First model version

README.md

@@ -1,3 +1,141 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+library_name: transformers
+tags:
+- dllm
+- diffusion
+- llm
+- text_generation
+---
+# LLaDA2.0-flash-CAP
+
+**LLaDA2.0-flash-CAP** is an enhanced version of LLaDA2.0-flash that incorporates **Confidence-Aware Parallel (CAP) Training** for significantly improved inference efficiency. Built upon the 100B-A6B Mixture-of-Experts (MoE) diffusion architecture, this model achieves faster parallel decoding while maintaining strong performance across diverse benchmarks.
+
+---
+
+## 📊 Performance Comparison
+### Efficiency vs. Quality Trade-off
+| Model | Average Score | Tokens/Forward (TPF) | Speedup |
+| :---: | :---: | :---: | :---: |
+| LLaDA2.0-flash | 78.57 | 3.19 | 1.0× |
+| **LLaDA2.0-flash-CAP** | **76.85** | **4.65** | **1.46×** |
+
+
+_Evaluated on 12 diverse benchmarks covering knowledge, reasoning, coding, and mathematics._
+
+### Key Insights
++ **1.46× faster generation** with only a 1.72% performance trade-off
++ Ideal for latency-sensitive applications requiring real-time responses
++ Maintains competitive accuracy across all task categories
+
+---
+
+## 🔬 What is CAP Training?
+**Confidence-Aware Parallel (CAP) Training** is a novel training technique designed to enhance parallel decoding efficiency in diffusion language models.
+
+### Technical Overview
+The training objective combines two complementary losses:
+
+```math
+L(θ) = L_SFT(θ) + λL_conf(θ)
+```
+
+Where:
+
++ **L_SFT**: Supervised fine-tuning loss ensuring prediction correctness
++ **L_conf**: Confidence loss that minimizes entropy only for correctly predicted tokens
++ **λ**: Hyperparameter balancing the two objectives
+
+### Why CAP Works
+1. **Sharpens Correct Predictions**: While standard training ensures correctness, it provides diminishing incentive to increase confidence on already-correct tokens. CAP explicitly optimizes for high-confidence predictions.
+2. **Enables Aggressive Parallelism**: Higher confidence allows the model to decode multiple tokens simultaneously with greater reliability, reducing the total number of forward passes needed.
+3. **Selective Optimization**: By focusing only on correct predictions, CAP avoids penalizing the model's exploration of uncertain outputs.
+
+---
+
+## 📦 Model Variants
+| Model ID | Description | Hugging Face Link |
+| --- | --- | --- |
+| `inclusionAI/LLaDA2.0-flash-CAP` | CAP-enhanced model optimized for fast inference | [🤗 Model Card](https://huggingface.co/inclusionAI/LLaDA2.0-flash-CAP) |
+| `inclusionAI/LLaDA2.0-flash` | Base instruction-tuned model | [🤗 Model Card](https://huggingface.co/inclusionAI/LLaDA2.0-flash) |
+
+
+---
+
+## 🔍 Model Overview
+**LLaDA2.0-flash-CAP** inherits the architecture of LLaDA2.0-flash:
+
++ **Type**: Mixture-of-Experts (MoE) Diffusion Language Model
++ **Total Parameters (Non-Embedding)**: 100B
++ **Number of Layers**: 32
++ **Attention Heads**: 32
++ **Context Length**: 32,768 tokens
++ **Position Embedding**: Rotary (RoPE)
++ **Vocabulary Size**: 157,184
++ **Training Enhancement**: Confidence-Aware Parallel (CAP) Training
+
+---
+
+## 💻 Usage
+### 🤗 Hugging Face Transformers
+```python
+import torch
+import torch.nn.functional as F
+from transformers import AutoModelForCausalLM
+from transformers import AutoTokenizer
+
+model_path = "/path/to/LLaDA2.0-flash-CAP"
+device = "cuda:0"
+model = AutoModelForCausalLM.from_pretrained(
+    model_path, trust_remote_code=True, device_map=device
+)
+model = model.to(torch.bfloat16)
+model.eval()
+tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+prompt = "Why does Camus think that Sisyphus is happy?"
+input_ids = tokenizer.apply_chat_template(
+    [{"role": "user", "content": prompt}],
+    add_generation_prompt=True,
+    tokenize=True,
+    return_tensors="pt",
+)
+generated_tokens = model.generate(
+    inputs=input_ids,
+    eos_early_stop=True,
+    gen_length=512,
+    block_length=32,
+    steps=32,
+    temperature=0.0,
+)
+generated_answer = tokenizer.decode(
+    generated_tokens[0],
+    skip_special_tokens=True,
+)
+print(generated_answer)
+```
+
+
+
+### Best Practices
+To achieve optimal performance, we recommend the following settings:
+
+1. **Sampling Parameters**:  
+We suggest using `Temperature=0.0`, `block_length=32`, and `steps=32`. Using a higher temperature value may occasionally result in language mixing and a slight decrease in model performance.
+2. **Adequate Output Length**:  
+We recommend using an output length of 32768 tokens for most queries.
+
+---
+
+## 🌐 License
+This project is licensed under the terms of the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).
+
+---
+
+## 🤝 Contact & Collaboration
+For questions, collaborations, or feedback, please reach out via [Hugging Face](https://huggingface.co/inclusionAI/LLaDA2.0-flash-CAP) or open an issue in the [repository](https://github.com/inclusionAI).
+
+👉 Join us in advancing open, efficient, and intelligent language models!
+
+---
+

