717_cast_fp16)[name = string("op_27722_cast_fp16")];
            fp16 var_27723_to_fp16 = const()[name = string("op_27723_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_347_cast_fp16 = mul(x = var_27722_cast_fp16, y = var_27723_to_fp16)[name = string("scores_pos_347_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_347_cast_fp16 = add(x = scores_content_347_cast_fp16, y = scores_pos_347_cast_fp16)[name = string("logits_347_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_27726_cast_fp16 = softmax(axis = var_26846, x = logits_347_cast_fp16)[name = string("op_27726_cast_fp16")];
            bool var_27728_transpose_x_1 = const()[name = string("op_27728_transpose_x_1"), val = bool(false)];
            bool var_27728_transpose_y_1 = const()[name = string("op_27728_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_27728_cast_fp16 = matmul(transpose_x = var_27728_transpose_x_1, transpose_y = var_27728_transpose_y_1, x = var_27726_cast_fp16, y = v_head_695_cast_fp16)[name = string("op_27728_cast_fp16")];
            tensor<int32, [1]> var_27729_axes_1 = const()[name = string("op_27729_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_27729_cast_fp16 = squeeze(axes = var_27729_axes_1, x = var_27728_cast_fp16)[name = string("op_27729_cast_fp16")];
            string dense_output_1745_pad_type_1 = const()[name = string("dense_output_1745_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1745_strides_1 = const()[name = string("dense_output_1745_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1745_pad_1 = const()[name = string("dense_output_1745_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1745_dilations_1 = const()[name = string("dense_output_1745_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1745_groups_1 = const()[name = string("dense_output_1745_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632300416))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632431552))))[name = string("layers_21_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1745_cast_fp16 = conv(dilations = dense_output_1745_dilations_1, groups = dense_output_1745_groups_1, pad = dense_output_1745_pad_1, pad_type = dense_output_1745_pad_type_1, strides = dense_output_1745_strides_1, weight = layers_21_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1745_cast_fp16")];
            string sparse_output_1745_pad_type_1 = const()[name = string("sparse_output_1745_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1745_strides_1 = const()[name = string("sparse_output_1745_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1745_pad_1 = const()[name = string("sparse_output_1745_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1745_dilations_1 = const()[name = string("sparse_output_1745_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1745_groups_1 = const()[name = string("sparse_output_1745_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632434816))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632432128))))[name = string("layers_21_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1745_cast_fp16 = conv(dilations = sparse_output_1745_dilations_1, groups = sparse_output_1745_groups_1, pad = sparse_output_1745_pad_1, pad_type = sparse_output_1745_pad_type_1, strides = sparse_output_1745_strides_1, weight = layers_21_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1745_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27744_cast_fp16 = add(x = dense_output_1745_cast_fp16, y = sparse_output_1745_cast_fp16)[name = string("op_27744_cast_fp16")];
            tensor<int32, [4]> var_27745 = const()[name = string("op_27745"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27747 = const()[name = string("op_27747"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27746_cast_fp16 = transpose(perm = var_27745, x = var_27744_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 51, 128]> q_head_349_cast_fp16 = reshape(shape = var_27747, x = var_27746_cast_fp16)[name = string("q_head_349_cast_fp16")];
            string dense_output_1747_pad_type_1 = const()[name = string("dense_output_1747_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1747_strides_1 = const()[name = string("dense_output_1747_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1747_pad_1 = const()[name = string("dense_output_1747_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1747_dilations_1 = const()[name = string("dense_output_1747_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1747_groups_1 = const()[name = string("dense_output_1747_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632451264))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632582400))))[name = string("layers_21_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1747_cast_fp16 = conv(dilations = dense_output_1747_dilations_1, groups = dense_output_1747_groups_1, pad = dense_output_1747_pad_1, pad_type = dense_output_1747_pad_type_1, strides = dense_output_1747_strides_1, weight = layers_21_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1747_cast_fp16")];
            string sparse_output_1747_pad_type_1 = const()[name = string("sparse_output_1747_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1747_strides_1 = const()[name = string("sparse_output_1747_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1747_pad_1 = const()[name = string("sparse_output_1747_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1747_dilations_1 = const()[name = string("sparse_output_1747_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1747_groups_1 = const()[name = string("sparse_output_1747_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632585664))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632582976))))[name = string("layers_21_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1747_cast_fp16 = conv(dilations = sparse_output_1747_dilations_1, groups = sparse_output_1747_groups_1, pad = sparse_output_1747_pad_1, pad_type = sparse_output_1747_pad_type_1, strides = sparse_output_1747_strides_1, weight = layers_21_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1747_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27763_cast_fp16 = add(x = dense_output_1747_cast_fp16, y = sparse_output_1747_cast_fp16)[name = string("op_27763_cast_fp16")];
            tensor<int32, [4]> var_27764 = const()[name = string("op_27764"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27766 = const()[name = string("op_27766"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27765_cast_fp16 = transpose(perm = var_27764, x = var_27763_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 51, 128]> k_head_697_cast_fp16 = reshape(shape = var_27766, x = var_27765_cast_fp16)[name = string("k_head_697_cast_fp16")];
            string dense_output_1749_pad_type_1 = const()[name = string("dense_output_1749_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1749_strides_1 = const()[name = string("dense_output_1749_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1749_pad_1 = const()[name = string("dense_output_1749_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1749_dilations_1 = const()[name = string("dense_output_1749_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1749_groups_1 = const()[name = string("dense_output_1749_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632602112))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632733248))))[name = string("layers_21_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1749_cast_fp16 = conv(dilations = dense_output_1749_dilations_1, groups = dense_output_1749_groups_1, pad = dense_output_1749_pad_1, pad_type = dense_output_1749_pad_type_1, strides = dense_output_1749_strides_1, weight = layers_21_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1749_cast_fp16")];
            string sparse_output_1749_pad_type_1 = const()[name = string("sparse_output_1749_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1749_strides_1 = const()[name = string("sparse_output_1749_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1749_pad_1 = const()[name = string("sparse_output_1749_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1749_dilations_1 = const()[name = string("sparse_output_1749_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1749_groups_1 = const()[name = string("sparse_output_1749_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632736512))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632733824))))[name = string("layers_21_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1749_cast_fp16 = conv(dilations = sparse_output_1749_dilations_1, groups = sparse_output_1749_groups_1, pad = sparse_output_1749_pad_1, pad_type = sparse_output_1749_pad_type_1, strides = sparse_output_1749_strides_1, weight = layers_21_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1749_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27782_cast_fp16 = add(x = dense_output_1749_cast_fp16, y = sparse_output_1749_cast_fp16)[name = string("op_27782_cast_fp16")];
            tensor<int32, [4]> var_27783 = const()[name = string("op_27783"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27785 = const()[name = string("op_27785"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27784_cast_fp16 = transpose(perm = var_27783, x = var_27782_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 51, 128]> v_head_697_cast_fp16 = reshape(shape = var_27785, x = var_27784_cast_fp16)[name = string("v_head_697_cast_fp16")];
            string dense_output_1751_pad_type_1 = const()[name = string("dense_output_1751_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1751_strides_1 = const()[name = string("dense_output_1751_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1751_pad_1 = const()[name = string("dense_output_1751_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1751_dilations_1 = const()[name = string("dense_output_1751_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1751_groups_1 = const()[name = string("dense_output_1751_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632752960))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632884096))))[name = string("layers_21_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1751_cast_fp16 = conv(dilations = dense_output_1751_dilations_1, groups = dense_output_1751_groups_1, pad = dense_output_1751_pad_1, pad_type = dense_output_1751_pad_type_1, strides = dense_output_1751_strides_1, weight = layers_21_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1751_cast_fp16")];
            string sparse_output_1751_pad_type_1 = const()[name = string("sparse_output_1751_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1751_strides_1 = const()[name = string("sparse_output_1751_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1751_pad_1 = const()[name = string("sparse_output_1751_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1751_dilations_1 = const()[name = string("sparse_output_1751_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1751_groups_1 = const()[name = string("sparse_output_1751_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632887360))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632884672))))[name = string("layers_21_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1751_cast_fp16 = conv(dilations = sparse_output_1751_dilations_1, groups = sparse_output_1751_groups_1, pad = sparse_output_1751_pad_1, pad_type = sparse_output_1751_pad_type_1, strides = sparse_output_1751_strides_1, weight = layers_21_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1751_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_27801_cast_fp16 = add(x = dense_output_1751_cast_fp16, y = sparse_output_1751_cast_fp16)[name = string("op_27801_cast_fp16")];
            tensor<int32, [4]> var_27802 = const()[name = string("op_27802"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27804 = const()[name = string("op_27804"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_27803_cast_fp16 = transpose(perm = var_27802, x = var_27801_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 101, 128]> p_head_697_cast_fp16 = reshape(shape = var_27804, x = var_27803_cast_fp16)[name = string("p_head_697_cast_fp16")];
            tensor<fp16, [128]> var_27806_to_fp16 = const()[name = string("op_27806_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632903808)))];
            tensor<fp16, [1, 51, 128]> var_27807_cast_fp16 = add(x = q_head_349_cast_fp16, y = var_27806_to_fp16)[name = string("op_27807_cast_fp16")];
            tensor<int32, [1]> q_u_349_axes_1 = const()[name = string("q_u_349_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_349_cast_fp16 = expand_dims(axes = q_u_349_axes_1, x = var_27807_cast_fp16)[name = string("q_u_349_cast_fp16")];
            tensor<fp16, [128]> var_27809_to_fp16 = const()[name = string("op_27809_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632904128)))];
            tensor<fp16, [1, 51, 128]> var_27810_cast_fp16 = add(x = q_head_349_cast_fp16, y = var_27809_to_fp16)[name = string("op_27810_cast_fp16")];
            tensor<int32, [1]> q_v_349_axes_1 = const()[name = string("q_v_349_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_349_cast_fp16 = expand_dims(axes = q_v_349_axes_1, x = var_27810_cast_fp16)[name = string("q_v_349_cast_fp16")];
            tensor<int32, [1]> k_head_699_axes_1 = const()[name = string("k_head_699_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_699_cast_fp16 = expand_dims(axes = k_head_699_axes_1, x = k_head_697_cast_fp16)[name = string("k_head_699_cast_fp16")];
            tensor<int32, [1]> v_head_699_axes_1 = const()[name = string("v_head_699_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_699_cast_fp16 = expand_dims(axes = v_head_699_axes_1, x = v_head_697_cast_fp16)[name = string("v_head_699_cast_fp16")];
            tensor<int32, [1]> p_head_699_axes_1 = const()[name = string("p_head_699_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_699_cast_fp16 = expand_dims(axes = p_head_699_axes_1, x = p_head_697_cast_fp16)[name = string("p_head_699_cast_fp16")];
            bool var_27816_transpose_x_3 = const()[name = string("op_27816_transpose_x_3"), val = bool(false)];
            bool var_27816_transpose_y_3 = const()[name = string("op_27816_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_27816_cast_fp16 = matmul(transpose_x = var_27816_transpose_x_3, transpose_y = var_27816_transpose_y_3, x = q_u_349_cast_fp16, y = k_head_699_cast_fp16)[name = string("op_27816_cast_fp16")];
            fp16 var_27817_to_fp16 = const()[name = string("op_27817_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_349_cast_fp16 = mul(x = var_27816_cast_fp16, y = var_27817_to_fp16)[name = string("scores_content_349_cast_fp16")];
            bool x_1833_transpose_x_3 = const()[name = string("x_1833_transpose_x_3"), val = bool(false)];
            bool x_1833_transpose_y_3 = const()[name = string("x_1833_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1833_cast_fp16 = matmul(transpose_x = x_1833_transpose_x_3, transpose_y = x_1833_transpose_y_3, x = q_v_349_cast_fp16, y = p_head_699_cast_fp16)[name = string("x_1833_cast_fp16")];
            tensor<int32, [8]> x_1835_pad_1 = const()[name = string("x_1835_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1835_mode_1 = const()[name = string("x_1835_mode_1"), val = string("constant")];
            fp16 const_2479_to_fp16 = const()[name = string("const_2479_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1835_cast_fp16 = pad(constant_val = const_2479_to_fp16, mode = x_1835_mode_1, pad = x_1835_pad_1, x = x_1833_cast_fp16)[name = string("x_1835_cast_fp16")];
            tensor<int32, [4]> var_27831 = const()[name = string("op_27831"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1837_cast_fp16 = reshape(shape = var_27831, x = x_1835_cast_fp16)[name = string("x_1837_cast_fp16")];
            tensor<int32, [4]> var_27835_begin_1 = const()[name = string("op_27835_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_27835_end_1 = const()[name = string("op_27835_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_27835_end_mask_1 = const()[name = string("op_27835_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_27835_cast_fp16 = slice_by_index(begin = var_27835_begin_1, end = var_27835_end_1, end_mask = var_27835_end_mask_1, x = x_1837_cast_fp16)[name = string("op_27835_cast_fp16")];
            tensor<int32, [4]> var_27837 = const()[name = string("op_27837"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_27838_cast_fp16 = reshape(shape = var_27837, x = var_27835_cast_fp16)[name = string("op_27838_cast_fp16")];
            tensor<int32, [4]> var_27843_begin_1 = const()[name = string("op_27843_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_27843_end_1 = const()[name = string("op_27843_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_27843_end_mask_1 = const()[name = string("op_27843_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_27843_cast_fp16 = slice_by_index(begin = var_27843_begin_1, end = var_27843_end_1, end_mask = var_27843_end_mask_1, x = var_27838_cast_fp16)[name = string("op_27843_cast_fp16")];
            fp16 var_27844_to_fp16 = const()[name = string("op_27844_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_349_cast_fp16 = mul(x = var_27843_cast_fp16, y = var_27844_to_fp16)[name = string("scores_pos_349_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_349_cast_fp16 = add(x = scores_content_349_cast_fp16, y = scores_pos_349_cast_fp16)[name = string("logits_349_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_27847_cast_fp16 = softmax(axis = var_26846, x = logits_349_cast_fp16)[name = string("op_27847_cast_fp16")];
            bool var_27849_transpose_x_1 = const()[name = string("op_27849_transpose_x_1"), val = bool(false)];
            bool var_27849_transpose_y_1 = const()[name = string("op_27849_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_27849_cast_fp16 = matmul(transpose_x = var_27849_transpose_x_1, transpose_y = var_27849_transpose_y_1, x = var_27847_cast_fp16, y = v_head_699_cast_fp16)[name = string("op_27849_cast_fp16")];
            tensor<int32, [1]> var_27850_axes_1 = const()[name = string("op_27850_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_27850_cast_fp16 = squeeze(axes = var_27850_axes_1, x = var_27849_cast_fp16)[name = string("op_27850_cast_fp16")];
            string dense_output_1753_pad_type_1 = const()[name = string("dense_output_1753_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1753_strides_1 = const()[name = string("dense_output_1753_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1753_pad_1 = const()[name = string("dense_output_1753_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1753_dilations_1 = const()[name = string("dense_output_1753_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1753_groups_1 = const()[name = string("dense_output_1753_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632904448))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633035584))))[name = string("layers_21_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1753_cast_fp16 = conv(dilations = dense_output_1753_dilations_1, groups = dense_output_1753_groups_1, pad = dense_output_1753_pad_1, pad_type = dense_output_1753_pad_type_1, strides = dense_output_1753_strides_1, weight = layers_21_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1753_cast_fp16")];
            string sparse_output_1753_pad_type_1 = const()[name = string("sparse_output_1753_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1753_strides_1 = const()[name = string("sparse_output_1753_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1753_pad_1 = const()[name = string("sparse_output_1753_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1753_dilations_1 = const()[name = string("sparse_output_1753_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1753_groups_1 = const()[name = string("sparse_output_1753_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633038848))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633036160))))[name = string("layers_21_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1753_cast_fp16 = conv(dilations = sparse_output_1753_dilations_1, groups = sparse_output_1753_groups_1, pad = sparse_output_1753_pad_1, pad_type = sparse_output_1753_pad_type_1, strides = sparse_output_1753_strides_1, weight = layers_21_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1753_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27865_cast_fp16 = add(x = dense_output_1753_cast_fp16, y = sparse_output_1753_cast_fp16)[name = string("op_27865_cast_fp16")];
            tensor<int32, [4]> var_27866 = const()[name = string("op_27866"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27868 = const()[name = string("op_27868"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27867_cast_fp16 = transpose(perm = var_27866, x = var_27865_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 51, 128]> q_head_351_cast_fp16 = reshape(shape = var_27868, x = var_27867_cast_fp16)[name = string("q_head_351_cast_fp16")];
            string dense_output_1755_pad_type_1 = const()[name = string("dense_output_1755_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1755_strides_1 = const()[name = string("dense_output_1755_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1755_pad_1 = const()[name = string("dense_output_1755_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1755_dilations_1 = const()[name = string("dense_output_1755_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1755_groups_1 = const()[name = string("dense_output_1755_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633055296))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633186432))))[name = string("layers_21_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1755_cast_fp16 = conv(dilations = dense_output_1755_dilations_1, groups = dense_output_1755_groups_1, pad = dense_output_1755_pad_1, pad_type = dense_output_1755_pad_type_1, strides = dense_output_1755_strides_1, weight = layers_21_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1755_cast_fp16")];
            string sparse_output_1755_pad_type_1 = const()[name = string("sparse_output_1755_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1755_strides_1 = const()[name = string("sparse_output_1755_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1755_pad_1 = const()[name = string("sparse_output_1755_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1755_dilations_1 = const()[name = string("sparse_output_1755_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1755_groups_1 = const()[name = string("sparse_output_1755_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633189696))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633187008))))[name = string("layers_21_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1755_cast_fp16 = conv(dilations = sparse_output_1755_dilations_1, groups = sparse_output_1755_groups_1, pad = sparse_output_1755_pad_1, pad_type = sparse_output_1755_pad_type_1, strides = sparse_output_1755_strides_1, weight = layers_21_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1755_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27884_cast_fp16 = add(x = dense_output_1755_cast_fp16, y = sparse_output_1755_cast_fp16)[name = string("op_27884_cast_fp16")];
            tensor<int32, [4]> var_27885 = const()[name = string("op_27885"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27887 = const()[name = string("op_27887"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27886_cast_fp16 = transpose(perm = var_27885, x = var_27884_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 51, 128]> k_head_701_cast_fp16 = reshape(shape = var_27887, x = var_27886_cast_fp16)[name = string("k_head_701_cast_fp16")];
            string dense_output_1757_pad_type_1 = const()[name = string("dense_output_1757_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1757_strides_1 = const()[name = string("dense_output_1757_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1757_pad_1 = const()[name = string("dense_output_1757_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1757_dilations_1 = const()[name = string("dense_output_1757_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1757_groups_1 = const()[name = string("dense_output_1757_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633206144))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633337280))))[name = string("layers_21_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1757_cast_fp16 = conv(dilations = dense_output_1757_dilations_1, groups = dense_output_1757_groups_1, pad = dense_output_1757_pad_1, pad_type = dense_output_1757_pad_type_1, strides = dense_output_1757_strides_1, weight = layers_21_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_995_cast_fp16)[name = string("dense_output_1757_cast_fp16")];
            string sparse_output_1757_pad_type_1 = const()[name = string("sparse_output_1757_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1757_strides_1 = const()[name = string("sparse_output_1757_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1757_pad_1 = const()[name = string("sparse_output_1757_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1757_dilations_1 = const()[name = string("sparse_output_1757_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1757_groups_1 = const()[name = string("sparse_output_1757_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633340544))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633337856))))[name = string("layers_21_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1757_cast_fp16 = conv(dilations = sparse_output_1757_dilations_1, groups = sparse_output_1757_groups_1, pad = sparse_output_1757_pad_1, pad_type = sparse_output_1757_pad_type_1, strides = sparse_output_1757_strides_1, weight = layers_21_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_995_cast_fp16)[name = string("sparse_output_1757_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_27903_cast_fp16 = add(x = dense_output_1757_cast_fp16, y = sparse_output_1757_cast_fp16)[name = string("op_27903_cast_fp16")];
            tensor<int32, [4]> var_27904 = const()[name = string("op_27904"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27906 = const()[name = string("op_27906"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_27905_cast_fp16 = transpose(perm = var_27904, x = var_27903_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 51, 128]> v_head_701_cast_fp16 = reshape(shape = var_27906, x = var_27905_cast_fp16)[name = string("v_head_701_cast_fp16")];
            string dense_output_1759_pad_type_1 = const()[name = string("dense_output_1759_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1759_strides_1 = const()[name = string("dense_output_1759_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1759_pad_1 = const()[name = string("dense_output_1759_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1759_dilations_1 = const()[name = string("dense_output_1759_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1759_groups_1 = const()[name = string("dense_output_1759_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633356992))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633488128))))[name = string("layers_21_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1759_cast_fp16 = conv(dilations = dense_output_1759_dilations_1, groups = dense_output_1759_groups_1, pad = dense_output_1759_pad_1, pad_type = dense_output_1759_pad_type_1, strides = dense_output_1759_strides_1, weight = layers_21_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1759_cast_fp16")];
            string sparse_output_1759_pad_type_1 = const()[name = string("sparse_output_1759_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1759_strides_1 = const()[name = string("sparse_output_1759_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1759_pad_1 = const()[name = string("sparse_output_1759_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1759_dilations_1 = const()[name = string("sparse_output_1759_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1759_groups_1 = const()[name = string("sparse_output_1759_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_21_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633491392))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633488704))))[name = string("layers_21_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1759_cast_fp16 = conv(dilations = sparse_output_1759_dilations_1, groups = sparse_output_1759_groups_1, pad = sparse_output_1759_pad_1, pad_type = sparse_output_1759_pad_type_1, strides = sparse_output_1759_strides_1, weight = layers_21_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1759_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_27922_cast_fp16 = add(x = dense_output_1759_cast_fp16, y = sparse_output_1759_cast_fp16)[name = string("op_27922_cast_fp16")];
            tensor<int32, [4]> var_27923 = const()[name = string("op_27923"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_27925 = const()[name = string("op_27925"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_27924_cast_fp16 = transpose(perm = var_27923, x = var_27922_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 101, 128]> p_head_701_cast_fp16 = reshape(shape = var_27925, x = var_27924_cast_fp16)[name = string("p_head_701_cast_fp16")];
            tensor<fp16, [128]> var_27927_to_fp16 = const()[name = string("op_27927_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633507840)))];
            tensor<fp16, [1, 51, 128]> var_27928_cast_fp16 = add(x = q_head_351_cast_fp16, y = var_27927_to_fp16)[name = string("op_27928_cast_fp16")];
            tensor<int32, [1]> q_u_351_axes_1 = const()[name = string("q_u_351_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_351_cast_fp16 = expand_dims(axes = q_u_351_axes_1, x = var_27928_cast_fp16)[name = string("q_u_351_cast_fp16")];
            tensor<fp16, [128]> var_27930_to_fp16 = const()[name = string("op_27930_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633508160)))];
            tensor<fp16, [1, 51, 128]> var_27931_cast_fp16 = add(x = q_head_351_cast_fp16, y = var_27930_to_fp16)[name = string("op_27931_cast_fp16")];
            tensor<int32, [1]> q_v_351_axes_1 = const()[name = string("q_v_351_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_351_cast_fp16 = expand_dims(axes = q_v_351_axes_1, x = var_27931_cast_fp16)[name = string("q_v_351_cast_fp16")];
            tensor<int32, [1]> k_head_703_axes_1 = const()[name = string("k_head_703_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_703_cast_fp16 = expand_dims(axes = k_head_703_axes_1, x = k_head_701_cast_fp16)[name = string("k_head_703_cast_fp16")];
            tensor<int32, [1]> v_head_703_axes_1 = const()[name = string("v_head_703_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_703_cast_fp16 = expand_dims(axes = v_head_703_axes_1, x = v_head_701_cast_fp16)[name = string("v_head_703_cast_fp16")];
            tensor<int32, [1]> p_head_703_axes_1 = const()[name = string("p_head_703_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_703_cast_fp16 = expand_dims(axes = p_head_703_axes_1, x = p_head_701_cast_fp16)[name = string("p_head_703_cast_fp16")];
            bool var_27937_transpose_x_3 = const()[name = string("op_27937_transpose_x_3"), val = bool(false)];
            bool var_27937_transpose_y_3 = const()[name = string("op_27937_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_27937_cast_fp16 = matmul(transpose_x = var_27937_transpose_x_3, transpose_y = var_27937_transpose_y_3, x = q_u_351_cast_fp16, y = k_head_703_cast_fp16)[name = string("op_27937_cast_fp16")];
            fp16 var_27938_to_fp16 = const()[name = string("op_27938_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_351_cast_fp16 = mul(x = var_27937_cast_fp16, y = var_27938_to_fp16)[name = string("scores_content_351_cast_fp16")];
            bool x_1841_transpose_x_3 = const()[name = string("x_1841_transpose_x_3"), val = bool(false)];
            bool x_1841_transpose_y_3 = const()[name = string("x_1841_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1841_cast_fp16 = matmul(transpose_x = x_1841_transpose_x_3, transpose_y = x_1841_transpose_y_3, x = q_v_351_cast_fp16, y = p_head_703_cast_fp16)[name = string("x_1841_cast_fp16")];
            tensor<int32, [8]> x_1843_pad_1 = const()[name = string("x_1843_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1843_mode_1 = const()[name = string("x_1843_mode_1"), val = string("constant")];
            fp16 const_2485_to_fp16 = const()[name = string("const_2485_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1843_cast_fp16 = pad(constant_val = const_2485_to_fp16, mode = x_1843_mode_1, pad = x_1843_pad_1, x = x_1841_cast_fp16)[name = string("x_1843_cast_fp16")];
            tensor<int32, [4]> var_27952 = const()[name = string("op_27952"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1845_cast_fp16 = reshape(shape = var_27952, x = x_1843_cast_fp16)[name = string("x_1845_cast_fp16")];
            tensor<int32, [4]> var_27956_begin_1 = const()[name = string("op_27956_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_27956_end_1 = const()[name = string("op_27956_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_27956_end_mask_1 = const()[name = string("op_27956_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_27956_cast_fp16 = slice_by_index(begin = var_27956_begin_1, end = var_27956_end_1, end_mask = var_27956_end_mask_1, x = x_1845_cast_fp16)[name = string("op_27956_cast_fp16")];
            tensor<int32, [4]> var_27958 = const()[name = string("op_27958"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_27959_cast_fp16 = reshape(shape = var_27958, x = var_27956_cast_fp16)[name = string("op_27959_cast_fp16")];
            tensor<int32, [4]> var_27964_begin_1 = const()[name = string("op_27964_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_27964_end_1 = const()[name = string("op_27964_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_27964_end_mask_1 = const()[name = string("op_27964_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_27964_cast_fp16 = slice_by_index(begin = var_27964_begin_1, end = var_27964_end_1, end_mask = var_27964_end_mask_1, x = var_27959_cast_fp16)[name = string("op_27964_cast_fp16")];
            fp16 var_27965_to_fp16 = const()[name = string("op_27965_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_351_cast_fp16 = mul(x = var_27964_cast_fp16, y = var_27965_to_fp16)[name = string("scores_pos_351_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_351_cast_fp16 = add(x = scores_content_351_cast_fp16, y = scores_pos_351_cast_fp16)[name = string("logits_351_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_27968_cast_fp16 = softmax(axis = var_26846, x = logits_351_cast_fp16)[name = string("op_27968_cast_fp16")];
            bool var_27970_transpose_x_1 = const()[name = string("op_27970_transpose_x_1"), val = bool(false)];
            bool var_27970_transpose_y_1 = const()[name = string("op_27970_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_27970_cast_fp16 = matmul(transpose_x = var_27970_transpose_x_1, transpose_y = var_27970_transpose_y_1, x = var_27968_cast_fp16, y = v_head_703_cast_fp16)[name = string("op_27970_cast_fp16")];
            tensor<int32, [1]> o_head_43_axes_1 = const()[name = string("o_head_43_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_43_cast_fp16 = squeeze(axes = o_head_43_axes_1, x = var_27970_cast_fp16)[name = string("o_head_43_cast_fp16")];
            bool out_43_interleave_1 = const()[name = string("out_43_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_43_cast_fp16 = concat(axis = var_26846, interleave = out_43_interleave_1, values = (var_27124_cast_fp16, var_27245_cast_fp16, var_27366_cast_fp16, var_27487_cast_fp16, var_27608_cast_fp16, var_27729_cast_fp16, var_27850_cast_fp16, o_head_43_cast_fp16))[name = string("out_43_cast_fp16")];
            tensor<int32, [3]> var_27974_perm_1 = const()[name = string("op_27974_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1003_axes_1 = const()[name = string("input_1003_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_27974_cast_fp16 = transpose(perm = var_27974_perm_1, x = out_43_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 1024, 51, 1]> input_1003_cast_fp16 = expand_dims(axes = input_1003_axes_1, x = var_27974_cast_fp16)[name = string("input_1003_cast_fp16")];
            string dense_output_1761_pad_type_1 = const()[name = string("dense_output_1761_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1761_strides_1 = const()[name = string("dense_output_1761_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1761_pad_1 = const()[name = string("dense_output_1761_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1761_dilations_1 = const()[name = string("dense_output_1761_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1761_groups_1 = const()[name = string("dense_output_1761_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633508480))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(634557120))))[name = string("layers_21_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1761_cast_fp16 = conv(dilations = dense_output_1761_dilations_1, groups = dense_output_1761_groups_1, pad = dense_output_1761_pad_1, pad_type = dense_output_1761_pad_type_1, strides = dense_output_1761_strides_1, weight = layers_21_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_1003_cast_fp16)[name = string("dense_output_1761_cast_fp16")];
            string sparse_output_1761_pad_type_1 = const()[name = string("sparse_output_1761_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1761_strides_1 = const()[name = string("sparse_output_1761_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1761_pad_1 = const()[name = string("sparse_output_1761_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1761_dilations_1 = const()[name = string("sparse_output_1761_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1761_groups_1 = const()[name = string("sparse_output_1761_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(634578752))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(634557696))))[name = string("layers_21_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1761_cast_fp16 = conv(dilations = sparse_output_1761_dilations_1, groups = sparse_output_1761_groups_1, pad = sparse_output_1761_pad_1, pad_type = sparse_output_1761_pad_type_1, strides = sparse_output_1761_strides_1, weight = layers_21_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_1003_cast_fp16)[name = string("sparse_output_1761_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_43_cast_fp16 = add(x = dense_output_1761_cast_fp16, y = sparse_output_1761_cast_fp16)[name = string("out_conv_43_cast_fp16")];
            tensor<int32, [1]> var_27991_axes_1 = const()[name = string("op_27991_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_27991_cast_fp16 = squeeze(axes = var_27991_axes_1, x = out_conv_43_cast_fp16)[name = string("op_27991_cast_fp16")];
            tensor<int32, [3]> var_27992_perm_1 = const()[name = string("op_27992_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_27992_cast_fp16 = transpose(perm = var_27992_perm_1, x = var_27991_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 51, 1024]> input_1005_cast_fp16 = add(x = input_993_cast_fp16, y = var_27992_cast_fp16)[name = string("input_1005_cast_fp16")];
            tensor<int32, [1]> x_1849_axes_1 = const()[name = string("x_1849_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_21_norm_conv_weight_to_fp16 = const()[name = string("layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(634709888)))];
            tensor<fp16, [1024]> layers_21_norm_conv_bias_to_fp16 = const()[name = string("layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(634712000)))];
            tensor<fp16, [1, 51, 1024]> x_1849_cast_fp16 = layer_norm(axes = x_1849_axes_1, beta = layers_21_norm_conv_bias_to_fp16, epsilon = var_26861_to_fp16, gamma = layers_21_norm_conv_weight_to_fp16, x = input_1005_cast_fp16)[name = string("x_1849_cast_fp16")];
            tensor<int32, [3]> var_28002_perm_1 = const()[name = string("op_28002_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1007_axes_1 = const()[name = string("input_1007_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_28002_cast_fp16 = transpose(perm = var_28002_perm_1, x = x_1849_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 1024, 51, 1]> input_1007_cast_fp16 = expand_dims(axes = input_1007_axes_1, x = var_28002_cast_fp16)[name = string("input_1007_cast_fp16")];
            string dense_output_1763_pad_type_1 = const()[name = string("dense_output_1763_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1763_strides_1 = const()[name = string("dense_output_1763_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1763_pad_1 = const()[name = string("dense_output_1763_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1763_dilations_1 = const()[name = string("dense_output_1763_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1763_groups_1 = const()[name = string("dense_output_1763_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_21_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(634714112))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(636811328))))[name = string("layers_21_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1763_cast_fp16 = conv(dilations = dense_output_1763_dilations_1, groups = dense_output_1763_groups_1, pad = dense_output_1763_pad_1, pad_type = dense_output_1763_pad_type_1, strides = dense_output_1763_strides_1, weight = layers_21_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_1007_cast_fp16)[name = string("dense_output_1763_cast_fp16")];
            string sparse_output_1763_pad_type_1 = const()[name = string("sparse_output_1763_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1763_strides_1 = const()[name = string("sparse_output_1763_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1763_pad_1 = const()[name = string("sparse_output_1763_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1763_dilations_1 = const()[name = string("sparse_output_1763_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1763_groups_1 = const()[name = string("sparse_output_1763_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_21_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(636853952))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(636811904))))[name = string("layers_21_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1763_cast_fp16 = conv(dilations = sparse_output_1763_dilations_1, groups = sparse_output_1763_groups_1, pad = sparse_output_1763_pad_1, pad_type = sparse_output_1763_pad_type_1, strides = sparse_output_1763_strides_1, weight = layers_21_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_1007_cast_fp16)[name = string("sparse_output_1763_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_1009_cast_fp16 = add(x = dense_output_1763_cast_fp16, y = sparse_output_1763_cast_fp16)[name = string("input_1009_cast_fp16")];
            int32 input_1011_split_num_splits_1 = const()[name = string("input_1011_split_num_splits_1"), val = int32(2)];
            int32 input_1011_split_axis_1 = const()[name = string("input_1011_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_1011_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_1011_split_cast_fp16_1 = split(axis = input_1011_split_axis_1, num_splits = input_1011_split_num_splits_1, x = input_1009_cast_fp16)[name = string("input_1011_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1011_split_1_sigmoid_cast_fp16 = sigmoid(x = input_1011_split_cast_fp16_1)[name = string("input_1011_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1011_cast_fp16 = mul(x = input_1011_split_cast_fp16_0, y = input_1011_split_1_sigmoid_cast_fp16)[name = string("input_1011_cast_fp16")];
            tensor<int32, [8]> input_1013_pad_1 = const()[name = string("input_1013_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_1013_mode_1 = const()[name = string("input_1013_mode_1"), val = string("constant")];
            fp16 const_2487_to_fp16 = const()[name = string("const_2487_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_1013_cast_fp16 = pad(constant_val = const_2487_to_fp16, mode = input_1013_mode_1, pad = input_1013_pad_1, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")];
            string dense_output_1765_pad_type_1 = const()[name = string("dense_output_1765_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1765_strides_1 = const()[name = string("dense_output_1765_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1765_pad_1 = const()[name = string("dense_output_1765_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1765_dilations_1 = const()[name = string("dense_output_1765_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1765_groups_1 = const()[name = string("dense_output_1765_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1765_cast_fp16 = conv(dilations = dense_output_1765_dilations_1, groups = dense_output_1765_groups_1, pad = dense_output_1765_pad_1, pad_type = dense_output_1765_pad_type_1, strides = dense_output_1765_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_1013_cast_fp16)[name = string("dense_output_1765_cast_fp16")];
            string sparse_output_1765_pad_type_1 = const()[name = string("sparse_output_1765_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1765_strides_1 = const()[name = string("sparse_output_1765_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1765_pad_1 = const()[name = string("sparse_output_1765_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1765_dilations_1 = const()[name = string("sparse_output_1765_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1765_groups_1 = const()[name = string("sparse_output_1765_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_21_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637116160))))[name = string("layers_21_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1765_cast_fp16 = conv(dilations = sparse_output_1765_dilations_1, groups = sparse_output_1765_groups_1, pad = sparse_output_1765_pad_1, pad_type = sparse_output_1765_pad_type_1, strides = sparse_output_1765_strides_1, weight = layers_21_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_1013_cast_fp16)[name = string("sparse_output_1765_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1015_cast_fp16 = add(x = dense_output_1765_cast_fp16, y = sparse_output_1765_cast_fp16)[name = string("input_1015_cast_fp16")];
            tensor<fp16, [1024]> layers_21_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_21_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637134656)))];
            tensor<fp16, [1024]> layers_21_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_21_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637136768)))];
            tensor<fp16, [1024]> layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_21_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637138880)))];
            tensor<fp16, [1024]> layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_21_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637140992)))];
            tensor<fp16, [1, 1024, 51, 1]> input_1017_cast_fp16 = batch_norm(beta = layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_26861_to_fp16, gamma = layers_21_conv_batch_norm_weight_to_fp16, mean = layers_21_conv_batch_norm_running_mean_to_fp16, variance = layers_21_conv_batch_norm_running_var_to_fp16, x = input_1015_cast_fp16)[name = string("input_1017_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1019_cast_fp16 = silu(x = input_1017_cast_fp16)[name = string("input_1019_cast_fp16")];
            string dense_output_1767_pad_type_1 = const()[name = string("dense_output_1767_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1767_strides_1 = const()[name = string("dense_output_1767_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1767_pad_1 = const()[name = string("dense_output_1767_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1767_dilations_1 = const()[name = string("dense_output_1767_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1767_groups_1 = const()[name = string("dense_output_1767_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637143104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(638191744))))[name = string("layers_21_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1767_cast_fp16 = conv(dilations = dense_output_1767_dilations_1, groups = dense_output_1767_groups_1, pad = dense_output_1767_pad_1, pad_type = dense_output_1767_pad_type_1, strides = dense_output_1767_strides_1, weight = layers_21_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_1019_cast_fp16)[name = string("dense_output_1767_cast_fp16")];
            string sparse_output_1767_pad_type_1 = const()[name = string("sparse_output_1767_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1767_strides_1 = const()[name = string("sparse_output_1767_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1767_pad_1 = const()[name = string("sparse_output_1767_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1767_dilations_1 = const()[name = string("sparse_output_1767_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1767_groups_1 = const()[name = string("sparse_output_1767_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_21_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(638213376))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(638192320))))[name = string("layers_21_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1767_cast_fp16 = conv(dilations = sparse_output_1767_dilations_1, groups = sparse_output_1767_groups_1, pad = sparse_output_1767_pad_1, pad_type = sparse_output_1767_pad_type_1, strides = sparse_output_1767_strides_1, weight = layers_21_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_1019_cast_fp16)[name = string("sparse_output_1767_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1851_cast_fp16 = add(x = dense_output_1767_cast_fp16, y = sparse_output_1767_cast_fp16)[name = string("x_1851_cast_fp16")];
