Qwen

/

Qwen3-30B-A3B-Instruct-2507

like 686

Follow

Qwen 60.5k

Hello,
I'm trying to run this model with vLLM utilizing dual chunk flash attention.
However, it showed following errors.

vLLM v0.11.1rc2

Value error, Invalid value 'DUAL_CHUNK_FLASH_ATTN' for VLLM_ATTENTION_BACKEND. Valid options: ['FLASH_ATTN', 'TRITON_ATTN', 'XFORMERS', 'ROCM_ATTN', 'ROCM_AITER_MLA', 'ROCM_AITER_FA', 'TORCH_SDPA', 'FLASHINFER', 'FLASHINFER_MLA', 'TRITON_MLA', 'CUTLASS_MLA', 'FLASHMLA', 'FLASHMLA_SPARSE', 'FLASH_ATTN_MLA', 'PALLAS', 'IPEX', 'NO_ATTENTION', 'FLEX_ATTENTION', 'TREE_ATTN', 'ROCM_AITER_UNIFIED_ATTN']. [type=value_error, input_value=ArgsKwargs((), {'model': ...rocessor_plugin': None}), input_type=ArgsKwargs]

vLLM v0.10.2

  File "/usr/local/lib/python3.12/dist-packages/vllm/model_executor/models/qwen3_moe.py", line 399, in <lambda>
    lambda prefix: Qwen3MoeDecoderLayer(config=config,
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/vllm/model_executor/models/qwen3_moe.py", line 312, in __init__
    self.self_attn = Qwen3MoeAttention(
                     ^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/vllm/model_executor/models/qwen3_moe.py", line 253, in __init__
    self.attn = Attention(
                ^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/vllm/attention/layer.py", line 182, in __init__
    self.impl = impl_cls(num_heads, head_size, scale, num_kv_heads,
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/vllm/attention/backends/dual_chunk_flash_attn.py", line 325, in __init__
    assert dual_chunk_attention_config is not None
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError

Would you help me to find the right config or vLLM version to utilize this model with DUAL_CHUNK_FLASH_ATTN?