config.json

@@ -0,0 +1,56 @@
+{
+  "architectures": [
+    "LLaDA2MoeModelLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_llada2_moe.LLaDA2MoeConfig",
+    "AutoModel": "modeling_llada2_moe.LLaDA2MoeModel",
+    "AutoModelForCausalLM": "modeling_llada2_moe.LLaDA2MoeModelLM"
+  },
+  "embedding_dropout": 0.0,
+  "first_k_dense_replace": 1,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 9216,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "llada2_moe",
+  "moe_intermediate_size": 1024,
+  "moe_router_enable_expert_bias": true,
+  "n_group": 8,
+  "norm_head": false,
+  "norm_softmax": false,
+  "norm_topk_prob": true,
+  "num_attention_heads": 32,
+  "num_experts": 256,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 4,
+  "num_shared_experts": 1,
+  "output_dropout": 0.0,
+  "output_router_logits": false,
+  "pad_token_id": 156892,
+  "partial_rotary_factor": 0.5,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 600000,
+  "rotary_dim": 64,
+  "routed_scaling_factor": 2.5,
+  "router_dtype": "fp32",
+  "score_function": "sigmoid",
+  "sliding_window": 4096,
+  "tie_word_embeddings": false,
+  "topk_group": 4,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.0",
+  "use_bias": false,
+  "use_cache": false,
+  "use_qkv_bias": false,
+  "use_rmsnorm": true,
+  "use_sliding_window": false,
+  "using_split_qkv_in_self_attention": false,
+  "vocab_size": 157184
+}

configuration_llada2_moe.py

@@ -0,0 +1,87 @@
+"""LLaDA2 MoE model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class LLaDA2MoeConfig(PretrainedConfig):
+    model_type = "llada2_moe"
+
+    def __init__(
+        self,
+        vocab_size=30592,
+        hidden_size=1024,
+        intermediate_size=None,
+        num_hidden_layers=24,
+        num_attention_heads=16,
+        num_key_value_heads=0,
+        hidden_act="silu",
+        use_qkv_bias=False,  # llada2 only
+        use_qk_norm=False,
+        use_bias=True,  # llada2 only
+        rms_norm_eps=1e-05,
+        norm_head=False,  # llada2 only
+        tie_word_embeddings=False,  # PretrainedConfig key, here change default value.
+        embedding_dropout=0.1,
+        attention_dropout=0.1,
+        output_dropout=0.1,
+        initializer_range=0.02,
+        max_position_embeddings=16384,
+        rope_theta=10000.0,
+        use_cache=True,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        rope_scaling=None,
+        pad_token_id=126081,
+        num_experts=16,
+        num_shared_experts=0,
+        num_experts_per_tok=2,
+        n_group=8,
+        topk_group=4,
+        routed_scaling_factor=2.5,
+        moe_intermediate_size=None,
+        first_k_dense_replace=0,
+        head_dim=None,
+        output_router_logits=False,
+        partial_rotary_factor=0.5,
+        **kwargs,
+    ):
+        self.num_hidden_layers = num_hidden_layers
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.use_qkv_bias = use_qkv_bias
+        self.use_bias = use_bias
+        self.norm_head = norm_head
+        self.rms_norm_eps = rms_norm_eps
+        self.embedding_dropout = embedding_dropout
+        self.attention_dropout = attention_dropout
+        self.output_dropout = output_dropout
+        self.initializer_range = initializer_range
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_theta = rope_theta
+        self.use_cache = use_cache
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.max_window_layers = max_window_layers
+        self.head_dim = head_dim or self.hidden_size // self.num_attention_heads
+        self.rope_scaling = rope_scaling
+
+        # MoE configs
+        self.num_experts = num_experts
+        self.num_shared_experts = num_shared_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.moe_intermediate_size = moe_intermediate_size
+        self.first_k_dense_replace = first_k_dense_replace
+        self.output_router_logits = output_router_logits
+        self.routed_scaling_factor = routed_scaling_factor
+        self.partial_rotary_factor = partial_rotary_factor
+
+        super().__init__(
+            pad_token_id=pad_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
+        )

model-00001-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a69b721654eb565d97d87b89ad7795338d23f6778658012528092774cedd3e
+size 4995484480