            tensor<int32, [1]> var_28058_axes_1 = const()[name = string("op_28058_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_28058_cast_fp16 = squeeze(axes = var_28058_axes_1, x = x_1851_cast_fp16)[name = string("op_28058_cast_fp16")];
            tensor<int32, [3]> var_28059_perm_1 = const()[name = string("op_28059_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_28059_cast_fp16 = transpose(perm = var_28059_perm_1, x = var_28058_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 51, 1024]> input_1021_cast_fp16 = add(x = input_1005_cast_fp16, y = var_28059_cast_fp16)[name = string("input_1021_cast_fp16")];
            tensor<int32, [1]> x_1853_axes_1 = const()[name = string("x_1853_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(638344512)))];
            tensor<fp16, [1024]> layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(638346624)))];
            tensor<fp16, [1, 51, 1024]> x_1853_cast_fp16 = layer_norm(axes = x_1853_axes_1, beta = layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_26861_to_fp16, gamma = layers_21_norm_feed_forward2_weight_to_fp16, x = input_1021_cast_fp16)[name = string("x_1853_cast_fp16")];
            tensor<int32, [4]> var_28069 = const()[name = string("op_28069"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1855_cast_fp16 = reshape(shape = var_28069, x = x_1853_cast_fp16)[name = string("x_1855_cast_fp16")];
            tensor<int32, [4]> input_1023_perm_1 = const()[name = string("input_1023_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1769_pad_type_1 = const()[name = string("dense_output_1769_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1769_strides_1 = const()[name = string("dense_output_1769_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1769_pad_1 = const()[name = string("dense_output_1769_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1769_dilations_1 = const()[name = string("dense_output_1769_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1769_groups_1 = const()[name = string("dense_output_1769_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_21_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(638348736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642543104))))[name = string("layers_21_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_1023_cast_fp16 = transpose(perm = input_1023_perm_1, x = x_1855_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1769_cast_fp16 = conv(dilations = dense_output_1769_dilations_1, groups = dense_output_1769_groups_1, pad = dense_output_1769_pad_1, pad_type = dense_output_1769_pad_type_1, strides = dense_output_1769_strides_1, weight = layers_21_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_1023_cast_fp16)[name = string("dense_output_1769_cast_fp16")];
            string sparse_output_1769_pad_type_1 = const()[name = string("sparse_output_1769_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1769_strides_1 = const()[name = string("sparse_output_1769_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1769_pad_1 = const()[name = string("sparse_output_1769_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1769_dilations_1 = const()[name = string("sparse_output_1769_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1769_groups_1 = const()[name = string("sparse_output_1769_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_21_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642627648))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642543680))))[name = string("layers_21_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1769_cast_fp16 = conv(dilations = sparse_output_1769_dilations_1, groups = sparse_output_1769_groups_1, pad = sparse_output_1769_pad_1, pad_type = sparse_output_1769_pad_type_1, strides = sparse_output_1769_strides_1, weight = layers_21_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_1023_cast_fp16)[name = string("sparse_output_1769_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1025_cast_fp16 = add(x = dense_output_1769_cast_fp16, y = sparse_output_1769_cast_fp16)[name = string("input_1025_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1027_cast_fp16 = silu(x = input_1025_cast_fp16)[name = string("input_1027_cast_fp16")];
            string dense_output_1771_pad_type_1 = const()[name = string("dense_output_1771_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1771_strides_1 = const()[name = string("dense_output_1771_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1771_pad_1 = const()[name = string("dense_output_1771_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1771_dilations_1 = const()[name = string("dense_output_1771_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1771_groups_1 = const()[name = string("dense_output_1771_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_21_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643152000))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647346368))))[name = string("layers_21_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1771_cast_fp16 = conv(dilations = dense_output_1771_dilations_1, groups = dense_output_1771_groups_1, pad = dense_output_1771_pad_1, pad_type = dense_output_1771_pad_type_1, strides = dense_output_1771_strides_1, weight = layers_21_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_1027_cast_fp16)[name = string("dense_output_1771_cast_fp16")];
            string sparse_output_1771_pad_type_1 = const()[name = string("sparse_output_1771_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1771_strides_1 = const()[name = string("sparse_output_1771_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1771_pad_1 = const()[name = string("sparse_output_1771_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1771_dilations_1 = const()[name = string("sparse_output_1771_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1771_groups_1 = const()[name = string("sparse_output_1771_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_21_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647430912))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647346944))))[name = string("layers_21_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1771_cast_fp16 = conv(dilations = sparse_output_1771_dilations_1, groups = sparse_output_1771_groups_1, pad = sparse_output_1771_pad_1, pad_type = sparse_output_1771_pad_type_1, strides = sparse_output_1771_strides_1, weight = layers_21_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_1027_cast_fp16)[name = string("sparse_output_1771_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1857_cast_fp16 = add(x = dense_output_1771_cast_fp16, y = sparse_output_1771_cast_fp16)[name = string("x_1857_cast_fp16")];
            tensor<int32, [4]> x_1859_perm_1 = const()[name = string("x_1859_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_28104 = const()[name = string("op_28104"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1859_cast_fp16 = transpose(perm = x_1859_perm_1, x = x_1857_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 51, 1024]> var_28105_cast_fp16 = reshape(shape = var_28104, x = x_1859_cast_fp16)[name = string("op_28105_cast_fp16")];
            fp16 var_28106_to_fp16 = const()[name = string("op_28106_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_28107_cast_fp16 = mul(x = var_28105_cast_fp16, y = var_28106_to_fp16)[name = string("op_28107_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_1029_cast_fp16 = add(x = input_1021_cast_fp16, y = var_28107_cast_fp16)[name = string("input_1029_cast_fp16")];
            tensor<int32, [1]> input_1031_axes_1 = const()[name = string("input_1031_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_21_norm_out_weight_to_fp16 = const()[name = string("layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647955264)))];
            tensor<fp16, [1024]> layers_21_norm_out_bias_to_fp16 = const()[name = string("layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647957376)))];
            tensor<fp16, [1, 51, 1024]> input_1031_cast_fp16 = layer_norm(axes = input_1031_axes_1, beta = layers_21_norm_out_bias_to_fp16, epsilon = var_26861_to_fp16, gamma = layers_21_norm_out_weight_to_fp16, x = input_1029_cast_fp16)[name = string("input_1031_cast_fp16")];
            int32 var_28115 = const()[name = string("op_28115"), val = int32(-1)];
            tensor<int32, [1]> x_1861_axes_1 = const()[name = string("x_1861_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647959488)))];
            tensor<fp16, [1024]> layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647961600)))];
            fp16 var_28130_to_fp16 = const()[name = string("op_28130_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1861_cast_fp16 = layer_norm(axes = x_1861_axes_1, beta = layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_28130_to_fp16, gamma = layers_22_norm_feed_forward1_weight_to_fp16, x = input_1031_cast_fp16)[name = string("x_1861_cast_fp16")];
            tensor<int32, [4]> var_28149 = const()[name = string("op_28149"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1863_cast_fp16 = reshape(shape = var_28149, x = x_1861_cast_fp16)[name = string("x_1863_cast_fp16")];
            tensor<int32, [4]> input_1033_perm_1 = const()[name = string("input_1033_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1773_pad_type_1 = const()[name = string("dense_output_1773_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1773_strides_1 = const()[name = string("dense_output_1773_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1773_pad_1 = const()[name = string("dense_output_1773_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1773_dilations_1 = const()[name = string("dense_output_1773_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1773_groups_1 = const()[name = string("dense_output_1773_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_22_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647963712))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652158080))))[name = string("layers_22_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_1033_cast_fp16 = transpose(perm = input_1033_perm_1, x = x_1863_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1773_cast_fp16 = conv(dilations = dense_output_1773_dilations_1, groups = dense_output_1773_groups_1, pad = dense_output_1773_pad_1, pad_type = dense_output_1773_pad_type_1, strides = dense_output_1773_strides_1, weight = layers_22_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_1033_cast_fp16)[name = string("dense_output_1773_cast_fp16")];
            string sparse_output_1773_pad_type_1 = const()[name = string("sparse_output_1773_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1773_strides_1 = const()[name = string("sparse_output_1773_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1773_pad_1 = const()[name = string("sparse_output_1773_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1773_dilations_1 = const()[name = string("sparse_output_1773_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1773_groups_1 = const()[name = string("sparse_output_1773_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_22_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652242624))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652158656))))[name = string("layers_22_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1773_cast_fp16 = conv(dilations = sparse_output_1773_dilations_1, groups = sparse_output_1773_groups_1, pad = sparse_output_1773_pad_1, pad_type = sparse_output_1773_pad_type_1, strides = sparse_output_1773_strides_1, weight = layers_22_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_1033_cast_fp16)[name = string("sparse_output_1773_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1035_cast_fp16 = add(x = dense_output_1773_cast_fp16, y = sparse_output_1773_cast_fp16)[name = string("input_1035_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1037_cast_fp16 = silu(x = input_1035_cast_fp16)[name = string("input_1037_cast_fp16")];
            string dense_output_1775_pad_type_1 = const()[name = string("dense_output_1775_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1775_strides_1 = const()[name = string("dense_output_1775_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1775_pad_1 = const()[name = string("dense_output_1775_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1775_dilations_1 = const()[name = string("dense_output_1775_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1775_groups_1 = const()[name = string("dense_output_1775_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_22_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652766976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(656961344))))[name = string("layers_22_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1775_cast_fp16 = conv(dilations = dense_output_1775_dilations_1, groups = dense_output_1775_groups_1, pad = dense_output_1775_pad_1, pad_type = dense_output_1775_pad_type_1, strides = dense_output_1775_strides_1, weight = layers_22_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_1037_cast_fp16)[name = string("dense_output_1775_cast_fp16")];
            string sparse_output_1775_pad_type_1 = const()[name = string("sparse_output_1775_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1775_strides_1 = const()[name = string("sparse_output_1775_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1775_pad_1 = const()[name = string("sparse_output_1775_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1775_dilations_1 = const()[name = string("sparse_output_1775_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1775_groups_1 = const()[name = string("sparse_output_1775_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_22_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657045888))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(656961920))))[name = string("layers_22_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1775_cast_fp16 = conv(dilations = sparse_output_1775_dilations_1, groups = sparse_output_1775_groups_1, pad = sparse_output_1775_pad_1, pad_type = sparse_output_1775_pad_type_1, strides = sparse_output_1775_strides_1, weight = layers_22_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_1037_cast_fp16)[name = string("sparse_output_1775_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1865_cast_fp16 = add(x = dense_output_1775_cast_fp16, y = sparse_output_1775_cast_fp16)[name = string("x_1865_cast_fp16")];
            tensor<int32, [4]> x_1867_perm_1 = const()[name = string("x_1867_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_28184 = const()[name = string("op_28184"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1867_cast_fp16 = transpose(perm = x_1867_perm_1, x = x_1865_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 51, 1024]> var_28185_cast_fp16 = reshape(shape = var_28184, x = x_1867_cast_fp16)[name = string("op_28185_cast_fp16")];
            fp16 var_28186_to_fp16 = const()[name = string("op_28186_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_28187_cast_fp16 = mul(x = var_28185_cast_fp16, y = var_28186_to_fp16)[name = string("op_28187_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_1039_cast_fp16 = add(x = input_1031_cast_fp16, y = var_28187_cast_fp16)[name = string("input_1039_cast_fp16")];
            tensor<int32, [1]> q_45_axes_1 = const()[name = string("q_45_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_22_norm_self_att_weight_to_fp16 = const()[name = string("layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657570240)))];
            tensor<fp16, [1024]> layers_22_norm_self_att_bias_to_fp16 = const()[name = string("layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657572352)))];
            tensor<fp16, [1, 51, 1024]> q_45_cast_fp16 = layer_norm(axes = q_45_axes_1, beta = layers_22_norm_self_att_bias_to_fp16, epsilon = var_28130_to_fp16, gamma = layers_22_norm_self_att_weight_to_fp16, x = input_1039_cast_fp16)[name = string("q_45_cast_fp16")];
            tensor<int32, [3]> var_28261 = const()[name = string("op_28261"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1041_axes_1 = const()[name = string("input_1041_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_28262_cast_fp16 = transpose(perm = var_28261, x = q_45_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 1024, 51, 1]> input_1041_cast_fp16 = expand_dims(axes = input_1041_axes_1, x = var_28262_cast_fp16)[name = string("input_1041_cast_fp16")];
            string dense_output_1777_pad_type_1 = const()[name = string("dense_output_1777_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1777_strides_1 = const()[name = string("dense_output_1777_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1777_pad_1 = const()[name = string("dense_output_1777_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1777_dilations_1 = const()[name = string("dense_output_1777_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1777_groups_1 = const()[name = string("dense_output_1777_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657574464))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657705600))))[name = string("layers_22_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1777_cast_fp16 = conv(dilations = dense_output_1777_dilations_1, groups = dense_output_1777_groups_1, pad = dense_output_1777_pad_1, pad_type = dense_output_1777_pad_type_1, strides = dense_output_1777_strides_1, weight = layers_22_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1777_cast_fp16")];
            string sparse_output_1777_pad_type_1 = const()[name = string("sparse_output_1777_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1777_strides_1 = const()[name = string("sparse_output_1777_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1777_pad_1 = const()[name = string("sparse_output_1777_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1777_dilations_1 = const()[name = string("sparse_output_1777_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1777_groups_1 = const()[name = string("sparse_output_1777_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657708864))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657706176))))[name = string("layers_22_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1777_cast_fp16 = conv(dilations = sparse_output_1777_dilations_1, groups = sparse_output_1777_groups_1, pad = sparse_output_1777_pad_1, pad_type = sparse_output_1777_pad_type_1, strides = sparse_output_1777_strides_1, weight = layers_22_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1777_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28287_cast_fp16 = add(x = dense_output_1777_cast_fp16, y = sparse_output_1777_cast_fp16)[name = string("op_28287_cast_fp16")];
            tensor<int32, [4]> var_28288 = const()[name = string("op_28288"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28290 = const()[name = string("op_28290"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28289_cast_fp16 = transpose(perm = var_28288, x = var_28287_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 51, 128]> q_head_353_cast_fp16 = reshape(shape = var_28290, x = var_28289_cast_fp16)[name = string("q_head_353_cast_fp16")];
            string dense_output_1779_pad_type_1 = const()[name = string("dense_output_1779_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1779_strides_1 = const()[name = string("dense_output_1779_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1779_pad_1 = const()[name = string("dense_output_1779_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1779_dilations_1 = const()[name = string("dense_output_1779_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1779_groups_1 = const()[name = string("dense_output_1779_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657725312))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657856448))))[name = string("layers_22_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1779_cast_fp16 = conv(dilations = dense_output_1779_dilations_1, groups = dense_output_1779_groups_1, pad = dense_output_1779_pad_1, pad_type = dense_output_1779_pad_type_1, strides = dense_output_1779_strides_1, weight = layers_22_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1779_cast_fp16")];
            string sparse_output_1779_pad_type_1 = const()[name = string("sparse_output_1779_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1779_strides_1 = const()[name = string("sparse_output_1779_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1779_pad_1 = const()[name = string("sparse_output_1779_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1779_dilations_1 = const()[name = string("sparse_output_1779_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1779_groups_1 = const()[name = string("sparse_output_1779_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657859712))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657857024))))[name = string("layers_22_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1779_cast_fp16 = conv(dilations = sparse_output_1779_dilations_1, groups = sparse_output_1779_groups_1, pad = sparse_output_1779_pad_1, pad_type = sparse_output_1779_pad_type_1, strides = sparse_output_1779_strides_1, weight = layers_22_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1779_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28306_cast_fp16 = add(x = dense_output_1779_cast_fp16, y = sparse_output_1779_cast_fp16)[name = string("op_28306_cast_fp16")];
            tensor<int32, [4]> var_28307 = const()[name = string("op_28307"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28309 = const()[name = string("op_28309"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28308_cast_fp16 = transpose(perm = var_28307, x = var_28306_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 51, 128]> k_head_705_cast_fp16 = reshape(shape = var_28309, x = var_28308_cast_fp16)[name = string("k_head_705_cast_fp16")];
            string dense_output_1781_pad_type_1 = const()[name = string("dense_output_1781_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1781_strides_1 = const()[name = string("dense_output_1781_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1781_pad_1 = const()[name = string("dense_output_1781_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1781_dilations_1 = const()[name = string("dense_output_1781_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1781_groups_1 = const()[name = string("dense_output_1781_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657876160))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658007296))))[name = string("layers_22_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1781_cast_fp16 = conv(dilations = dense_output_1781_dilations_1, groups = dense_output_1781_groups_1, pad = dense_output_1781_pad_1, pad_type = dense_output_1781_pad_type_1, strides = dense_output_1781_strides_1, weight = layers_22_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1781_cast_fp16")];
            string sparse_output_1781_pad_type_1 = const()[name = string("sparse_output_1781_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1781_strides_1 = const()[name = string("sparse_output_1781_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1781_pad_1 = const()[name = string("sparse_output_1781_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1781_dilations_1 = const()[name = string("sparse_output_1781_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1781_groups_1 = const()[name = string("sparse_output_1781_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658010560))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658007872))))[name = string("layers_22_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1781_cast_fp16 = conv(dilations = sparse_output_1781_dilations_1, groups = sparse_output_1781_groups_1, pad = sparse_output_1781_pad_1, pad_type = sparse_output_1781_pad_type_1, strides = sparse_output_1781_strides_1, weight = layers_22_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1781_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28325_cast_fp16 = add(x = dense_output_1781_cast_fp16, y = sparse_output_1781_cast_fp16)[name = string("op_28325_cast_fp16")];
            tensor<int32, [4]> var_28326 = const()[name = string("op_28326"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28328 = const()[name = string("op_28328"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28327_cast_fp16 = transpose(perm = var_28326, x = var_28325_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 51, 128]> v_head_705_cast_fp16 = reshape(shape = var_28328, x = var_28327_cast_fp16)[name = string("v_head_705_cast_fp16")];
            string dense_output_1783_pad_type_1 = const()[name = string("dense_output_1783_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1783_strides_1 = const()[name = string("dense_output_1783_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1783_pad_1 = const()[name = string("dense_output_1783_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1783_dilations_1 = const()[name = string("dense_output_1783_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1783_groups_1 = const()[name = string("dense_output_1783_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658027008))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658158144))))[name = string("layers_22_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1783_cast_fp16 = conv(dilations = dense_output_1783_dilations_1, groups = dense_output_1783_groups_1, pad = dense_output_1783_pad_1, pad_type = dense_output_1783_pad_type_1, strides = dense_output_1783_strides_1, weight = layers_22_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1783_cast_fp16")];
            string sparse_output_1783_pad_type_1 = const()[name = string("sparse_output_1783_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1783_strides_1 = const()[name = string("sparse_output_1783_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1783_pad_1 = const()[name = string("sparse_output_1783_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1783_dilations_1 = const()[name = string("sparse_output_1783_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1783_groups_1 = const()[name = string("sparse_output_1783_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658161408))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658158720))))[name = string("layers_22_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1783_cast_fp16 = conv(dilations = sparse_output_1783_dilations_1, groups = sparse_output_1783_groups_1, pad = sparse_output_1783_pad_1, pad_type = sparse_output_1783_pad_type_1, strides = sparse_output_1783_strides_1, weight = layers_22_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1783_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_28344_cast_fp16 = add(x = dense_output_1783_cast_fp16, y = sparse_output_1783_cast_fp16)[name = string("op_28344_cast_fp16")];
            tensor<int32, [4]> var_28345 = const()[name = string("op_28345"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28347 = const()[name = string("op_28347"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_28346_cast_fp16 = transpose(perm = var_28345, x = var_28344_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 101, 128]> p_head_705_cast_fp16 = reshape(shape = var_28347, x = var_28346_cast_fp16)[name = string("p_head_705_cast_fp16")];
            tensor<fp16, [128]> var_28349_to_fp16 = const()[name = string("op_28349_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658177856)))];
            tensor<fp16, [1, 51, 128]> var_28350_cast_fp16 = add(x = q_head_353_cast_fp16, y = var_28349_to_fp16)[name = string("op_28350_cast_fp16")];
            tensor<int32, [1]> q_u_353_axes_1 = const()[name = string("q_u_353_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_353_cast_fp16 = expand_dims(axes = q_u_353_axes_1, x = var_28350_cast_fp16)[name = string("q_u_353_cast_fp16")];
            tensor<fp16, [128]> var_28352_to_fp16 = const()[name = string("op_28352_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658178176)))];
            tensor<fp16, [1, 51, 128]> var_28353_cast_fp16 = add(x = q_head_353_cast_fp16, y = var_28352_to_fp16)[name = string("op_28353_cast_fp16")];
            tensor<int32, [1]> q_v_353_axes_1 = const()[name = string("q_v_353_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_353_cast_fp16 = expand_dims(axes = q_v_353_axes_1, x = var_28353_cast_fp16)[name = string("q_v_353_cast_fp16")];
            tensor<int32, [1]> k_head_707_axes_1 = const()[name = string("k_head_707_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_707_cast_fp16 = expand_dims(axes = k_head_707_axes_1, x = k_head_705_cast_fp16)[name = string("k_head_707_cast_fp16")];
            tensor<int32, [1]> v_head_707_axes_1 = const()[name = string("v_head_707_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_707_cast_fp16 = expand_dims(axes = v_head_707_axes_1, x = v_head_705_cast_fp16)[name = string("v_head_707_cast_fp16")];
            tensor<int32, [1]> p_head_707_axes_1 = const()[name = string("p_head_707_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_707_cast_fp16 = expand_dims(axes = p_head_707_axes_1, x = p_head_705_cast_fp16)[name = string("p_head_707_cast_fp16")];
            bool var_28359_transpose_x_3 = const()[name = string("op_28359_transpose_x_3"), val = bool(false)];
            bool var_28359_transpose_y_3 = const()[name = string("op_28359_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_28359_cast_fp16 = matmul(transpose_x = var_28359_transpose_x_3, transpose_y = var_28359_transpose_y_3, x = q_u_353_cast_fp16, y = k_head_707_cast_fp16)[name = string("op_28359_cast_fp16")];
            fp16 var_28360_to_fp16 = const()[name = string("op_28360_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_353_cast_fp16 = mul(x = var_28359_cast_fp16, y = var_28360_to_fp16)[name = string("scores_content_353_cast_fp16")];
            bool x_1869_transpose_x_3 = const()[name = string("x_1869_transpose_x_3"), val = bool(false)];
            bool x_1869_transpose_y_3 = const()[name = string("x_1869_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1869_cast_fp16 = matmul(transpose_x = x_1869_transpose_x_3, transpose_y = x_1869_transpose_y_3, x = q_v_353_cast_fp16, y = p_head_707_cast_fp16)[name = string("x_1869_cast_fp16")];
            tensor<int32, [8]> x_1871_pad_1 = const()[name = string("x_1871_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1871_mode_1 = const()[name = string("x_1871_mode_1"), val = string("constant")];
            fp16 const_2497_to_fp16 = const()[name = string("const_2497_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1871_cast_fp16 = pad(constant_val = const_2497_to_fp16, mode = x_1871_mode_1, pad = x_1871_pad_1, x = x_1869_cast_fp16)[name = string("x_1871_cast_fp16")];
            tensor<int32, [4]> var_28374 = const()[name = string("op_28374"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1873_cast_fp16 = reshape(shape = var_28374, x = x_1871_cast_fp16)[name = string("x_1873_cast_fp16")];
            tensor<int32, [4]> var_28378_begin_1 = const()[name = string("op_28378_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_28378_end_1 = const()[name = string("op_28378_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_28378_end_mask_1 = const()[name = string("op_28378_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_28378_cast_fp16 = slice_by_index(begin = var_28378_begin_1, end = var_28378_end_1, end_mask = var_28378_end_mask_1, x = x_1873_cast_fp16)[name = string("op_28378_cast_fp16")];
            tensor<int32, [4]> var_28380 = const()[name = string("op_28380"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_28381_cast_fp16 = reshape(shape = var_28380, x = var_28378_cast_fp16)[name = string("op_28381_cast_fp16")];
            tensor<int32, [4]> var_28386_begin_1 = const()[name = string("op_28386_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_28386_end_1 = const()[name = string("op_28386_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_28386_end_mask_1 = const()[name = string("op_28386_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_28386_cast_fp16 = slice_by_index(begin = var_28386_begin_1, end = var_28386_end_1, end_mask = var_28386_end_mask_1, x = var_28381_cast_fp16)[name = string("op_28386_cast_fp16")];
            fp16 var_28387_to_fp16 = const()[name = string("op_28387_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_353_cast_fp16 = mul(x = var_28386_cast_fp16, y = var_28387_to_fp16)[name = string("scores_pos_353_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_353_cast_fp16 = add(x = scores_content_353_cast_fp16, y = scores_pos_353_cast_fp16)[name = string("logits_353_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_28390_cast_fp16 = softmax(axis = var_28115, x = logits_353_cast_fp16)[name = string("op_28390_cast_fp16")];
            bool var_28392_transpose_x_1 = const()[name = string("op_28392_transpose_x_1"), val = bool(false)];
            bool var_28392_transpose_y_1 = const()[name = string("op_28392_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_28392_cast_fp16 = matmul(transpose_x = var_28392_transpose_x_1, transpose_y = var_28392_transpose_y_1, x = var_28390_cast_fp16, y = v_head_707_cast_fp16)[name = string("op_28392_cast_fp16")];
            tensor<int32, [1]> var_28393_axes_1 = const()[name = string("op_28393_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_28393_cast_fp16 = squeeze(axes = var_28393_axes_1, x = var_28392_cast_fp16)[name = string("op_28393_cast_fp16")];
            string dense_output_1785_pad_type_1 = const()[name = string("dense_output_1785_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1785_strides_1 = const()[name = string("dense_output_1785_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1785_pad_1 = const()[name = string("dense_output_1785_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1785_dilations_1 = const()[name = string("dense_output_1785_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1785_groups_1 = const()[name = string("dense_output_1785_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658178496))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658309632))))[name = string("layers_22_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1785_cast_fp16 = conv(dilations = dense_output_1785_dilations_1, groups = dense_output_1785_groups_1, pad = dense_output_1785_pad_1, pad_type = dense_output_1785_pad_type_1, strides = dense_output_1785_strides_1, weight = layers_22_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1785_cast_fp16")];
            string sparse_output_1785_pad_type_1 = const()[name = string("sparse_output_1785_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1785_strides_1 = const()[name = string("sparse_output_1785_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1785_pad_1 = const()[name = string("sparse_output_1785_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1785_dilations_1 = const()[name = string("sparse_output_1785_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1785_groups_1 = const()[name = string("sparse_output_1785_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658312896))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658310208))))[name = string("layers_22_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1785_cast_fp16 = conv(dilations = sparse_output_1785_dilations_1, groups = sparse_output_1785_groups_1, pad = sparse_output_1785_pad_1, pad_type = sparse_output_1785_pad_type_1, strides = sparse_output_1785_strides_1, weight = layers_22_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1785_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28408_cast_fp16 = add(x = dense_output_1785_cast_fp16, y = sparse_output_1785_cast_fp16)[name = string("op_28408_cast_fp16")];
            tensor<int32, [4]> var_28409 = const()[name = string("op_28409"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28411 = const()[name = string("op_28411"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28410_cast_fp16 = transpose(perm = var_28409, x = var_28408_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 51, 128]> q_head_355_cast_fp16 = reshape(shape = var_28411, x = var_28410_cast_fp16)[name = string("q_head_355_cast_fp16")];
            string dense_output_1787_pad_type_1 = const()[name = string("dense_output_1787_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1787_strides_1 = const()[name = string("dense_output_1787_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1787_pad_1 = const()[name = string("dense_output_1787_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1787_dilations_1 = const()[name = string("dense_output_1787_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1787_groups_1 = const()[name = string("dense_output_1787_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658329344))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658460480))))[name = string("layers_22_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1787_cast_fp16 = conv(dilations = dense_output_1787_dilations_1, groups = dense_output_1787_groups_1, pad = dense_output_1787_pad_1, pad_type = dense_output_1787_pad_type_1, strides = dense_output_1787_strides_1, weight = layers_22_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1787_cast_fp16")];
            string sparse_output_1787_pad_type_1 = const()[name = string("sparse_output_1787_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1787_strides_1 = const()[name = string("sparse_output_1787_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1787_pad_1 = const()[name = string("sparse_output_1787_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1787_dilations_1 = const()[name = string("sparse_output_1787_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1787_groups_1 = const()[name = string("sparse_output_1787_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658463744))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658461056))))[name = string("layers_22_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1787_cast_fp16 = conv(dilations = sparse_output_1787_dilations_1, groups = sparse_output_1787_groups_1, pad = sparse_output_1787_pad_1, pad_type = sparse_output_1787_pad_type_1, strides = sparse_output_1787_strides_1, weight = layers_22_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1787_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28427_cast_fp16 = add(x = dense_output_1787_cast_fp16, y = sparse_output_1787_cast_fp16)[name = string("op_28427_cast_fp16")];
            tensor<int32, [4]> var_28428 = const()[name = string("op_28428"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28430 = const()[name = string("op_28430"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28429_cast_fp16 = transpose(perm = var_28428, x = var_28427_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 51, 128]> k_head_709_cast_fp16 = reshape(shape = var_28430, x = var_28429_cast_fp16)[name = string("k_head_709_cast_fp16")];
            string dense_output_1789_pad_type_1 = const()[name = string("dense_output_1789_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1789_strides_1 = const()[name = string("dense_output_1789_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1789_pad_1 = const()[name = string("dense_output_1789_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1789_dilations_1 = const()[name = string("dense_output_1789_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1789_groups_1 = const()[name = string("dense_output_1789_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658480192))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658611328))))[name = string("layers_22_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1789_cast_fp16 = conv(dilations = dense_output_1789_dilations_1, groups = dense_output_1789_groups_1, pad = dense_output_1789_pad_1, pad_type = dense_output_1789_pad_type_1, strides = dense_output_1789_strides_1, weight = layers_22_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1789_cast_fp16")];
            string sparse_output_1789_pad_type_1 = const()[name = string("sparse_output_1789_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1789_strides_1 = const()[name = string("sparse_output_1789_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1789_pad_1 = const()[name = string("sparse_output_1789_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1789_dilations_1 = const()[name = string("sparse_output_1789_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1789_groups_1 = const()[name = string("sparse_output_1789_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658614592))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658611904))))[name = string("layers_22_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1789_cast_fp16 = conv(dilations = sparse_output_1789_dilations_1, groups = sparse_output_1789_groups_1, pad = sparse_output_1789_pad_1, pad_type = sparse_output_1789_pad_type_1, strides = sparse_output_1789_strides_1, weight = layers_22_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1789_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28446_cast_fp16 = add(x = dense_output_1789_cast_fp16, y = sparse_output_1789_cast_fp16)[name = string("op_28446_cast_fp16")];
            tensor<int32, [4]> var_28447 = const()[name = string("op_28447"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28449 = const()[name = string("op_28449"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28448_cast_fp16 = transpose(perm = var_28447, x = var_28446_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 51, 128]> v_head_709_cast_fp16 = reshape(shape = var_28449, x = var_28448_cast_fp16)[name = string("v_head_709_cast_fp16")];
            string dense_output_1791_pad_type_1 = const()[name = string("dense_output_1791_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1791_strides_1 = const()[name = string("dense_output_1791_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1791_pad_1 = const()[name = string("dense_output_1791_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1791_dilations_1 = const()[name = string("dense_output_1791_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1791_groups_1 = const()[name = string("dense_output_1791_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658631040))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658762176))))[name = string("layers_22_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1791_cast_fp16 = conv(dilations = dense_output_1791_dilations_1, groups = dense_output_1791_groups_1, pad = dense_output_1791_pad_1, pad_type = dense_output_1791_pad_type_1, strides = dense_output_1791_strides_1, weight = layers_22_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1791_cast_fp16")];
            string sparse_output_1791_pad_type_1 = const()[name = string("sparse_output_1791_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1791_strides_1 = const()[name = string("sparse_output_1791_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1791_pad_1 = const()[name = string("sparse_output_1791_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1791_dilations_1 = const()[name = string("sparse_output_1791_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1791_groups_1 = const()[name = string("sparse_output_1791_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658765440))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658762752))))[name = string("layers_22_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1791_cast_fp16 = conv(dilations = sparse_output_1791_dilations_1, groups = sparse_output_1791_groups_1, pad = sparse_output_1791_pad_1, pad_type = sparse_output_1791_pad_type_1, strides = sparse_output_1791_strides_1, weight = layers_22_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1791_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_28465_cast_fp16 = add(x = dense_output_1791_cast_fp16, y = sparse_output_1791_cast_fp16)[name = string("op_28465_cast_fp16")];
            tensor<int32, [4]> var_28466 = const()[name = string("op_28466"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28468 = const()[name = string("op_28468"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_28467_cast_fp16 = transpose(perm = var_28466, x = var_28465_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 101, 128]> p_head_709_cast_fp16 = reshape(shape = var_28468, x = var_28467_cast_fp16)[name = string("p_head_709_cast_fp16")];
            tensor<fp16, [128]> var_28470_to_fp16 = const()[name = string("op_28470_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658781888)))];
            tensor<fp16, [1, 51, 128]> var_28471_cast_fp16 = add(x = q_head_355_cast_fp16, y = var_28470_to_fp16)[name = string("op_28471_cast_fp16")];
            tensor<int32, [1]> q_u_355_axes_1 = const()[name = string("q_u_355_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_355_cast_fp16 = expand_dims(axes = q_u_355_axes_1, x = var_28471_cast_fp16)[name = string("q_u_355_cast_fp16")];
            tensor<fp16, [128]> var_28473_to_fp16 = const()[name = string("op_28473_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658782208)))];
            tensor<fp16, [1, 51, 128]> var_28474_cast_fp16 = add(x = q_head_355_cast_fp16, y = var_28473_to_fp16)[name = string("op_28474_cast_fp16")];
            tensor<int32, [1]> q_v_355_axes_1 = const()[name = string("q_v_355_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_355_cast_fp16 = expand_dims(axes = q_v_355_axes_1, x = var_28474_cast_fp16)[name = string("q_v_355_cast_fp16")];
            tensor<int32, [1]> k_head_711_axes_1 = const()[name = string("k_head_711_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_711_cast_fp16 = expand_dims(axes = k_head_711_axes_1, x = k_head_709_cast_fp16)[name = string("k_head_711_cast_fp16")];
            tensor<int32, [1]> v_head_711_axes_1 = const()[name = string("v_head_711_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_711_cast_fp16 = expand_dims(axes = v_head_711_axes_1, x = v_head_709_cast_fp16)[name = string("v_head_711_cast_fp16")];
            tensor<int32, [1]> p_head_711_axes_1 = const()[name = string("p_head_711_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_711_cast_fp16 = expand_dims(axes = p_head_711_axes_1, x = p_head_709_cast_fp16)[name = string("p_head_711_cast_fp16")];
            bool var_28480_transpose_x_3 = const()[name = string("op_28480_transpose_x_3"), val = bool(false)];
            bool var_28480_transpose_y_3 = const()[name = string("op_28480_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_28480_cast_fp16 = matmul(transpose_x = var_28480_transpose_x_3, transpose_y = var_28480_transpose_y_3, x = q_u_355_cast_fp16, y = k_head_711_cast_fp16)[name = string("op_28480_cast_fp16")];
            fp16 var_28481_to_fp16 = const()[name = string("op_28481_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_355_cast_fp16 = mul(x = var_28480_cast_fp16, y = var_28481_to_fp16)[name = string("scores_content_355_cast_fp16")];
            bool x_1877_transpose_x_3 = const()[name = string("x_1877_transpose_x_3"), val = bool(false)];
            bool x_1877_transpose_y_3 = const()[name = string("x_1877_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1877_cast_fp16 = matmul(transpose_x = x_1877_transpose_x_3, transpose_y = x_1877_transpose_y_3, x = q_v_355_cast_fp16, y = p_head_711_cast_fp16)[name = string("x_1877_cast_fp16")];
            tensor<int32, [8]> x_1879_pad_1 = const()[name = string("x_1879_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1879_mode_1 = const()[name = string("x_1879_mode_1"), val = string("constant")];
            fp16 const_2503_to_fp16 = const()[name = string("const_2503_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1879_cast_fp16 = pad(constant_val = const_2503_to_fp16, mode = x_1879_mode_1, pad = x_1879_pad_1, x = x_1877_cast_fp16)[name = string("x_1879_cast_fp16")];
            tensor<int32, [4]> var_28495 = const()[name = string("op_28495"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1881_cast_fp16 = reshape(shape = var_28495, x = x_1879_cast_fp16)[name = string("x_1881_cast_fp16")];
            tensor<int32, [4]> var_28499_begin_1 = const()[name = string("op_28499_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_28499_end_1 = const()[name = string("op_28499_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_28499_end_mask_1 = const()[name = string("op_28499_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_28499_cast_fp16 = slice_by_index(begin = var_28499_begin_1, end = var_28499_end_1, end_mask = var_28499_end_mask_1, x = x_1881_cast_fp16)[name = string("op_28499_cast_fp16")];
            tensor<int32, [4]> var_28501 = const()[name = string("op_28501"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_28502_cast_fp16 = reshape(shape = var_28501, x = var_28499_cast_fp16)[name = string("op_28502_cast_fp16")];
            tensor<int32, [4]> var_28507_begin_1 = const()[name = string("op_28507_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_28507_end_1 = const()[name = string("op_28507_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_28507_end_mask_1 = const()[name = string("op_28507_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_28507_cast_fp16 = slice_by_index(begin = var_28507_begin_1, end = var_28507_end_1, end_mask = var_28507_end_mask_1, x = var_28502_cast_fp16)[name = string("op_28507_cast_fp16")];