model-00002-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2383dc18cd853afe26d35eaa04823df41e091d064e25b30c5c9a361e43b5f044
+size 4993409784

model-00003-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36795163f3e7a7272e8b99b2d7ae168e5d07015cfa7dbd9d4536017ca6297596
+size 4993409464

model-00004-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d453ecf6ba690b46cb5d67a42053b80b020e8befb2fafb24f175986e6ed24145
+size 4999684432

model-00005-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ff3526a3b9ccee49147448539175b2fd07d6ab5a6c3baa284bc821ff254391e
+size 4993409648

model-00006-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04cd3ac2a284af7b29212fd27d9bbec5f306e6298a12cc2cbea4daa2bfc372cc
+size 4993409864

model-00007-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67d32d6a5d8685f368657323fe52f0b4281e05f8e0d87e86e8549f8a3f6b7da1
+size 4993409424

model-00008-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61f4df4da00fcc8be046d53d8e36d40c99ac8c830bf419373c39353961b553cc
+size 4999684472

model-00009-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3dd7933cd4d2560cad02d069af99c9d59212e48c75ee2a0364f7202a50232cd
+size 4993409576

model-00010-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6cf0d073bc68e70edaaac58e85e3349336e1d0d82801720c75a37976e9b6d2b
+size 4993409824

model-00011-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2edf764d30fe4de9d15e640418a4a87f32fe9ed1f4d2585d261cc401e58df907
+size 4993409512

model-00012-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:345de49ff058831f927ac611607e1278045d3bfd694d5f4f1daf467f60d2b7b5
+size 4999684504

model-00013-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a1cbc4d3b694186024588857cfd6917d14a92d7575a53ae4cab2bf844dfd4dc
+size 4993410040

model-00014-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a704b28f8b7b1227e8bc62b860e53d1813beece7db563998c6f9e0d9e1f4f6b9
+size 4993410376

model-00015-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2fe0a734958f744a2d7b5a2c3faa11f899fff475fb0701260fff18d36d0fe4b
+size 4993410232

model-00016-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59c84495781bfab04c73558091aefebfc95727d9b7fcfd210dc7bb0164927177
+size 4999685104

model-00017-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8dacf8af93a32160685dea1b20e7499bd23b2a0d370a99ad1b896d09c1b02f2
+size 4993410040

model-00018-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8740773e62a31401e1bd29a4a635bc5ce35b4ea924ade07073a45e8069d9f392
+size 4993410312

model-00019-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a405f2fbd957855c94ae6b2303c9e7c6fb634cfa3df5333d9a2a638dfd0a50dc
+size 4993410328

model-00020-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1b10d2a846b634b1ad4a854f0fcf21cef44d3ce0b0fb0acac77cd32953b5855
+size 4993410080

model-00021-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2c43057a1658cf9d65c1a7895ccee505fca7c163c54e9319a1da348fc737b77
+size 4999685016

model-00022-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08fbd68f6e439629354c4838b9fc047c3139878b553b47bd68278bdf5ad7e0e2
+size 4993410272

model-00023-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad41131311ee87fa0b638759333d9ee60d66ef027b4695e6add9e41baaffa9a6
+size 4993410416

model-00024-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7851edcf9ea6a5bc833fe4301513154c053eefea2efecafaa9edf5563ea6f219
+size 4993410032

model-00025-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:931bba910813da3e18f30c9647ad27d1fce891ce83b02f9db24c104208947060
+size 4999685048

model-00026-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71895f3407434d64c85008014dcc9507f19ea71a51822221a6aa9750f3cf41ea
+size 4993410200

model-00027-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bfb3e4bc12458fa4ba5aa570170cbd051b94801a405046205c2bc71441a08e2
+size 4993410432

model-00028-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ef90ef0e57ba2389d3046fc7aa4e8fde9c3843a41639949a6bb8785b269d7ce
+size 4993410056

model-00029-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7bdc968de461947117519cdfd6896e6690f0d7ca8733d30b2a642e254eead8dc
+size 4999685088

model-00030-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cbbf061a8a63ffe3e4d0c2d20ec3d72a475d1721db7123ac9ade0676504ac0d3
+size 4993410128

model-00031-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68501a88c3cbb8616a951f6cd17c67ba9eeccc640f515d4297ad6324f0dc6b6d
+size 4993410392

model-00032-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:505b26ead99b47f2be748e3f1febae299c54d46f78c614b93203366cffe0d61f
+size 4993410168

model-00033-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dc2530db28c60078b8578532193f9d81b29ac1544e659200419023b69c6631e
+size 4999685104

model-00034-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fbc289edaaa517a2a7e730e19a4f66abfa74a1d1d7e0fd243b9c0012395edab
+size 4993410064

model-00035-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4312862c80a9ffbcddf6e3a5ea4cc61da97edd30ff57e48788ab0b085f186e61
+size 4993410336