            fp16 var_28508_to_fp16 = const()[name = string("op_28508_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_355_cast_fp16 = mul(x = var_28507_cast_fp16, y = var_28508_to_fp16)[name = string("scores_pos_355_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_355_cast_fp16 = add(x = scores_content_355_cast_fp16, y = scores_pos_355_cast_fp16)[name = string("logits_355_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_28511_cast_fp16 = softmax(axis = var_28115, x = logits_355_cast_fp16)[name = string("op_28511_cast_fp16")];
            bool var_28513_transpose_x_1 = const()[name = string("op_28513_transpose_x_1"), val = bool(false)];
            bool var_28513_transpose_y_1 = const()[name = string("op_28513_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_28513_cast_fp16 = matmul(transpose_x = var_28513_transpose_x_1, transpose_y = var_28513_transpose_y_1, x = var_28511_cast_fp16, y = v_head_711_cast_fp16)[name = string("op_28513_cast_fp16")];
            tensor<int32, [1]> var_28514_axes_1 = const()[name = string("op_28514_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_28514_cast_fp16 = squeeze(axes = var_28514_axes_1, x = var_28513_cast_fp16)[name = string("op_28514_cast_fp16")];
            string dense_output_1793_pad_type_1 = const()[name = string("dense_output_1793_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1793_strides_1 = const()[name = string("dense_output_1793_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1793_pad_1 = const()[name = string("dense_output_1793_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1793_dilations_1 = const()[name = string("dense_output_1793_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1793_groups_1 = const()[name = string("dense_output_1793_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658782528))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658913664))))[name = string("layers_22_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1793_cast_fp16 = conv(dilations = dense_output_1793_dilations_1, groups = dense_output_1793_groups_1, pad = dense_output_1793_pad_1, pad_type = dense_output_1793_pad_type_1, strides = dense_output_1793_strides_1, weight = layers_22_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1793_cast_fp16")];
            string sparse_output_1793_pad_type_1 = const()[name = string("sparse_output_1793_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1793_strides_1 = const()[name = string("sparse_output_1793_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1793_pad_1 = const()[name = string("sparse_output_1793_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1793_dilations_1 = const()[name = string("sparse_output_1793_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1793_groups_1 = const()[name = string("sparse_output_1793_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658916928))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658914240))))[name = string("layers_22_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1793_cast_fp16 = conv(dilations = sparse_output_1793_dilations_1, groups = sparse_output_1793_groups_1, pad = sparse_output_1793_pad_1, pad_type = sparse_output_1793_pad_type_1, strides = sparse_output_1793_strides_1, weight = layers_22_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1793_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28529_cast_fp16 = add(x = dense_output_1793_cast_fp16, y = sparse_output_1793_cast_fp16)[name = string("op_28529_cast_fp16")];
            tensor<int32, [4]> var_28530 = const()[name = string("op_28530"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28532 = const()[name = string("op_28532"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28531_cast_fp16 = transpose(perm = var_28530, x = var_28529_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 51, 128]> q_head_357_cast_fp16 = reshape(shape = var_28532, x = var_28531_cast_fp16)[name = string("q_head_357_cast_fp16")];
            string dense_output_1795_pad_type_1 = const()[name = string("dense_output_1795_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1795_strides_1 = const()[name = string("dense_output_1795_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1795_pad_1 = const()[name = string("dense_output_1795_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1795_dilations_1 = const()[name = string("dense_output_1795_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1795_groups_1 = const()[name = string("dense_output_1795_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658933376))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659064512))))[name = string("layers_22_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1795_cast_fp16 = conv(dilations = dense_output_1795_dilations_1, groups = dense_output_1795_groups_1, pad = dense_output_1795_pad_1, pad_type = dense_output_1795_pad_type_1, strides = dense_output_1795_strides_1, weight = layers_22_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1795_cast_fp16")];
            string sparse_output_1795_pad_type_1 = const()[name = string("sparse_output_1795_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1795_strides_1 = const()[name = string("sparse_output_1795_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1795_pad_1 = const()[name = string("sparse_output_1795_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1795_dilations_1 = const()[name = string("sparse_output_1795_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1795_groups_1 = const()[name = string("sparse_output_1795_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659067776))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659065088))))[name = string("layers_22_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1795_cast_fp16 = conv(dilations = sparse_output_1795_dilations_1, groups = sparse_output_1795_groups_1, pad = sparse_output_1795_pad_1, pad_type = sparse_output_1795_pad_type_1, strides = sparse_output_1795_strides_1, weight = layers_22_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1795_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28548_cast_fp16 = add(x = dense_output_1795_cast_fp16, y = sparse_output_1795_cast_fp16)[name = string("op_28548_cast_fp16")];
            tensor<int32, [4]> var_28549 = const()[name = string("op_28549"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28551 = const()[name = string("op_28551"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28550_cast_fp16 = transpose(perm = var_28549, x = var_28548_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 51, 128]> k_head_713_cast_fp16 = reshape(shape = var_28551, x = var_28550_cast_fp16)[name = string("k_head_713_cast_fp16")];
            string dense_output_1797_pad_type_1 = const()[name = string("dense_output_1797_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1797_strides_1 = const()[name = string("dense_output_1797_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1797_pad_1 = const()[name = string("dense_output_1797_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1797_dilations_1 = const()[name = string("dense_output_1797_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1797_groups_1 = const()[name = string("dense_output_1797_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659084224))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659215360))))[name = string("layers_22_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1797_cast_fp16 = conv(dilations = dense_output_1797_dilations_1, groups = dense_output_1797_groups_1, pad = dense_output_1797_pad_1, pad_type = dense_output_1797_pad_type_1, strides = dense_output_1797_strides_1, weight = layers_22_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1797_cast_fp16")];
            string sparse_output_1797_pad_type_1 = const()[name = string("sparse_output_1797_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1797_strides_1 = const()[name = string("sparse_output_1797_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1797_pad_1 = const()[name = string("sparse_output_1797_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1797_dilations_1 = const()[name = string("sparse_output_1797_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1797_groups_1 = const()[name = string("sparse_output_1797_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659218624))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659215936))))[name = string("layers_22_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1797_cast_fp16 = conv(dilations = sparse_output_1797_dilations_1, groups = sparse_output_1797_groups_1, pad = sparse_output_1797_pad_1, pad_type = sparse_output_1797_pad_type_1, strides = sparse_output_1797_strides_1, weight = layers_22_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1797_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28567_cast_fp16 = add(x = dense_output_1797_cast_fp16, y = sparse_output_1797_cast_fp16)[name = string("op_28567_cast_fp16")];
            tensor<int32, [4]> var_28568 = const()[name = string("op_28568"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28570 = const()[name = string("op_28570"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28569_cast_fp16 = transpose(perm = var_28568, x = var_28567_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 51, 128]> v_head_713_cast_fp16 = reshape(shape = var_28570, x = var_28569_cast_fp16)[name = string("v_head_713_cast_fp16")];
            string dense_output_1799_pad_type_1 = const()[name = string("dense_output_1799_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1799_strides_1 = const()[name = string("dense_output_1799_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1799_pad_1 = const()[name = string("dense_output_1799_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1799_dilations_1 = const()[name = string("dense_output_1799_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1799_groups_1 = const()[name = string("dense_output_1799_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659235072))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659366208))))[name = string("layers_22_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1799_cast_fp16 = conv(dilations = dense_output_1799_dilations_1, groups = dense_output_1799_groups_1, pad = dense_output_1799_pad_1, pad_type = dense_output_1799_pad_type_1, strides = dense_output_1799_strides_1, weight = layers_22_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1799_cast_fp16")];
            string sparse_output_1799_pad_type_1 = const()[name = string("sparse_output_1799_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1799_strides_1 = const()[name = string("sparse_output_1799_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1799_pad_1 = const()[name = string("sparse_output_1799_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1799_dilations_1 = const()[name = string("sparse_output_1799_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1799_groups_1 = const()[name = string("sparse_output_1799_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659369472))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659366784))))[name = string("layers_22_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1799_cast_fp16 = conv(dilations = sparse_output_1799_dilations_1, groups = sparse_output_1799_groups_1, pad = sparse_output_1799_pad_1, pad_type = sparse_output_1799_pad_type_1, strides = sparse_output_1799_strides_1, weight = layers_22_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1799_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_28586_cast_fp16 = add(x = dense_output_1799_cast_fp16, y = sparse_output_1799_cast_fp16)[name = string("op_28586_cast_fp16")];
            tensor<int32, [4]> var_28587 = const()[name = string("op_28587"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28589 = const()[name = string("op_28589"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_28588_cast_fp16 = transpose(perm = var_28587, x = var_28586_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 101, 128]> p_head_713_cast_fp16 = reshape(shape = var_28589, x = var_28588_cast_fp16)[name = string("p_head_713_cast_fp16")];
            tensor<fp16, [128]> var_28591_to_fp16 = const()[name = string("op_28591_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659385920)))];
            tensor<fp16, [1, 51, 128]> var_28592_cast_fp16 = add(x = q_head_357_cast_fp16, y = var_28591_to_fp16)[name = string("op_28592_cast_fp16")];
            tensor<int32, [1]> q_u_357_axes_1 = const()[name = string("q_u_357_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_357_cast_fp16 = expand_dims(axes = q_u_357_axes_1, x = var_28592_cast_fp16)[name = string("q_u_357_cast_fp16")];
            tensor<fp16, [128]> var_28594_to_fp16 = const()[name = string("op_28594_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659386240)))];
            tensor<fp16, [1, 51, 128]> var_28595_cast_fp16 = add(x = q_head_357_cast_fp16, y = var_28594_to_fp16)[name = string("op_28595_cast_fp16")];
            tensor<int32, [1]> q_v_357_axes_1 = const()[name = string("q_v_357_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_357_cast_fp16 = expand_dims(axes = q_v_357_axes_1, x = var_28595_cast_fp16)[name = string("q_v_357_cast_fp16")];
            tensor<int32, [1]> k_head_715_axes_1 = const()[name = string("k_head_715_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_715_cast_fp16 = expand_dims(axes = k_head_715_axes_1, x = k_head_713_cast_fp16)[name = string("k_head_715_cast_fp16")];
            tensor<int32, [1]> v_head_715_axes_1 = const()[name = string("v_head_715_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_715_cast_fp16 = expand_dims(axes = v_head_715_axes_1, x = v_head_713_cast_fp16)[name = string("v_head_715_cast_fp16")];
            tensor<int32, [1]> p_head_715_axes_1 = const()[name = string("p_head_715_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_715_cast_fp16 = expand_dims(axes = p_head_715_axes_1, x = p_head_713_cast_fp16)[name = string("p_head_715_cast_fp16")];
            bool var_28601_transpose_x_3 = const()[name = string("op_28601_transpose_x_3"), val = bool(false)];
            bool var_28601_transpose_y_3 = const()[name = string("op_28601_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_28601_cast_fp16 = matmul(transpose_x = var_28601_transpose_x_3, transpose_y = var_28601_transpose_y_3, x = q_u_357_cast_fp16, y = k_head_715_cast_fp16)[name = string("op_28601_cast_fp16")];
            fp16 var_28602_to_fp16 = const()[name = string("op_28602_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_357_cast_fp16 = mul(x = var_28601_cast_fp16, y = var_28602_to_fp16)[name = string("scores_content_357_cast_fp16")];
            bool x_1885_transpose_x_3 = const()[name = string("x_1885_transpose_x_3"), val = bool(false)];
            bool x_1885_transpose_y_3 = const()[name = string("x_1885_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1885_cast_fp16 = matmul(transpose_x = x_1885_transpose_x_3, transpose_y = x_1885_transpose_y_3, x = q_v_357_cast_fp16, y = p_head_715_cast_fp16)[name = string("x_1885_cast_fp16")];
            tensor<int32, [8]> x_1887_pad_1 = const()[name = string("x_1887_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1887_mode_1 = const()[name = string("x_1887_mode_1"), val = string("constant")];
            fp16 const_2509_to_fp16 = const()[name = string("const_2509_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1887_cast_fp16 = pad(constant_val = const_2509_to_fp16, mode = x_1887_mode_1, pad = x_1887_pad_1, x = x_1885_cast_fp16)[name = string("x_1887_cast_fp16")];
            tensor<int32, [4]> var_28616 = const()[name = string("op_28616"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1889_cast_fp16 = reshape(shape = var_28616, x = x_1887_cast_fp16)[name = string("x_1889_cast_fp16")];
            tensor<int32, [4]> var_28620_begin_1 = const()[name = string("op_28620_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_28620_end_1 = const()[name = string("op_28620_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_28620_end_mask_1 = const()[name = string("op_28620_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_28620_cast_fp16 = slice_by_index(begin = var_28620_begin_1, end = var_28620_end_1, end_mask = var_28620_end_mask_1, x = x_1889_cast_fp16)[name = string("op_28620_cast_fp16")];
            tensor<int32, [4]> var_28622 = const()[name = string("op_28622"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_28623_cast_fp16 = reshape(shape = var_28622, x = var_28620_cast_fp16)[name = string("op_28623_cast_fp16")];
            tensor<int32, [4]> var_28628_begin_1 = const()[name = string("op_28628_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_28628_end_1 = const()[name = string("op_28628_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_28628_end_mask_1 = const()[name = string("op_28628_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_28628_cast_fp16 = slice_by_index(begin = var_28628_begin_1, end = var_28628_end_1, end_mask = var_28628_end_mask_1, x = var_28623_cast_fp16)[name = string("op_28628_cast_fp16")];
            fp16 var_28629_to_fp16 = const()[name = string("op_28629_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_357_cast_fp16 = mul(x = var_28628_cast_fp16, y = var_28629_to_fp16)[name = string("scores_pos_357_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_357_cast_fp16 = add(x = scores_content_357_cast_fp16, y = scores_pos_357_cast_fp16)[name = string("logits_357_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_28632_cast_fp16 = softmax(axis = var_28115, x = logits_357_cast_fp16)[name = string("op_28632_cast_fp16")];
            bool var_28634_transpose_x_1 = const()[name = string("op_28634_transpose_x_1"), val = bool(false)];
            bool var_28634_transpose_y_1 = const()[name = string("op_28634_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_28634_cast_fp16 = matmul(transpose_x = var_28634_transpose_x_1, transpose_y = var_28634_transpose_y_1, x = var_28632_cast_fp16, y = v_head_715_cast_fp16)[name = string("op_28634_cast_fp16")];
            tensor<int32, [1]> var_28635_axes_1 = const()[name = string("op_28635_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_28635_cast_fp16 = squeeze(axes = var_28635_axes_1, x = var_28634_cast_fp16)[name = string("op_28635_cast_fp16")];
            string dense_output_1801_pad_type_1 = const()[name = string("dense_output_1801_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1801_strides_1 = const()[name = string("dense_output_1801_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1801_pad_1 = const()[name = string("dense_output_1801_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1801_dilations_1 = const()[name = string("dense_output_1801_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1801_groups_1 = const()[name = string("dense_output_1801_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659386560))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659517696))))[name = string("layers_22_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1801_cast_fp16 = conv(dilations = dense_output_1801_dilations_1, groups = dense_output_1801_groups_1, pad = dense_output_1801_pad_1, pad_type = dense_output_1801_pad_type_1, strides = dense_output_1801_strides_1, weight = layers_22_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1801_cast_fp16")];
            string sparse_output_1801_pad_type_1 = const()[name = string("sparse_output_1801_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1801_strides_1 = const()[name = string("sparse_output_1801_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1801_pad_1 = const()[name = string("sparse_output_1801_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1801_dilations_1 = const()[name = string("sparse_output_1801_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1801_groups_1 = const()[name = string("sparse_output_1801_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659520960))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659518272))))[name = string("layers_22_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1801_cast_fp16 = conv(dilations = sparse_output_1801_dilations_1, groups = sparse_output_1801_groups_1, pad = sparse_output_1801_pad_1, pad_type = sparse_output_1801_pad_type_1, strides = sparse_output_1801_strides_1, weight = layers_22_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1801_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28650_cast_fp16 = add(x = dense_output_1801_cast_fp16, y = sparse_output_1801_cast_fp16)[name = string("op_28650_cast_fp16")];
            tensor<int32, [4]> var_28651 = const()[name = string("op_28651"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28653 = const()[name = string("op_28653"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28652_cast_fp16 = transpose(perm = var_28651, x = var_28650_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 51, 128]> q_head_359_cast_fp16 = reshape(shape = var_28653, x = var_28652_cast_fp16)[name = string("q_head_359_cast_fp16")];
            string dense_output_1803_pad_type_1 = const()[name = string("dense_output_1803_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1803_strides_1 = const()[name = string("dense_output_1803_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1803_pad_1 = const()[name = string("dense_output_1803_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1803_dilations_1 = const()[name = string("dense_output_1803_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1803_groups_1 = const()[name = string("dense_output_1803_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659537408))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659668544))))[name = string("layers_22_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1803_cast_fp16 = conv(dilations = dense_output_1803_dilations_1, groups = dense_output_1803_groups_1, pad = dense_output_1803_pad_1, pad_type = dense_output_1803_pad_type_1, strides = dense_output_1803_strides_1, weight = layers_22_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1803_cast_fp16")];
            string sparse_output_1803_pad_type_1 = const()[name = string("sparse_output_1803_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1803_strides_1 = const()[name = string("sparse_output_1803_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1803_pad_1 = const()[name = string("sparse_output_1803_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1803_dilations_1 = const()[name = string("sparse_output_1803_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1803_groups_1 = const()[name = string("sparse_output_1803_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659671808))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659669120))))[name = string("layers_22_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1803_cast_fp16 = conv(dilations = sparse_output_1803_dilations_1, groups = sparse_output_1803_groups_1, pad = sparse_output_1803_pad_1, pad_type = sparse_output_1803_pad_type_1, strides = sparse_output_1803_strides_1, weight = layers_22_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1803_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28669_cast_fp16 = add(x = dense_output_1803_cast_fp16, y = sparse_output_1803_cast_fp16)[name = string("op_28669_cast_fp16")];
            tensor<int32, [4]> var_28670 = const()[name = string("op_28670"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28672 = const()[name = string("op_28672"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28671_cast_fp16 = transpose(perm = var_28670, x = var_28669_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 51, 128]> k_head_717_cast_fp16 = reshape(shape = var_28672, x = var_28671_cast_fp16)[name = string("k_head_717_cast_fp16")];
            string dense_output_1805_pad_type_1 = const()[name = string("dense_output_1805_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1805_strides_1 = const()[name = string("dense_output_1805_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1805_pad_1 = const()[name = string("dense_output_1805_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1805_dilations_1 = const()[name = string("dense_output_1805_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1805_groups_1 = const()[name = string("dense_output_1805_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659688256))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659819392))))[name = string("layers_22_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1805_cast_fp16 = conv(dilations = dense_output_1805_dilations_1, groups = dense_output_1805_groups_1, pad = dense_output_1805_pad_1, pad_type = dense_output_1805_pad_type_1, strides = dense_output_1805_strides_1, weight = layers_22_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1805_cast_fp16")];
            string sparse_output_1805_pad_type_1 = const()[name = string("sparse_output_1805_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1805_strides_1 = const()[name = string("sparse_output_1805_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1805_pad_1 = const()[name = string("sparse_output_1805_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1805_dilations_1 = const()[name = string("sparse_output_1805_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1805_groups_1 = const()[name = string("sparse_output_1805_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659822656))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659819968))))[name = string("layers_22_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1805_cast_fp16 = conv(dilations = sparse_output_1805_dilations_1, groups = sparse_output_1805_groups_1, pad = sparse_output_1805_pad_1, pad_type = sparse_output_1805_pad_type_1, strides = sparse_output_1805_strides_1, weight = layers_22_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1805_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28688_cast_fp16 = add(x = dense_output_1805_cast_fp16, y = sparse_output_1805_cast_fp16)[name = string("op_28688_cast_fp16")];
            tensor<int32, [4]> var_28689 = const()[name = string("op_28689"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28691 = const()[name = string("op_28691"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28690_cast_fp16 = transpose(perm = var_28689, x = var_28688_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 51, 128]> v_head_717_cast_fp16 = reshape(shape = var_28691, x = var_28690_cast_fp16)[name = string("v_head_717_cast_fp16")];
            string dense_output_1807_pad_type_1 = const()[name = string("dense_output_1807_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1807_strides_1 = const()[name = string("dense_output_1807_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1807_pad_1 = const()[name = string("dense_output_1807_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1807_dilations_1 = const()[name = string("dense_output_1807_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1807_groups_1 = const()[name = string("dense_output_1807_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659839104))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659970240))))[name = string("layers_22_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1807_cast_fp16 = conv(dilations = dense_output_1807_dilations_1, groups = dense_output_1807_groups_1, pad = dense_output_1807_pad_1, pad_type = dense_output_1807_pad_type_1, strides = dense_output_1807_strides_1, weight = layers_22_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1807_cast_fp16")];
            string sparse_output_1807_pad_type_1 = const()[name = string("sparse_output_1807_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1807_strides_1 = const()[name = string("sparse_output_1807_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1807_pad_1 = const()[name = string("sparse_output_1807_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1807_dilations_1 = const()[name = string("sparse_output_1807_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1807_groups_1 = const()[name = string("sparse_output_1807_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659973504))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659970816))))[name = string("layers_22_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1807_cast_fp16 = conv(dilations = sparse_output_1807_dilations_1, groups = sparse_output_1807_groups_1, pad = sparse_output_1807_pad_1, pad_type = sparse_output_1807_pad_type_1, strides = sparse_output_1807_strides_1, weight = layers_22_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1807_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_28707_cast_fp16 = add(x = dense_output_1807_cast_fp16, y = sparse_output_1807_cast_fp16)[name = string("op_28707_cast_fp16")];
            tensor<int32, [4]> var_28708 = const()[name = string("op_28708"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28710 = const()[name = string("op_28710"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_28709_cast_fp16 = transpose(perm = var_28708, x = var_28707_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 101, 128]> p_head_717_cast_fp16 = reshape(shape = var_28710, x = var_28709_cast_fp16)[name = string("p_head_717_cast_fp16")];
            tensor<fp16, [128]> var_28712_to_fp16 = const()[name = string("op_28712_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659989952)))];
            tensor<fp16, [1, 51, 128]> var_28713_cast_fp16 = add(x = q_head_359_cast_fp16, y = var_28712_to_fp16)[name = string("op_28713_cast_fp16")];
            tensor<int32, [1]> q_u_359_axes_1 = const()[name = string("q_u_359_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_359_cast_fp16 = expand_dims(axes = q_u_359_axes_1, x = var_28713_cast_fp16)[name = string("q_u_359_cast_fp16")];
            tensor<fp16, [128]> var_28715_to_fp16 = const()[name = string("op_28715_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659990272)))];
            tensor<fp16, [1, 51, 128]> var_28716_cast_fp16 = add(x = q_head_359_cast_fp16, y = var_28715_to_fp16)[name = string("op_28716_cast_fp16")];
            tensor<int32, [1]> q_v_359_axes_1 = const()[name = string("q_v_359_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_359_cast_fp16 = expand_dims(axes = q_v_359_axes_1, x = var_28716_cast_fp16)[name = string("q_v_359_cast_fp16")];
            tensor<int32, [1]> k_head_719_axes_1 = const()[name = string("k_head_719_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_719_cast_fp16 = expand_dims(axes = k_head_719_axes_1, x = k_head_717_cast_fp16)[name = string("k_head_719_cast_fp16")];
            tensor<int32, [1]> v_head_719_axes_1 = const()[name = string("v_head_719_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_719_cast_fp16 = expand_dims(axes = v_head_719_axes_1, x = v_head_717_cast_fp16)[name = string("v_head_719_cast_fp16")];
            tensor<int32, [1]> p_head_719_axes_1 = const()[name = string("p_head_719_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_719_cast_fp16 = expand_dims(axes = p_head_719_axes_1, x = p_head_717_cast_fp16)[name = string("p_head_719_cast_fp16")];
            bool var_28722_transpose_x_3 = const()[name = string("op_28722_transpose_x_3"), val = bool(false)];
            bool var_28722_transpose_y_3 = const()[name = string("op_28722_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_28722_cast_fp16 = matmul(transpose_x = var_28722_transpose_x_3, transpose_y = var_28722_transpose_y_3, x = q_u_359_cast_fp16, y = k_head_719_cast_fp16)[name = string("op_28722_cast_fp16")];
            fp16 var_28723_to_fp16 = const()[name = string("op_28723_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_359_cast_fp16 = mul(x = var_28722_cast_fp16, y = var_28723_to_fp16)[name = string("scores_content_359_cast_fp16")];
            bool x_1893_transpose_x_3 = const()[name = string("x_1893_transpose_x_3"), val = bool(false)];
            bool x_1893_transpose_y_3 = const()[name = string("x_1893_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1893_cast_fp16 = matmul(transpose_x = x_1893_transpose_x_3, transpose_y = x_1893_transpose_y_3, x = q_v_359_cast_fp16, y = p_head_719_cast_fp16)[name = string("x_1893_cast_fp16")];
            tensor<int32, [8]> x_1895_pad_1 = const()[name = string("x_1895_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1895_mode_1 = const()[name = string("x_1895_mode_1"), val = string("constant")];
            fp16 const_2515_to_fp16 = const()[name = string("const_2515_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1895_cast_fp16 = pad(constant_val = const_2515_to_fp16, mode = x_1895_mode_1, pad = x_1895_pad_1, x = x_1893_cast_fp16)[name = string("x_1895_cast_fp16")];
            tensor<int32, [4]> var_28737 = const()[name = string("op_28737"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1897_cast_fp16 = reshape(shape = var_28737, x = x_1895_cast_fp16)[name = string("x_1897_cast_fp16")];
            tensor<int32, [4]> var_28741_begin_1 = const()[name = string("op_28741_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_28741_end_1 = const()[name = string("op_28741_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_28741_end_mask_1 = const()[name = string("op_28741_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_28741_cast_fp16 = slice_by_index(begin = var_28741_begin_1, end = var_28741_end_1, end_mask = var_28741_end_mask_1, x = x_1897_cast_fp16)[name = string("op_28741_cast_fp16")];
            tensor<int32, [4]> var_28743 = const()[name = string("op_28743"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_28744_cast_fp16 = reshape(shape = var_28743, x = var_28741_cast_fp16)[name = string("op_28744_cast_fp16")];
            tensor<int32, [4]> var_28749_begin_1 = const()[name = string("op_28749_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_28749_end_1 = const()[name = string("op_28749_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_28749_end_mask_1 = const()[name = string("op_28749_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_28749_cast_fp16 = slice_by_index(begin = var_28749_begin_1, end = var_28749_end_1, end_mask = var_28749_end_mask_1, x = var_28744_cast_fp16)[name = string("op_28749_cast_fp16")];
            fp16 var_28750_to_fp16 = const()[name = string("op_28750_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_359_cast_fp16 = mul(x = var_28749_cast_fp16, y = var_28750_to_fp16)[name = string("scores_pos_359_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_359_cast_fp16 = add(x = scores_content_359_cast_fp16, y = scores_pos_359_cast_fp16)[name = string("logits_359_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_28753_cast_fp16 = softmax(axis = var_28115, x = logits_359_cast_fp16)[name = string("op_28753_cast_fp16")];
            bool var_28755_transpose_x_1 = const()[name = string("op_28755_transpose_x_1"), val = bool(false)];
            bool var_28755_transpose_y_1 = const()[name = string("op_28755_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_28755_cast_fp16 = matmul(transpose_x = var_28755_transpose_x_1, transpose_y = var_28755_transpose_y_1, x = var_28753_cast_fp16, y = v_head_719_cast_fp16)[name = string("op_28755_cast_fp16")];
            tensor<int32, [1]> var_28756_axes_1 = const()[name = string("op_28756_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_28756_cast_fp16 = squeeze(axes = var_28756_axes_1, x = var_28755_cast_fp16)[name = string("op_28756_cast_fp16")];
            string dense_output_1809_pad_type_1 = const()[name = string("dense_output_1809_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1809_strides_1 = const()[name = string("dense_output_1809_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1809_pad_1 = const()[name = string("dense_output_1809_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1809_dilations_1 = const()[name = string("dense_output_1809_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1809_groups_1 = const()[name = string("dense_output_1809_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(659990592))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660121728))))[name = string("layers_22_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1809_cast_fp16 = conv(dilations = dense_output_1809_dilations_1, groups = dense_output_1809_groups_1, pad = dense_output_1809_pad_1, pad_type = dense_output_1809_pad_type_1, strides = dense_output_1809_strides_1, weight = layers_22_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1809_cast_fp16")];
            string sparse_output_1809_pad_type_1 = const()[name = string("sparse_output_1809_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1809_strides_1 = const()[name = string("sparse_output_1809_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1809_pad_1 = const()[name = string("sparse_output_1809_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1809_dilations_1 = const()[name = string("sparse_output_1809_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1809_groups_1 = const()[name = string("sparse_output_1809_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660124992))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660122304))))[name = string("layers_22_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1809_cast_fp16 = conv(dilations = sparse_output_1809_dilations_1, groups = sparse_output_1809_groups_1, pad = sparse_output_1809_pad_1, pad_type = sparse_output_1809_pad_type_1, strides = sparse_output_1809_strides_1, weight = layers_22_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1809_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28771_cast_fp16 = add(x = dense_output_1809_cast_fp16, y = sparse_output_1809_cast_fp16)[name = string("op_28771_cast_fp16")];
            tensor<int32, [4]> var_28772 = const()[name = string("op_28772"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28774 = const()[name = string("op_28774"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28773_cast_fp16 = transpose(perm = var_28772, x = var_28771_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 51, 128]> q_head_361_cast_fp16 = reshape(shape = var_28774, x = var_28773_cast_fp16)[name = string("q_head_361_cast_fp16")];
            string dense_output_1811_pad_type_1 = const()[name = string("dense_output_1811_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1811_strides_1 = const()[name = string("dense_output_1811_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1811_pad_1 = const()[name = string("dense_output_1811_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1811_dilations_1 = const()[name = string("dense_output_1811_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1811_groups_1 = const()[name = string("dense_output_1811_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660141440))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660272576))))[name = string("layers_22_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1811_cast_fp16 = conv(dilations = dense_output_1811_dilations_1, groups = dense_output_1811_groups_1, pad = dense_output_1811_pad_1, pad_type = dense_output_1811_pad_type_1, strides = dense_output_1811_strides_1, weight = layers_22_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1811_cast_fp16")];
            string sparse_output_1811_pad_type_1 = const()[name = string("sparse_output_1811_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1811_strides_1 = const()[name = string("sparse_output_1811_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1811_pad_1 = const()[name = string("sparse_output_1811_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1811_dilations_1 = const()[name = string("sparse_output_1811_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1811_groups_1 = const()[name = string("sparse_output_1811_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660275840))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660273152))))[name = string("layers_22_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1811_cast_fp16 = conv(dilations = sparse_output_1811_dilations_1, groups = sparse_output_1811_groups_1, pad = sparse_output_1811_pad_1, pad_type = sparse_output_1811_pad_type_1, strides = sparse_output_1811_strides_1, weight = layers_22_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1811_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28790_cast_fp16 = add(x = dense_output_1811_cast_fp16, y = sparse_output_1811_cast_fp16)[name = string("op_28790_cast_fp16")];
            tensor<int32, [4]> var_28791 = const()[name = string("op_28791"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28793 = const()[name = string("op_28793"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28792_cast_fp16 = transpose(perm = var_28791, x = var_28790_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 51, 128]> k_head_721_cast_fp16 = reshape(shape = var_28793, x = var_28792_cast_fp16)[name = string("k_head_721_cast_fp16")];
            string dense_output_1813_pad_type_1 = const()[name = string("dense_output_1813_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1813_strides_1 = const()[name = string("dense_output_1813_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1813_pad_1 = const()[name = string("dense_output_1813_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1813_dilations_1 = const()[name = string("dense_output_1813_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1813_groups_1 = const()[name = string("dense_output_1813_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660292288))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660423424))))[name = string("layers_22_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1813_cast_fp16 = conv(dilations = dense_output_1813_dilations_1, groups = dense_output_1813_groups_1, pad = dense_output_1813_pad_1, pad_type = dense_output_1813_pad_type_1, strides = dense_output_1813_strides_1, weight = layers_22_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1813_cast_fp16")];
            string sparse_output_1813_pad_type_1 = const()[name = string("sparse_output_1813_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1813_strides_1 = const()[name = string("sparse_output_1813_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1813_pad_1 = const()[name = string("sparse_output_1813_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1813_dilations_1 = const()[name = string("sparse_output_1813_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1813_groups_1 = const()[name = string("sparse_output_1813_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660426688))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660424000))))[name = string("layers_22_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1813_cast_fp16 = conv(dilations = sparse_output_1813_dilations_1, groups = sparse_output_1813_groups_1, pad = sparse_output_1813_pad_1, pad_type = sparse_output_1813_pad_type_1, strides = sparse_output_1813_strides_1, weight = layers_22_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1813_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28809_cast_fp16 = add(x = dense_output_1813_cast_fp16, y = sparse_output_1813_cast_fp16)[name = string("op_28809_cast_fp16")];
            tensor<int32, [4]> var_28810 = const()[name = string("op_28810"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28812 = const()[name = string("op_28812"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28811_cast_fp16 = transpose(perm = var_28810, x = var_28809_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 51, 128]> v_head_721_cast_fp16 = reshape(shape = var_28812, x = var_28811_cast_fp16)[name = string("v_head_721_cast_fp16")];
            string dense_output_1815_pad_type_1 = const()[name = string("dense_output_1815_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1815_strides_1 = const()[name = string("dense_output_1815_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1815_pad_1 = const()[name = string("dense_output_1815_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1815_dilations_1 = const()[name = string("dense_output_1815_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1815_groups_1 = const()[name = string("dense_output_1815_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660443136))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660574272))))[name = string("layers_22_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1815_cast_fp16 = conv(dilations = dense_output_1815_dilations_1, groups = dense_output_1815_groups_1, pad = dense_output_1815_pad_1, pad_type = dense_output_1815_pad_type_1, strides = dense_output_1815_strides_1, weight = layers_22_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1815_cast_fp16")];
            string sparse_output_1815_pad_type_1 = const()[name = string("sparse_output_1815_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1815_strides_1 = const()[name = string("sparse_output_1815_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1815_pad_1 = const()[name = string("sparse_output_1815_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1815_dilations_1 = const()[name = string("sparse_output_1815_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1815_groups_1 = const()[name = string("sparse_output_1815_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660577536))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660574848))))[name = string("layers_22_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1815_cast_fp16 = conv(dilations = sparse_output_1815_dilations_1, groups = sparse_output_1815_groups_1, pad = sparse_output_1815_pad_1, pad_type = sparse_output_1815_pad_type_1, strides = sparse_output_1815_strides_1, weight = layers_22_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1815_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_28828_cast_fp16 = add(x = dense_output_1815_cast_fp16, y = sparse_output_1815_cast_fp16)[name = string("op_28828_cast_fp16")];
            tensor<int32, [4]> var_28829 = const()[name = string("op_28829"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28831 = const()[name = string("op_28831"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_28830_cast_fp16 = transpose(perm = var_28829, x = var_28828_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 101, 128]> p_head_721_cast_fp16 = reshape(shape = var_28831, x = var_28830_cast_fp16)[name = string("p_head_721_cast_fp16")];
            tensor<fp16, [128]> var_28833_to_fp16 = const()[name = string("op_28833_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660593984)))];
            tensor<fp16, [1, 51, 128]> var_28834_cast_fp16 = add(x = q_head_361_cast_fp16, y = var_28833_to_fp16)[name = string("op_28834_cast_fp16")];
            tensor<int32, [1]> q_u_361_axes_1 = const()[name = string("q_u_361_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_361_cast_fp16 = expand_dims(axes = q_u_361_axes_1, x = var_28834_cast_fp16)[name = string("q_u_361_cast_fp16")];
            tensor<fp16, [128]> var_28836_to_fp16 = const()[name = string("op_28836_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660594304)))];
            tensor<fp16, [1, 51, 128]> var_28837_cast_fp16 = add(x = q_head_361_cast_fp16, y = var_28836_to_fp16)[name = string("op_28837_cast_fp16")];
            tensor<int32, [1]> q_v_361_axes_1 = const()[name = string("q_v_361_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_361_cast_fp16 = expand_dims(axes = q_v_361_axes_1, x = var_28837_cast_fp16)[name = string("q_v_361_cast_fp16")];
            tensor<int32, [1]> k_head_723_axes_1 = const()[name = string("k_head_723_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_723_cast_fp16 = expand_dims(axes = k_head_723_axes_1, x = k_head_721_cast_fp16)[name = string("k_head_723_cast_fp16")];
            tensor<int32, [1]> v_head_723_axes_1 = const()[name = string("v_head_723_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_723_cast_fp16 = expand_dims(axes = v_head_723_axes_1, x = v_head_721_cast_fp16)[name = string("v_head_723_cast_fp16")];
            tensor<int32, [1]> p_head_723_axes_1 = const()[name = string("p_head_723_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_723_cast_fp16 = expand_dims(axes = p_head_723_axes_1, x = p_head_721_cast_fp16)[name = string("p_head_723_cast_fp16")];
            bool var_28843_transpose_x_3 = const()[name = string("op_28843_transpose_x_3"), val = bool(false)];
            bool var_28843_transpose_y_3 = const()[name = string("op_28843_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_28843_cast_fp16 = matmul(transpose_x = var_28843_transpose_x_3, transpose_y = var_28843_transpose_y_3, x = q_u_361_cast_fp16, y = k_head_723_cast_fp16)[name = string("op_28843_cast_fp16")];
            fp16 var_28844_to_fp16 = const()[name = string("op_28844_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_361_cast_fp16 = mul(x = var_28843_cast_fp16, y = var_28844_to_fp16)[name = string("scores_content_361_cast_fp16")];
            bool x_1901_transpose_x_3 = const()[name = string("x_1901_transpose_x_3"), val = bool(false)];
            bool x_1901_transpose_y_3 = const()[name = string("x_1901_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1901_cast_fp16 = matmul(transpose_x = x_1901_transpose_x_3, transpose_y = x_1901_transpose_y_3, x = q_v_361_cast_fp16, y = p_head_723_cast_fp16)[name = string("x_1901_cast_fp16")];
            tensor<int32, [8]> x_1903_pad_1 = const()[name = string("x_1903_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1903_mode_1 = const()[name = string("x_1903_mode_1"), val = string("constant")];