model-00036-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc721a1a7cccc79d437180c840cf9f511ca877003d98837b2cfa038c18957547
+size 4993410288

model-00037-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e32be857d2bf35d0b94d249c409462bd9a64d0592016054d5c0db992ba1d783
+size 4999685104

model-00038-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa00f7cb3dda9422d61a2950e2f73bf1e3aa5126ece721029994fe6ca3c43953
+size 4993410024

model-00039-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58a80b2cc4a612cb87761d0dbbae74853349a758b8a75043733a24a3ca6f21f6
+size 4993410296

model-00040-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8655edc8ae867a805933e2636f694a2d7843a0e411adba813585d4d68c648820
+size 4993410368

model-00041-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6111dd16362a80b1df3e8e70715b965a366f959bd8f5a1f12bcfe9998ccc82b
+size 4699812800

model-00042-of-00042.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f55d21f4f9523b17cd9e3402f83741f9b08fa94ddfe65c72f9c802f3afe4e58
+size 1287670696

modeling_llada2_moe.py

@@ -0,0 +1,1804 @@
+# Copyright 2025 Antgroup and The HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch LLaDA2MoE model."""
+
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import (
+    _prepare_4d_causal_attention_mask,
+    _prepare_4d_causal_attention_mask_for_sdpa,
+)
+from transformers.modeling_outputs import (
+    MoeModelOutputWithPast,
+    MoeCausalLMOutputWithPast,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import (
+    ALL_LAYERNORM_LAYERS,
+    is_torch_greater_or_equal_than_1_13,
+)
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.import_utils import is_torch_fx_available
+from .configuration_llada2_moe import LLaDA2MoeConfig
+from transformers.generation.utils import GenerationMixin
+
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
+# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
+# It means that the function will not be traced through and simply appear as a node in the graph.
+if is_torch_fx_available():
+    if not is_torch_greater_or_equal_than_1_13:
+        import torch.fx
+
+    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LLaDA2MoeConfig"
+
+
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(
+        torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0)
+    )
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+class LLaDA2MoeRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        LLaDA2MoeRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+ALL_LAYERNORM_LAYERS.append(LLaDA2MoeRMSNorm)
+
+
+class LLaDA2MoeRotaryEmbedding(nn.Module):
+    def __init__(self, config: LLaDA2MoeConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get(
+                "rope_type", config.rope_scaling.get("type")
+            )
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = (
+            self.inv_freq[None, :, None]
+            .float()
+            .expand(position_ids.shape[0], -1, 1)
+            .to(x.device)
+        )
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = (
+            x.device.type
+            if isinstance(x.device.type, str) and x.device.type != "mps"
+            else "cpu"
+        )
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (
+                inv_freq_expanded.float() @ position_ids_expanded.float()
+            ).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    # Keep half or full tensor for later concatenation
+    rotary_dim = cos.shape[-1]
+    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
+    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
+
+    # Apply rotary embeddings on the first half or full tensor
+    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
+    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)
+
+    # Concatenate back to full shape
+    q_embed = torch.cat([q_embed, q_pass], dim=-1)
+    k_embed = torch.cat([k_embed, k_pass], dim=-1)
+    return q_embed, k_embed
+
+
+class LLaDA2MoeMLP(nn.Module):
+    def __init__(self, config: LLaDA2MoeConfig, intermediate_size: int):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class LLaDA2MoeGate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.num_experts = config.num_experts
+
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+
+        # topk selection algorithm
+        self.gating_dim = config.hidden_size
+        self.weight = nn.Parameter(torch.empty((self.num_experts, self.gating_dim)))
+        self.routed_scaling_factor = config.routed_scaling_factor
+
+        self.register_buffer("expert_bias", torch.zeros(self.num_experts))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        import torch.nn.init as init
+
+        init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+
+    def group_limited_topk(
+        self,
+        scores: torch.Tensor,
+    ):
+        num_tokens, _ = scores.size()
+        # Organize the experts into groups
+        group_scores = (
+            scores.view(num_tokens, self.n_group, -1).topk(2, dim=-1)[0].sum(dim=-1)
+        )
+        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+
+        # Mask the experts based on selection groups
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(num_tokens, self.n_group, self.num_experts // self.n_group)
+            .reshape(num_tokens, -1)
+        )
+
+        masked_scores = scores.masked_fill(~score_mask.bool(), float("-inf"))
+        probs, top_indices = torch.topk(masked_scores, k=self.top_k, dim=-1)
+
+        return probs, top_indices
+
+    def forward(self, hidden_states):
+        # compute gating score
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        logits = F.linear(
+            hidden_states.type(torch.float32), self.weight.type(torch.float32)
+        )
+
+        scores = torch.sigmoid(logits.float()).type_as(logits)
+
+        scores_for_routing = scores + self.expert_bias
+        _, topk_idx = self.group_limited_topk(scores_for_routing)
+
+        scores = torch.gather(scores, dim=1, index=topk_idx).type_as(logits)
+
+        topk_weight = (
+            scores / (scores.sum(dim=-1, keepdim=True) + 1e-20)
+            if self.top_k > 1
+            else scores
+        )
+        topk_weight = topk_weight * self.routed_scaling_factor
+
+        return topk_idx, topk_weight, logits
+
+
+class LLaDA2MoeSparseMoeBlock(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config: LLaDA2MoeConfig):
+        super().__init__()
+        self.config = config
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self._setup_experts()
+        self.gate = LLaDA2MoeGate(config)
+        if config.num_shared_experts is not None:
+            self.shared_experts = LLaDA2MoeMLP(
+                config=config,
+                intermediate_size=config.moe_intermediate_size
+                * config.num_shared_experts,
+            )
+
+    def _setup_experts(self):
+        self.experts = nn.ModuleList(
+            [
+                LLaDA2MoeMLP(
+                    config=self.config,
+                    intermediate_size=self.config.moe_intermediate_size,
+                )
+                for _ in range(self.config.num_experts)
+            ]
+        )
+
+    def forward(self, hidden_states):
+        identity = hidden_states
+        bsz, seq_len, h = hidden_states.shape
+        topk_idx, topk_weight, router_logits = self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        flat_topk_idx = topk_idx.view(-1)
+        if self.training:
+            hidden_states = hidden_states.repeat_interleave(
+                self.num_experts_per_tok, dim=0
+            )
+            y = torch.empty_like(hidden_states)
+            for i, expert in enumerate(self.experts):
+                y[flat_topk_idx == i] = expert(hidden_states[flat_topk_idx == i])
+            y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
+            y = y.to(hidden_states.dtype).view(bsz, seq_len, h)
+        else:
+            y = self.moe_infer(hidden_states, topk_idx, topk_weight).view(
+                bsz, seq_len, h
+            )
+        if self.config.num_shared_experts is not None:
+            y = y + self.shared_experts(identity)
+        return y, (
+            router_logits.view(bsz, seq_len, -1),
+            topk_idx.view(bsz, seq_len, -1),
+        )
+
+    @torch.no_grad()
+    def moe_infer(self, x, topk_ids, topk_weight):
+        cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
+        cnts.scatter_(1, topk_ids, 1)
+        tokens_per_expert = cnts.sum(dim=0)
+        idxs = topk_ids.view(-1).argsort()
+        sorted_tokens = x[idxs // topk_ids.shape[1]]
+        tokens_per_expert = tokens_per_expert.cpu().numpy()
+        outputs = []
+        start_idx = 0
+        for i, num_tokens in enumerate(tokens_per_expert):
+            end_idx = start_idx + num_tokens
+            if num_tokens == 0:
+                continue
+            expert = self.experts[i]
+            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
+            expert_out = expert(tokens_for_this_expert)
+            outputs.append(expert_out.to(x.device))
+            start_idx = end_idx
+
+        outs = torch.cat(outputs, dim=0) if len(outputs) else sorted_tokens.new_empty(0)
+        new_x = torch.empty_like(outs)
+        new_x[idxs] = outs
+        final_out = (
+            new_x.view(*topk_ids.shape, -1)
+            .type(topk_weight.dtype)
+            .mul_(topk_weight.unsqueeze(dim=-1))
+            .sum(dim=1)
+            .type(new_x.dtype)
+        )
+        return final_out
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(
+        batch, num_key_value_heads, n_rep, slen, head_dim
+    )
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaAttention with Llama->LLaDA2Moe
+class LLaDA2MoeAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: LLaDA2MoeConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = config.head_dim or self.hidden_size // self.num_heads
+        partial_rotary_factor = (
+            config.partial_rotary_factor
+            if hasattr(config, "partial_rotary_factor")
+            else 1.0
+        )
+        self.rope_dim = int(self.head_dim * partial_rotary_factor)
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.is_causal = False
+
+        self.query_key_value = nn.Linear(
+            self.hidden_size,
+            (self.num_heads + 2 * self.num_key_value_heads) * self.head_dim,
+            bias=config.use_qkv_bias,
+        )
+
+        self.query_layernorm = LLaDA2MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.key_layernorm = LLaDA2MoeRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.dense = nn.Linear(
+            self.num_heads * self.head_dim, self.hidden_size, bias=config.use_bias
+        )
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return (
+            tensor.view(bsz, seq_len, self.num_heads, self.head_dim)
+            .transpose(1, 2)
+            .contiguous()
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        position_embeddings: Optional[
+            Tuple[torch.Tensor, torch.Tensor]
+        ] = None,  # necessary, but kept here for BC
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.query_key_value(hidden_states)
+        qkv = qkv.view(
+            bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim
+        )
+
+        query_states, key_states, value_states = qkv.split(
+            [self.num_heads, self.num_key_value_heads, self.num_key_value_heads], dim=-2
+        )
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        query_states = self.query_layernorm(query_states)
+        key_states = self.key_layernorm(key_states)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin, position_ids
+        )
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(
+            query_states, key_states.transpose(2, 3)
+        ) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+        # attention_mask = None
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(
+            attn_weights, dim=-1, dtype=torch.float32
+        ).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(
+            attn_weights, p=self.attention_dropout, training=self.training
+        )
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+
+        attn_output = self.dense(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2 with Llama->LLaDA2Moe
+class LLaDA2MoeFlashAttention2(LLaDA2MoeAttention):
+    """
+    LLaDA2Moe flash attention module. This module inherits from `LLaDA2MoeAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        position_embeddings: Optional[
+            Tuple[torch.Tensor, torch.Tensor]
+        ] = None,  # necessary, but kept here for BC
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # LLaDA2MoeFlashAttention2 attention does not support output_attentions
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+
+            # overwrite attention_mask with padding_mask
+            attention_mask = kwargs.pop("padding_mask")