            fp16 const_2521_to_fp16 = const()[name = string("const_2521_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1903_cast_fp16 = pad(constant_val = const_2521_to_fp16, mode = x_1903_mode_1, pad = x_1903_pad_1, x = x_1901_cast_fp16)[name = string("x_1903_cast_fp16")];
            tensor<int32, [4]> var_28858 = const()[name = string("op_28858"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1905_cast_fp16 = reshape(shape = var_28858, x = x_1903_cast_fp16)[name = string("x_1905_cast_fp16")];
            tensor<int32, [4]> var_28862_begin_1 = const()[name = string("op_28862_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_28862_end_1 = const()[name = string("op_28862_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_28862_end_mask_1 = const()[name = string("op_28862_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_28862_cast_fp16 = slice_by_index(begin = var_28862_begin_1, end = var_28862_end_1, end_mask = var_28862_end_mask_1, x = x_1905_cast_fp16)[name = string("op_28862_cast_fp16")];
            tensor<int32, [4]> var_28864 = const()[name = string("op_28864"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_28865_cast_fp16 = reshape(shape = var_28864, x = var_28862_cast_fp16)[name = string("op_28865_cast_fp16")];
            tensor<int32, [4]> var_28870_begin_1 = const()[name = string("op_28870_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_28870_end_1 = const()[name = string("op_28870_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_28870_end_mask_1 = const()[name = string("op_28870_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_28870_cast_fp16 = slice_by_index(begin = var_28870_begin_1, end = var_28870_end_1, end_mask = var_28870_end_mask_1, x = var_28865_cast_fp16)[name = string("op_28870_cast_fp16")];
            fp16 var_28871_to_fp16 = const()[name = string("op_28871_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_361_cast_fp16 = mul(x = var_28870_cast_fp16, y = var_28871_to_fp16)[name = string("scores_pos_361_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_361_cast_fp16 = add(x = scores_content_361_cast_fp16, y = scores_pos_361_cast_fp16)[name = string("logits_361_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_28874_cast_fp16 = softmax(axis = var_28115, x = logits_361_cast_fp16)[name = string("op_28874_cast_fp16")];
            bool var_28876_transpose_x_1 = const()[name = string("op_28876_transpose_x_1"), val = bool(false)];
            bool var_28876_transpose_y_1 = const()[name = string("op_28876_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_28876_cast_fp16 = matmul(transpose_x = var_28876_transpose_x_1, transpose_y = var_28876_transpose_y_1, x = var_28874_cast_fp16, y = v_head_723_cast_fp16)[name = string("op_28876_cast_fp16")];
            tensor<int32, [1]> var_28877_axes_1 = const()[name = string("op_28877_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_28877_cast_fp16 = squeeze(axes = var_28877_axes_1, x = var_28876_cast_fp16)[name = string("op_28877_cast_fp16")];
            string dense_output_1817_pad_type_1 = const()[name = string("dense_output_1817_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1817_strides_1 = const()[name = string("dense_output_1817_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1817_pad_1 = const()[name = string("dense_output_1817_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1817_dilations_1 = const()[name = string("dense_output_1817_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1817_groups_1 = const()[name = string("dense_output_1817_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660594624))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660725760))))[name = string("layers_22_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1817_cast_fp16 = conv(dilations = dense_output_1817_dilations_1, groups = dense_output_1817_groups_1, pad = dense_output_1817_pad_1, pad_type = dense_output_1817_pad_type_1, strides = dense_output_1817_strides_1, weight = layers_22_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1817_cast_fp16")];
            string sparse_output_1817_pad_type_1 = const()[name = string("sparse_output_1817_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1817_strides_1 = const()[name = string("sparse_output_1817_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1817_pad_1 = const()[name = string("sparse_output_1817_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1817_dilations_1 = const()[name = string("sparse_output_1817_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1817_groups_1 = const()[name = string("sparse_output_1817_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660729024))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660726336))))[name = string("layers_22_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1817_cast_fp16 = conv(dilations = sparse_output_1817_dilations_1, groups = sparse_output_1817_groups_1, pad = sparse_output_1817_pad_1, pad_type = sparse_output_1817_pad_type_1, strides = sparse_output_1817_strides_1, weight = layers_22_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1817_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28892_cast_fp16 = add(x = dense_output_1817_cast_fp16, y = sparse_output_1817_cast_fp16)[name = string("op_28892_cast_fp16")];
            tensor<int32, [4]> var_28893 = const()[name = string("op_28893"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28895 = const()[name = string("op_28895"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28894_cast_fp16 = transpose(perm = var_28893, x = var_28892_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 51, 128]> q_head_363_cast_fp16 = reshape(shape = var_28895, x = var_28894_cast_fp16)[name = string("q_head_363_cast_fp16")];
            string dense_output_1819_pad_type_1 = const()[name = string("dense_output_1819_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1819_strides_1 = const()[name = string("dense_output_1819_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1819_pad_1 = const()[name = string("dense_output_1819_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1819_dilations_1 = const()[name = string("dense_output_1819_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1819_groups_1 = const()[name = string("dense_output_1819_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660745472))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660876608))))[name = string("layers_22_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1819_cast_fp16 = conv(dilations = dense_output_1819_dilations_1, groups = dense_output_1819_groups_1, pad = dense_output_1819_pad_1, pad_type = dense_output_1819_pad_type_1, strides = dense_output_1819_strides_1, weight = layers_22_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1819_cast_fp16")];
            string sparse_output_1819_pad_type_1 = const()[name = string("sparse_output_1819_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1819_strides_1 = const()[name = string("sparse_output_1819_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1819_pad_1 = const()[name = string("sparse_output_1819_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1819_dilations_1 = const()[name = string("sparse_output_1819_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1819_groups_1 = const()[name = string("sparse_output_1819_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660879872))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660877184))))[name = string("layers_22_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1819_cast_fp16 = conv(dilations = sparse_output_1819_dilations_1, groups = sparse_output_1819_groups_1, pad = sparse_output_1819_pad_1, pad_type = sparse_output_1819_pad_type_1, strides = sparse_output_1819_strides_1, weight = layers_22_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1819_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28911_cast_fp16 = add(x = dense_output_1819_cast_fp16, y = sparse_output_1819_cast_fp16)[name = string("op_28911_cast_fp16")];
            tensor<int32, [4]> var_28912 = const()[name = string("op_28912"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28914 = const()[name = string("op_28914"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28913_cast_fp16 = transpose(perm = var_28912, x = var_28911_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 51, 128]> k_head_725_cast_fp16 = reshape(shape = var_28914, x = var_28913_cast_fp16)[name = string("k_head_725_cast_fp16")];
            string dense_output_1821_pad_type_1 = const()[name = string("dense_output_1821_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1821_strides_1 = const()[name = string("dense_output_1821_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1821_pad_1 = const()[name = string("dense_output_1821_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1821_dilations_1 = const()[name = string("dense_output_1821_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1821_groups_1 = const()[name = string("dense_output_1821_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660896320))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661027456))))[name = string("layers_22_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1821_cast_fp16 = conv(dilations = dense_output_1821_dilations_1, groups = dense_output_1821_groups_1, pad = dense_output_1821_pad_1, pad_type = dense_output_1821_pad_type_1, strides = dense_output_1821_strides_1, weight = layers_22_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1821_cast_fp16")];
            string sparse_output_1821_pad_type_1 = const()[name = string("sparse_output_1821_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1821_strides_1 = const()[name = string("sparse_output_1821_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1821_pad_1 = const()[name = string("sparse_output_1821_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1821_dilations_1 = const()[name = string("sparse_output_1821_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1821_groups_1 = const()[name = string("sparse_output_1821_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661030720))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661028032))))[name = string("layers_22_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1821_cast_fp16 = conv(dilations = sparse_output_1821_dilations_1, groups = sparse_output_1821_groups_1, pad = sparse_output_1821_pad_1, pad_type = sparse_output_1821_pad_type_1, strides = sparse_output_1821_strides_1, weight = layers_22_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1821_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_28930_cast_fp16 = add(x = dense_output_1821_cast_fp16, y = sparse_output_1821_cast_fp16)[name = string("op_28930_cast_fp16")];
            tensor<int32, [4]> var_28931 = const()[name = string("op_28931"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28933 = const()[name = string("op_28933"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_28932_cast_fp16 = transpose(perm = var_28931, x = var_28930_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 51, 128]> v_head_725_cast_fp16 = reshape(shape = var_28933, x = var_28932_cast_fp16)[name = string("v_head_725_cast_fp16")];
            string dense_output_1823_pad_type_1 = const()[name = string("dense_output_1823_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1823_strides_1 = const()[name = string("dense_output_1823_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1823_pad_1 = const()[name = string("dense_output_1823_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1823_dilations_1 = const()[name = string("dense_output_1823_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1823_groups_1 = const()[name = string("dense_output_1823_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661047168))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661178304))))[name = string("layers_22_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1823_cast_fp16 = conv(dilations = dense_output_1823_dilations_1, groups = dense_output_1823_groups_1, pad = dense_output_1823_pad_1, pad_type = dense_output_1823_pad_type_1, strides = dense_output_1823_strides_1, weight = layers_22_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1823_cast_fp16")];
            string sparse_output_1823_pad_type_1 = const()[name = string("sparse_output_1823_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1823_strides_1 = const()[name = string("sparse_output_1823_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1823_pad_1 = const()[name = string("sparse_output_1823_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1823_dilations_1 = const()[name = string("sparse_output_1823_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1823_groups_1 = const()[name = string("sparse_output_1823_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661181568))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661178880))))[name = string("layers_22_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1823_cast_fp16 = conv(dilations = sparse_output_1823_dilations_1, groups = sparse_output_1823_groups_1, pad = sparse_output_1823_pad_1, pad_type = sparse_output_1823_pad_type_1, strides = sparse_output_1823_strides_1, weight = layers_22_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1823_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_28949_cast_fp16 = add(x = dense_output_1823_cast_fp16, y = sparse_output_1823_cast_fp16)[name = string("op_28949_cast_fp16")];
            tensor<int32, [4]> var_28950 = const()[name = string("op_28950"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_28952 = const()[name = string("op_28952"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_28951_cast_fp16 = transpose(perm = var_28950, x = var_28949_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 101, 128]> p_head_725_cast_fp16 = reshape(shape = var_28952, x = var_28951_cast_fp16)[name = string("p_head_725_cast_fp16")];
            tensor<fp16, [128]> var_28954_to_fp16 = const()[name = string("op_28954_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661198016)))];
            tensor<fp16, [1, 51, 128]> var_28955_cast_fp16 = add(x = q_head_363_cast_fp16, y = var_28954_to_fp16)[name = string("op_28955_cast_fp16")];
            tensor<int32, [1]> q_u_363_axes_1 = const()[name = string("q_u_363_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_363_cast_fp16 = expand_dims(axes = q_u_363_axes_1, x = var_28955_cast_fp16)[name = string("q_u_363_cast_fp16")];
            tensor<fp16, [128]> var_28957_to_fp16 = const()[name = string("op_28957_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661198336)))];
            tensor<fp16, [1, 51, 128]> var_28958_cast_fp16 = add(x = q_head_363_cast_fp16, y = var_28957_to_fp16)[name = string("op_28958_cast_fp16")];
            tensor<int32, [1]> q_v_363_axes_1 = const()[name = string("q_v_363_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_363_cast_fp16 = expand_dims(axes = q_v_363_axes_1, x = var_28958_cast_fp16)[name = string("q_v_363_cast_fp16")];
            tensor<int32, [1]> k_head_727_axes_1 = const()[name = string("k_head_727_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_727_cast_fp16 = expand_dims(axes = k_head_727_axes_1, x = k_head_725_cast_fp16)[name = string("k_head_727_cast_fp16")];
            tensor<int32, [1]> v_head_727_axes_1 = const()[name = string("v_head_727_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_727_cast_fp16 = expand_dims(axes = v_head_727_axes_1, x = v_head_725_cast_fp16)[name = string("v_head_727_cast_fp16")];
            tensor<int32, [1]> p_head_727_axes_1 = const()[name = string("p_head_727_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_727_cast_fp16 = expand_dims(axes = p_head_727_axes_1, x = p_head_725_cast_fp16)[name = string("p_head_727_cast_fp16")];
            bool var_28964_transpose_x_3 = const()[name = string("op_28964_transpose_x_3"), val = bool(false)];
            bool var_28964_transpose_y_3 = const()[name = string("op_28964_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_28964_cast_fp16 = matmul(transpose_x = var_28964_transpose_x_3, transpose_y = var_28964_transpose_y_3, x = q_u_363_cast_fp16, y = k_head_727_cast_fp16)[name = string("op_28964_cast_fp16")];
            fp16 var_28965_to_fp16 = const()[name = string("op_28965_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_363_cast_fp16 = mul(x = var_28964_cast_fp16, y = var_28965_to_fp16)[name = string("scores_content_363_cast_fp16")];
            bool x_1909_transpose_x_3 = const()[name = string("x_1909_transpose_x_3"), val = bool(false)];
            bool x_1909_transpose_y_3 = const()[name = string("x_1909_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1909_cast_fp16 = matmul(transpose_x = x_1909_transpose_x_3, transpose_y = x_1909_transpose_y_3, x = q_v_363_cast_fp16, y = p_head_727_cast_fp16)[name = string("x_1909_cast_fp16")];
            tensor<int32, [8]> x_1911_pad_1 = const()[name = string("x_1911_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1911_mode_1 = const()[name = string("x_1911_mode_1"), val = string("constant")];
            fp16 const_2527_to_fp16 = const()[name = string("const_2527_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1911_cast_fp16 = pad(constant_val = const_2527_to_fp16, mode = x_1911_mode_1, pad = x_1911_pad_1, x = x_1909_cast_fp16)[name = string("x_1911_cast_fp16")];
            tensor<int32, [4]> var_28979 = const()[name = string("op_28979"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1913_cast_fp16 = reshape(shape = var_28979, x = x_1911_cast_fp16)[name = string("x_1913_cast_fp16")];
            tensor<int32, [4]> var_28983_begin_1 = const()[name = string("op_28983_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_28983_end_1 = const()[name = string("op_28983_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_28983_end_mask_1 = const()[name = string("op_28983_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_28983_cast_fp16 = slice_by_index(begin = var_28983_begin_1, end = var_28983_end_1, end_mask = var_28983_end_mask_1, x = x_1913_cast_fp16)[name = string("op_28983_cast_fp16")];
            tensor<int32, [4]> var_28985 = const()[name = string("op_28985"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_28986_cast_fp16 = reshape(shape = var_28985, x = var_28983_cast_fp16)[name = string("op_28986_cast_fp16")];
            tensor<int32, [4]> var_28991_begin_1 = const()[name = string("op_28991_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_28991_end_1 = const()[name = string("op_28991_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_28991_end_mask_1 = const()[name = string("op_28991_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_28991_cast_fp16 = slice_by_index(begin = var_28991_begin_1, end = var_28991_end_1, end_mask = var_28991_end_mask_1, x = var_28986_cast_fp16)[name = string("op_28991_cast_fp16")];
            fp16 var_28992_to_fp16 = const()[name = string("op_28992_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_363_cast_fp16 = mul(x = var_28991_cast_fp16, y = var_28992_to_fp16)[name = string("scores_pos_363_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_363_cast_fp16 = add(x = scores_content_363_cast_fp16, y = scores_pos_363_cast_fp16)[name = string("logits_363_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_28995_cast_fp16 = softmax(axis = var_28115, x = logits_363_cast_fp16)[name = string("op_28995_cast_fp16")];
            bool var_28997_transpose_x_1 = const()[name = string("op_28997_transpose_x_1"), val = bool(false)];
            bool var_28997_transpose_y_1 = const()[name = string("op_28997_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_28997_cast_fp16 = matmul(transpose_x = var_28997_transpose_x_1, transpose_y = var_28997_transpose_y_1, x = var_28995_cast_fp16, y = v_head_727_cast_fp16)[name = string("op_28997_cast_fp16")];
            tensor<int32, [1]> var_28998_axes_1 = const()[name = string("op_28998_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_28998_cast_fp16 = squeeze(axes = var_28998_axes_1, x = var_28997_cast_fp16)[name = string("op_28998_cast_fp16")];
            string dense_output_1825_pad_type_1 = const()[name = string("dense_output_1825_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1825_strides_1 = const()[name = string("dense_output_1825_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1825_pad_1 = const()[name = string("dense_output_1825_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1825_dilations_1 = const()[name = string("dense_output_1825_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1825_groups_1 = const()[name = string("dense_output_1825_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661198656))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661329792))))[name = string("layers_22_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1825_cast_fp16 = conv(dilations = dense_output_1825_dilations_1, groups = dense_output_1825_groups_1, pad = dense_output_1825_pad_1, pad_type = dense_output_1825_pad_type_1, strides = dense_output_1825_strides_1, weight = layers_22_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1825_cast_fp16")];
            string sparse_output_1825_pad_type_1 = const()[name = string("sparse_output_1825_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1825_strides_1 = const()[name = string("sparse_output_1825_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1825_pad_1 = const()[name = string("sparse_output_1825_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1825_dilations_1 = const()[name = string("sparse_output_1825_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1825_groups_1 = const()[name = string("sparse_output_1825_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661333056))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661330368))))[name = string("layers_22_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1825_cast_fp16 = conv(dilations = sparse_output_1825_dilations_1, groups = sparse_output_1825_groups_1, pad = sparse_output_1825_pad_1, pad_type = sparse_output_1825_pad_type_1, strides = sparse_output_1825_strides_1, weight = layers_22_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1825_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29013_cast_fp16 = add(x = dense_output_1825_cast_fp16, y = sparse_output_1825_cast_fp16)[name = string("op_29013_cast_fp16")];
            tensor<int32, [4]> var_29014 = const()[name = string("op_29014"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29016 = const()[name = string("op_29016"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29015_cast_fp16 = transpose(perm = var_29014, x = var_29013_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 51, 128]> q_head_365_cast_fp16 = reshape(shape = var_29016, x = var_29015_cast_fp16)[name = string("q_head_365_cast_fp16")];
            string dense_output_1827_pad_type_1 = const()[name = string("dense_output_1827_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1827_strides_1 = const()[name = string("dense_output_1827_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1827_pad_1 = const()[name = string("dense_output_1827_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1827_dilations_1 = const()[name = string("dense_output_1827_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1827_groups_1 = const()[name = string("dense_output_1827_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661349504))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661480640))))[name = string("layers_22_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1827_cast_fp16 = conv(dilations = dense_output_1827_dilations_1, groups = dense_output_1827_groups_1, pad = dense_output_1827_pad_1, pad_type = dense_output_1827_pad_type_1, strides = dense_output_1827_strides_1, weight = layers_22_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1827_cast_fp16")];
            string sparse_output_1827_pad_type_1 = const()[name = string("sparse_output_1827_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1827_strides_1 = const()[name = string("sparse_output_1827_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1827_pad_1 = const()[name = string("sparse_output_1827_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1827_dilations_1 = const()[name = string("sparse_output_1827_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1827_groups_1 = const()[name = string("sparse_output_1827_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661483904))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661481216))))[name = string("layers_22_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1827_cast_fp16 = conv(dilations = sparse_output_1827_dilations_1, groups = sparse_output_1827_groups_1, pad = sparse_output_1827_pad_1, pad_type = sparse_output_1827_pad_type_1, strides = sparse_output_1827_strides_1, weight = layers_22_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1827_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29032_cast_fp16 = add(x = dense_output_1827_cast_fp16, y = sparse_output_1827_cast_fp16)[name = string("op_29032_cast_fp16")];
            tensor<int32, [4]> var_29033 = const()[name = string("op_29033"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29035 = const()[name = string("op_29035"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29034_cast_fp16 = transpose(perm = var_29033, x = var_29032_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 51, 128]> k_head_729_cast_fp16 = reshape(shape = var_29035, x = var_29034_cast_fp16)[name = string("k_head_729_cast_fp16")];
            string dense_output_1829_pad_type_1 = const()[name = string("dense_output_1829_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1829_strides_1 = const()[name = string("dense_output_1829_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1829_pad_1 = const()[name = string("dense_output_1829_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1829_dilations_1 = const()[name = string("dense_output_1829_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1829_groups_1 = const()[name = string("dense_output_1829_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661500352))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661631488))))[name = string("layers_22_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1829_cast_fp16 = conv(dilations = dense_output_1829_dilations_1, groups = dense_output_1829_groups_1, pad = dense_output_1829_pad_1, pad_type = dense_output_1829_pad_type_1, strides = dense_output_1829_strides_1, weight = layers_22_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1829_cast_fp16")];
            string sparse_output_1829_pad_type_1 = const()[name = string("sparse_output_1829_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1829_strides_1 = const()[name = string("sparse_output_1829_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1829_pad_1 = const()[name = string("sparse_output_1829_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1829_dilations_1 = const()[name = string("sparse_output_1829_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1829_groups_1 = const()[name = string("sparse_output_1829_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661634752))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661632064))))[name = string("layers_22_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1829_cast_fp16 = conv(dilations = sparse_output_1829_dilations_1, groups = sparse_output_1829_groups_1, pad = sparse_output_1829_pad_1, pad_type = sparse_output_1829_pad_type_1, strides = sparse_output_1829_strides_1, weight = layers_22_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1829_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29051_cast_fp16 = add(x = dense_output_1829_cast_fp16, y = sparse_output_1829_cast_fp16)[name = string("op_29051_cast_fp16")];
            tensor<int32, [4]> var_29052 = const()[name = string("op_29052"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29054 = const()[name = string("op_29054"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29053_cast_fp16 = transpose(perm = var_29052, x = var_29051_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 51, 128]> v_head_729_cast_fp16 = reshape(shape = var_29054, x = var_29053_cast_fp16)[name = string("v_head_729_cast_fp16")];
            string dense_output_1831_pad_type_1 = const()[name = string("dense_output_1831_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1831_strides_1 = const()[name = string("dense_output_1831_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1831_pad_1 = const()[name = string("dense_output_1831_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1831_dilations_1 = const()[name = string("dense_output_1831_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1831_groups_1 = const()[name = string("dense_output_1831_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661651200))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661782336))))[name = string("layers_22_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1831_cast_fp16 = conv(dilations = dense_output_1831_dilations_1, groups = dense_output_1831_groups_1, pad = dense_output_1831_pad_1, pad_type = dense_output_1831_pad_type_1, strides = dense_output_1831_strides_1, weight = layers_22_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1831_cast_fp16")];
            string sparse_output_1831_pad_type_1 = const()[name = string("sparse_output_1831_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1831_strides_1 = const()[name = string("sparse_output_1831_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1831_pad_1 = const()[name = string("sparse_output_1831_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1831_dilations_1 = const()[name = string("sparse_output_1831_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1831_groups_1 = const()[name = string("sparse_output_1831_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661785600))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661782912))))[name = string("layers_22_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1831_cast_fp16 = conv(dilations = sparse_output_1831_dilations_1, groups = sparse_output_1831_groups_1, pad = sparse_output_1831_pad_1, pad_type = sparse_output_1831_pad_type_1, strides = sparse_output_1831_strides_1, weight = layers_22_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1831_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_29070_cast_fp16 = add(x = dense_output_1831_cast_fp16, y = sparse_output_1831_cast_fp16)[name = string("op_29070_cast_fp16")];
            tensor<int32, [4]> var_29071 = const()[name = string("op_29071"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29073 = const()[name = string("op_29073"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_29072_cast_fp16 = transpose(perm = var_29071, x = var_29070_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 101, 128]> p_head_729_cast_fp16 = reshape(shape = var_29073, x = var_29072_cast_fp16)[name = string("p_head_729_cast_fp16")];
            tensor<fp16, [128]> var_29075_to_fp16 = const()[name = string("op_29075_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661802048)))];
            tensor<fp16, [1, 51, 128]> var_29076_cast_fp16 = add(x = q_head_365_cast_fp16, y = var_29075_to_fp16)[name = string("op_29076_cast_fp16")];
            tensor<int32, [1]> q_u_365_axes_1 = const()[name = string("q_u_365_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_365_cast_fp16 = expand_dims(axes = q_u_365_axes_1, x = var_29076_cast_fp16)[name = string("q_u_365_cast_fp16")];
            tensor<fp16, [128]> var_29078_to_fp16 = const()[name = string("op_29078_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661802368)))];
            tensor<fp16, [1, 51, 128]> var_29079_cast_fp16 = add(x = q_head_365_cast_fp16, y = var_29078_to_fp16)[name = string("op_29079_cast_fp16")];
            tensor<int32, [1]> q_v_365_axes_1 = const()[name = string("q_v_365_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_365_cast_fp16 = expand_dims(axes = q_v_365_axes_1, x = var_29079_cast_fp16)[name = string("q_v_365_cast_fp16")];
            tensor<int32, [1]> k_head_731_axes_1 = const()[name = string("k_head_731_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_731_cast_fp16 = expand_dims(axes = k_head_731_axes_1, x = k_head_729_cast_fp16)[name = string("k_head_731_cast_fp16")];
            tensor<int32, [1]> v_head_731_axes_1 = const()[name = string("v_head_731_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_731_cast_fp16 = expand_dims(axes = v_head_731_axes_1, x = v_head_729_cast_fp16)[name = string("v_head_731_cast_fp16")];
            tensor<int32, [1]> p_head_731_axes_1 = const()[name = string("p_head_731_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_731_cast_fp16 = expand_dims(axes = p_head_731_axes_1, x = p_head_729_cast_fp16)[name = string("p_head_731_cast_fp16")];
            bool var_29085_transpose_x_3 = const()[name = string("op_29085_transpose_x_3"), val = bool(false)];
            bool var_29085_transpose_y_3 = const()[name = string("op_29085_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_29085_cast_fp16 = matmul(transpose_x = var_29085_transpose_x_3, transpose_y = var_29085_transpose_y_3, x = q_u_365_cast_fp16, y = k_head_731_cast_fp16)[name = string("op_29085_cast_fp16")];
            fp16 var_29086_to_fp16 = const()[name = string("op_29086_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_365_cast_fp16 = mul(x = var_29085_cast_fp16, y = var_29086_to_fp16)[name = string("scores_content_365_cast_fp16")];
            bool x_1917_transpose_x_3 = const()[name = string("x_1917_transpose_x_3"), val = bool(false)];
            bool x_1917_transpose_y_3 = const()[name = string("x_1917_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1917_cast_fp16 = matmul(transpose_x = x_1917_transpose_x_3, transpose_y = x_1917_transpose_y_3, x = q_v_365_cast_fp16, y = p_head_731_cast_fp16)[name = string("x_1917_cast_fp16")];
            tensor<int32, [8]> x_1919_pad_1 = const()[name = string("x_1919_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1919_mode_1 = const()[name = string("x_1919_mode_1"), val = string("constant")];
            fp16 const_2533_to_fp16 = const()[name = string("const_2533_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1919_cast_fp16 = pad(constant_val = const_2533_to_fp16, mode = x_1919_mode_1, pad = x_1919_pad_1, x = x_1917_cast_fp16)[name = string("x_1919_cast_fp16")];
            tensor<int32, [4]> var_29100 = const()[name = string("op_29100"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1921_cast_fp16 = reshape(shape = var_29100, x = x_1919_cast_fp16)[name = string("x_1921_cast_fp16")];
            tensor<int32, [4]> var_29104_begin_1 = const()[name = string("op_29104_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_29104_end_1 = const()[name = string("op_29104_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_29104_end_mask_1 = const()[name = string("op_29104_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_29104_cast_fp16 = slice_by_index(begin = var_29104_begin_1, end = var_29104_end_1, end_mask = var_29104_end_mask_1, x = x_1921_cast_fp16)[name = string("op_29104_cast_fp16")];
            tensor<int32, [4]> var_29106 = const()[name = string("op_29106"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_29107_cast_fp16 = reshape(shape = var_29106, x = var_29104_cast_fp16)[name = string("op_29107_cast_fp16")];
            tensor<int32, [4]> var_29112_begin_1 = const()[name = string("op_29112_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_29112_end_1 = const()[name = string("op_29112_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_29112_end_mask_1 = const()[name = string("op_29112_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_29112_cast_fp16 = slice_by_index(begin = var_29112_begin_1, end = var_29112_end_1, end_mask = var_29112_end_mask_1, x = var_29107_cast_fp16)[name = string("op_29112_cast_fp16")];
            fp16 var_29113_to_fp16 = const()[name = string("op_29113_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_365_cast_fp16 = mul(x = var_29112_cast_fp16, y = var_29113_to_fp16)[name = string("scores_pos_365_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_365_cast_fp16 = add(x = scores_content_365_cast_fp16, y = scores_pos_365_cast_fp16)[name = string("logits_365_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_29116_cast_fp16 = softmax(axis = var_28115, x = logits_365_cast_fp16)[name = string("op_29116_cast_fp16")];
            bool var_29118_transpose_x_1 = const()[name = string("op_29118_transpose_x_1"), val = bool(false)];
            bool var_29118_transpose_y_1 = const()[name = string("op_29118_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_29118_cast_fp16 = matmul(transpose_x = var_29118_transpose_x_1, transpose_y = var_29118_transpose_y_1, x = var_29116_cast_fp16, y = v_head_731_cast_fp16)[name = string("op_29118_cast_fp16")];
            tensor<int32, [1]> var_29119_axes_1 = const()[name = string("op_29119_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_29119_cast_fp16 = squeeze(axes = var_29119_axes_1, x = var_29118_cast_fp16)[name = string("op_29119_cast_fp16")];
            string dense_output_1833_pad_type_1 = const()[name = string("dense_output_1833_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1833_strides_1 = const()[name = string("dense_output_1833_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1833_pad_1 = const()[name = string("dense_output_1833_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1833_dilations_1 = const()[name = string("dense_output_1833_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1833_groups_1 = const()[name = string("dense_output_1833_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661802688))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661933824))))[name = string("layers_22_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1833_cast_fp16 = conv(dilations = dense_output_1833_dilations_1, groups = dense_output_1833_groups_1, pad = dense_output_1833_pad_1, pad_type = dense_output_1833_pad_type_1, strides = dense_output_1833_strides_1, weight = layers_22_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1833_cast_fp16")];
            string sparse_output_1833_pad_type_1 = const()[name = string("sparse_output_1833_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1833_strides_1 = const()[name = string("sparse_output_1833_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1833_pad_1 = const()[name = string("sparse_output_1833_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1833_dilations_1 = const()[name = string("sparse_output_1833_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1833_groups_1 = const()[name = string("sparse_output_1833_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661937088))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661934400))))[name = string("layers_22_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1833_cast_fp16 = conv(dilations = sparse_output_1833_dilations_1, groups = sparse_output_1833_groups_1, pad = sparse_output_1833_pad_1, pad_type = sparse_output_1833_pad_type_1, strides = sparse_output_1833_strides_1, weight = layers_22_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1833_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29134_cast_fp16 = add(x = dense_output_1833_cast_fp16, y = sparse_output_1833_cast_fp16)[name = string("op_29134_cast_fp16")];
            tensor<int32, [4]> var_29135 = const()[name = string("op_29135"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29137 = const()[name = string("op_29137"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29136_cast_fp16 = transpose(perm = var_29135, x = var_29134_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 51, 128]> q_head_367_cast_fp16 = reshape(shape = var_29137, x = var_29136_cast_fp16)[name = string("q_head_367_cast_fp16")];
            string dense_output_1835_pad_type_1 = const()[name = string("dense_output_1835_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1835_strides_1 = const()[name = string("dense_output_1835_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1835_pad_1 = const()[name = string("dense_output_1835_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1835_dilations_1 = const()[name = string("dense_output_1835_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1835_groups_1 = const()[name = string("dense_output_1835_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661953536))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662084672))))[name = string("layers_22_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1835_cast_fp16 = conv(dilations = dense_output_1835_dilations_1, groups = dense_output_1835_groups_1, pad = dense_output_1835_pad_1, pad_type = dense_output_1835_pad_type_1, strides = dense_output_1835_strides_1, weight = layers_22_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1835_cast_fp16")];
            string sparse_output_1835_pad_type_1 = const()[name = string("sparse_output_1835_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1835_strides_1 = const()[name = string("sparse_output_1835_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1835_pad_1 = const()[name = string("sparse_output_1835_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1835_dilations_1 = const()[name = string("sparse_output_1835_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1835_groups_1 = const()[name = string("sparse_output_1835_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662087936))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662085248))))[name = string("layers_22_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1835_cast_fp16 = conv(dilations = sparse_output_1835_dilations_1, groups = sparse_output_1835_groups_1, pad = sparse_output_1835_pad_1, pad_type = sparse_output_1835_pad_type_1, strides = sparse_output_1835_strides_1, weight = layers_22_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1835_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29153_cast_fp16 = add(x = dense_output_1835_cast_fp16, y = sparse_output_1835_cast_fp16)[name = string("op_29153_cast_fp16")];
            tensor<int32, [4]> var_29154 = const()[name = string("op_29154"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29156 = const()[name = string("op_29156"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29155_cast_fp16 = transpose(perm = var_29154, x = var_29153_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 51, 128]> k_head_733_cast_fp16 = reshape(shape = var_29156, x = var_29155_cast_fp16)[name = string("k_head_733_cast_fp16")];
            string dense_output_1837_pad_type_1 = const()[name = string("dense_output_1837_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1837_strides_1 = const()[name = string("dense_output_1837_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1837_pad_1 = const()[name = string("dense_output_1837_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1837_dilations_1 = const()[name = string("dense_output_1837_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1837_groups_1 = const()[name = string("dense_output_1837_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662104384))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662235520))))[name = string("layers_22_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1837_cast_fp16 = conv(dilations = dense_output_1837_dilations_1, groups = dense_output_1837_groups_1, pad = dense_output_1837_pad_1, pad_type = dense_output_1837_pad_type_1, strides = dense_output_1837_strides_1, weight = layers_22_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_1041_cast_fp16)[name = string("dense_output_1837_cast_fp16")];
            string sparse_output_1837_pad_type_1 = const()[name = string("sparse_output_1837_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1837_strides_1 = const()[name = string("sparse_output_1837_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1837_pad_1 = const()[name = string("sparse_output_1837_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1837_dilations_1 = const()[name = string("sparse_output_1837_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1837_groups_1 = const()[name = string("sparse_output_1837_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662238784))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662236096))))[name = string("layers_22_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1837_cast_fp16 = conv(dilations = sparse_output_1837_dilations_1, groups = sparse_output_1837_groups_1, pad = sparse_output_1837_pad_1, pad_type = sparse_output_1837_pad_type_1, strides = sparse_output_1837_strides_1, weight = layers_22_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_1041_cast_fp16)[name = string("sparse_output_1837_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29172_cast_fp16 = add(x = dense_output_1837_cast_fp16, y = sparse_output_1837_cast_fp16)[name = string("op_29172_cast_fp16")];
            tensor<int32, [4]> var_29173 = const()[name = string("op_29173"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29175 = const()[name = string("op_29175"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29174_cast_fp16 = transpose(perm = var_29173, x = var_29172_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 51, 128]> v_head_733_cast_fp16 = reshape(shape = var_29175, x = var_29174_cast_fp16)[name = string("v_head_733_cast_fp16")];
            string dense_output_1839_pad_type_1 = const()[name = string("dense_output_1839_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1839_strides_1 = const()[name = string("dense_output_1839_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1839_pad_1 = const()[name = string("dense_output_1839_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1839_dilations_1 = const()[name = string("dense_output_1839_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1839_groups_1 = const()[name = string("dense_output_1839_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662255232))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662386368))))[name = string("layers_22_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1839_cast_fp16 = conv(dilations = dense_output_1839_dilations_1, groups = dense_output_1839_groups_1, pad = dense_output_1839_pad_1, pad_type = dense_output_1839_pad_type_1, strides = dense_output_1839_strides_1, weight = layers_22_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1839_cast_fp16")];
            string sparse_output_1839_pad_type_1 = const()[name = string("sparse_output_1839_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1839_strides_1 = const()[name = string("sparse_output_1839_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1839_pad_1 = const()[name = string("sparse_output_1839_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1839_dilations_1 = const()[name = string("sparse_output_1839_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1839_groups_1 = const()[name = string("sparse_output_1839_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_22_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662389632))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662386944))))[name = string("layers_22_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1839_cast_fp16 = conv(dilations = sparse_output_1839_dilations_1, groups = sparse_output_1839_groups_1, pad = sparse_output_1839_pad_1, pad_type = sparse_output_1839_pad_type_1, strides = sparse_output_1839_strides_1, weight = layers_22_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1839_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_29191_cast_fp16 = add(x = dense_output_1839_cast_fp16, y = sparse_output_1839_cast_fp16)[name = string("op_29191_cast_fp16")];
            tensor<int32, [4]> var_29192 = const()[name = string("op_29192"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29194 = const()[name = string("op_29194"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_29193_cast_fp16 = transpose(perm = var_29192, x = var_29191_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 101, 128]> p_head_733_cast_fp16 = reshape(shape = var_29194, x = var_29193_cast_fp16)[name = string("p_head_733_cast_fp16")];
            tensor<fp16, [128]> var_29196_to_fp16 = const()[name = string("op_29196_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662406080)))];
            tensor<fp16, [1, 51, 128]> var_29197_cast_fp16 = add(x = q_head_367_cast_fp16, y = var_29196_to_fp16)[name = string("op_29197_cast_fp16")];
            tensor<int32, [1]> q_u_367_axes_1 = const()[name = string("q_u_367_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_367_cast_fp16 = expand_dims(axes = q_u_367_axes_1, x = var_29197_cast_fp16)[name = string("q_u_367_cast_fp16")];
            tensor<fp16, [128]> var_29199_to_fp16 = const()[name = string("op_29199_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662406400)))];
            tensor<fp16, [1, 51, 128]> var_29200_cast_fp16 = add(x = q_head_367_cast_fp16, y = var_29199_to_fp16)[name = string("op_29200_cast_fp16")];
            tensor<int32, [1]> q_v_367_axes_1 = const()[name = string("q_v_367_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_367_cast_fp16 = expand_dims(axes = q_v_367_axes_1, x = var_29200_cast_fp16)[name = string("q_v_367_cast_fp16")];
            tensor<int32, [1]> k_head_735_axes_1 = const()[name = string("k_head_735_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_735_cast_fp16 = expand_dims(axes = k_head_735_axes_1, x = k_head_733_cast_fp16)[name = string("k_head_735_cast_fp16")];
            tensor<int32, [1]> v_head_735_axes_1 = const()[name = string("v_head_735_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_735_cast_fp16 = expand_dims(axes = v_head_735_axes_1, x = v_head_733_cast_fp16)[name = string("v_head_735_cast_fp16")];