+
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+
+        qkv = self.query_key_value(hidden_states)
+        qkv = qkv.view(
+            bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim
+        )
+
+        query_states, key_states, value_states = qkv.split(
+            [self.num_heads, self.num_key_value_heads, self.num_key_value_heads], dim=-2
+        )
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        query_states = self.query_layernorm(query_states)
+        key_states = self.key_layernorm(key_states)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin, position_ids
+        )
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.attention_dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently cast in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slow down training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (LLaDA2MoeRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            # Handle the case where the model is quantized
+            if hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            elif torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            else:
+                target_dtype = self.query_key_value.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
+        attn_output = self.dense(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    def _flash_attention_forward(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        query_length,
+        dropout=0.0,
+        softmax_scale=None,
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+            query_length (`int`):
+                The length of the query sequence in terms of tokens. This represents the number of tokens in the
+                `query_states` tensor along the sequence dimension. It is used to determine the effective sequence
+                length for attention computations.
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LLaDA2MoeFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # attention_mask = None
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            (
+                query_states,
+                key_states,
+                value_states,
+                indices_q,
+                cu_seq_lens,
+                max_seq_lens,
+            ) = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(
+                attn_output_unpad, indices_q, batch_size, query_length
+            )
+        else:
+            attn_output = flash_attn_func(
+                query_states,
+                key_states,
+                value_states,
+                dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+        return attn_output
+
+    def _upad_input(
+        self, query_layer, key_layer, value_layer, attention_mask, query_length
+    ):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
+            indices_k,
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
+            indices_k,
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim),
+                indices_k,
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(
+                query_layer, attention_mask
+            )
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->LLaDA2Moe
+class LLaDA2MoeSdpaAttention(LLaDA2MoeAttention):
+    """
+    LLaDA2Moe attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `LLaDA2MoeAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from LLaDA2MoeAttention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        position_embeddings: Optional[
+            Tuple[torch.Tensor, torch.Tensor]
+        ] = None,  # necessary, but kept here for BC
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "LLaDA2MoeModel is using LLaDA2MoeSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.query_key_value(hidden_states)
+        qkv = qkv.view(
+            bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim
+        )
+
+        query_states, key_states, value_states = qkv.split(
+            [self.num_heads, self.num_key_value_heads, self.num_key_value_heads], dim=-2
+        )
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        query_states = self.query_layernorm(query_states)
+        key_states = self.key_layernorm(key_states)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = position_embeddings
+
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin, position_ids
+        )
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        # attention_mask = None
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+            is_causal=self.is_causal and attention_mask is None and q_len > 1,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+
+        attn_output = self.dense(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+ATTENTION_CLASSES = {
+    "eager": LLaDA2MoeAttention,
+    "flash_attention_2": LLaDA2MoeFlashAttention2,
+    "sdpa": LLaDA2MoeSdpaAttention,
+}
+
+
+class LLaDA2MoeDecoderLayer(nn.Module):
+    def __init__(self, config: LLaDA2MoeConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.attention = ATTENTION_CLASSES[config._attn_implementation](
+            config=config, layer_idx=layer_idx
+        )
+
+        self.mlp = (
+            LLaDA2MoeSparseMoeBlock(config)
+            if (
+                config.num_experts is not None
+                and layer_idx >= config.first_k_dense_replace
+            )
+            else LLaDA2MoeMLP(config=config, intermediate_size=config.intermediate_size)
+        )
+        self.input_layernorm = LLaDA2MoeRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_attention_layernorm = LLaDA2MoeRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        position_embeddings: Optional[
+            Tuple[torch.Tensor, torch.Tensor]
+        ] = None,  # necessary, but kept here for BC
+        **kwargs,
+    ) -> Tuple[
+        torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
+    ]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*):
+                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
+                query_sequence_length, key_sequence_length)` if default attention is used.
+            position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+                config.n_positions - 1]`.
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*):
+                cached past key and value projection states
+            output_attentions (`bool`, *optional*):
+                Whether to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_router_logits (`bool`, *optional*):