            tensor<int32, [1]> p_head_735_axes_1 = const()[name = string("p_head_735_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_735_cast_fp16 = expand_dims(axes = p_head_735_axes_1, x = p_head_733_cast_fp16)[name = string("p_head_735_cast_fp16")];
            bool var_29206_transpose_x_3 = const()[name = string("op_29206_transpose_x_3"), val = bool(false)];
            bool var_29206_transpose_y_3 = const()[name = string("op_29206_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_29206_cast_fp16 = matmul(transpose_x = var_29206_transpose_x_3, transpose_y = var_29206_transpose_y_3, x = q_u_367_cast_fp16, y = k_head_735_cast_fp16)[name = string("op_29206_cast_fp16")];
            fp16 var_29207_to_fp16 = const()[name = string("op_29207_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_367_cast_fp16 = mul(x = var_29206_cast_fp16, y = var_29207_to_fp16)[name = string("scores_content_367_cast_fp16")];
            bool x_1925_transpose_x_3 = const()[name = string("x_1925_transpose_x_3"), val = bool(false)];
            bool x_1925_transpose_y_3 = const()[name = string("x_1925_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1925_cast_fp16 = matmul(transpose_x = x_1925_transpose_x_3, transpose_y = x_1925_transpose_y_3, x = q_v_367_cast_fp16, y = p_head_735_cast_fp16)[name = string("x_1925_cast_fp16")];
            tensor<int32, [8]> x_1927_pad_1 = const()[name = string("x_1927_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1927_mode_1 = const()[name = string("x_1927_mode_1"), val = string("constant")];
            fp16 const_2539_to_fp16 = const()[name = string("const_2539_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1927_cast_fp16 = pad(constant_val = const_2539_to_fp16, mode = x_1927_mode_1, pad = x_1927_pad_1, x = x_1925_cast_fp16)[name = string("x_1927_cast_fp16")];
            tensor<int32, [4]> var_29221 = const()[name = string("op_29221"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1929_cast_fp16 = reshape(shape = var_29221, x = x_1927_cast_fp16)[name = string("x_1929_cast_fp16")];
            tensor<int32, [4]> var_29225_begin_1 = const()[name = string("op_29225_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_29225_end_1 = const()[name = string("op_29225_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_29225_end_mask_1 = const()[name = string("op_29225_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_29225_cast_fp16 = slice_by_index(begin = var_29225_begin_1, end = var_29225_end_1, end_mask = var_29225_end_mask_1, x = x_1929_cast_fp16)[name = string("op_29225_cast_fp16")];
            tensor<int32, [4]> var_29227 = const()[name = string("op_29227"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_29228_cast_fp16 = reshape(shape = var_29227, x = var_29225_cast_fp16)[name = string("op_29228_cast_fp16")];
            tensor<int32, [4]> var_29233_begin_1 = const()[name = string("op_29233_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_29233_end_1 = const()[name = string("op_29233_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_29233_end_mask_1 = const()[name = string("op_29233_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_29233_cast_fp16 = slice_by_index(begin = var_29233_begin_1, end = var_29233_end_1, end_mask = var_29233_end_mask_1, x = var_29228_cast_fp16)[name = string("op_29233_cast_fp16")];
            fp16 var_29234_to_fp16 = const()[name = string("op_29234_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_367_cast_fp16 = mul(x = var_29233_cast_fp16, y = var_29234_to_fp16)[name = string("scores_pos_367_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_367_cast_fp16 = add(x = scores_content_367_cast_fp16, y = scores_pos_367_cast_fp16)[name = string("logits_367_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_29237_cast_fp16 = softmax(axis = var_28115, x = logits_367_cast_fp16)[name = string("op_29237_cast_fp16")];
            bool var_29239_transpose_x_1 = const()[name = string("op_29239_transpose_x_1"), val = bool(false)];
            bool var_29239_transpose_y_1 = const()[name = string("op_29239_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_29239_cast_fp16 = matmul(transpose_x = var_29239_transpose_x_1, transpose_y = var_29239_transpose_y_1, x = var_29237_cast_fp16, y = v_head_735_cast_fp16)[name = string("op_29239_cast_fp16")];
            tensor<int32, [1]> o_head_45_axes_1 = const()[name = string("o_head_45_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_45_cast_fp16 = squeeze(axes = o_head_45_axes_1, x = var_29239_cast_fp16)[name = string("o_head_45_cast_fp16")];
            bool out_45_interleave_1 = const()[name = string("out_45_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_45_cast_fp16 = concat(axis = var_28115, interleave = out_45_interleave_1, values = (var_28393_cast_fp16, var_28514_cast_fp16, var_28635_cast_fp16, var_28756_cast_fp16, var_28877_cast_fp16, var_28998_cast_fp16, var_29119_cast_fp16, o_head_45_cast_fp16))[name = string("out_45_cast_fp16")];
            tensor<int32, [3]> var_29243_perm_1 = const()[name = string("op_29243_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1049_axes_1 = const()[name = string("input_1049_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_29243_cast_fp16 = transpose(perm = var_29243_perm_1, x = out_45_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 1024, 51, 1]> input_1049_cast_fp16 = expand_dims(axes = input_1049_axes_1, x = var_29243_cast_fp16)[name = string("input_1049_cast_fp16")];
            string dense_output_1841_pad_type_1 = const()[name = string("dense_output_1841_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1841_strides_1 = const()[name = string("dense_output_1841_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1841_pad_1 = const()[name = string("dense_output_1841_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1841_dilations_1 = const()[name = string("dense_output_1841_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1841_groups_1 = const()[name = string("dense_output_1841_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662406720))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663455360))))[name = string("layers_22_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1841_cast_fp16 = conv(dilations = dense_output_1841_dilations_1, groups = dense_output_1841_groups_1, pad = dense_output_1841_pad_1, pad_type = dense_output_1841_pad_type_1, strides = dense_output_1841_strides_1, weight = layers_22_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_1049_cast_fp16)[name = string("dense_output_1841_cast_fp16")];
            string sparse_output_1841_pad_type_1 = const()[name = string("sparse_output_1841_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1841_strides_1 = const()[name = string("sparse_output_1841_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1841_pad_1 = const()[name = string("sparse_output_1841_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1841_dilations_1 = const()[name = string("sparse_output_1841_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1841_groups_1 = const()[name = string("sparse_output_1841_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663476992))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663455936))))[name = string("layers_22_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1841_cast_fp16 = conv(dilations = sparse_output_1841_dilations_1, groups = sparse_output_1841_groups_1, pad = sparse_output_1841_pad_1, pad_type = sparse_output_1841_pad_type_1, strides = sparse_output_1841_strides_1, weight = layers_22_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_1049_cast_fp16)[name = string("sparse_output_1841_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_45_cast_fp16 = add(x = dense_output_1841_cast_fp16, y = sparse_output_1841_cast_fp16)[name = string("out_conv_45_cast_fp16")];
            tensor<int32, [1]> var_29260_axes_1 = const()[name = string("op_29260_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_29260_cast_fp16 = squeeze(axes = var_29260_axes_1, x = out_conv_45_cast_fp16)[name = string("op_29260_cast_fp16")];
            tensor<int32, [3]> var_29261_perm_1 = const()[name = string("op_29261_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_29261_cast_fp16 = transpose(perm = var_29261_perm_1, x = var_29260_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 51, 1024]> input_1051_cast_fp16 = add(x = input_1039_cast_fp16, y = var_29261_cast_fp16)[name = string("input_1051_cast_fp16")];
            tensor<int32, [1]> x_1933_axes_1 = const()[name = string("x_1933_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_22_norm_conv_weight_to_fp16 = const()[name = string("layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663608128)))];
            tensor<fp16, [1024]> layers_22_norm_conv_bias_to_fp16 = const()[name = string("layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663610240)))];
            tensor<fp16, [1, 51, 1024]> x_1933_cast_fp16 = layer_norm(axes = x_1933_axes_1, beta = layers_22_norm_conv_bias_to_fp16, epsilon = var_28130_to_fp16, gamma = layers_22_norm_conv_weight_to_fp16, x = input_1051_cast_fp16)[name = string("x_1933_cast_fp16")];
            tensor<int32, [3]> var_29271_perm_1 = const()[name = string("op_29271_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1053_axes_1 = const()[name = string("input_1053_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_29271_cast_fp16 = transpose(perm = var_29271_perm_1, x = x_1933_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 1024, 51, 1]> input_1053_cast_fp16 = expand_dims(axes = input_1053_axes_1, x = var_29271_cast_fp16)[name = string("input_1053_cast_fp16")];
            string dense_output_1843_pad_type_1 = const()[name = string("dense_output_1843_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1843_strides_1 = const()[name = string("dense_output_1843_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1843_pad_1 = const()[name = string("dense_output_1843_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1843_dilations_1 = const()[name = string("dense_output_1843_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1843_groups_1 = const()[name = string("dense_output_1843_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_22_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(663612352))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(665709568))))[name = string("layers_22_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1843_cast_fp16 = conv(dilations = dense_output_1843_dilations_1, groups = dense_output_1843_groups_1, pad = dense_output_1843_pad_1, pad_type = dense_output_1843_pad_type_1, strides = dense_output_1843_strides_1, weight = layers_22_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_1053_cast_fp16)[name = string("dense_output_1843_cast_fp16")];
            string sparse_output_1843_pad_type_1 = const()[name = string("sparse_output_1843_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1843_strides_1 = const()[name = string("sparse_output_1843_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1843_pad_1 = const()[name = string("sparse_output_1843_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1843_dilations_1 = const()[name = string("sparse_output_1843_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1843_groups_1 = const()[name = string("sparse_output_1843_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_22_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(665752192))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(665710144))))[name = string("layers_22_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1843_cast_fp16 = conv(dilations = sparse_output_1843_dilations_1, groups = sparse_output_1843_groups_1, pad = sparse_output_1843_pad_1, pad_type = sparse_output_1843_pad_type_1, strides = sparse_output_1843_strides_1, weight = layers_22_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_1053_cast_fp16)[name = string("sparse_output_1843_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_1055_cast_fp16 = add(x = dense_output_1843_cast_fp16, y = sparse_output_1843_cast_fp16)[name = string("input_1055_cast_fp16")];
            int32 input_1057_split_num_splits_1 = const()[name = string("input_1057_split_num_splits_1"), val = int32(2)];
            int32 input_1057_split_axis_1 = const()[name = string("input_1057_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_1057_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_1057_split_cast_fp16_1 = split(axis = input_1057_split_axis_1, num_splits = input_1057_split_num_splits_1, x = input_1055_cast_fp16)[name = string("input_1057_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1057_split_1_sigmoid_cast_fp16 = sigmoid(x = input_1057_split_cast_fp16_1)[name = string("input_1057_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1057_cast_fp16 = mul(x = input_1057_split_cast_fp16_0, y = input_1057_split_1_sigmoid_cast_fp16)[name = string("input_1057_cast_fp16")];
            tensor<int32, [8]> input_1059_pad_1 = const()[name = string("input_1059_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_1059_mode_1 = const()[name = string("input_1059_mode_1"), val = string("constant")];
            fp16 const_2541_to_fp16 = const()[name = string("const_2541_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_1059_cast_fp16 = pad(constant_val = const_2541_to_fp16, mode = input_1059_mode_1, pad = input_1059_pad_1, x = input_1057_cast_fp16)[name = string("input_1059_cast_fp16")];
            string dense_output_1845_pad_type_1 = const()[name = string("dense_output_1845_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1845_strides_1 = const()[name = string("dense_output_1845_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1845_pad_1 = const()[name = string("dense_output_1845_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1845_dilations_1 = const()[name = string("dense_output_1845_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1845_groups_1 = const()[name = string("dense_output_1845_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1845_cast_fp16 = conv(dilations = dense_output_1845_dilations_1, groups = dense_output_1845_groups_1, pad = dense_output_1845_pad_1, pad_type = dense_output_1845_pad_type_1, strides = dense_output_1845_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_1059_cast_fp16)[name = string("dense_output_1845_cast_fp16")];
            string sparse_output_1845_pad_type_1 = const()[name = string("sparse_output_1845_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1845_strides_1 = const()[name = string("sparse_output_1845_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1845_pad_1 = const()[name = string("sparse_output_1845_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1845_dilations_1 = const()[name = string("sparse_output_1845_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1845_groups_1 = const()[name = string("sparse_output_1845_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_22_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666014400))))[name = string("layers_22_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1845_cast_fp16 = conv(dilations = sparse_output_1845_dilations_1, groups = sparse_output_1845_groups_1, pad = sparse_output_1845_pad_1, pad_type = sparse_output_1845_pad_type_1, strides = sparse_output_1845_strides_1, weight = layers_22_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_1059_cast_fp16)[name = string("sparse_output_1845_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1061_cast_fp16 = add(x = dense_output_1845_cast_fp16, y = sparse_output_1845_cast_fp16)[name = string("input_1061_cast_fp16")];
            tensor<fp16, [1024]> layers_22_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_22_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666032896)))];
            tensor<fp16, [1024]> layers_22_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_22_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666035008)))];
            tensor<fp16, [1024]> layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_22_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666037120)))];
            tensor<fp16, [1024]> layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_22_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666039232)))];
            tensor<fp16, [1, 1024, 51, 1]> input_1063_cast_fp16 = batch_norm(beta = layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_28130_to_fp16, gamma = layers_22_conv_batch_norm_weight_to_fp16, mean = layers_22_conv_batch_norm_running_mean_to_fp16, variance = layers_22_conv_batch_norm_running_var_to_fp16, x = input_1061_cast_fp16)[name = string("input_1063_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1065_cast_fp16 = silu(x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")];
            string dense_output_1847_pad_type_1 = const()[name = string("dense_output_1847_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1847_strides_1 = const()[name = string("dense_output_1847_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1847_pad_1 = const()[name = string("dense_output_1847_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1847_dilations_1 = const()[name = string("dense_output_1847_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1847_groups_1 = const()[name = string("dense_output_1847_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(666041344))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667089984))))[name = string("layers_22_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1847_cast_fp16 = conv(dilations = dense_output_1847_dilations_1, groups = dense_output_1847_groups_1, pad = dense_output_1847_pad_1, pad_type = dense_output_1847_pad_type_1, strides = dense_output_1847_strides_1, weight = layers_22_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_1065_cast_fp16)[name = string("dense_output_1847_cast_fp16")];
            string sparse_output_1847_pad_type_1 = const()[name = string("sparse_output_1847_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1847_strides_1 = const()[name = string("sparse_output_1847_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1847_pad_1 = const()[name = string("sparse_output_1847_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1847_dilations_1 = const()[name = string("sparse_output_1847_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1847_groups_1 = const()[name = string("sparse_output_1847_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_22_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667111616))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667090560))))[name = string("layers_22_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1847_cast_fp16 = conv(dilations = sparse_output_1847_dilations_1, groups = sparse_output_1847_groups_1, pad = sparse_output_1847_pad_1, pad_type = sparse_output_1847_pad_type_1, strides = sparse_output_1847_strides_1, weight = layers_22_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_1065_cast_fp16)[name = string("sparse_output_1847_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_1935_cast_fp16 = add(x = dense_output_1847_cast_fp16, y = sparse_output_1847_cast_fp16)[name = string("x_1935_cast_fp16")];
            tensor<int32, [1]> var_29327_axes_1 = const()[name = string("op_29327_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_29327_cast_fp16 = squeeze(axes = var_29327_axes_1, x = x_1935_cast_fp16)[name = string("op_29327_cast_fp16")];
            tensor<int32, [3]> var_29328_perm_1 = const()[name = string("op_29328_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_29328_cast_fp16 = transpose(perm = var_29328_perm_1, x = var_29327_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 51, 1024]> input_1067_cast_fp16 = add(x = input_1051_cast_fp16, y = var_29328_cast_fp16)[name = string("input_1067_cast_fp16")];
            tensor<int32, [1]> x_1937_axes_1 = const()[name = string("x_1937_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667242752)))];
            tensor<fp16, [1024]> layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667244864)))];
            tensor<fp16, [1, 51, 1024]> x_1937_cast_fp16 = layer_norm(axes = x_1937_axes_1, beta = layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_28130_to_fp16, gamma = layers_22_norm_feed_forward2_weight_to_fp16, x = input_1067_cast_fp16)[name = string("x_1937_cast_fp16")];
            tensor<int32, [4]> var_29338 = const()[name = string("op_29338"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1939_cast_fp16 = reshape(shape = var_29338, x = x_1937_cast_fp16)[name = string("x_1939_cast_fp16")];
            tensor<int32, [4]> input_1069_perm_1 = const()[name = string("input_1069_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1849_pad_type_1 = const()[name = string("dense_output_1849_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1849_strides_1 = const()[name = string("dense_output_1849_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1849_pad_1 = const()[name = string("dense_output_1849_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1849_dilations_1 = const()[name = string("dense_output_1849_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1849_groups_1 = const()[name = string("dense_output_1849_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_22_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667246976))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(671441344))))[name = string("layers_22_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_1069_cast_fp16 = transpose(perm = input_1069_perm_1, x = x_1939_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1849_cast_fp16 = conv(dilations = dense_output_1849_dilations_1, groups = dense_output_1849_groups_1, pad = dense_output_1849_pad_1, pad_type = dense_output_1849_pad_type_1, strides = dense_output_1849_strides_1, weight = layers_22_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_1069_cast_fp16)[name = string("dense_output_1849_cast_fp16")];
            string sparse_output_1849_pad_type_1 = const()[name = string("sparse_output_1849_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1849_strides_1 = const()[name = string("sparse_output_1849_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1849_pad_1 = const()[name = string("sparse_output_1849_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1849_dilations_1 = const()[name = string("sparse_output_1849_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1849_groups_1 = const()[name = string("sparse_output_1849_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_22_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(671525888))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(671441920))))[name = string("layers_22_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1849_cast_fp16 = conv(dilations = sparse_output_1849_dilations_1, groups = sparse_output_1849_groups_1, pad = sparse_output_1849_pad_1, pad_type = sparse_output_1849_pad_type_1, strides = sparse_output_1849_strides_1, weight = layers_22_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_1069_cast_fp16)[name = string("sparse_output_1849_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1071_cast_fp16 = add(x = dense_output_1849_cast_fp16, y = sparse_output_1849_cast_fp16)[name = string("input_1071_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1073_cast_fp16 = silu(x = input_1071_cast_fp16)[name = string("input_1073_cast_fp16")];
            string dense_output_1851_pad_type_1 = const()[name = string("dense_output_1851_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1851_strides_1 = const()[name = string("dense_output_1851_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1851_pad_1 = const()[name = string("dense_output_1851_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1851_dilations_1 = const()[name = string("dense_output_1851_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1851_groups_1 = const()[name = string("dense_output_1851_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_22_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672050240))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676244608))))[name = string("layers_22_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1851_cast_fp16 = conv(dilations = dense_output_1851_dilations_1, groups = dense_output_1851_groups_1, pad = dense_output_1851_pad_1, pad_type = dense_output_1851_pad_type_1, strides = dense_output_1851_strides_1, weight = layers_22_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_1073_cast_fp16)[name = string("dense_output_1851_cast_fp16")];
            string sparse_output_1851_pad_type_1 = const()[name = string("sparse_output_1851_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1851_strides_1 = const()[name = string("sparse_output_1851_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1851_pad_1 = const()[name = string("sparse_output_1851_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1851_dilations_1 = const()[name = string("sparse_output_1851_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1851_groups_1 = const()[name = string("sparse_output_1851_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_22_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676329152))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676245184))))[name = string("layers_22_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1851_cast_fp16 = conv(dilations = sparse_output_1851_dilations_1, groups = sparse_output_1851_groups_1, pad = sparse_output_1851_pad_1, pad_type = sparse_output_1851_pad_type_1, strides = sparse_output_1851_strides_1, weight = layers_22_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_1073_cast_fp16)[name = string("sparse_output_1851_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1941_cast_fp16 = add(x = dense_output_1851_cast_fp16, y = sparse_output_1851_cast_fp16)[name = string("x_1941_cast_fp16")];
            tensor<int32, [4]> x_1943_perm_1 = const()[name = string("x_1943_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_29373 = const()[name = string("op_29373"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1943_cast_fp16 = transpose(perm = x_1943_perm_1, x = x_1941_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 51, 1024]> var_29374_cast_fp16 = reshape(shape = var_29373, x = x_1943_cast_fp16)[name = string("op_29374_cast_fp16")];
            fp16 var_29375_to_fp16 = const()[name = string("op_29375_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_29376_cast_fp16 = mul(x = var_29374_cast_fp16, y = var_29375_to_fp16)[name = string("op_29376_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_1075_cast_fp16 = add(x = input_1067_cast_fp16, y = var_29376_cast_fp16)[name = string("input_1075_cast_fp16")];
            tensor<int32, [1]> input_1077_axes_1 = const()[name = string("input_1077_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_22_norm_out_weight_to_fp16 = const()[name = string("layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676853504)))];
            tensor<fp16, [1024]> layers_22_norm_out_bias_to_fp16 = const()[name = string("layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676855616)))];
            tensor<fp16, [1, 51, 1024]> input_1077_cast_fp16 = layer_norm(axes = input_1077_axes_1, beta = layers_22_norm_out_bias_to_fp16, epsilon = var_28130_to_fp16, gamma = layers_22_norm_out_weight_to_fp16, x = input_1075_cast_fp16)[name = string("input_1077_cast_fp16")];
            int32 var_29384 = const()[name = string("op_29384"), val = int32(-1)];
            tensor<int32, [1]> x_1945_axes_1 = const()[name = string("x_1945_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676857728)))];
            tensor<fp16, [1024]> layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676859840)))];
            fp16 var_29399_to_fp16 = const()[name = string("op_29399_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 51, 1024]> x_1945_cast_fp16 = layer_norm(axes = x_1945_axes_1, beta = layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_29399_to_fp16, gamma = layers_23_norm_feed_forward1_weight_to_fp16, x = input_1077_cast_fp16)[name = string("x_1945_cast_fp16")];
            tensor<int32, [4]> var_29418 = const()[name = string("op_29418"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1947_cast_fp16 = reshape(shape = var_29418, x = x_1945_cast_fp16)[name = string("x_1947_cast_fp16")];
            tensor<int32, [4]> input_1079_perm_1 = const()[name = string("input_1079_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1853_pad_type_1 = const()[name = string("dense_output_1853_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1853_strides_1 = const()[name = string("dense_output_1853_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1853_pad_1 = const()[name = string("dense_output_1853_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1853_dilations_1 = const()[name = string("dense_output_1853_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1853_groups_1 = const()[name = string("dense_output_1853_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_23_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676861952))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681056320))))[name = string("layers_23_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_1079_cast_fp16 = transpose(perm = input_1079_perm_1, x = x_1947_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1853_cast_fp16 = conv(dilations = dense_output_1853_dilations_1, groups = dense_output_1853_groups_1, pad = dense_output_1853_pad_1, pad_type = dense_output_1853_pad_type_1, strides = dense_output_1853_strides_1, weight = layers_23_feed_forward1_linear1_dense_conv_weight_to_fp16_palettized, x = input_1079_cast_fp16)[name = string("dense_output_1853_cast_fp16")];
            string sparse_output_1853_pad_type_1 = const()[name = string("sparse_output_1853_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1853_strides_1 = const()[name = string("sparse_output_1853_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1853_pad_1 = const()[name = string("sparse_output_1853_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1853_dilations_1 = const()[name = string("sparse_output_1853_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1853_groups_1 = const()[name = string("sparse_output_1853_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_23_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681140864))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681056896))))[name = string("layers_23_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1853_cast_fp16 = conv(dilations = sparse_output_1853_dilations_1, groups = sparse_output_1853_groups_1, pad = sparse_output_1853_pad_1, pad_type = sparse_output_1853_pad_type_1, strides = sparse_output_1853_strides_1, weight = layers_23_feed_forward1_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_1079_cast_fp16)[name = string("sparse_output_1853_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1081_cast_fp16 = add(x = dense_output_1853_cast_fp16, y = sparse_output_1853_cast_fp16)[name = string("input_1081_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1083_cast_fp16 = silu(x = input_1081_cast_fp16)[name = string("input_1083_cast_fp16")];
            string dense_output_1855_pad_type_1 = const()[name = string("dense_output_1855_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1855_strides_1 = const()[name = string("dense_output_1855_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1855_pad_1 = const()[name = string("dense_output_1855_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1855_dilations_1 = const()[name = string("dense_output_1855_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1855_groups_1 = const()[name = string("dense_output_1855_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_23_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(681665216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685859584))))[name = string("layers_23_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_1855_cast_fp16 = conv(dilations = dense_output_1855_dilations_1, groups = dense_output_1855_groups_1, pad = dense_output_1855_pad_1, pad_type = dense_output_1855_pad_type_1, strides = dense_output_1855_strides_1, weight = layers_23_feed_forward1_linear2_dense_conv_weight_to_fp16_palettized, x = input_1083_cast_fp16)[name = string("dense_output_1855_cast_fp16")];
            string sparse_output_1855_pad_type_1 = const()[name = string("sparse_output_1855_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1855_strides_1 = const()[name = string("sparse_output_1855_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1855_pad_1 = const()[name = string("sparse_output_1855_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1855_dilations_1 = const()[name = string("sparse_output_1855_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1855_groups_1 = const()[name = string("sparse_output_1855_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_23_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685944128))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685860160))))[name = string("layers_23_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_1855_cast_fp16 = conv(dilations = sparse_output_1855_dilations_1, groups = sparse_output_1855_groups_1, pad = sparse_output_1855_pad_1, pad_type = sparse_output_1855_pad_type_1, strides = sparse_output_1855_strides_1, weight = layers_23_feed_forward1_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_1083_cast_fp16)[name = string("sparse_output_1855_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_1949_cast_fp16 = add(x = dense_output_1855_cast_fp16, y = sparse_output_1855_cast_fp16)[name = string("x_1949_cast_fp16")];
            tensor<int32, [4]> x_1951_perm_1 = const()[name = string("x_1951_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_29453 = const()[name = string("op_29453"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_1951_cast_fp16 = transpose(perm = x_1951_perm_1, x = x_1949_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 51, 1024]> var_29454_cast_fp16 = reshape(shape = var_29453, x = x_1951_cast_fp16)[name = string("op_29454_cast_fp16")];
            fp16 var_29455_to_fp16 = const()[name = string("op_29455_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_29456_cast_fp16 = mul(x = var_29454_cast_fp16, y = var_29455_to_fp16)[name = string("op_29456_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_1085_cast_fp16 = add(x = input_1077_cast_fp16, y = var_29456_cast_fp16)[name = string("input_1085_cast_fp16")];
            tensor<int32, [1]> q_axes_1 = const()[name = string("q_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_23_norm_self_att_weight_to_fp16 = const()[name = string("layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686468480)))];
            tensor<fp16, [1024]> layers_23_norm_self_att_bias_to_fp16 = const()[name = string("layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686470592)))];
            tensor<fp16, [1, 51, 1024]> q_cast_fp16 = layer_norm(axes = q_axes_1, beta = layers_23_norm_self_att_bias_to_fp16, epsilon = var_29399_to_fp16, gamma = layers_23_norm_self_att_weight_to_fp16, x = input_1085_cast_fp16)[name = string("q_cast_fp16")];
            tensor<int32, [3]> var_29530 = const()[name = string("op_29530"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1087_axes_1 = const()[name = string("input_1087_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_29531_cast_fp16 = transpose(perm = var_29530, x = q_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 1024, 51, 1]> input_1087_cast_fp16 = expand_dims(axes = input_1087_axes_1, x = var_29531_cast_fp16)[name = string("input_1087_cast_fp16")];
            string dense_output_1857_pad_type_1 = const()[name = string("dense_output_1857_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1857_strides_1 = const()[name = string("dense_output_1857_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1857_pad_1 = const()[name = string("dense_output_1857_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1857_dilations_1 = const()[name = string("dense_output_1857_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1857_groups_1 = const()[name = string("dense_output_1857_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686472704))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686603840))))[name = string("layers_23_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1857_cast_fp16 = conv(dilations = dense_output_1857_dilations_1, groups = dense_output_1857_groups_1, pad = dense_output_1857_pad_1, pad_type = dense_output_1857_pad_type_1, strides = dense_output_1857_strides_1, weight = layers_23_self_attn_linear_q_0_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1857_cast_fp16")];
            string sparse_output_1857_pad_type_1 = const()[name = string("sparse_output_1857_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1857_strides_1 = const()[name = string("sparse_output_1857_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1857_pad_1 = const()[name = string("sparse_output_1857_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1857_dilations_1 = const()[name = string("sparse_output_1857_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1857_groups_1 = const()[name = string("sparse_output_1857_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686607104))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686604416))))[name = string("layers_23_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1857_cast_fp16 = conv(dilations = sparse_output_1857_dilations_1, groups = sparse_output_1857_groups_1, pad = sparse_output_1857_pad_1, pad_type = sparse_output_1857_pad_type_1, strides = sparse_output_1857_strides_1, weight = layers_23_self_attn_linear_q_0_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1857_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29556_cast_fp16 = add(x = dense_output_1857_cast_fp16, y = sparse_output_1857_cast_fp16)[name = string("op_29556_cast_fp16")];
            tensor<int32, [4]> var_29557 = const()[name = string("op_29557"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29559 = const()[name = string("op_29559"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29558_cast_fp16 = transpose(perm = var_29557, x = var_29556_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 51, 128]> q_head_369_cast_fp16 = reshape(shape = var_29559, x = var_29558_cast_fp16)[name = string("q_head_369_cast_fp16")];
            string dense_output_1859_pad_type_1 = const()[name = string("dense_output_1859_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1859_strides_1 = const()[name = string("dense_output_1859_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1859_pad_1 = const()[name = string("dense_output_1859_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1859_dilations_1 = const()[name = string("dense_output_1859_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1859_groups_1 = const()[name = string("dense_output_1859_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686623552))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686754688))))[name = string("layers_23_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1859_cast_fp16 = conv(dilations = dense_output_1859_dilations_1, groups = dense_output_1859_groups_1, pad = dense_output_1859_pad_1, pad_type = dense_output_1859_pad_type_1, strides = dense_output_1859_strides_1, weight = layers_23_self_attn_linear_k_0_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1859_cast_fp16")];
            string sparse_output_1859_pad_type_1 = const()[name = string("sparse_output_1859_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1859_strides_1 = const()[name = string("sparse_output_1859_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1859_pad_1 = const()[name = string("sparse_output_1859_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1859_dilations_1 = const()[name = string("sparse_output_1859_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1859_groups_1 = const()[name = string("sparse_output_1859_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686757952))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686755264))))[name = string("layers_23_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1859_cast_fp16 = conv(dilations = sparse_output_1859_dilations_1, groups = sparse_output_1859_groups_1, pad = sparse_output_1859_pad_1, pad_type = sparse_output_1859_pad_type_1, strides = sparse_output_1859_strides_1, weight = layers_23_self_attn_linear_k_0_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1859_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29575_cast_fp16 = add(x = dense_output_1859_cast_fp16, y = sparse_output_1859_cast_fp16)[name = string("op_29575_cast_fp16")];
            tensor<int32, [4]> var_29576 = const()[name = string("op_29576"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29578 = const()[name = string("op_29578"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29577_cast_fp16 = transpose(perm = var_29576, x = var_29575_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 51, 128]> k_head_737_cast_fp16 = reshape(shape = var_29578, x = var_29577_cast_fp16)[name = string("k_head_737_cast_fp16")];
            string dense_output_1861_pad_type_1 = const()[name = string("dense_output_1861_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1861_strides_1 = const()[name = string("dense_output_1861_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1861_pad_1 = const()[name = string("dense_output_1861_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1861_dilations_1 = const()[name = string("dense_output_1861_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1861_groups_1 = const()[name = string("dense_output_1861_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686774400))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686905536))))[name = string("layers_23_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1861_cast_fp16 = conv(dilations = dense_output_1861_dilations_1, groups = dense_output_1861_groups_1, pad = dense_output_1861_pad_1, pad_type = dense_output_1861_pad_type_1, strides = dense_output_1861_strides_1, weight = layers_23_self_attn_linear_v_0_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1861_cast_fp16")];
            string sparse_output_1861_pad_type_1 = const()[name = string("sparse_output_1861_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1861_strides_1 = const()[name = string("sparse_output_1861_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1861_pad_1 = const()[name = string("sparse_output_1861_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1861_dilations_1 = const()[name = string("sparse_output_1861_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1861_groups_1 = const()[name = string("sparse_output_1861_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686908800))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686906112))))[name = string("layers_23_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1861_cast_fp16 = conv(dilations = sparse_output_1861_dilations_1, groups = sparse_output_1861_groups_1, pad = sparse_output_1861_pad_1, pad_type = sparse_output_1861_pad_type_1, strides = sparse_output_1861_strides_1, weight = layers_23_self_attn_linear_v_0_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1861_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29594_cast_fp16 = add(x = dense_output_1861_cast_fp16, y = sparse_output_1861_cast_fp16)[name = string("op_29594_cast_fp16")];
            tensor<int32, [4]> var_29595 = const()[name = string("op_29595"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29597 = const()[name = string("op_29597"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29596_cast_fp16 = transpose(perm = var_29595, x = var_29594_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 51, 128]> v_head_737_cast_fp16 = reshape(shape = var_29597, x = var_29596_cast_fp16)[name = string("v_head_737_cast_fp16")];
            string dense_output_1863_pad_type_1 = const()[name = string("dense_output_1863_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1863_strides_1 = const()[name = string("dense_output_1863_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1863_pad_1 = const()[name = string("dense_output_1863_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1863_dilations_1 = const()[name = string("dense_output_1863_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1863_groups_1 = const()[name = string("dense_output_1863_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686925248))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687056384))))[name = string("layers_23_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1863_cast_fp16 = conv(dilations = dense_output_1863_dilations_1, groups = dense_output_1863_groups_1, pad = dense_output_1863_pad_1, pad_type = dense_output_1863_pad_type_1, strides = dense_output_1863_strides_1, weight = layers_23_self_attn_linear_pos_0_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1863_cast_fp16")];
            string sparse_output_1863_pad_type_1 = const()[name = string("sparse_output_1863_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1863_strides_1 = const()[name = string("sparse_output_1863_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1863_pad_1 = const()[name = string("sparse_output_1863_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1863_dilations_1 = const()[name = string("sparse_output_1863_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1863_groups_1 = const()[name = string("sparse_output_1863_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687059648))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687056960))))[name = string("layers_23_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1863_cast_fp16 = conv(dilations = sparse_output_1863_dilations_1, groups = sparse_output_1863_groups_1, pad = sparse_output_1863_pad_1, pad_type = sparse_output_1863_pad_type_1, strides = sparse_output_1863_strides_1, weight = layers_23_self_attn_linear_pos_0_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1863_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_29613_cast_fp16 = add(x = dense_output_1863_cast_fp16, y = sparse_output_1863_cast_fp16)[name = string("op_29613_cast_fp16")];
            tensor<int32, [4]> var_29614 = const()[name = string("op_29614"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29616 = const()[name = string("op_29616"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_29615_cast_fp16 = transpose(perm = var_29614, x = var_29613_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 101, 128]> p_head_737_cast_fp16 = reshape(shape = var_29616, x = var_29615_cast_fp16)[name = string("p_head_737_cast_fp16")];
            tensor<fp16, [128]> var_29618_to_fp16 = const()[name = string("op_29618_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687076096)))];
            tensor<fp16, [1, 51, 128]> var_29619_cast_fp16 = add(x = q_head_369_cast_fp16, y = var_29618_to_fp16)[name = string("op_29619_cast_fp16")];
            tensor<int32, [1]> q_u_369_axes_1 = const()[name = string("q_u_369_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_369_cast_fp16 = expand_dims(axes = q_u_369_axes_1, x = var_29619_cast_fp16)[name = string("q_u_369_cast_fp16")];
            tensor<fp16, [128]> var_29621_to_fp16 = const()[name = string("op_29621_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687076416)))];
            tensor<fp16, [1, 51, 128]> var_29622_cast_fp16 = add(x = q_head_369_cast_fp16, y = var_29621_to_fp16)[name = string("op_29622_cast_fp16")];
            tensor<int32, [1]> q_v_369_axes_1 = const()[name = string("q_v_369_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_369_cast_fp16 = expand_dims(axes = q_v_369_axes_1, x = var_29622_cast_fp16)[name = string("q_v_369_cast_fp16")];
            tensor<int32, [1]> k_head_739_axes_1 = const()[name = string("k_head_739_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_739_cast_fp16 = expand_dims(axes = k_head_739_axes_1, x = k_head_737_cast_fp16)[name = string("k_head_739_cast_fp16")];
            tensor<int32, [1]> v_head_739_axes_1 = const()[name = string("v_head_739_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_739_cast_fp16 = expand_dims(axes = v_head_739_axes_1, x = v_head_737_cast_fp16)[name = string("v_head_739_cast_fp16")];
            tensor<int32, [1]> p_head_739_axes_1 = const()[name = string("p_head_739_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_739_cast_fp16 = expand_dims(axes = p_head_739_axes_1, x = p_head_737_cast_fp16)[name = string("p_head_739_cast_fp16")];
            bool var_29628_transpose_x_3 = const()[name = string("op_29628_transpose_x_3"), val = bool(false)];
            bool var_29628_transpose_y_3 = const()[name = string("op_29628_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_29628_cast_fp16 = matmul(transpose_x = var_29628_transpose_x_3, transpose_y = var_29628_transpose_y_3, x = q_u_369_cast_fp16, y = k_head_739_cast_fp16)[name = string("op_29628_cast_fp16")];
            fp16 var_29629_to_fp16 = const()[name = string("op_29629_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_369_cast_fp16 = mul(x = var_29628_cast_fp16, y = var_29629_to_fp16)[name = string("scores_content_369_cast_fp16")];
            bool x_1953_transpose_x_3 = const()[name = string("x_1953_transpose_x_3"), val = bool(false)];
            bool x_1953_transpose_y_3 = const()[name = string("x_1953_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1953_cast_fp16 = matmul(transpose_x = x_1953_transpose_x_3, transpose_y = x_1953_transpose_y_3, x = q_v_369_cast_fp16, y = p_head_739_cast_fp16)[name = string("x_1953_cast_fp16")];