+                Whether or not to return the logits of all the routers. They are useful for computing the router loss,
+                and should not be returned during inference.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+        """
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.attention(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            position_embeddings=position_embeddings,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        if isinstance(hidden_states, tuple):
+            hidden_states, router_logits = hidden_states
+        else:
+            router_logits = None
+        hidden_states = residual + hidden_states.to(residual.device)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+LLADA2MOE_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`LLaDA2MoeConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaDA2Moe Model outputting raw hidden-states without any specific head on top.",
+    LLADA2MOE_START_DOCSTRING,
+)
+class LLaDA2MoePreTrainedModel(PreTrainedModel):
+    config_class = LLaDA2MoeConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LLaDA2MoeDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+LLADA2MOE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaDA2Moe Model outputting raw hidden-states without any specific head on top.",
+    LLADA2MOE_START_DOCSTRING,
+)
+class LLaDA2MoeModel(LLaDA2MoePreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LLaDA2MoeDecoderLayer`]
+
+    Args:
+        config: LLaDA2MoeConfig
+    """
+
+    def __init__(self, config: LLaDA2MoeConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.word_embeddings = nn.Embedding(
+            config.vocab_size, config.hidden_size, self.padding_idx
+        )
+        self.layers = nn.ModuleList(
+            [
+                LLaDA2MoeDecoderLayer(config, layer_idx)
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+        self._use_sdpa = config._attn_implementation == "sdpa"
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self.norm = LLaDA2MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = LLaDA2MoeRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.word_embeddings = value
+
+    @add_start_docstrings_to_model_forward(LLADA2MOE_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoeModelOutputWithPast]:
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        output_router_logits = (
+            output_router_logits
+            if output_router_logits is not None
+            else self.config.output_router_logits
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`transformers."
+                )
+                use_cache = False
+
+        past_key_values_length = 0
+        if use_cache:
+            use_legacy_cache = not isinstance(past_key_values, Cache)
+            if use_legacy_cache:
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            past_key_values_length = past_key_values.get_usable_length(seq_length)
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length,
+                seq_length + past_key_values_length,
+                dtype=torch.long,
+                device=device,
+            )
+            position_ids = position_ids.unsqueeze(0)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        # TODO flash attention 2 can not support custom attention mask
+        # if self._use_flash_attention_2:
+        #     # 2d mask is passed through the layers
+        #     attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        if self._use_sdpa and not output_attentions:
+            # output_attentions=True can not be supported when using SDPA, and we fall back on
+            # the manual implementation that requires a 4D causal mask in all cases.
+            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
+            )
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
+            )
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_logits = () if output_router_logits else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    output_router_logits,
+                    use_cache,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    position_embeddings=position_embeddings,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            if output_router_logits and layer_outputs[-1] is not None:
+                all_router_logits += (layer_outputs[-1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = None
+        if use_cache:
+            next_cache = (
+                next_decoder_cache.to_legacy_cache()
+                if use_legacy_cache
+                else next_decoder_cache
+            )
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_cache,
+                    all_hidden_states,
+                    all_self_attns,
+                    all_router_logits,
+                ]
+                if v is not None
+            )
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            router_logits=all_router_logits,
+        )
+
+
+class LLaDA2MoeModelLM(LLaDA2MoePreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: LLaDA2MoeConfig):
+        super().__init__(config)
+        self.model = LLaDA2MoeModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.word_embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(LLADA2MOE_INPUTS_DOCSTRING)
+    @replace_return_docstrings(
+        output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+    )
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer
+
+        >>> model = LLaDA2MoeForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        output_router_logits = (
+            output_router_logits
+            if output_router_logits is not None
+            else self.config.output_router_logits
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            output_router_logits=output_router_logits,
+            return_dict=return_dict,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        aux_loss = None
+
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            if output_router_logits:
+                output = (aux_loss,) + output
+            return (loss,) + output if loss is not None else output
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        token_type_ids=None,
+        **kwargs,
+    ):
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+                max_cache_length = (
+                    past_key_values.get_max_length()
+                    if hasattr(past_key_values, "get_max_length")