            tensor<int32, [8]> x_1955_pad_1 = const()[name = string("x_1955_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1955_mode_1 = const()[name = string("x_1955_mode_1"), val = string("constant")];
            fp16 const_2551_to_fp16 = const()[name = string("const_2551_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1955_cast_fp16 = pad(constant_val = const_2551_to_fp16, mode = x_1955_mode_1, pad = x_1955_pad_1, x = x_1953_cast_fp16)[name = string("x_1955_cast_fp16")];
            tensor<int32, [4]> var_29643 = const()[name = string("op_29643"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1957_cast_fp16 = reshape(shape = var_29643, x = x_1955_cast_fp16)[name = string("x_1957_cast_fp16")];
            tensor<int32, [4]> var_29647_begin_1 = const()[name = string("op_29647_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_29647_end_1 = const()[name = string("op_29647_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_29647_end_mask_1 = const()[name = string("op_29647_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_29647_cast_fp16 = slice_by_index(begin = var_29647_begin_1, end = var_29647_end_1, end_mask = var_29647_end_mask_1, x = x_1957_cast_fp16)[name = string("op_29647_cast_fp16")];
            tensor<int32, [4]> var_29649 = const()[name = string("op_29649"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_29650_cast_fp16 = reshape(shape = var_29649, x = var_29647_cast_fp16)[name = string("op_29650_cast_fp16")];
            tensor<int32, [4]> var_29655_begin_1 = const()[name = string("op_29655_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_29655_end_1 = const()[name = string("op_29655_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_29655_end_mask_1 = const()[name = string("op_29655_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_29655_cast_fp16 = slice_by_index(begin = var_29655_begin_1, end = var_29655_end_1, end_mask = var_29655_end_mask_1, x = var_29650_cast_fp16)[name = string("op_29655_cast_fp16")];
            fp16 var_29656_to_fp16 = const()[name = string("op_29656_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_369_cast_fp16 = mul(x = var_29655_cast_fp16, y = var_29656_to_fp16)[name = string("scores_pos_369_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_369_cast_fp16 = add(x = scores_content_369_cast_fp16, y = scores_pos_369_cast_fp16)[name = string("logits_369_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_29659_cast_fp16 = softmax(axis = var_29384, x = logits_369_cast_fp16)[name = string("op_29659_cast_fp16")];
            bool var_29661_transpose_x_1 = const()[name = string("op_29661_transpose_x_1"), val = bool(false)];
            bool var_29661_transpose_y_1 = const()[name = string("op_29661_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_29661_cast_fp16 = matmul(transpose_x = var_29661_transpose_x_1, transpose_y = var_29661_transpose_y_1, x = var_29659_cast_fp16, y = v_head_739_cast_fp16)[name = string("op_29661_cast_fp16")];
            tensor<int32, [1]> var_29662_axes_1 = const()[name = string("op_29662_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_29662_cast_fp16 = squeeze(axes = var_29662_axes_1, x = var_29661_cast_fp16)[name = string("op_29662_cast_fp16")];
            string dense_output_1865_pad_type_1 = const()[name = string("dense_output_1865_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1865_strides_1 = const()[name = string("dense_output_1865_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1865_pad_1 = const()[name = string("dense_output_1865_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1865_dilations_1 = const()[name = string("dense_output_1865_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1865_groups_1 = const()[name = string("dense_output_1865_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687076736))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687207872))))[name = string("layers_23_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1865_cast_fp16 = conv(dilations = dense_output_1865_dilations_1, groups = dense_output_1865_groups_1, pad = dense_output_1865_pad_1, pad_type = dense_output_1865_pad_type_1, strides = dense_output_1865_strides_1, weight = layers_23_self_attn_linear_q_1_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1865_cast_fp16")];
            string sparse_output_1865_pad_type_1 = const()[name = string("sparse_output_1865_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1865_strides_1 = const()[name = string("sparse_output_1865_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1865_pad_1 = const()[name = string("sparse_output_1865_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1865_dilations_1 = const()[name = string("sparse_output_1865_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1865_groups_1 = const()[name = string("sparse_output_1865_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687211136))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687208448))))[name = string("layers_23_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1865_cast_fp16 = conv(dilations = sparse_output_1865_dilations_1, groups = sparse_output_1865_groups_1, pad = sparse_output_1865_pad_1, pad_type = sparse_output_1865_pad_type_1, strides = sparse_output_1865_strides_1, weight = layers_23_self_attn_linear_q_1_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1865_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29677_cast_fp16 = add(x = dense_output_1865_cast_fp16, y = sparse_output_1865_cast_fp16)[name = string("op_29677_cast_fp16")];
            tensor<int32, [4]> var_29678 = const()[name = string("op_29678"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29680 = const()[name = string("op_29680"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29679_cast_fp16 = transpose(perm = var_29678, x = var_29677_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 51, 128]> q_head_371_cast_fp16 = reshape(shape = var_29680, x = var_29679_cast_fp16)[name = string("q_head_371_cast_fp16")];
            string dense_output_1867_pad_type_1 = const()[name = string("dense_output_1867_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1867_strides_1 = const()[name = string("dense_output_1867_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1867_pad_1 = const()[name = string("dense_output_1867_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1867_dilations_1 = const()[name = string("dense_output_1867_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1867_groups_1 = const()[name = string("dense_output_1867_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687227584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687358720))))[name = string("layers_23_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1867_cast_fp16 = conv(dilations = dense_output_1867_dilations_1, groups = dense_output_1867_groups_1, pad = dense_output_1867_pad_1, pad_type = dense_output_1867_pad_type_1, strides = dense_output_1867_strides_1, weight = layers_23_self_attn_linear_k_1_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1867_cast_fp16")];
            string sparse_output_1867_pad_type_1 = const()[name = string("sparse_output_1867_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1867_strides_1 = const()[name = string("sparse_output_1867_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1867_pad_1 = const()[name = string("sparse_output_1867_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1867_dilations_1 = const()[name = string("sparse_output_1867_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1867_groups_1 = const()[name = string("sparse_output_1867_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687361984))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687359296))))[name = string("layers_23_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1867_cast_fp16 = conv(dilations = sparse_output_1867_dilations_1, groups = sparse_output_1867_groups_1, pad = sparse_output_1867_pad_1, pad_type = sparse_output_1867_pad_type_1, strides = sparse_output_1867_strides_1, weight = layers_23_self_attn_linear_k_1_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1867_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29696_cast_fp16 = add(x = dense_output_1867_cast_fp16, y = sparse_output_1867_cast_fp16)[name = string("op_29696_cast_fp16")];
            tensor<int32, [4]> var_29697 = const()[name = string("op_29697"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29699 = const()[name = string("op_29699"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29698_cast_fp16 = transpose(perm = var_29697, x = var_29696_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 51, 128]> k_head_741_cast_fp16 = reshape(shape = var_29699, x = var_29698_cast_fp16)[name = string("k_head_741_cast_fp16")];
            string dense_output_1869_pad_type_1 = const()[name = string("dense_output_1869_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1869_strides_1 = const()[name = string("dense_output_1869_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1869_pad_1 = const()[name = string("dense_output_1869_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1869_dilations_1 = const()[name = string("dense_output_1869_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1869_groups_1 = const()[name = string("dense_output_1869_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687378432))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687509568))))[name = string("layers_23_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1869_cast_fp16 = conv(dilations = dense_output_1869_dilations_1, groups = dense_output_1869_groups_1, pad = dense_output_1869_pad_1, pad_type = dense_output_1869_pad_type_1, strides = dense_output_1869_strides_1, weight = layers_23_self_attn_linear_v_1_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1869_cast_fp16")];
            string sparse_output_1869_pad_type_1 = const()[name = string("sparse_output_1869_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1869_strides_1 = const()[name = string("sparse_output_1869_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1869_pad_1 = const()[name = string("sparse_output_1869_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1869_dilations_1 = const()[name = string("sparse_output_1869_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1869_groups_1 = const()[name = string("sparse_output_1869_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687512832))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687510144))))[name = string("layers_23_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1869_cast_fp16 = conv(dilations = sparse_output_1869_dilations_1, groups = sparse_output_1869_groups_1, pad = sparse_output_1869_pad_1, pad_type = sparse_output_1869_pad_type_1, strides = sparse_output_1869_strides_1, weight = layers_23_self_attn_linear_v_1_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1869_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29715_cast_fp16 = add(x = dense_output_1869_cast_fp16, y = sparse_output_1869_cast_fp16)[name = string("op_29715_cast_fp16")];
            tensor<int32, [4]> var_29716 = const()[name = string("op_29716"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29718 = const()[name = string("op_29718"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29717_cast_fp16 = transpose(perm = var_29716, x = var_29715_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 51, 128]> v_head_741_cast_fp16 = reshape(shape = var_29718, x = var_29717_cast_fp16)[name = string("v_head_741_cast_fp16")];
            string dense_output_1871_pad_type_1 = const()[name = string("dense_output_1871_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1871_strides_1 = const()[name = string("dense_output_1871_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1871_pad_1 = const()[name = string("dense_output_1871_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1871_dilations_1 = const()[name = string("dense_output_1871_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1871_groups_1 = const()[name = string("dense_output_1871_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687529280))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687660416))))[name = string("layers_23_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1871_cast_fp16 = conv(dilations = dense_output_1871_dilations_1, groups = dense_output_1871_groups_1, pad = dense_output_1871_pad_1, pad_type = dense_output_1871_pad_type_1, strides = dense_output_1871_strides_1, weight = layers_23_self_attn_linear_pos_1_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1871_cast_fp16")];
            string sparse_output_1871_pad_type_1 = const()[name = string("sparse_output_1871_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1871_strides_1 = const()[name = string("sparse_output_1871_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1871_pad_1 = const()[name = string("sparse_output_1871_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1871_dilations_1 = const()[name = string("sparse_output_1871_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1871_groups_1 = const()[name = string("sparse_output_1871_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687663680))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687660992))))[name = string("layers_23_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1871_cast_fp16 = conv(dilations = sparse_output_1871_dilations_1, groups = sparse_output_1871_groups_1, pad = sparse_output_1871_pad_1, pad_type = sparse_output_1871_pad_type_1, strides = sparse_output_1871_strides_1, weight = layers_23_self_attn_linear_pos_1_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1871_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_29734_cast_fp16 = add(x = dense_output_1871_cast_fp16, y = sparse_output_1871_cast_fp16)[name = string("op_29734_cast_fp16")];
            tensor<int32, [4]> var_29735 = const()[name = string("op_29735"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29737 = const()[name = string("op_29737"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_29736_cast_fp16 = transpose(perm = var_29735, x = var_29734_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 101, 128]> p_head_741_cast_fp16 = reshape(shape = var_29737, x = var_29736_cast_fp16)[name = string("p_head_741_cast_fp16")];
            tensor<fp16, [128]> var_29739_to_fp16 = const()[name = string("op_29739_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687680128)))];
            tensor<fp16, [1, 51, 128]> var_29740_cast_fp16 = add(x = q_head_371_cast_fp16, y = var_29739_to_fp16)[name = string("op_29740_cast_fp16")];
            tensor<int32, [1]> q_u_371_axes_1 = const()[name = string("q_u_371_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_371_cast_fp16 = expand_dims(axes = q_u_371_axes_1, x = var_29740_cast_fp16)[name = string("q_u_371_cast_fp16")];
            tensor<fp16, [128]> var_29742_to_fp16 = const()[name = string("op_29742_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687680448)))];
            tensor<fp16, [1, 51, 128]> var_29743_cast_fp16 = add(x = q_head_371_cast_fp16, y = var_29742_to_fp16)[name = string("op_29743_cast_fp16")];
            tensor<int32, [1]> q_v_371_axes_1 = const()[name = string("q_v_371_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_371_cast_fp16 = expand_dims(axes = q_v_371_axes_1, x = var_29743_cast_fp16)[name = string("q_v_371_cast_fp16")];
            tensor<int32, [1]> k_head_743_axes_1 = const()[name = string("k_head_743_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_743_cast_fp16 = expand_dims(axes = k_head_743_axes_1, x = k_head_741_cast_fp16)[name = string("k_head_743_cast_fp16")];
            tensor<int32, [1]> v_head_743_axes_1 = const()[name = string("v_head_743_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_743_cast_fp16 = expand_dims(axes = v_head_743_axes_1, x = v_head_741_cast_fp16)[name = string("v_head_743_cast_fp16")];
            tensor<int32, [1]> p_head_743_axes_1 = const()[name = string("p_head_743_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_743_cast_fp16 = expand_dims(axes = p_head_743_axes_1, x = p_head_741_cast_fp16)[name = string("p_head_743_cast_fp16")];
            bool var_29749_transpose_x_3 = const()[name = string("op_29749_transpose_x_3"), val = bool(false)];
            bool var_29749_transpose_y_3 = const()[name = string("op_29749_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_29749_cast_fp16 = matmul(transpose_x = var_29749_transpose_x_3, transpose_y = var_29749_transpose_y_3, x = q_u_371_cast_fp16, y = k_head_743_cast_fp16)[name = string("op_29749_cast_fp16")];
            fp16 var_29750_to_fp16 = const()[name = string("op_29750_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_371_cast_fp16 = mul(x = var_29749_cast_fp16, y = var_29750_to_fp16)[name = string("scores_content_371_cast_fp16")];
            bool x_1961_transpose_x_3 = const()[name = string("x_1961_transpose_x_3"), val = bool(false)];
            bool x_1961_transpose_y_3 = const()[name = string("x_1961_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1961_cast_fp16 = matmul(transpose_x = x_1961_transpose_x_3, transpose_y = x_1961_transpose_y_3, x = q_v_371_cast_fp16, y = p_head_743_cast_fp16)[name = string("x_1961_cast_fp16")];
            tensor<int32, [8]> x_1963_pad_1 = const()[name = string("x_1963_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1963_mode_1 = const()[name = string("x_1963_mode_1"), val = string("constant")];
            fp16 const_2557_to_fp16 = const()[name = string("const_2557_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1963_cast_fp16 = pad(constant_val = const_2557_to_fp16, mode = x_1963_mode_1, pad = x_1963_pad_1, x = x_1961_cast_fp16)[name = string("x_1963_cast_fp16")];
            tensor<int32, [4]> var_29764 = const()[name = string("op_29764"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1965_cast_fp16 = reshape(shape = var_29764, x = x_1963_cast_fp16)[name = string("x_1965_cast_fp16")];
            tensor<int32, [4]> var_29768_begin_1 = const()[name = string("op_29768_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_29768_end_1 = const()[name = string("op_29768_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_29768_end_mask_1 = const()[name = string("op_29768_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_29768_cast_fp16 = slice_by_index(begin = var_29768_begin_1, end = var_29768_end_1, end_mask = var_29768_end_mask_1, x = x_1965_cast_fp16)[name = string("op_29768_cast_fp16")];
            tensor<int32, [4]> var_29770 = const()[name = string("op_29770"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_29771_cast_fp16 = reshape(shape = var_29770, x = var_29768_cast_fp16)[name = string("op_29771_cast_fp16")];
            tensor<int32, [4]> var_29776_begin_1 = const()[name = string("op_29776_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_29776_end_1 = const()[name = string("op_29776_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_29776_end_mask_1 = const()[name = string("op_29776_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_29776_cast_fp16 = slice_by_index(begin = var_29776_begin_1, end = var_29776_end_1, end_mask = var_29776_end_mask_1, x = var_29771_cast_fp16)[name = string("op_29776_cast_fp16")];
            fp16 var_29777_to_fp16 = const()[name = string("op_29777_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_371_cast_fp16 = mul(x = var_29776_cast_fp16, y = var_29777_to_fp16)[name = string("scores_pos_371_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_371_cast_fp16 = add(x = scores_content_371_cast_fp16, y = scores_pos_371_cast_fp16)[name = string("logits_371_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_29780_cast_fp16 = softmax(axis = var_29384, x = logits_371_cast_fp16)[name = string("op_29780_cast_fp16")];
            bool var_29782_transpose_x_1 = const()[name = string("op_29782_transpose_x_1"), val = bool(false)];
            bool var_29782_transpose_y_1 = const()[name = string("op_29782_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_29782_cast_fp16 = matmul(transpose_x = var_29782_transpose_x_1, transpose_y = var_29782_transpose_y_1, x = var_29780_cast_fp16, y = v_head_743_cast_fp16)[name = string("op_29782_cast_fp16")];
            tensor<int32, [1]> var_29783_axes_1 = const()[name = string("op_29783_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_29783_cast_fp16 = squeeze(axes = var_29783_axes_1, x = var_29782_cast_fp16)[name = string("op_29783_cast_fp16")];
            string dense_output_1873_pad_type_1 = const()[name = string("dense_output_1873_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1873_strides_1 = const()[name = string("dense_output_1873_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1873_pad_1 = const()[name = string("dense_output_1873_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1873_dilations_1 = const()[name = string("dense_output_1873_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1873_groups_1 = const()[name = string("dense_output_1873_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687680768))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687811904))))[name = string("layers_23_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1873_cast_fp16 = conv(dilations = dense_output_1873_dilations_1, groups = dense_output_1873_groups_1, pad = dense_output_1873_pad_1, pad_type = dense_output_1873_pad_type_1, strides = dense_output_1873_strides_1, weight = layers_23_self_attn_linear_q_2_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1873_cast_fp16")];
            string sparse_output_1873_pad_type_1 = const()[name = string("sparse_output_1873_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1873_strides_1 = const()[name = string("sparse_output_1873_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1873_pad_1 = const()[name = string("sparse_output_1873_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1873_dilations_1 = const()[name = string("sparse_output_1873_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1873_groups_1 = const()[name = string("sparse_output_1873_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687815168))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687812480))))[name = string("layers_23_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1873_cast_fp16 = conv(dilations = sparse_output_1873_dilations_1, groups = sparse_output_1873_groups_1, pad = sparse_output_1873_pad_1, pad_type = sparse_output_1873_pad_type_1, strides = sparse_output_1873_strides_1, weight = layers_23_self_attn_linear_q_2_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1873_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29798_cast_fp16 = add(x = dense_output_1873_cast_fp16, y = sparse_output_1873_cast_fp16)[name = string("op_29798_cast_fp16")];
            tensor<int32, [4]> var_29799 = const()[name = string("op_29799"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29801 = const()[name = string("op_29801"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29800_cast_fp16 = transpose(perm = var_29799, x = var_29798_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 51, 128]> q_head_373_cast_fp16 = reshape(shape = var_29801, x = var_29800_cast_fp16)[name = string("q_head_373_cast_fp16")];
            string dense_output_1875_pad_type_1 = const()[name = string("dense_output_1875_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1875_strides_1 = const()[name = string("dense_output_1875_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1875_pad_1 = const()[name = string("dense_output_1875_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1875_dilations_1 = const()[name = string("dense_output_1875_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1875_groups_1 = const()[name = string("dense_output_1875_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687831616))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687962752))))[name = string("layers_23_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1875_cast_fp16 = conv(dilations = dense_output_1875_dilations_1, groups = dense_output_1875_groups_1, pad = dense_output_1875_pad_1, pad_type = dense_output_1875_pad_type_1, strides = dense_output_1875_strides_1, weight = layers_23_self_attn_linear_k_2_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1875_cast_fp16")];
            string sparse_output_1875_pad_type_1 = const()[name = string("sparse_output_1875_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1875_strides_1 = const()[name = string("sparse_output_1875_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1875_pad_1 = const()[name = string("sparse_output_1875_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1875_dilations_1 = const()[name = string("sparse_output_1875_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1875_groups_1 = const()[name = string("sparse_output_1875_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687966016))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687963328))))[name = string("layers_23_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1875_cast_fp16 = conv(dilations = sparse_output_1875_dilations_1, groups = sparse_output_1875_groups_1, pad = sparse_output_1875_pad_1, pad_type = sparse_output_1875_pad_type_1, strides = sparse_output_1875_strides_1, weight = layers_23_self_attn_linear_k_2_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1875_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29817_cast_fp16 = add(x = dense_output_1875_cast_fp16, y = sparse_output_1875_cast_fp16)[name = string("op_29817_cast_fp16")];
            tensor<int32, [4]> var_29818 = const()[name = string("op_29818"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29820 = const()[name = string("op_29820"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29819_cast_fp16 = transpose(perm = var_29818, x = var_29817_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 51, 128]> k_head_745_cast_fp16 = reshape(shape = var_29820, x = var_29819_cast_fp16)[name = string("k_head_745_cast_fp16")];
            string dense_output_1877_pad_type_1 = const()[name = string("dense_output_1877_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1877_strides_1 = const()[name = string("dense_output_1877_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1877_pad_1 = const()[name = string("dense_output_1877_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1877_dilations_1 = const()[name = string("dense_output_1877_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1877_groups_1 = const()[name = string("dense_output_1877_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687982464))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688113600))))[name = string("layers_23_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1877_cast_fp16 = conv(dilations = dense_output_1877_dilations_1, groups = dense_output_1877_groups_1, pad = dense_output_1877_pad_1, pad_type = dense_output_1877_pad_type_1, strides = dense_output_1877_strides_1, weight = layers_23_self_attn_linear_v_2_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1877_cast_fp16")];
            string sparse_output_1877_pad_type_1 = const()[name = string("sparse_output_1877_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1877_strides_1 = const()[name = string("sparse_output_1877_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1877_pad_1 = const()[name = string("sparse_output_1877_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1877_dilations_1 = const()[name = string("sparse_output_1877_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1877_groups_1 = const()[name = string("sparse_output_1877_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688116864))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688114176))))[name = string("layers_23_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1877_cast_fp16 = conv(dilations = sparse_output_1877_dilations_1, groups = sparse_output_1877_groups_1, pad = sparse_output_1877_pad_1, pad_type = sparse_output_1877_pad_type_1, strides = sparse_output_1877_strides_1, weight = layers_23_self_attn_linear_v_2_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1877_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29836_cast_fp16 = add(x = dense_output_1877_cast_fp16, y = sparse_output_1877_cast_fp16)[name = string("op_29836_cast_fp16")];
            tensor<int32, [4]> var_29837 = const()[name = string("op_29837"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29839 = const()[name = string("op_29839"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29838_cast_fp16 = transpose(perm = var_29837, x = var_29836_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 51, 128]> v_head_745_cast_fp16 = reshape(shape = var_29839, x = var_29838_cast_fp16)[name = string("v_head_745_cast_fp16")];
            string dense_output_1879_pad_type_1 = const()[name = string("dense_output_1879_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1879_strides_1 = const()[name = string("dense_output_1879_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1879_pad_1 = const()[name = string("dense_output_1879_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1879_dilations_1 = const()[name = string("dense_output_1879_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1879_groups_1 = const()[name = string("dense_output_1879_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688133312))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688264448))))[name = string("layers_23_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1879_cast_fp16 = conv(dilations = dense_output_1879_dilations_1, groups = dense_output_1879_groups_1, pad = dense_output_1879_pad_1, pad_type = dense_output_1879_pad_type_1, strides = dense_output_1879_strides_1, weight = layers_23_self_attn_linear_pos_2_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1879_cast_fp16")];
            string sparse_output_1879_pad_type_1 = const()[name = string("sparse_output_1879_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1879_strides_1 = const()[name = string("sparse_output_1879_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1879_pad_1 = const()[name = string("sparse_output_1879_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1879_dilations_1 = const()[name = string("sparse_output_1879_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1879_groups_1 = const()[name = string("sparse_output_1879_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688267712))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688265024))))[name = string("layers_23_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1879_cast_fp16 = conv(dilations = sparse_output_1879_dilations_1, groups = sparse_output_1879_groups_1, pad = sparse_output_1879_pad_1, pad_type = sparse_output_1879_pad_type_1, strides = sparse_output_1879_strides_1, weight = layers_23_self_attn_linear_pos_2_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1879_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_29855_cast_fp16 = add(x = dense_output_1879_cast_fp16, y = sparse_output_1879_cast_fp16)[name = string("op_29855_cast_fp16")];
            tensor<int32, [4]> var_29856 = const()[name = string("op_29856"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29858 = const()[name = string("op_29858"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_29857_cast_fp16 = transpose(perm = var_29856, x = var_29855_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 101, 128]> p_head_745_cast_fp16 = reshape(shape = var_29858, x = var_29857_cast_fp16)[name = string("p_head_745_cast_fp16")];
            tensor<fp16, [128]> var_29860_to_fp16 = const()[name = string("op_29860_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688284160)))];
            tensor<fp16, [1, 51, 128]> var_29861_cast_fp16 = add(x = q_head_373_cast_fp16, y = var_29860_to_fp16)[name = string("op_29861_cast_fp16")];
            tensor<int32, [1]> q_u_373_axes_1 = const()[name = string("q_u_373_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_373_cast_fp16 = expand_dims(axes = q_u_373_axes_1, x = var_29861_cast_fp16)[name = string("q_u_373_cast_fp16")];
            tensor<fp16, [128]> var_29863_to_fp16 = const()[name = string("op_29863_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688284480)))];
            tensor<fp16, [1, 51, 128]> var_29864_cast_fp16 = add(x = q_head_373_cast_fp16, y = var_29863_to_fp16)[name = string("op_29864_cast_fp16")];
            tensor<int32, [1]> q_v_373_axes_1 = const()[name = string("q_v_373_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_373_cast_fp16 = expand_dims(axes = q_v_373_axes_1, x = var_29864_cast_fp16)[name = string("q_v_373_cast_fp16")];
            tensor<int32, [1]> k_head_747_axes_1 = const()[name = string("k_head_747_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_747_cast_fp16 = expand_dims(axes = k_head_747_axes_1, x = k_head_745_cast_fp16)[name = string("k_head_747_cast_fp16")];
            tensor<int32, [1]> v_head_747_axes_1 = const()[name = string("v_head_747_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_747_cast_fp16 = expand_dims(axes = v_head_747_axes_1, x = v_head_745_cast_fp16)[name = string("v_head_747_cast_fp16")];
            tensor<int32, [1]> p_head_747_axes_1 = const()[name = string("p_head_747_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_747_cast_fp16 = expand_dims(axes = p_head_747_axes_1, x = p_head_745_cast_fp16)[name = string("p_head_747_cast_fp16")];
            bool var_29870_transpose_x_3 = const()[name = string("op_29870_transpose_x_3"), val = bool(false)];
            bool var_29870_transpose_y_3 = const()[name = string("op_29870_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_29870_cast_fp16 = matmul(transpose_x = var_29870_transpose_x_3, transpose_y = var_29870_transpose_y_3, x = q_u_373_cast_fp16, y = k_head_747_cast_fp16)[name = string("op_29870_cast_fp16")];
            fp16 var_29871_to_fp16 = const()[name = string("op_29871_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_373_cast_fp16 = mul(x = var_29870_cast_fp16, y = var_29871_to_fp16)[name = string("scores_content_373_cast_fp16")];
            bool x_1969_transpose_x_3 = const()[name = string("x_1969_transpose_x_3"), val = bool(false)];
            bool x_1969_transpose_y_3 = const()[name = string("x_1969_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1969_cast_fp16 = matmul(transpose_x = x_1969_transpose_x_3, transpose_y = x_1969_transpose_y_3, x = q_v_373_cast_fp16, y = p_head_747_cast_fp16)[name = string("x_1969_cast_fp16")];
            tensor<int32, [8]> x_1971_pad_1 = const()[name = string("x_1971_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1971_mode_1 = const()[name = string("x_1971_mode_1"), val = string("constant")];
            fp16 const_2563_to_fp16 = const()[name = string("const_2563_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1971_cast_fp16 = pad(constant_val = const_2563_to_fp16, mode = x_1971_mode_1, pad = x_1971_pad_1, x = x_1969_cast_fp16)[name = string("x_1971_cast_fp16")];
            tensor<int32, [4]> var_29885 = const()[name = string("op_29885"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1973_cast_fp16 = reshape(shape = var_29885, x = x_1971_cast_fp16)[name = string("x_1973_cast_fp16")];
            tensor<int32, [4]> var_29889_begin_1 = const()[name = string("op_29889_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_29889_end_1 = const()[name = string("op_29889_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_29889_end_mask_1 = const()[name = string("op_29889_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_29889_cast_fp16 = slice_by_index(begin = var_29889_begin_1, end = var_29889_end_1, end_mask = var_29889_end_mask_1, x = x_1973_cast_fp16)[name = string("op_29889_cast_fp16")];
            tensor<int32, [4]> var_29891 = const()[name = string("op_29891"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_29892_cast_fp16 = reshape(shape = var_29891, x = var_29889_cast_fp16)[name = string("op_29892_cast_fp16")];
            tensor<int32, [4]> var_29897_begin_1 = const()[name = string("op_29897_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_29897_end_1 = const()[name = string("op_29897_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_29897_end_mask_1 = const()[name = string("op_29897_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_29897_cast_fp16 = slice_by_index(begin = var_29897_begin_1, end = var_29897_end_1, end_mask = var_29897_end_mask_1, x = var_29892_cast_fp16)[name = string("op_29897_cast_fp16")];
            fp16 var_29898_to_fp16 = const()[name = string("op_29898_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_373_cast_fp16 = mul(x = var_29897_cast_fp16, y = var_29898_to_fp16)[name = string("scores_pos_373_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_373_cast_fp16 = add(x = scores_content_373_cast_fp16, y = scores_pos_373_cast_fp16)[name = string("logits_373_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_29901_cast_fp16 = softmax(axis = var_29384, x = logits_373_cast_fp16)[name = string("op_29901_cast_fp16")];
            bool var_29903_transpose_x_1 = const()[name = string("op_29903_transpose_x_1"), val = bool(false)];
            bool var_29903_transpose_y_1 = const()[name = string("op_29903_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_29903_cast_fp16 = matmul(transpose_x = var_29903_transpose_x_1, transpose_y = var_29903_transpose_y_1, x = var_29901_cast_fp16, y = v_head_747_cast_fp16)[name = string("op_29903_cast_fp16")];
            tensor<int32, [1]> var_29904_axes_1 = const()[name = string("op_29904_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_29904_cast_fp16 = squeeze(axes = var_29904_axes_1, x = var_29903_cast_fp16)[name = string("op_29904_cast_fp16")];
            string dense_output_1881_pad_type_1 = const()[name = string("dense_output_1881_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1881_strides_1 = const()[name = string("dense_output_1881_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1881_pad_1 = const()[name = string("dense_output_1881_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1881_dilations_1 = const()[name = string("dense_output_1881_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1881_groups_1 = const()[name = string("dense_output_1881_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688284800))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688415936))))[name = string("layers_23_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1881_cast_fp16 = conv(dilations = dense_output_1881_dilations_1, groups = dense_output_1881_groups_1, pad = dense_output_1881_pad_1, pad_type = dense_output_1881_pad_type_1, strides = dense_output_1881_strides_1, weight = layers_23_self_attn_linear_q_3_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1881_cast_fp16")];
            string sparse_output_1881_pad_type_1 = const()[name = string("sparse_output_1881_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1881_strides_1 = const()[name = string("sparse_output_1881_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1881_pad_1 = const()[name = string("sparse_output_1881_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1881_dilations_1 = const()[name = string("sparse_output_1881_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1881_groups_1 = const()[name = string("sparse_output_1881_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688419200))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688416512))))[name = string("layers_23_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1881_cast_fp16 = conv(dilations = sparse_output_1881_dilations_1, groups = sparse_output_1881_groups_1, pad = sparse_output_1881_pad_1, pad_type = sparse_output_1881_pad_type_1, strides = sparse_output_1881_strides_1, weight = layers_23_self_attn_linear_q_3_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1881_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29919_cast_fp16 = add(x = dense_output_1881_cast_fp16, y = sparse_output_1881_cast_fp16)[name = string("op_29919_cast_fp16")];
            tensor<int32, [4]> var_29920 = const()[name = string("op_29920"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29922 = const()[name = string("op_29922"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29921_cast_fp16 = transpose(perm = var_29920, x = var_29919_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 51, 128]> q_head_375_cast_fp16 = reshape(shape = var_29922, x = var_29921_cast_fp16)[name = string("q_head_375_cast_fp16")];
            string dense_output_1883_pad_type_1 = const()[name = string("dense_output_1883_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1883_strides_1 = const()[name = string("dense_output_1883_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1883_pad_1 = const()[name = string("dense_output_1883_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1883_dilations_1 = const()[name = string("dense_output_1883_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1883_groups_1 = const()[name = string("dense_output_1883_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688435648))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688566784))))[name = string("layers_23_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1883_cast_fp16 = conv(dilations = dense_output_1883_dilations_1, groups = dense_output_1883_groups_1, pad = dense_output_1883_pad_1, pad_type = dense_output_1883_pad_type_1, strides = dense_output_1883_strides_1, weight = layers_23_self_attn_linear_k_3_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1883_cast_fp16")];
            string sparse_output_1883_pad_type_1 = const()[name = string("sparse_output_1883_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1883_strides_1 = const()[name = string("sparse_output_1883_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1883_pad_1 = const()[name = string("sparse_output_1883_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1883_dilations_1 = const()[name = string("sparse_output_1883_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1883_groups_1 = const()[name = string("sparse_output_1883_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688570048))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688567360))))[name = string("layers_23_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1883_cast_fp16 = conv(dilations = sparse_output_1883_dilations_1, groups = sparse_output_1883_groups_1, pad = sparse_output_1883_pad_1, pad_type = sparse_output_1883_pad_type_1, strides = sparse_output_1883_strides_1, weight = layers_23_self_attn_linear_k_3_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1883_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29938_cast_fp16 = add(x = dense_output_1883_cast_fp16, y = sparse_output_1883_cast_fp16)[name = string("op_29938_cast_fp16")];
            tensor<int32, [4]> var_29939 = const()[name = string("op_29939"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29941 = const()[name = string("op_29941"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29940_cast_fp16 = transpose(perm = var_29939, x = var_29938_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 51, 128]> k_head_749_cast_fp16 = reshape(shape = var_29941, x = var_29940_cast_fp16)[name = string("k_head_749_cast_fp16")];
            string dense_output_1885_pad_type_1 = const()[name = string("dense_output_1885_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1885_strides_1 = const()[name = string("dense_output_1885_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1885_pad_1 = const()[name = string("dense_output_1885_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1885_dilations_1 = const()[name = string("dense_output_1885_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1885_groups_1 = const()[name = string("dense_output_1885_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688586496))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688717632))))[name = string("layers_23_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1885_cast_fp16 = conv(dilations = dense_output_1885_dilations_1, groups = dense_output_1885_groups_1, pad = dense_output_1885_pad_1, pad_type = dense_output_1885_pad_type_1, strides = dense_output_1885_strides_1, weight = layers_23_self_attn_linear_v_3_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1885_cast_fp16")];
            string sparse_output_1885_pad_type_1 = const()[name = string("sparse_output_1885_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1885_strides_1 = const()[name = string("sparse_output_1885_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1885_pad_1 = const()[name = string("sparse_output_1885_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1885_dilations_1 = const()[name = string("sparse_output_1885_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1885_groups_1 = const()[name = string("sparse_output_1885_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688720896))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688718208))))[name = string("layers_23_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1885_cast_fp16 = conv(dilations = sparse_output_1885_dilations_1, groups = sparse_output_1885_groups_1, pad = sparse_output_1885_pad_1, pad_type = sparse_output_1885_pad_type_1, strides = sparse_output_1885_strides_1, weight = layers_23_self_attn_linear_v_3_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1885_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_29957_cast_fp16 = add(x = dense_output_1885_cast_fp16, y = sparse_output_1885_cast_fp16)[name = string("op_29957_cast_fp16")];
            tensor<int32, [4]> var_29958 = const()[name = string("op_29958"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29960 = const()[name = string("op_29960"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_29959_cast_fp16 = transpose(perm = var_29958, x = var_29957_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 51, 128]> v_head_749_cast_fp16 = reshape(shape = var_29960, x = var_29959_cast_fp16)[name = string("v_head_749_cast_fp16")];
            string dense_output_1887_pad_type_1 = const()[name = string("dense_output_1887_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1887_strides_1 = const()[name = string("dense_output_1887_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1887_pad_1 = const()[name = string("dense_output_1887_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1887_dilations_1 = const()[name = string("dense_output_1887_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1887_groups_1 = const()[name = string("dense_output_1887_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688737344))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688868480))))[name = string("layers_23_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1887_cast_fp16 = conv(dilations = dense_output_1887_dilations_1, groups = dense_output_1887_groups_1, pad = dense_output_1887_pad_1, pad_type = dense_output_1887_pad_type_1, strides = dense_output_1887_strides_1, weight = layers_23_self_attn_linear_pos_3_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1887_cast_fp16")];
            string sparse_output_1887_pad_type_1 = const()[name = string("sparse_output_1887_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1887_strides_1 = const()[name = string("sparse_output_1887_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1887_pad_1 = const()[name = string("sparse_output_1887_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1887_dilations_1 = const()[name = string("sparse_output_1887_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1887_groups_1 = const()[name = string("sparse_output_1887_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688871744))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688869056))))[name = string("layers_23_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1887_cast_fp16 = conv(dilations = sparse_output_1887_dilations_1, groups = sparse_output_1887_groups_1, pad = sparse_output_1887_pad_1, pad_type = sparse_output_1887_pad_type_1, strides = sparse_output_1887_strides_1, weight = layers_23_self_attn_linear_pos_3_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1887_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_29976_cast_fp16 = add(x = dense_output_1887_cast_fp16, y = sparse_output_1887_cast_fp16)[name = string("op_29976_cast_fp16")];
            tensor<int32, [4]> var_29977 = const()[name = string("op_29977"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_29979 = const()[name = string("op_29979"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_29978_cast_fp16 = transpose(perm = var_29977, x = var_29976_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 101, 128]> p_head_749_cast_fp16 = reshape(shape = var_29979, x = var_29978_cast_fp16)[name = string("p_head_749_cast_fp16")];
            tensor<fp16, [128]> var_29981_to_fp16 = const()[name = string("op_29981_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688888192)))];
            tensor<fp16, [1, 51, 128]> var_29982_cast_fp16 = add(x = q_head_375_cast_fp16, y = var_29981_to_fp16)[name = string("op_29982_cast_fp16")];
            tensor<int32, [1]> q_u_375_axes_1 = const()[name = string("q_u_375_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_375_cast_fp16 = expand_dims(axes = q_u_375_axes_1, x = var_29982_cast_fp16)[name = string("q_u_375_cast_fp16")];
            tensor<fp16, [128]> var_29984_to_fp16 = const()[name = string("op_29984_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688888512)))];
            tensor<fp16, [1, 51, 128]> var_29985_cast_fp16 = add(x = q_head_375_cast_fp16, y = var_29984_to_fp16)[name = string("op_29985_cast_fp16")];
            tensor<int32, [1]> q_v_375_axes_1 = const()[name = string("q_v_375_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_375_cast_fp16 = expand_dims(axes = q_v_375_axes_1, x = var_29985_cast_fp16)[name = string("q_v_375_cast_fp16")];
            tensor<int32, [1]> k_head_751_axes_1 = const()[name = string("k_head_751_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_751_cast_fp16 = expand_dims(axes = k_head_751_axes_1, x = k_head_749_cast_fp16)[name = string("k_head_751_cast_fp16")];
            tensor<int32, [1]> v_head_751_axes_1 = const()[name = string("v_head_751_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_751_cast_fp16 = expand_dims(axes = v_head_751_axes_1, x = v_head_749_cast_fp16)[name = string("v_head_751_cast_fp16")];
            tensor<int32, [1]> p_head_751_axes_1 = const()[name = string("p_head_751_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_751_cast_fp16 = expand_dims(axes = p_head_751_axes_1, x = p_head_749_cast_fp16)[name = string("p_head_751_cast_fp16")];
            bool var_29991_transpose_x_3 = const()[name = string("op_29991_transpose_x_3"), val = bool(false)];
            bool var_29991_transpose_y_3 = const()[name = string("op_29991_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_29991_cast_fp16 = matmul(transpose_x = var_29991_transpose_x_3, transpose_y = var_29991_transpose_y_3, x = q_u_375_cast_fp16, y = k_head_751_cast_fp16)[name = string("op_29991_cast_fp16")];
            fp16 var_29992_to_fp16 = const()[name = string("op_29992_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_375_cast_fp16 = mul(x = var_29991_cast_fp16, y = var_29992_to_fp16)[name = string("scores_content_375_cast_fp16")];
            bool x_1977_transpose_x_3 = const()[name = string("x_1977_transpose_x_3"), val = bool(false)];
            bool x_1977_transpose_y_3 = const()[name = string("x_1977_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1977_cast_fp16 = matmul(transpose_x = x_1977_transpose_x_3, transpose_y = x_1977_transpose_y_3, x = q_v_375_cast_fp16, y = p_head_751_cast_fp16)[name = string("x_1977_cast_fp16")];
            tensor<int32, [8]> x_1979_pad_1 = const()[name = string("x_1979_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1979_mode_1 = const()[name = string("x_1979_mode_1"), val = string("constant")];
            fp16 const_2569_to_fp16 = const()[name = string("const_2569_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1979_cast_fp16 = pad(constant_val = const_2569_to_fp16, mode = x_1979_mode_1, pad = x_1979_pad_1, x = x_1977_cast_fp16)[name = string("x_1979_cast_fp16")];
            tensor<int32, [4]> var_30006 = const()[name = string("op_30006"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1981_cast_fp16 = reshape(shape = var_30006, x = x_1979_cast_fp16)[name = string("x_1981_cast_fp16")];
            tensor<int32, [4]> var_30010_begin_1 = const()[name = string("op_30010_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_30010_end_1 = const()[name = string("op_30010_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_30010_end_mask_1 = const()[name = string("op_30010_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_30010_cast_fp16 = slice_by_index(begin = var_30010_begin_1, end = var_30010_end_1, end_mask = var_30010_end_mask_1, x = x_1981_cast_fp16)[name = string("op_30010_cast_fp16")];
            tensor<int32, [4]> var_30012 = const()[name = string("op_30012"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_30013_cast_fp16 = reshape(shape = var_30012, x = var_30010_cast_fp16)[name = string("op_30013_cast_fp16")];
            tensor<int32, [4]> var_30018_begin_1 = const()[name = string("op_30018_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_30018_end_1 = const()[name = string("op_30018_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_30018_end_mask_1 = const()[name = string("op_30018_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_30018_cast_fp16 = slice_by_index(begin = var_30018_begin_1, end = var_30018_end_1, end_mask = var_30018_end_mask_1, x = var_30013_cast_fp16)[name = string("op_30018_cast_fp16")];
            fp16 var_30019_to_fp16 = const()[name = string("op_30019_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_375_cast_fp16 = mul(x = var_30018_cast_fp16, y = var_30019_to_fp16)[name = string("scores_pos_375_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_375_cast_fp16 = add(x = scores_content_375_cast_fp16, y = scores_pos_375_cast_fp16)[name = string("logits_375_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_30022_cast_fp16 = softmax(axis = var_29384, x = logits_375_cast_fp16)[name = string("op_30022_cast_fp16")];
            bool var_30024_transpose_x_1 = const()[name = string("op_30024_transpose_x_1"), val = bool(false)];
            bool var_30024_transpose_y_1 = const()[name = string("op_30024_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_30024_cast_fp16 = matmul(transpose_x = var_30024_transpose_x_1, transpose_y = var_30024_transpose_y_1, x = var_30022_cast_fp16, y = v_head_751_cast_fp16)[name = string("op_30024_cast_fp16")];
            tensor<int32, [1]> var_30025_axes_1 = const()[name = string("op_30025_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_30025_cast_fp16 = squeeze(axes = var_30025_axes_1, x = var_30024_cast_fp16)[name = string("op_30025_cast_fp16")];
            string dense_output_1889_pad_type_1 = const()[name = string("dense_output_1889_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1889_strides_1 = const()[name = string("dense_output_1889_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1889_pad_1 = const()[name = string("dense_output_1889_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1889_dilations_1 = const()[name = string("dense_output_1889_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1889_groups_1 = const()[name = string("dense_output_1889_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(688888832))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689019968))))[name = string("layers_23_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1889_cast_fp16 = conv(dilations = dense_output_1889_dilations_1, groups = dense_output_1889_groups_1, pad = dense_output_1889_pad_1, pad_type = dense_output_1889_pad_type_1, strides = dense_output_1889_strides_1, weight = layers_23_self_attn_linear_q_4_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1889_cast_fp16")];
            string sparse_output_1889_pad_type_1 = const()[name = string("sparse_output_1889_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1889_strides_1 = const()[name = string("sparse_output_1889_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1889_pad_1 = const()[name = string("sparse_output_1889_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1889_dilations_1 = const()[name = string("sparse_output_1889_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1889_groups_1 = const()[name = string("sparse_output_1889_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689023232))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689020544))))[name = string("layers_23_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1889_cast_fp16 = conv(dilations = sparse_output_1889_dilations_1, groups = sparse_output_1889_groups_1, pad = sparse_output_1889_pad_1, pad_type = sparse_output_1889_pad_type_1, strides = sparse_output_1889_strides_1, weight = layers_23_self_attn_linear_q_4_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1889_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30040_cast_fp16 = add(x = dense_output_1889_cast_fp16, y = sparse_output_1889_cast_fp16)[name = string("op_30040_cast_fp16")];
            tensor<int32, [4]> var_30041 = const()[name = string("op_30041"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30043 = const()[name = string("op_30043"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30042_cast_fp16 = transpose(perm = var_30041, x = var_30040_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 51, 128]> q_head_377_cast_fp16 = reshape(shape = var_30043, x = var_30042_cast_fp16)[name = string("q_head_377_cast_fp16")];
            string dense_output_1891_pad_type_1 = const()[name = string("dense_output_1891_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1891_strides_1 = const()[name = string("dense_output_1891_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1891_pad_1 = const()[name = string("dense_output_1891_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1891_dilations_1 = const()[name = string("dense_output_1891_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1891_groups_1 = const()[name = string("dense_output_1891_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689039680))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689170816))))[name = string("layers_23_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1891_cast_fp16 = conv(dilations = dense_output_1891_dilations_1, groups = dense_output_1891_groups_1, pad = dense_output_1891_pad_1, pad_type = dense_output_1891_pad_type_1, strides = dense_output_1891_strides_1, weight = layers_23_self_attn_linear_k_4_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1891_cast_fp16")];
            string sparse_output_1891_pad_type_1 = const()[name = string("sparse_output_1891_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1891_strides_1 = const()[name = string("sparse_output_1891_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1891_pad_1 = const()[name = string("sparse_output_1891_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1891_dilations_1 = const()[name = string("sparse_output_1891_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1891_groups_1 = const()[name = string("sparse_output_1891_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689174080))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689171392))))[name = string("layers_23_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1891_cast_fp16 = conv(dilations = sparse_output_1891_dilations_1, groups = sparse_output_1891_groups_1, pad = sparse_output_1891_pad_1, pad_type = sparse_output_1891_pad_type_1, strides = sparse_output_1891_strides_1, weight = layers_23_self_attn_linear_k_4_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1891_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30059_cast_fp16 = add(x = dense_output_1891_cast_fp16, y = sparse_output_1891_cast_fp16)[name = string("op_30059_cast_fp16")];
            tensor<int32, [4]> var_30060 = const()[name = string("op_30060"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30062 = const()[name = string("op_30062"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30061_cast_fp16 = transpose(perm = var_30060, x = var_30059_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 51, 128]> k_head_753_cast_fp16 = reshape(shape = var_30062, x = var_30061_cast_fp16)[name = string("k_head_753_cast_fp16")];
            string dense_output_1893_pad_type_1 = const()[name = string("dense_output_1893_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1893_strides_1 = const()[name = string("dense_output_1893_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1893_pad_1 = const()[name = string("dense_output_1893_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1893_dilations_1 = const()[name = string("dense_output_1893_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1893_groups_1 = const()[name = string("dense_output_1893_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689190528))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689321664))))[name = string("layers_23_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1893_cast_fp16 = conv(dilations = dense_output_1893_dilations_1, groups = dense_output_1893_groups_1, pad = dense_output_1893_pad_1, pad_type = dense_output_1893_pad_type_1, strides = dense_output_1893_strides_1, weight = layers_23_self_attn_linear_v_4_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1893_cast_fp16")];
            string sparse_output_1893_pad_type_1 = const()[name = string("sparse_output_1893_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1893_strides_1 = const()[name = string("sparse_output_1893_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1893_pad_1 = const()[name = string("sparse_output_1893_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1893_dilations_1 = const()[name = string("sparse_output_1893_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1893_groups_1 = const()[name = string("sparse_output_1893_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689324928))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689322240))))[name = string("layers_23_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1893_cast_fp16 = conv(dilations = sparse_output_1893_dilations_1, groups = sparse_output_1893_groups_1, pad = sparse_output_1893_pad_1, pad_type = sparse_output_1893_pad_type_1, strides = sparse_output_1893_strides_1, weight = layers_23_self_attn_linear_v_4_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1893_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30078_cast_fp16 = add(x = dense_output_1893_cast_fp16, y = sparse_output_1893_cast_fp16)[name = string("op_30078_cast_fp16")];
            tensor<int32, [4]> var_30079 = const()[name = string("op_30079"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30081 = const()[name = string("op_30081"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30080_cast_fp16 = transpose(perm = var_30079, x = var_30078_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 51, 128]> v_head_753_cast_fp16 = reshape(shape = var_30081, x = var_30080_cast_fp16)[name = string("v_head_753_cast_fp16")];
            string dense_output_1895_pad_type_1 = const()[name = string("dense_output_1895_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1895_strides_1 = const()[name = string("dense_output_1895_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1895_pad_1 = const()[name = string("dense_output_1895_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1895_dilations_1 = const()[name = string("dense_output_1895_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1895_groups_1 = const()[name = string("dense_output_1895_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689341376))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689472512))))[name = string("layers_23_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1895_cast_fp16 = conv(dilations = dense_output_1895_dilations_1, groups = dense_output_1895_groups_1, pad = dense_output_1895_pad_1, pad_type = dense_output_1895_pad_type_1, strides = dense_output_1895_strides_1, weight = layers_23_self_attn_linear_pos_4_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1895_cast_fp16")];
            string sparse_output_1895_pad_type_1 = const()[name = string("sparse_output_1895_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1895_strides_1 = const()[name = string("sparse_output_1895_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1895_pad_1 = const()[name = string("sparse_output_1895_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1895_dilations_1 = const()[name = string("sparse_output_1895_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1895_groups_1 = const()[name = string("sparse_output_1895_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689475776))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689473088))))[name = string("layers_23_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1895_cast_fp16 = conv(dilations = sparse_output_1895_dilations_1, groups = sparse_output_1895_groups_1, pad = sparse_output_1895_pad_1, pad_type = sparse_output_1895_pad_type_1, strides = sparse_output_1895_strides_1, weight = layers_23_self_attn_linear_pos_4_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1895_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_30097_cast_fp16 = add(x = dense_output_1895_cast_fp16, y = sparse_output_1895_cast_fp16)[name = string("op_30097_cast_fp16")];
            tensor<int32, [4]> var_30098 = const()[name = string("op_30098"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30100 = const()[name = string("op_30100"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_30099_cast_fp16 = transpose(perm = var_30098, x = var_30097_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 101, 128]> p_head_753_cast_fp16 = reshape(shape = var_30100, x = var_30099_cast_fp16)[name = string("p_head_753_cast_fp16")];
            tensor<fp16, [128]> var_30102_to_fp16 = const()[name = string("op_30102_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689492224)))];
            tensor<fp16, [1, 51, 128]> var_30103_cast_fp16 = add(x = q_head_377_cast_fp16, y = var_30102_to_fp16)[name = string("op_30103_cast_fp16")];
            tensor<int32, [1]> q_u_377_axes_1 = const()[name = string("q_u_377_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_377_cast_fp16 = expand_dims(axes = q_u_377_axes_1, x = var_30103_cast_fp16)[name = string("q_u_377_cast_fp16")];
            tensor<fp16, [128]> var_30105_to_fp16 = const()[name = string("op_30105_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689492544)))];
            tensor<fp16, [1, 51, 128]> var_30106_cast_fp16 = add(x = q_head_377_cast_fp16, y = var_30105_to_fp16)[name = string("op_30106_cast_fp16")];
            tensor<int32, [1]> q_v_377_axes_1 = const()[name = string("q_v_377_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_377_cast_fp16 = expand_dims(axes = q_v_377_axes_1, x = var_30106_cast_fp16)[name = string("q_v_377_cast_fp16")];
            tensor<int32, [1]> k_head_755_axes_1 = const()[name = string("k_head_755_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_755_cast_fp16 = expand_dims(axes = k_head_755_axes_1, x = k_head_753_cast_fp16)[name = string("k_head_755_cast_fp16")];
            tensor<int32, [1]> v_head_755_axes_1 = const()[name = string("v_head_755_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_755_cast_fp16 = expand_dims(axes = v_head_755_axes_1, x = v_head_753_cast_fp16)[name = string("v_head_755_cast_fp16")];
            tensor<int32, [1]> p_head_755_axes_1 = const()[name = string("p_head_755_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_755_cast_fp16 = expand_dims(axes = p_head_755_axes_1, x = p_head_753_cast_fp16)[name = string("p_head_755_cast_fp16")];
            bool var_30112_transpose_x_3 = const()[name = string("op_30112_transpose_x_3"), val = bool(false)];
            bool var_30112_transpose_y_3 = const()[name = string("op_30112_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_30112_cast_fp16 = matmul(transpose_x = var_30112_transpose_x_3, transpose_y = var_30112_transpose_y_3, x = q_u_377_cast_fp16, y = k_head_755_cast_fp16)[name = string("op_30112_cast_fp16")];
            fp16 var_30113_to_fp16 = const()[name = string("op_30113_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_377_cast_fp16 = mul(x = var_30112_cast_fp16, y = var_30113_to_fp16)[name = string("scores_content_377_cast_fp16")];
            bool x_1985_transpose_x_3 = const()[name = string("x_1985_transpose_x_3"), val = bool(false)];
            bool x_1985_transpose_y_3 = const()[name = string("x_1985_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1985_cast_fp16 = matmul(transpose_x = x_1985_transpose_x_3, transpose_y = x_1985_transpose_y_3, x = q_v_377_cast_fp16, y = p_head_755_cast_fp16)[name = string("x_1985_cast_fp16")];
            tensor<int32, [8]> x_1987_pad_1 = const()[name = string("x_1987_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1987_mode_1 = const()[name = string("x_1987_mode_1"), val = string("constant")];
            fp16 const_2575_to_fp16 = const()[name = string("const_2575_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1987_cast_fp16 = pad(constant_val = const_2575_to_fp16, mode = x_1987_mode_1, pad = x_1987_pad_1, x = x_1985_cast_fp16)[name = string("x_1987_cast_fp16")];
            tensor<int32, [4]> var_30127 = const()[name = string("op_30127"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1989_cast_fp16 = reshape(shape = var_30127, x = x_1987_cast_fp16)[name = string("x_1989_cast_fp16")];
            tensor<int32, [4]> var_30131_begin_1 = const()[name = string("op_30131_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_30131_end_1 = const()[name = string("op_30131_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_30131_end_mask_1 = const()[name = string("op_30131_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_30131_cast_fp16 = slice_by_index(begin = var_30131_begin_1, end = var_30131_end_1, end_mask = var_30131_end_mask_1, x = x_1989_cast_fp16)[name = string("op_30131_cast_fp16")];
            tensor<int32, [4]> var_30133 = const()[name = string("op_30133"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_30134_cast_fp16 = reshape(shape = var_30133, x = var_30131_cast_fp16)[name = string("op_30134_cast_fp16")];
            tensor<int32, [4]> var_30139_begin_1 = const()[name = string("op_30139_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_30139_end_1 = const()[name = string("op_30139_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_30139_end_mask_1 = const()[name = string("op_30139_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_30139_cast_fp16 = slice_by_index(begin = var_30139_begin_1, end = var_30139_end_1, end_mask = var_30139_end_mask_1, x = var_30134_cast_fp16)[name = string("op_30139_cast_fp16")];
            fp16 var_30140_to_fp16 = const()[name = string("op_30140_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_377_cast_fp16 = mul(x = var_30139_cast_fp16, y = var_30140_to_fp16)[name = string("scores_pos_377_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_377_cast_fp16 = add(x = scores_content_377_cast_fp16, y = scores_pos_377_cast_fp16)[name = string("logits_377_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_30143_cast_fp16 = softmax(axis = var_29384, x = logits_377_cast_fp16)[name = string("op_30143_cast_fp16")];
            bool var_30145_transpose_x_1 = const()[name = string("op_30145_transpose_x_1"), val = bool(false)];
            bool var_30145_transpose_y_1 = const()[name = string("op_30145_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_30145_cast_fp16 = matmul(transpose_x = var_30145_transpose_x_1, transpose_y = var_30145_transpose_y_1, x = var_30143_cast_fp16, y = v_head_755_cast_fp16)[name = string("op_30145_cast_fp16")];
            tensor<int32, [1]> var_30146_axes_1 = const()[name = string("op_30146_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_30146_cast_fp16 = squeeze(axes = var_30146_axes_1, x = var_30145_cast_fp16)[name = string("op_30146_cast_fp16")];
            string dense_output_1897_pad_type_1 = const()[name = string("dense_output_1897_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1897_strides_1 = const()[name = string("dense_output_1897_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1897_pad_1 = const()[name = string("dense_output_1897_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1897_dilations_1 = const()[name = string("dense_output_1897_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1897_groups_1 = const()[name = string("dense_output_1897_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689492864))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689624000))))[name = string("layers_23_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1897_cast_fp16 = conv(dilations = dense_output_1897_dilations_1, groups = dense_output_1897_groups_1, pad = dense_output_1897_pad_1, pad_type = dense_output_1897_pad_type_1, strides = dense_output_1897_strides_1, weight = layers_23_self_attn_linear_q_5_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1897_cast_fp16")];
            string sparse_output_1897_pad_type_1 = const()[name = string("sparse_output_1897_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1897_strides_1 = const()[name = string("sparse_output_1897_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1897_pad_1 = const()[name = string("sparse_output_1897_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1897_dilations_1 = const()[name = string("sparse_output_1897_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1897_groups_1 = const()[name = string("sparse_output_1897_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689627264))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689624576))))[name = string("layers_23_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1897_cast_fp16 = conv(dilations = sparse_output_1897_dilations_1, groups = sparse_output_1897_groups_1, pad = sparse_output_1897_pad_1, pad_type = sparse_output_1897_pad_type_1, strides = sparse_output_1897_strides_1, weight = layers_23_self_attn_linear_q_5_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1897_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30161_cast_fp16 = add(x = dense_output_1897_cast_fp16, y = sparse_output_1897_cast_fp16)[name = string("op_30161_cast_fp16")];
            tensor<int32, [4]> var_30162 = const()[name = string("op_30162"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30164 = const()[name = string("op_30164"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30163_cast_fp16 = transpose(perm = var_30162, x = var_30161_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 51, 128]> q_head_379_cast_fp16 = reshape(shape = var_30164, x = var_30163_cast_fp16)[name = string("q_head_379_cast_fp16")];
            string dense_output_1899_pad_type_1 = const()[name = string("dense_output_1899_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1899_strides_1 = const()[name = string("dense_output_1899_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1899_pad_1 = const()[name = string("dense_output_1899_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1899_dilations_1 = const()[name = string("dense_output_1899_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1899_groups_1 = const()[name = string("dense_output_1899_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689643712))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689774848))))[name = string("layers_23_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1899_cast_fp16 = conv(dilations = dense_output_1899_dilations_1, groups = dense_output_1899_groups_1, pad = dense_output_1899_pad_1, pad_type = dense_output_1899_pad_type_1, strides = dense_output_1899_strides_1, weight = layers_23_self_attn_linear_k_5_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1899_cast_fp16")];
            string sparse_output_1899_pad_type_1 = const()[name = string("sparse_output_1899_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1899_strides_1 = const()[name = string("sparse_output_1899_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1899_pad_1 = const()[name = string("sparse_output_1899_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1899_dilations_1 = const()[name = string("sparse_output_1899_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1899_groups_1 = const()[name = string("sparse_output_1899_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689778112))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689775424))))[name = string("layers_23_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1899_cast_fp16 = conv(dilations = sparse_output_1899_dilations_1, groups = sparse_output_1899_groups_1, pad = sparse_output_1899_pad_1, pad_type = sparse_output_1899_pad_type_1, strides = sparse_output_1899_strides_1, weight = layers_23_self_attn_linear_k_5_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1899_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30180_cast_fp16 = add(x = dense_output_1899_cast_fp16, y = sparse_output_1899_cast_fp16)[name = string("op_30180_cast_fp16")];
            tensor<int32, [4]> var_30181 = const()[name = string("op_30181"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30183 = const()[name = string("op_30183"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30182_cast_fp16 = transpose(perm = var_30181, x = var_30180_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 51, 128]> k_head_757_cast_fp16 = reshape(shape = var_30183, x = var_30182_cast_fp16)[name = string("k_head_757_cast_fp16")];
            string dense_output_1901_pad_type_1 = const()[name = string("dense_output_1901_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1901_strides_1 = const()[name = string("dense_output_1901_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1901_pad_1 = const()[name = string("dense_output_1901_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1901_dilations_1 = const()[name = string("dense_output_1901_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1901_groups_1 = const()[name = string("dense_output_1901_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689794560))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689925696))))[name = string("layers_23_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1901_cast_fp16 = conv(dilations = dense_output_1901_dilations_1, groups = dense_output_1901_groups_1, pad = dense_output_1901_pad_1, pad_type = dense_output_1901_pad_type_1, strides = dense_output_1901_strides_1, weight = layers_23_self_attn_linear_v_5_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1901_cast_fp16")];
            string sparse_output_1901_pad_type_1 = const()[name = string("sparse_output_1901_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1901_strides_1 = const()[name = string("sparse_output_1901_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1901_pad_1 = const()[name = string("sparse_output_1901_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1901_dilations_1 = const()[name = string("sparse_output_1901_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1901_groups_1 = const()[name = string("sparse_output_1901_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689928960))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689926272))))[name = string("layers_23_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1901_cast_fp16 = conv(dilations = sparse_output_1901_dilations_1, groups = sparse_output_1901_groups_1, pad = sparse_output_1901_pad_1, pad_type = sparse_output_1901_pad_type_1, strides = sparse_output_1901_strides_1, weight = layers_23_self_attn_linear_v_5_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1901_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30199_cast_fp16 = add(x = dense_output_1901_cast_fp16, y = sparse_output_1901_cast_fp16)[name = string("op_30199_cast_fp16")];
            tensor<int32, [4]> var_30200 = const()[name = string("op_30200"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30202 = const()[name = string("op_30202"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30201_cast_fp16 = transpose(perm = var_30200, x = var_30199_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 51, 128]> v_head_757_cast_fp16 = reshape(shape = var_30202, x = var_30201_cast_fp16)[name = string("v_head_757_cast_fp16")];
            string dense_output_1903_pad_type_1 = const()[name = string("dense_output_1903_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1903_strides_1 = const()[name = string("dense_output_1903_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1903_pad_1 = const()[name = string("dense_output_1903_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1903_dilations_1 = const()[name = string("dense_output_1903_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1903_groups_1 = const()[name = string("dense_output_1903_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689945408))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690076544))))[name = string("layers_23_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1903_cast_fp16 = conv(dilations = dense_output_1903_dilations_1, groups = dense_output_1903_groups_1, pad = dense_output_1903_pad_1, pad_type = dense_output_1903_pad_type_1, strides = dense_output_1903_strides_1, weight = layers_23_self_attn_linear_pos_5_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1903_cast_fp16")];
            string sparse_output_1903_pad_type_1 = const()[name = string("sparse_output_1903_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1903_strides_1 = const()[name = string("sparse_output_1903_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1903_pad_1 = const()[name = string("sparse_output_1903_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1903_dilations_1 = const()[name = string("sparse_output_1903_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1903_groups_1 = const()[name = string("sparse_output_1903_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690079808))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690077120))))[name = string("layers_23_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1903_cast_fp16 = conv(dilations = sparse_output_1903_dilations_1, groups = sparse_output_1903_groups_1, pad = sparse_output_1903_pad_1, pad_type = sparse_output_1903_pad_type_1, strides = sparse_output_1903_strides_1, weight = layers_23_self_attn_linear_pos_5_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1903_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_30218_cast_fp16 = add(x = dense_output_1903_cast_fp16, y = sparse_output_1903_cast_fp16)[name = string("op_30218_cast_fp16")];
            tensor<int32, [4]> var_30219 = const()[name = string("op_30219"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30221 = const()[name = string("op_30221"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_30220_cast_fp16 = transpose(perm = var_30219, x = var_30218_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 101, 128]> p_head_757_cast_fp16 = reshape(shape = var_30221, x = var_30220_cast_fp16)[name = string("p_head_757_cast_fp16")];
            tensor<fp16, [128]> var_30223_to_fp16 = const()[name = string("op_30223_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690096256)))];
            tensor<fp16, [1, 51, 128]> var_30224_cast_fp16 = add(x = q_head_379_cast_fp16, y = var_30223_to_fp16)[name = string("op_30224_cast_fp16")];
            tensor<int32, [1]> q_u_379_axes_1 = const()[name = string("q_u_379_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_379_cast_fp16 = expand_dims(axes = q_u_379_axes_1, x = var_30224_cast_fp16)[name = string("q_u_379_cast_fp16")];
            tensor<fp16, [128]> var_30226_to_fp16 = const()[name = string("op_30226_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690096576)))];
            tensor<fp16, [1, 51, 128]> var_30227_cast_fp16 = add(x = q_head_379_cast_fp16, y = var_30226_to_fp16)[name = string("op_30227_cast_fp16")];
            tensor<int32, [1]> q_v_379_axes_1 = const()[name = string("q_v_379_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_379_cast_fp16 = expand_dims(axes = q_v_379_axes_1, x = var_30227_cast_fp16)[name = string("q_v_379_cast_fp16")];
            tensor<int32, [1]> k_head_759_axes_1 = const()[name = string("k_head_759_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_759_cast_fp16 = expand_dims(axes = k_head_759_axes_1, x = k_head_757_cast_fp16)[name = string("k_head_759_cast_fp16")];
            tensor<int32, [1]> v_head_759_axes_1 = const()[name = string("v_head_759_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_759_cast_fp16 = expand_dims(axes = v_head_759_axes_1, x = v_head_757_cast_fp16)[name = string("v_head_759_cast_fp16")];
            tensor<int32, [1]> p_head_759_axes_1 = const()[name = string("p_head_759_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_759_cast_fp16 = expand_dims(axes = p_head_759_axes_1, x = p_head_757_cast_fp16)[name = string("p_head_759_cast_fp16")];
            bool var_30233_transpose_x_3 = const()[name = string("op_30233_transpose_x_3"), val = bool(false)];
            bool var_30233_transpose_y_3 = const()[name = string("op_30233_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_30233_cast_fp16 = matmul(transpose_x = var_30233_transpose_x_3, transpose_y = var_30233_transpose_y_3, x = q_u_379_cast_fp16, y = k_head_759_cast_fp16)[name = string("op_30233_cast_fp16")];
            fp16 var_30234_to_fp16 = const()[name = string("op_30234_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_379_cast_fp16 = mul(x = var_30233_cast_fp16, y = var_30234_to_fp16)[name = string("scores_content_379_cast_fp16")];
            bool x_1993_transpose_x_3 = const()[name = string("x_1993_transpose_x_3"), val = bool(false)];
            bool x_1993_transpose_y_3 = const()[name = string("x_1993_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_1993_cast_fp16 = matmul(transpose_x = x_1993_transpose_x_3, transpose_y = x_1993_transpose_y_3, x = q_v_379_cast_fp16, y = p_head_759_cast_fp16)[name = string("x_1993_cast_fp16")];
            tensor<int32, [8]> x_1995_pad_1 = const()[name = string("x_1995_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_1995_mode_1 = const()[name = string("x_1995_mode_1"), val = string("constant")];
            fp16 const_2581_to_fp16 = const()[name = string("const_2581_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_1995_cast_fp16 = pad(constant_val = const_2581_to_fp16, mode = x_1995_mode_1, pad = x_1995_pad_1, x = x_1993_cast_fp16)[name = string("x_1995_cast_fp16")];
            tensor<int32, [4]> var_30248 = const()[name = string("op_30248"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_1997_cast_fp16 = reshape(shape = var_30248, x = x_1995_cast_fp16)[name = string("x_1997_cast_fp16")];
            tensor<int32, [4]> var_30252_begin_1 = const()[name = string("op_30252_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_30252_end_1 = const()[name = string("op_30252_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_30252_end_mask_1 = const()[name = string("op_30252_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_30252_cast_fp16 = slice_by_index(begin = var_30252_begin_1, end = var_30252_end_1, end_mask = var_30252_end_mask_1, x = x_1997_cast_fp16)[name = string("op_30252_cast_fp16")];
            tensor<int32, [4]> var_30254 = const()[name = string("op_30254"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_30255_cast_fp16 = reshape(shape = var_30254, x = var_30252_cast_fp16)[name = string("op_30255_cast_fp16")];
            tensor<int32, [4]> var_30260_begin_1 = const()[name = string("op_30260_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_30260_end_1 = const()[name = string("op_30260_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_30260_end_mask_1 = const()[name = string("op_30260_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_30260_cast_fp16 = slice_by_index(begin = var_30260_begin_1, end = var_30260_end_1, end_mask = var_30260_end_mask_1, x = var_30255_cast_fp16)[name = string("op_30260_cast_fp16")];
            fp16 var_30261_to_fp16 = const()[name = string("op_30261_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_379_cast_fp16 = mul(x = var_30260_cast_fp16, y = var_30261_to_fp16)[name = string("scores_pos_379_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_379_cast_fp16 = add(x = scores_content_379_cast_fp16, y = scores_pos_379_cast_fp16)[name = string("logits_379_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_30264_cast_fp16 = softmax(axis = var_29384, x = logits_379_cast_fp16)[name = string("op_30264_cast_fp16")];
            bool var_30266_transpose_x_1 = const()[name = string("op_30266_transpose_x_1"), val = bool(false)];
            bool var_30266_transpose_y_1 = const()[name = string("op_30266_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_30266_cast_fp16 = matmul(transpose_x = var_30266_transpose_x_1, transpose_y = var_30266_transpose_y_1, x = var_30264_cast_fp16, y = v_head_759_cast_fp16)[name = string("op_30266_cast_fp16")];
            tensor<int32, [1]> var_30267_axes_1 = const()[name = string("op_30267_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_30267_cast_fp16 = squeeze(axes = var_30267_axes_1, x = var_30266_cast_fp16)[name = string("op_30267_cast_fp16")];
            string dense_output_1905_pad_type_1 = const()[name = string("dense_output_1905_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1905_strides_1 = const()[name = string("dense_output_1905_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1905_pad_1 = const()[name = string("dense_output_1905_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1905_dilations_1 = const()[name = string("dense_output_1905_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1905_groups_1 = const()[name = string("dense_output_1905_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690096896))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690228032))))[name = string("layers_23_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1905_cast_fp16 = conv(dilations = dense_output_1905_dilations_1, groups = dense_output_1905_groups_1, pad = dense_output_1905_pad_1, pad_type = dense_output_1905_pad_type_1, strides = dense_output_1905_strides_1, weight = layers_23_self_attn_linear_q_6_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1905_cast_fp16")];
            string sparse_output_1905_pad_type_1 = const()[name = string("sparse_output_1905_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1905_strides_1 = const()[name = string("sparse_output_1905_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1905_pad_1 = const()[name = string("sparse_output_1905_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1905_dilations_1 = const()[name = string("sparse_output_1905_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1905_groups_1 = const()[name = string("sparse_output_1905_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690231296))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690228608))))[name = string("layers_23_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1905_cast_fp16 = conv(dilations = sparse_output_1905_dilations_1, groups = sparse_output_1905_groups_1, pad = sparse_output_1905_pad_1, pad_type = sparse_output_1905_pad_type_1, strides = sparse_output_1905_strides_1, weight = layers_23_self_attn_linear_q_6_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1905_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30282_cast_fp16 = add(x = dense_output_1905_cast_fp16, y = sparse_output_1905_cast_fp16)[name = string("op_30282_cast_fp16")];
            tensor<int32, [4]> var_30283 = const()[name = string("op_30283"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30285 = const()[name = string("op_30285"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30284_cast_fp16 = transpose(perm = var_30283, x = var_30282_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 51, 128]> q_head_381_cast_fp16 = reshape(shape = var_30285, x = var_30284_cast_fp16)[name = string("q_head_381_cast_fp16")];
            string dense_output_1907_pad_type_1 = const()[name = string("dense_output_1907_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1907_strides_1 = const()[name = string("dense_output_1907_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1907_pad_1 = const()[name = string("dense_output_1907_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1907_dilations_1 = const()[name = string("dense_output_1907_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1907_groups_1 = const()[name = string("dense_output_1907_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690247744))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690378880))))[name = string("layers_23_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1907_cast_fp16 = conv(dilations = dense_output_1907_dilations_1, groups = dense_output_1907_groups_1, pad = dense_output_1907_pad_1, pad_type = dense_output_1907_pad_type_1, strides = dense_output_1907_strides_1, weight = layers_23_self_attn_linear_k_6_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1907_cast_fp16")];
            string sparse_output_1907_pad_type_1 = const()[name = string("sparse_output_1907_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1907_strides_1 = const()[name = string("sparse_output_1907_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1907_pad_1 = const()[name = string("sparse_output_1907_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1907_dilations_1 = const()[name = string("sparse_output_1907_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1907_groups_1 = const()[name = string("sparse_output_1907_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690382144))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690379456))))[name = string("layers_23_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1907_cast_fp16 = conv(dilations = sparse_output_1907_dilations_1, groups = sparse_output_1907_groups_1, pad = sparse_output_1907_pad_1, pad_type = sparse_output_1907_pad_type_1, strides = sparse_output_1907_strides_1, weight = layers_23_self_attn_linear_k_6_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1907_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30301_cast_fp16 = add(x = dense_output_1907_cast_fp16, y = sparse_output_1907_cast_fp16)[name = string("op_30301_cast_fp16")];
            tensor<int32, [4]> var_30302 = const()[name = string("op_30302"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30304 = const()[name = string("op_30304"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30303_cast_fp16 = transpose(perm = var_30302, x = var_30301_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 51, 128]> k_head_761_cast_fp16 = reshape(shape = var_30304, x = var_30303_cast_fp16)[name = string("k_head_761_cast_fp16")];
            string dense_output_1909_pad_type_1 = const()[name = string("dense_output_1909_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1909_strides_1 = const()[name = string("dense_output_1909_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1909_pad_1 = const()[name = string("dense_output_1909_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1909_dilations_1 = const()[name = string("dense_output_1909_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1909_groups_1 = const()[name = string("dense_output_1909_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690398592))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690529728))))[name = string("layers_23_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1909_cast_fp16 = conv(dilations = dense_output_1909_dilations_1, groups = dense_output_1909_groups_1, pad = dense_output_1909_pad_1, pad_type = dense_output_1909_pad_type_1, strides = dense_output_1909_strides_1, weight = layers_23_self_attn_linear_v_6_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1909_cast_fp16")];
            string sparse_output_1909_pad_type_1 = const()[name = string("sparse_output_1909_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1909_strides_1 = const()[name = string("sparse_output_1909_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1909_pad_1 = const()[name = string("sparse_output_1909_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1909_dilations_1 = const()[name = string("sparse_output_1909_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1909_groups_1 = const()[name = string("sparse_output_1909_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690532992))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690530304))))[name = string("layers_23_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1909_cast_fp16 = conv(dilations = sparse_output_1909_dilations_1, groups = sparse_output_1909_groups_1, pad = sparse_output_1909_pad_1, pad_type = sparse_output_1909_pad_type_1, strides = sparse_output_1909_strides_1, weight = layers_23_self_attn_linear_v_6_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1909_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30320_cast_fp16 = add(x = dense_output_1909_cast_fp16, y = sparse_output_1909_cast_fp16)[name = string("op_30320_cast_fp16")];
            tensor<int32, [4]> var_30321 = const()[name = string("op_30321"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30323 = const()[name = string("op_30323"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30322_cast_fp16 = transpose(perm = var_30321, x = var_30320_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 51, 128]> v_head_761_cast_fp16 = reshape(shape = var_30323, x = var_30322_cast_fp16)[name = string("v_head_761_cast_fp16")];
            string dense_output_1911_pad_type_1 = const()[name = string("dense_output_1911_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1911_strides_1 = const()[name = string("dense_output_1911_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1911_pad_1 = const()[name = string("dense_output_1911_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1911_dilations_1 = const()[name = string("dense_output_1911_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1911_groups_1 = const()[name = string("dense_output_1911_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690549440))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690680576))))[name = string("layers_23_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1911_cast_fp16 = conv(dilations = dense_output_1911_dilations_1, groups = dense_output_1911_groups_1, pad = dense_output_1911_pad_1, pad_type = dense_output_1911_pad_type_1, strides = dense_output_1911_strides_1, weight = layers_23_self_attn_linear_pos_6_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1911_cast_fp16")];
            string sparse_output_1911_pad_type_1 = const()[name = string("sparse_output_1911_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1911_strides_1 = const()[name = string("sparse_output_1911_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1911_pad_1 = const()[name = string("sparse_output_1911_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1911_dilations_1 = const()[name = string("sparse_output_1911_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1911_groups_1 = const()[name = string("sparse_output_1911_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690683840))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690681152))))[name = string("layers_23_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1911_cast_fp16 = conv(dilations = sparse_output_1911_dilations_1, groups = sparse_output_1911_groups_1, pad = sparse_output_1911_pad_1, pad_type = sparse_output_1911_pad_type_1, strides = sparse_output_1911_strides_1, weight = layers_23_self_attn_linear_pos_6_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1911_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_30339_cast_fp16 = add(x = dense_output_1911_cast_fp16, y = sparse_output_1911_cast_fp16)[name = string("op_30339_cast_fp16")];
            tensor<int32, [4]> var_30340 = const()[name = string("op_30340"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30342 = const()[name = string("op_30342"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_30341_cast_fp16 = transpose(perm = var_30340, x = var_30339_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 101, 128]> p_head_761_cast_fp16 = reshape(shape = var_30342, x = var_30341_cast_fp16)[name = string("p_head_761_cast_fp16")];
            tensor<fp16, [128]> var_30344_to_fp16 = const()[name = string("op_30344_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690700288)))];
            tensor<fp16, [1, 51, 128]> var_30345_cast_fp16 = add(x = q_head_381_cast_fp16, y = var_30344_to_fp16)[name = string("op_30345_cast_fp16")];
            tensor<int32, [1]> q_u_381_axes_1 = const()[name = string("q_u_381_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_381_cast_fp16 = expand_dims(axes = q_u_381_axes_1, x = var_30345_cast_fp16)[name = string("q_u_381_cast_fp16")];
            tensor<fp16, [128]> var_30347_to_fp16 = const()[name = string("op_30347_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690700608)))];
            tensor<fp16, [1, 51, 128]> var_30348_cast_fp16 = add(x = q_head_381_cast_fp16, y = var_30347_to_fp16)[name = string("op_30348_cast_fp16")];
            tensor<int32, [1]> q_v_381_axes_1 = const()[name = string("q_v_381_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_381_cast_fp16 = expand_dims(axes = q_v_381_axes_1, x = var_30348_cast_fp16)[name = string("q_v_381_cast_fp16")];
            tensor<int32, [1]> k_head_763_axes_1 = const()[name = string("k_head_763_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_763_cast_fp16 = expand_dims(axes = k_head_763_axes_1, x = k_head_761_cast_fp16)[name = string("k_head_763_cast_fp16")];
            tensor<int32, [1]> v_head_763_axes_1 = const()[name = string("v_head_763_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_763_cast_fp16 = expand_dims(axes = v_head_763_axes_1, x = v_head_761_cast_fp16)[name = string("v_head_763_cast_fp16")];
            tensor<int32, [1]> p_head_763_axes_1 = const()[name = string("p_head_763_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_763_cast_fp16 = expand_dims(axes = p_head_763_axes_1, x = p_head_761_cast_fp16)[name = string("p_head_763_cast_fp16")];
            bool var_30354_transpose_x_3 = const()[name = string("op_30354_transpose_x_3"), val = bool(false)];
            bool var_30354_transpose_y_3 = const()[name = string("op_30354_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_30354_cast_fp16 = matmul(transpose_x = var_30354_transpose_x_3, transpose_y = var_30354_transpose_y_3, x = q_u_381_cast_fp16, y = k_head_763_cast_fp16)[name = string("op_30354_cast_fp16")];
            fp16 var_30355_to_fp16 = const()[name = string("op_30355_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_381_cast_fp16 = mul(x = var_30354_cast_fp16, y = var_30355_to_fp16)[name = string("scores_content_381_cast_fp16")];
            bool x_2001_transpose_x_3 = const()[name = string("x_2001_transpose_x_3"), val = bool(false)];
            bool x_2001_transpose_y_3 = const()[name = string("x_2001_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_2001_cast_fp16 = matmul(transpose_x = x_2001_transpose_x_3, transpose_y = x_2001_transpose_y_3, x = q_v_381_cast_fp16, y = p_head_763_cast_fp16)[name = string("x_2001_cast_fp16")];
            tensor<int32, [8]> x_2003_pad_1 = const()[name = string("x_2003_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_2003_mode_1 = const()[name = string("x_2003_mode_1"), val = string("constant")];
            fp16 const_2587_to_fp16 = const()[name = string("const_2587_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_2003_cast_fp16 = pad(constant_val = const_2587_to_fp16, mode = x_2003_mode_1, pad = x_2003_pad_1, x = x_2001_cast_fp16)[name = string("x_2003_cast_fp16")];
            tensor<int32, [4]> var_30369 = const()[name = string("op_30369"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_2005_cast_fp16 = reshape(shape = var_30369, x = x_2003_cast_fp16)[name = string("x_2005_cast_fp16")];
            tensor<int32, [4]> var_30373_begin_1 = const()[name = string("op_30373_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_30373_end_1 = const()[name = string("op_30373_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_30373_end_mask_1 = const()[name = string("op_30373_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_30373_cast_fp16 = slice_by_index(begin = var_30373_begin_1, end = var_30373_end_1, end_mask = var_30373_end_mask_1, x = x_2005_cast_fp16)[name = string("op_30373_cast_fp16")];
            tensor<int32, [4]> var_30375 = const()[name = string("op_30375"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_30376_cast_fp16 = reshape(shape = var_30375, x = var_30373_cast_fp16)[name = string("op_30376_cast_fp16")];
            tensor<int32, [4]> var_30381_begin_1 = const()[name = string("op_30381_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_30381_end_1 = const()[name = string("op_30381_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_30381_end_mask_1 = const()[name = string("op_30381_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_30381_cast_fp16 = slice_by_index(begin = var_30381_begin_1, end = var_30381_end_1, end_mask = var_30381_end_mask_1, x = var_30376_cast_fp16)[name = string("op_30381_cast_fp16")];
            fp16 var_30382_to_fp16 = const()[name = string("op_30382_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_381_cast_fp16 = mul(x = var_30381_cast_fp16, y = var_30382_to_fp16)[name = string("scores_pos_381_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_381_cast_fp16 = add(x = scores_content_381_cast_fp16, y = scores_pos_381_cast_fp16)[name = string("logits_381_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_30385_cast_fp16 = softmax(axis = var_29384, x = logits_381_cast_fp16)[name = string("op_30385_cast_fp16")];
            bool var_30387_transpose_x_1 = const()[name = string("op_30387_transpose_x_1"), val = bool(false)];
            bool var_30387_transpose_y_1 = const()[name = string("op_30387_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_30387_cast_fp16 = matmul(transpose_x = var_30387_transpose_x_1, transpose_y = var_30387_transpose_y_1, x = var_30385_cast_fp16, y = v_head_763_cast_fp16)[name = string("op_30387_cast_fp16")];
            tensor<int32, [1]> var_30388_axes_1 = const()[name = string("op_30388_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> var_30388_cast_fp16 = squeeze(axes = var_30388_axes_1, x = var_30387_cast_fp16)[name = string("op_30388_cast_fp16")];
            string dense_output_1913_pad_type_1 = const()[name = string("dense_output_1913_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1913_strides_1 = const()[name = string("dense_output_1913_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1913_pad_1 = const()[name = string("dense_output_1913_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1913_dilations_1 = const()[name = string("dense_output_1913_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1913_groups_1 = const()[name = string("dense_output_1913_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690700928))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690832064))))[name = string("layers_23_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1913_cast_fp16 = conv(dilations = dense_output_1913_dilations_1, groups = dense_output_1913_groups_1, pad = dense_output_1913_pad_1, pad_type = dense_output_1913_pad_type_1, strides = dense_output_1913_strides_1, weight = layers_23_self_attn_linear_q_7_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1913_cast_fp16")];
            string sparse_output_1913_pad_type_1 = const()[name = string("sparse_output_1913_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1913_strides_1 = const()[name = string("sparse_output_1913_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1913_pad_1 = const()[name = string("sparse_output_1913_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1913_dilations_1 = const()[name = string("sparse_output_1913_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1913_groups_1 = const()[name = string("sparse_output_1913_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690835328))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690832640))))[name = string("layers_23_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1913_cast_fp16 = conv(dilations = sparse_output_1913_dilations_1, groups = sparse_output_1913_groups_1, pad = sparse_output_1913_pad_1, pad_type = sparse_output_1913_pad_type_1, strides = sparse_output_1913_strides_1, weight = layers_23_self_attn_linear_q_7_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1913_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30403_cast_fp16 = add(x = dense_output_1913_cast_fp16, y = sparse_output_1913_cast_fp16)[name = string("op_30403_cast_fp16")];
            tensor<int32, [4]> var_30404 = const()[name = string("op_30404"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30406 = const()[name = string("op_30406"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30405_cast_fp16 = transpose(perm = var_30404, x = var_30403_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 51, 128]> q_head_cast_fp16 = reshape(shape = var_30406, x = var_30405_cast_fp16)[name = string("q_head_cast_fp16")];
            string dense_output_1915_pad_type_1 = const()[name = string("dense_output_1915_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1915_strides_1 = const()[name = string("dense_output_1915_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1915_pad_1 = const()[name = string("dense_output_1915_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1915_dilations_1 = const()[name = string("dense_output_1915_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1915_groups_1 = const()[name = string("dense_output_1915_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690851776))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690982912))))[name = string("layers_23_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1915_cast_fp16 = conv(dilations = dense_output_1915_dilations_1, groups = dense_output_1915_groups_1, pad = dense_output_1915_pad_1, pad_type = dense_output_1915_pad_type_1, strides = dense_output_1915_strides_1, weight = layers_23_self_attn_linear_k_7_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1915_cast_fp16")];
            string sparse_output_1915_pad_type_1 = const()[name = string("sparse_output_1915_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1915_strides_1 = const()[name = string("sparse_output_1915_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1915_pad_1 = const()[name = string("sparse_output_1915_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1915_dilations_1 = const()[name = string("sparse_output_1915_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1915_groups_1 = const()[name = string("sparse_output_1915_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690986176))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(690983488))))[name = string("layers_23_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1915_cast_fp16 = conv(dilations = sparse_output_1915_dilations_1, groups = sparse_output_1915_groups_1, pad = sparse_output_1915_pad_1, pad_type = sparse_output_1915_pad_type_1, strides = sparse_output_1915_strides_1, weight = layers_23_self_attn_linear_k_7_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1915_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30422_cast_fp16 = add(x = dense_output_1915_cast_fp16, y = sparse_output_1915_cast_fp16)[name = string("op_30422_cast_fp16")];
            tensor<int32, [4]> var_30423 = const()[name = string("op_30423"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30425 = const()[name = string("op_30425"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30424_cast_fp16 = transpose(perm = var_30423, x = var_30422_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 51, 128]> k_head_765_cast_fp16 = reshape(shape = var_30425, x = var_30424_cast_fp16)[name = string("k_head_765_cast_fp16")];
            string dense_output_1917_pad_type_1 = const()[name = string("dense_output_1917_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1917_strides_1 = const()[name = string("dense_output_1917_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1917_pad_1 = const()[name = string("dense_output_1917_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1917_dilations_1 = const()[name = string("dense_output_1917_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1917_groups_1 = const()[name = string("dense_output_1917_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691002624))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691133760))))[name = string("layers_23_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 51, 1]> dense_output_1917_cast_fp16 = conv(dilations = dense_output_1917_dilations_1, groups = dense_output_1917_groups_1, pad = dense_output_1917_pad_1, pad_type = dense_output_1917_pad_type_1, strides = dense_output_1917_strides_1, weight = layers_23_self_attn_linear_v_7_dense_conv_weight_to_fp16_palettized, x = input_1087_cast_fp16)[name = string("dense_output_1917_cast_fp16")];
            string sparse_output_1917_pad_type_1 = const()[name = string("sparse_output_1917_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1917_strides_1 = const()[name = string("sparse_output_1917_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1917_pad_1 = const()[name = string("sparse_output_1917_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1917_dilations_1 = const()[name = string("sparse_output_1917_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1917_groups_1 = const()[name = string("sparse_output_1917_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691137024))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691134336))))[name = string("layers_23_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 51, 1]> sparse_output_1917_cast_fp16 = conv(dilations = sparse_output_1917_dilations_1, groups = sparse_output_1917_groups_1, pad = sparse_output_1917_pad_1, pad_type = sparse_output_1917_pad_type_1, strides = sparse_output_1917_strides_1, weight = layers_23_self_attn_linear_v_7_sparse_conv_weight_to_fp16_sparsified, x = input_1087_cast_fp16)[name = string("sparse_output_1917_cast_fp16")];
            tensor<fp16, [1, 128, 51, 1]> var_30441_cast_fp16 = add(x = dense_output_1917_cast_fp16, y = sparse_output_1917_cast_fp16)[name = string("op_30441_cast_fp16")];
            tensor<int32, [4]> var_30442 = const()[name = string("op_30442"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30444 = const()[name = string("op_30444"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 51, 1, 128]> var_30443_cast_fp16 = transpose(perm = var_30442, x = var_30441_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 51, 128]> v_head_765_cast_fp16 = reshape(shape = var_30444, x = var_30443_cast_fp16)[name = string("v_head_765_cast_fp16")];
            string dense_output_1919_pad_type_1 = const()[name = string("dense_output_1919_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1919_strides_1 = const()[name = string("dense_output_1919_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1919_pad_1 = const()[name = string("dense_output_1919_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1919_dilations_1 = const()[name = string("dense_output_1919_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1919_groups_1 = const()[name = string("dense_output_1919_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691153472))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691284608))))[name = string("layers_23_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 101, 1]> dense_output_1919_cast_fp16 = conv(dilations = dense_output_1919_dilations_1, groups = dense_output_1919_groups_1, pad = dense_output_1919_pad_1, pad_type = dense_output_1919_pad_type_1, strides = dense_output_1919_strides_1, weight = layers_23_self_attn_linear_pos_7_dense_conv_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("dense_output_1919_cast_fp16")];
            string sparse_output_1919_pad_type_1 = const()[name = string("sparse_output_1919_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1919_strides_1 = const()[name = string("sparse_output_1919_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1919_pad_1 = const()[name = string("sparse_output_1919_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1919_dilations_1 = const()[name = string("sparse_output_1919_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1919_groups_1 = const()[name = string("sparse_output_1919_groups_1"), val = int32(1)];
            tensor<fp16, [128, 1024, 1, 1]> layers_23_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [128, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691287872))), nonzero_data = tensor<fp16, [1311]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691285184))))[name = string("layers_23_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 128, 101, 1]> sparse_output_1919_cast_fp16 = conv(dilations = sparse_output_1919_dilations_1, groups = sparse_output_1919_groups_1, pad = sparse_output_1919_pad_1, pad_type = sparse_output_1919_pad_type_1, strides = sparse_output_1919_strides_1, weight = layers_23_self_attn_linear_pos_7_sparse_conv_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = string("sparse_output_1919_cast_fp16")];
            tensor<fp16, [1, 128, 101, 1]> var_30460_cast_fp16 = add(x = dense_output_1919_cast_fp16, y = sparse_output_1919_cast_fp16)[name = string("op_30460_cast_fp16")];
            tensor<int32, [4]> var_30461 = const()[name = string("op_30461"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> var_30463 = const()[name = string("op_30463"), val = tensor<int32, [3]>([1, -1, 128])];
            tensor<fp16, [1, 101, 1, 128]> var_30462_cast_fp16 = transpose(perm = var_30461, x = var_30460_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 101, 128]> p_head_765_cast_fp16 = reshape(shape = var_30463, x = var_30462_cast_fp16)[name = string("p_head_765_cast_fp16")];
            tensor<fp16, [128]> var_30465_to_fp16 = const()[name = string("op_30465_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691304320)))];
            tensor<fp16, [1, 51, 128]> var_30466_cast_fp16 = add(x = q_head_cast_fp16, y = var_30465_to_fp16)[name = string("op_30466_cast_fp16")];
            tensor<int32, [1]> q_u_axes_1 = const()[name = string("q_u_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_u_cast_fp16 = expand_dims(axes = q_u_axes_1, x = var_30466_cast_fp16)[name = string("q_u_cast_fp16")];
            tensor<fp16, [128]> var_30468_to_fp16 = const()[name = string("op_30468_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691304640)))];
            tensor<fp16, [1, 51, 128]> var_30469_cast_fp16 = add(x = q_head_cast_fp16, y = var_30468_to_fp16)[name = string("op_30469_cast_fp16")];
            tensor<int32, [1]> q_v_axes_1 = const()[name = string("q_v_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> q_v_cast_fp16 = expand_dims(axes = q_v_axes_1, x = var_30469_cast_fp16)[name = string("q_v_cast_fp16")];
            tensor<int32, [1]> k_head_axes_1 = const()[name = string("k_head_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> k_head_cast_fp16 = expand_dims(axes = k_head_axes_1, x = k_head_765_cast_fp16)[name = string("k_head_cast_fp16")];
            tensor<int32, [1]> v_head_axes_1 = const()[name = string("v_head_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 51, 128]> v_head_cast_fp16 = expand_dims(axes = v_head_axes_1, x = v_head_765_cast_fp16)[name = string("v_head_cast_fp16")];
            tensor<int32, [1]> p_head_axes_1 = const()[name = string("p_head_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 101, 128]> p_head_cast_fp16 = expand_dims(axes = p_head_axes_1, x = p_head_765_cast_fp16)[name = string("p_head_cast_fp16")];
            bool var_30475_transpose_x_3 = const()[name = string("op_30475_transpose_x_3"), val = bool(false)];
            bool var_30475_transpose_y_3 = const()[name = string("op_30475_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 51]> var_30475_cast_fp16 = matmul(transpose_x = var_30475_transpose_x_3, transpose_y = var_30475_transpose_y_3, x = q_u_cast_fp16, y = k_head_cast_fp16)[name = string("op_30475_cast_fp16")];
            fp16 var_30476_to_fp16 = const()[name = string("op_30476_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_content_cast_fp16 = mul(x = var_30475_cast_fp16, y = var_30476_to_fp16)[name = string("scores_content_cast_fp16")];
            bool x_2009_transpose_x_3 = const()[name = string("x_2009_transpose_x_3"), val = bool(false)];
            bool x_2009_transpose_y_3 = const()[name = string("x_2009_transpose_y_3"), val = bool(true)];
            tensor<fp16, [1, 1, 51, 101]> x_2009_cast_fp16 = matmul(transpose_x = x_2009_transpose_x_3, transpose_y = x_2009_transpose_y_3, x = q_v_cast_fp16, y = p_head_cast_fp16)[name = string("x_2009_cast_fp16")];
            tensor<int32, [8]> x_2011_pad_1 = const()[name = string("x_2011_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string x_2011_mode_1 = const()[name = string("x_2011_mode_1"), val = string("constant")];
            fp16 const_2593_to_fp16 = const()[name = string("const_2593_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1, 51, 102]> x_2011_cast_fp16 = pad(constant_val = const_2593_to_fp16, mode = x_2011_mode_1, pad = x_2011_pad_1, x = x_2009_cast_fp16)[name = string("x_2011_cast_fp16")];
            tensor<int32, [4]> var_30490 = const()[name = string("op_30490"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<fp16, [1, 1, 102, 51]> x_2013_cast_fp16 = reshape(shape = var_30490, x = x_2011_cast_fp16)[name = string("x_2013_cast_fp16")];
            tensor<int32, [4]> var_30494_begin_1 = const()[name = string("op_30494_begin_1"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_30494_end_1 = const()[name = string("op_30494_end_1"), val = tensor<int32, [4]>([1, 1, 102, 51])];
            tensor<bool, [4]> var_30494_end_mask_1 = const()[name = string("op_30494_end_mask_1"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 101, 51]> var_30494_cast_fp16 = slice_by_index(begin = var_30494_begin_1, end = var_30494_end_1, end_mask = var_30494_end_mask_1, x = x_2013_cast_fp16)[name = string("op_30494_cast_fp16")];
            tensor<int32, [4]> var_30496 = const()[name = string("op_30496"), val = tensor<int32, [4]>([1, 1, 51, 101])];
            tensor<fp16, [1, 1, 51, 101]> var_30497_cast_fp16 = reshape(shape = var_30496, x = var_30494_cast_fp16)[name = string("op_30497_cast_fp16")];
            tensor<int32, [4]> var_30502_begin_1 = const()[name = string("op_30502_begin_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_30502_end_1 = const()[name = string("op_30502_end_1"), val = tensor<int32, [4]>([1, 1, 51, 51])];
            tensor<bool, [4]> var_30502_end_mask_1 = const()[name = string("op_30502_end_mask_1"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 51, 51]> var_30502_cast_fp16 = slice_by_index(begin = var_30502_begin_1, end = var_30502_end_1, end_mask = var_30502_end_mask_1, x = var_30497_cast_fp16)[name = string("op_30502_cast_fp16")];
            fp16 var_30503_to_fp16 = const()[name = string("op_30503_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 1, 51, 51]> scores_pos_cast_fp16 = mul(x = var_30502_cast_fp16, y = var_30503_to_fp16)[name = string("scores_pos_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> logits_cast_fp16 = add(x = scores_content_cast_fp16, y = scores_pos_cast_fp16)[name = string("logits_cast_fp16")];
            tensor<fp16, [1, 1, 51, 51]> var_30506_cast_fp16 = softmax(axis = var_29384, x = logits_cast_fp16)[name = string("op_30506_cast_fp16")];
            bool var_30508_transpose_x_1 = const()[name = string("op_30508_transpose_x_1"), val = bool(false)];
            bool var_30508_transpose_y_1 = const()[name = string("op_30508_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1, 1, 51, 128]> var_30508_cast_fp16 = matmul(transpose_x = var_30508_transpose_x_1, transpose_y = var_30508_transpose_y_1, x = var_30506_cast_fp16, y = v_head_cast_fp16)[name = string("op_30508_cast_fp16")];
            tensor<int32, [1]> o_head_axes_1 = const()[name = string("o_head_axes_1"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 51, 128]> o_head_cast_fp16 = squeeze(axes = o_head_axes_1, x = var_30508_cast_fp16)[name = string("o_head_cast_fp16")];
            bool out_47_interleave_1 = const()[name = string("out_47_interleave_1"), val = bool(false)];
            tensor<fp16, [1, 51, 1024]> out_47_cast_fp16 = concat(axis = var_29384, interleave = out_47_interleave_1, values = (var_29662_cast_fp16, var_29783_cast_fp16, var_29904_cast_fp16, var_30025_cast_fp16, var_30146_cast_fp16, var_30267_cast_fp16, var_30388_cast_fp16, o_head_cast_fp16))[name = string("out_47_cast_fp16")];
            tensor<int32, [3]> var_30512_perm_1 = const()[name = string("op_30512_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1095_axes_1 = const()[name = string("input_1095_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_30512_cast_fp16 = transpose(perm = var_30512_perm_1, x = out_47_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 1024, 51, 1]> input_1095_cast_fp16 = expand_dims(axes = input_1095_axes_1, x = var_30512_cast_fp16)[name = string("input_1095_cast_fp16")];
            string dense_output_1921_pad_type_1 = const()[name = string("dense_output_1921_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1921_strides_1 = const()[name = string("dense_output_1921_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1921_pad_1 = const()[name = string("dense_output_1921_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1921_dilations_1 = const()[name = string("dense_output_1921_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1921_groups_1 = const()[name = string("dense_output_1921_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_linear_out_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691304960))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692353600))))[name = string("layers_23_self_attn_linear_out_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1921_cast_fp16 = conv(dilations = dense_output_1921_dilations_1, groups = dense_output_1921_groups_1, pad = dense_output_1921_pad_1, pad_type = dense_output_1921_pad_type_1, strides = dense_output_1921_strides_1, weight = layers_23_self_attn_linear_out_dense_conv_weight_to_fp16_palettized, x = input_1095_cast_fp16)[name = string("dense_output_1921_cast_fp16")];
            string sparse_output_1921_pad_type_1 = const()[name = string("sparse_output_1921_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1921_strides_1 = const()[name = string("sparse_output_1921_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1921_pad_1 = const()[name = string("sparse_output_1921_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1921_dilations_1 = const()[name = string("sparse_output_1921_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1921_groups_1 = const()[name = string("sparse_output_1921_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692375232))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692354176))))[name = string("layers_23_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1921_cast_fp16 = conv(dilations = sparse_output_1921_dilations_1, groups = sparse_output_1921_groups_1, pad = sparse_output_1921_pad_1, pad_type = sparse_output_1921_pad_type_1, strides = sparse_output_1921_strides_1, weight = layers_23_self_attn_linear_out_sparse_conv_weight_to_fp16_sparsified, x = input_1095_cast_fp16)[name = string("sparse_output_1921_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> out_conv_cast_fp16 = add(x = dense_output_1921_cast_fp16, y = sparse_output_1921_cast_fp16)[name = string("out_conv_cast_fp16")];
            tensor<int32, [1]> var_30529_axes_1 = const()[name = string("op_30529_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_30529_cast_fp16 = squeeze(axes = var_30529_axes_1, x = out_conv_cast_fp16)[name = string("op_30529_cast_fp16")];
            tensor<int32, [3]> var_30530_perm_1 = const()[name = string("op_30530_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_30530_cast_fp16 = transpose(perm = var_30530_perm_1, x = var_30529_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 51, 1024]> input_1097_cast_fp16 = add(x = input_1085_cast_fp16, y = var_30530_cast_fp16)[name = string("input_1097_cast_fp16")];
            tensor<int32, [1]> x_2017_axes_1 = const()[name = string("x_2017_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_23_norm_conv_weight_to_fp16 = const()[name = string("layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692506368)))];
            tensor<fp16, [1024]> layers_23_norm_conv_bias_to_fp16 = const()[name = string("layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692508480)))];
            tensor<fp16, [1, 51, 1024]> x_2017_cast_fp16 = layer_norm(axes = x_2017_axes_1, beta = layers_23_norm_conv_bias_to_fp16, epsilon = var_29399_to_fp16, gamma = layers_23_norm_conv_weight_to_fp16, x = input_1097_cast_fp16)[name = string("x_2017_cast_fp16")];
            tensor<int32, [3]> var_30540_perm_1 = const()[name = string("op_30540_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1099_axes_1 = const()[name = string("input_1099_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_30540_cast_fp16 = transpose(perm = var_30540_perm_1, x = x_2017_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 1024, 51, 1]> input_1099_cast_fp16 = expand_dims(axes = input_1099_axes_1, x = var_30540_cast_fp16)[name = string("input_1099_cast_fp16")];
            string dense_output_1923_pad_type_1 = const()[name = string("dense_output_1923_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1923_strides_1 = const()[name = string("dense_output_1923_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1923_pad_1 = const()[name = string("dense_output_1923_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1923_dilations_1 = const()[name = string("dense_output_1923_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1923_groups_1 = const()[name = string("dense_output_1923_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_23_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692510592))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694607808))))[name = string("layers_23_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 2048, 51, 1]> dense_output_1923_cast_fp16 = conv(dilations = dense_output_1923_dilations_1, groups = dense_output_1923_groups_1, pad = dense_output_1923_pad_1, pad_type = dense_output_1923_pad_type_1, strides = dense_output_1923_strides_1, weight = layers_23_conv_pointwise_conv1_dense_conv_weight_to_fp16_palettized, x = input_1099_cast_fp16)[name = string("dense_output_1923_cast_fp16")];
            string sparse_output_1923_pad_type_1 = const()[name = string("sparse_output_1923_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1923_strides_1 = const()[name = string("sparse_output_1923_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1923_pad_1 = const()[name = string("sparse_output_1923_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1923_dilations_1 = const()[name = string("sparse_output_1923_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1923_groups_1 = const()[name = string("sparse_output_1923_groups_1"), val = int32(1)];
            tensor<fp16, [2048, 1024, 1, 1]> layers_23_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694650432))), nonzero_data = tensor<fp16, [20972]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694608384))))[name = string("layers_23_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 2048, 51, 1]> sparse_output_1923_cast_fp16 = conv(dilations = sparse_output_1923_dilations_1, groups = sparse_output_1923_groups_1, pad = sparse_output_1923_pad_1, pad_type = sparse_output_1923_pad_type_1, strides = sparse_output_1923_strides_1, weight = layers_23_conv_pointwise_conv1_sparse_conv_weight_to_fp16_sparsified, x = input_1099_cast_fp16)[name = string("sparse_output_1923_cast_fp16")];
            tensor<fp16, [1, 2048, 51, 1]> input_1101_cast_fp16 = add(x = dense_output_1923_cast_fp16, y = sparse_output_1923_cast_fp16)[name = string("input_1101_cast_fp16")];
            int32 input_1103_split_num_splits_1 = const()[name = string("input_1103_split_num_splits_1"), val = int32(2)];
            int32 input_1103_split_axis_1 = const()[name = string("input_1103_split_axis_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> input_1103_split_cast_fp16_0, tensor<fp16, [1, 1024, 51, 1]> input_1103_split_cast_fp16_1 = split(axis = input_1103_split_axis_1, num_splits = input_1103_split_num_splits_1, x = input_1101_cast_fp16)[name = string("input_1103_split_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1103_split_1_sigmoid_cast_fp16 = sigmoid(x = input_1103_split_cast_fp16_1)[name = string("input_1103_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1103_cast_fp16 = mul(x = input_1103_split_cast_fp16_0, y = input_1103_split_1_sigmoid_cast_fp16)[name = string("input_1103_cast_fp16")];
            tensor<int32, [8]> input_1105_pad_1 = const()[name = string("input_1105_pad_1"), val = tensor<int32, [8]>([0, 0, 0, 0, 4, 4, 0, 0])];
            string input_1105_mode_1 = const()[name = string("input_1105_mode_1"), val = string("constant")];
            fp16 const_2595_to_fp16 = const()[name = string("const_2595_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 1024, 59, 1]> input_1105_cast_fp16 = pad(constant_val = const_2595_to_fp16, mode = input_1105_mode_1, pad = input_1105_pad_1, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")];
            string dense_output_1925_pad_type_1 = const()[name = string("dense_output_1925_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1925_strides_1 = const()[name = string("dense_output_1925_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1925_pad_1 = const()[name = string("dense_output_1925_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1925_dilations_1 = const()[name = string("dense_output_1925_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1925_groups_1 = const()[name = string("dense_output_1925_groups_1"), val = int32(1)];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1925_cast_fp16 = conv(dilations = dense_output_1925_dilations_1, groups = dense_output_1925_groups_1, pad = dense_output_1925_pad_1, pad_type = dense_output_1925_pad_type_1, strides = dense_output_1925_strides_1, weight = layers_0_conv_depthwise_conv_dense_conv_weight_to_fp16_sparsified, x = input_1105_cast_fp16)[name = string("dense_output_1925_cast_fp16")];
            string sparse_output_1925_pad_type_1 = const()[name = string("sparse_output_1925_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1925_strides_1 = const()[name = string("sparse_output_1925_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1925_pad_1 = const()[name = string("sparse_output_1925_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1925_dilations_1 = const()[name = string("sparse_output_1925_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1925_groups_1 = const()[name = string("sparse_output_1925_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 9, 1]> layers_23_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 9, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24373056))), nonzero_data = tensor<fp16, [9216]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694912640))))[name = string("layers_23_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1925_cast_fp16 = conv(dilations = sparse_output_1925_dilations_1, groups = sparse_output_1925_groups_1, pad = sparse_output_1925_pad_1, pad_type = sparse_output_1925_pad_type_1, strides = sparse_output_1925_strides_1, weight = layers_23_conv_depthwise_conv_sparse_conv_weight_to_fp16_sparsified, x = input_1105_cast_fp16)[name = string("sparse_output_1925_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1107_cast_fp16 = add(x = dense_output_1925_cast_fp16, y = sparse_output_1925_cast_fp16)[name = string("input_1107_cast_fp16")];
            tensor<fp16, [1024]> layers_23_conv_batch_norm_running_mean_to_fp16 = const()[name = string("layers_23_conv_batch_norm_running_mean_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694931136)))];
            tensor<fp16, [1024]> layers_23_conv_batch_norm_running_var_to_fp16 = const()[name = string("layers_23_conv_batch_norm_running_var_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694933248)))];
            tensor<fp16, [1024]> layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("layers_23_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694935360)))];
            tensor<fp16, [1024]> layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("layers_23_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694937472)))];
            tensor<fp16, [1, 1024, 51, 1]> input_1109_cast_fp16 = batch_norm(beta = layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_29399_to_fp16, gamma = layers_23_conv_batch_norm_weight_to_fp16, mean = layers_23_conv_batch_norm_running_mean_to_fp16, variance = layers_23_conv_batch_norm_running_var_to_fp16, x = input_1107_cast_fp16)[name = string("input_1109_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> input_1111_cast_fp16 = silu(x = input_1109_cast_fp16)[name = string("input_1111_cast_fp16")];
            string dense_output_1927_pad_type_1 = const()[name = string("dense_output_1927_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1927_strides_1 = const()[name = string("dense_output_1927_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1927_pad_1 = const()[name = string("dense_output_1927_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1927_dilations_1 = const()[name = string("dense_output_1927_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1927_groups_1 = const()[name = string("dense_output_1927_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(694939584))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695988224))))[name = string("layers_23_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 51, 1]> dense_output_1927_cast_fp16 = conv(dilations = dense_output_1927_dilations_1, groups = dense_output_1927_groups_1, pad = dense_output_1927_pad_1, pad_type = dense_output_1927_pad_type_1, strides = dense_output_1927_strides_1, weight = layers_23_conv_pointwise_conv2_dense_conv_weight_to_fp16_palettized, x = input_1111_cast_fp16)[name = string("dense_output_1927_cast_fp16")];
            string sparse_output_1927_pad_type_1 = const()[name = string("sparse_output_1927_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1927_strides_1 = const()[name = string("sparse_output_1927_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1927_pad_1 = const()[name = string("sparse_output_1927_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1927_dilations_1 = const()[name = string("sparse_output_1927_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1927_groups_1 = const()[name = string("sparse_output_1927_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 1024, 1, 1]> layers_23_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696009856))), nonzero_data = tensor<fp16, [10486]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695988800))))[name = string("layers_23_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 51, 1]> sparse_output_1927_cast_fp16 = conv(dilations = sparse_output_1927_dilations_1, groups = sparse_output_1927_groups_1, pad = sparse_output_1927_pad_1, pad_type = sparse_output_1927_pad_type_1, strides = sparse_output_1927_strides_1, weight = layers_23_conv_pointwise_conv2_sparse_conv_weight_to_fp16_sparsified, x = input_1111_cast_fp16)[name = string("sparse_output_1927_cast_fp16")];
            tensor<fp16, [1, 1024, 51, 1]> x_2019_cast_fp16 = add(x = dense_output_1927_cast_fp16, y = sparse_output_1927_cast_fp16)[name = string("x_2019_cast_fp16")];
            tensor<int32, [1]> var_30596_axes_1 = const()[name = string("op_30596_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 51]> var_30596_cast_fp16 = squeeze(axes = var_30596_axes_1, x = x_2019_cast_fp16)[name = string("op_30596_cast_fp16")];
            tensor<int32, [3]> var_30597_perm_1 = const()[name = string("op_30597_perm_1"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 51, 1024]> var_30597_cast_fp16 = transpose(perm = var_30597_perm_1, x = var_30596_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 51, 1024]> input_1113_cast_fp16 = add(x = input_1097_cast_fp16, y = var_30597_cast_fp16)[name = string("input_1113_cast_fp16")];
            tensor<int32, [1]> x_2021_axes_1 = const()[name = string("x_2021_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696140992)))];
            tensor<fp16, [1024]> layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696143104)))];
            tensor<fp16, [1, 51, 1024]> x_2021_cast_fp16 = layer_norm(axes = x_2021_axes_1, beta = layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_29399_to_fp16, gamma = layers_23_norm_feed_forward2_weight_to_fp16, x = input_1113_cast_fp16)[name = string("x_2021_cast_fp16")];
            tensor<int32, [4]> var_30607 = const()[name = string("op_30607"), val = tensor<int32, [4]>([1, 51, 1, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_2023_cast_fp16 = reshape(shape = var_30607, x = x_2021_cast_fp16)[name = string("x_2023_cast_fp16")];
            tensor<int32, [4]> input_1115_perm_1 = const()[name = string("input_1115_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string dense_output_1929_pad_type_1 = const()[name = string("dense_output_1929_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_1929_strides_1 = const()[name = string("dense_output_1929_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_1929_pad_1 = const()[name = string("dense_output_1929_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_1929_dilations_1 = const()[name = string("dense_output_1929_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_1929_groups_1 = const()[name = string("dense_output_1929_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_23_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696145216))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700339584))))[name = string("layers_23_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> input_1115_cast_fp16 = transpose(perm = input_1115_perm_1, x = x_2023_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 4096, 1, 51]> dense_output_1929_cast_fp16 = conv(dilations = dense_output_1929_dilations_1, groups = dense_output_1929_groups_1, pad = dense_output_1929_pad_1, pad_type = dense_output_1929_pad_type_1, strides = dense_output_1929_strides_1, weight = layers_23_feed_forward2_linear1_dense_conv_weight_to_fp16_palettized, x = input_1115_cast_fp16)[name = string("dense_output_1929_cast_fp16")];
            string sparse_output_1929_pad_type_1 = const()[name = string("sparse_output_1929_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_1929_strides_1 = const()[name = string("sparse_output_1929_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_1929_pad_1 = const()[name = string("sparse_output_1929_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_1929_dilations_1 = const()[name = string("sparse_output_1929_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_1929_groups_1 = const()[name = string("sparse_output_1929_groups_1"), val = int32(1)];
            tensor<fp16, [4096, 1024, 1, 1]> layers_23_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [4096, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700424128))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700340160))))[name = string("layers_23_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 4096, 1, 51]> sparse_output_1929_cast_fp16 = conv(dilations = sparse_output_1929_dilations_1, groups = sparse_output_1929_groups_1, pad = sparse_output_1929_pad_1, pad_type = sparse_output_1929_pad_type_1, strides = sparse_output_1929_strides_1, weight = layers_23_feed_forward2_linear1_sparse_conv_weight_to_fp16_sparsified, x = input_1115_cast_fp16)[name = string("sparse_output_1929_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1117_cast_fp16 = add(x = dense_output_1929_cast_fp16, y = sparse_output_1929_cast_fp16)[name = string("input_1117_cast_fp16")];
            tensor<fp16, [1, 4096, 1, 51]> input_1119_cast_fp16 = silu(x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")];
            string dense_output_pad_type_1 = const()[name = string("dense_output_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> dense_output_strides_1 = const()[name = string("dense_output_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> dense_output_pad_1 = const()[name = string("dense_output_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> dense_output_dilations_1 = const()[name = string("dense_output_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 dense_output_groups_1 = const()[name = string("dense_output_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_23_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700948480))), lut = tensor<fp16, [1, 1, 1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705142848))))[name = string("layers_23_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized")];
            tensor<fp16, [1, 1024, 1, 51]> dense_output_cast_fp16 = conv(dilations = dense_output_dilations_1, groups = dense_output_groups_1, pad = dense_output_pad_1, pad_type = dense_output_pad_type_1, strides = dense_output_strides_1, weight = layers_23_feed_forward2_linear2_dense_conv_weight_to_fp16_palettized, x = input_1119_cast_fp16)[name = string("dense_output_cast_fp16")];
            string sparse_output_pad_type_1 = const()[name = string("sparse_output_pad_type_1"), val = string("valid")];
            tensor<int32, [2]> sparse_output_strides_1 = const()[name = string("sparse_output_strides_1"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> sparse_output_pad_1 = const()[name = string("sparse_output_pad_1"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> sparse_output_dilations_1 = const()[name = string("sparse_output_dilations_1"), val = tensor<int32, [2]>([1, 1])];
            int32 sparse_output_groups_1 = const()[name = string("sparse_output_groups_1"), val = int32(1)];
            tensor<fp16, [1024, 4096, 1, 1]> layers_23_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified = constexpr_sparse_to_dense(mask = tensor<uint1, [1024, 4096, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705227392))), nonzero_data = tensor<fp16, [41944]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705143424))))[name = string("layers_23_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified")];
            tensor<fp16, [1, 1024, 1, 51]> sparse_output_cast_fp16 = conv(dilations = sparse_output_dilations_1, groups = sparse_output_groups_1, pad = sparse_output_pad_1, pad_type = sparse_output_pad_type_1, strides = sparse_output_strides_1, weight = layers_23_feed_forward2_linear2_sparse_conv_weight_to_fp16_sparsified, x = input_1119_cast_fp16)[name = string("sparse_output_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 51]> x_2025_cast_fp16 = add(x = dense_output_cast_fp16, y = sparse_output_cast_fp16)[name = string("x_2025_cast_fp16")];
            tensor<int32, [4]> x_perm_1 = const()[name = string("x_perm_1"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [3]> var_30642 = const()[name = string("op_30642"), val = tensor<int32, [3]>([1, 51, 1024])];
            tensor<fp16, [1, 51, 1, 1024]> x_cast_fp16 = transpose(perm = x_perm_1, x = x_2025_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 51, 1024]> var_30643_cast_fp16 = reshape(shape = var_30642, x = x_cast_fp16)[name = string("op_30643_cast_fp16")];
            fp16 var_30644_to_fp16 = const()[name = string("op_30644_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 51, 1024]> var_30645_cast_fp16 = mul(x = var_30643_cast_fp16, y = var_30644_to_fp16)[name = string("op_30645_cast_fp16")];
            tensor<fp16, [1, 51, 1024]> input_cast_fp16 = add(x = input_1113_cast_fp16, y = var_30645_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_30650_axes_1 = const()[name = string("op_30650_axes_1"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> layers_23_norm_out_weight_to_fp16 = const()[name = string("layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705751744)))];
            tensor<fp16, [1024]> layers_23_norm_out_bias_to_fp16 = const()[name = string("layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705753856)))];
            tensor<fp16, [1, 51, 1024]> output = layer_norm(axes = var_30650_axes_1, beta = layers_23_norm_out_bias_to_fp16, epsilon = var_29399_to_fp16, gamma = layers_23_norm_out_weight_to_fp16, x = input_cast_fp16)[name = string("op_30650_cast_fp16")];
        } -> (output);
}