+                    else past_key_values.get_max_cache_shape()
+                )
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusivelly passed as part of the cache (e.g. when passing input_embeds as input)
+            if (
+                attention_mask is not None
+                and attention_mask.shape[1] > input_ids.shape[1]
+            ):
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(
+                    past_state.index_select(0, beam_idx.to(past_state.device))
+                    for past_state in layer_past
+                ),
+            )
+        return reordered_past
+
+    @staticmethod
+    def _top_k_logits(logits, k):
+        if k is None or k <= 0:
+            return logits
+        else:
+            values, _ = torch.topk(logits, k)
+            min_values = values[..., -1, None]
+            return torch.where(
+                logits < min_values, torch.full_like(logits, float("-inf")), logits
+            )
+
+    @staticmethod
+    def _top_p_logits(logits, p):
+        if p is None or p >= 1.0:
+            return logits
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+        sorted_mask = cumulative_probs > p
+        sorted_mask[..., 1:] = sorted_mask[..., :-1].clone()
+        sorted_mask[..., 0] = False
+        mask_indices = torch.scatter(
+            torch.full_like(logits, False, dtype=torch.bool),
+            -1,
+            sorted_indices,
+            sorted_mask,
+        )
+        return logits.masked_fill(mask_indices, float("-inf"))
+
+    def _sample_with_temperature_topk_topp(
+        self, logits, temperature=1.0, top_k=0, top_p=1.0
+    ):
+        orig_shape = logits.shape[:-1]
+        vocab_size = logits.shape[-1]
+        logits = logits.reshape(-1, vocab_size)
+        if temperature > 0 and temperature != 1.0:
+            logits = logits / temperature
+        logits = self._top_k_logits(logits, top_k)
+        logits = self._top_p_logits(logits, top_p)
+        probs = F.softmax(logits, dim=-1)
+        token = torch.multinomial(probs, num_samples=1)
+        token_prob = torch.gather(probs, -1, token)
+        return token.view(*orig_shape), token_prob.view(*orig_shape)
+
+    @staticmethod
+    def _get_num_transfer_tokens(block_length, steps):
+        if steps == 0:
+            return torch.tensor([], dtype=torch.int64)
+        base = block_length // steps
+        remainder = block_length % steps
+        num_transfer_tokens = torch.full((steps,), base, dtype=torch.int64)
+        num_transfer_tokens[:remainder] += 1
+        return num_transfer_tokens
+
+    @torch.no_grad()
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        temperature: int = 0.0,
+        block_length: int = 32,
+        steps: int = 32,
+        gen_length: int = 2048,
+        top_p: Optional[int] = None,
+        top_k: Optional[int] = None,
+        eos_early_stop: bool = False,
+        minimal_topk: int = 1,
+        threshold: float = 0.95,
+        eos_id: int = 156892,
+        mask_id: int = 156895,
+    ):
+        r"""
+        Generates tokens using a block-wise, iterative refinement strategy.
+
+        This method operates differently from standard autoregressive generation. It first creates a template of the
+        full desired length, filled with a special `mask_id`. It then processes this template in segments (`blocks`)
+        and iteratively "denoises" or "refines" the `mask_id` tokens into actual tokens over a series of `steps` for
+        each block. A custom block-diagonal causal attention mask ensures that generation within a block can attend to
+        all previous blocks but not future ones.
+
+        <Tip warning={true}>
+
+        This is a specialized generation method. The quality and speed of the output are highly dependent on the interplay
+        between `block_length`, `steps`, and `threshold`. It aims to achieve faster generation through parallel
+        decoding within blocks, which is a departure from the token-by-token generation of standard `.generate()` methods.
+
+        </Tip>
+
+        Parameters:
+            inputs (`torch.Tensor`):
+                The token sequence used as a prompt for the generation.
+            temperature (`float`, *optional*, defaults to 0.0):
+                The value used to module the next token probabilities. A value of 0.0 corresponds to greedy decoding.
+            block_length (`int`, *optional*, defaults to 32):
+                The size of each generation block. The model generates text in parallel within these blocks. This is a
+                key parameter for controlling the granularity of the generation process.
+            steps (`int`, *optional*, defaults to 32):
+                The number of iterative refinement (or "denoising") steps to perform for each block. Within each block,
+                the model will try to replace `mask_id` tokens with real tokens for this many iterations.
+            gen_length (`int`, *optional*, defaults to 2048):
+                The maximum number of tokens to generate, excluding the prompt.
+            top_p (`float`, *optional*):
+                If set to a float value between 0 and 1, only the most probable tokens with probabilities that add up to
+                `top_p` or higher are kept for generation (nucleus sampling).
+            top_k (`int`, *optional*):
+                The number of highest probability vocabulary tokens to keep for top-k-filtering.
+            eos_early_stop (`bool`, *optional*, defaults to `False`):
+                If `True`, generation will stop as soon as a valid End-Of-Sequence token is generated and confirmed,
+                even if `gen_length` has not been reached.
+            minimal_topk (`int`, *optional*, defaults to 1):
+                A parameter used to dynamically adjust the number of refinement `steps`. The effective number of steps
+                is capped at `gen_length // minimal_topk`.
+            threshold (`float`, *optional*, defaults to 0.95):
+                The confidence probability threshold for accepting a sampled token. During each refinement step, a
+                sampled token is only kept if its probability is above this threshold. If not enough tokens meet the
+                threshold, the ones with the highest confidence are chosen.
+            eos_id (`int`, *optional*, defaults to 156892):
+                The token ID for the end-of-sequence token. Used for `eos_early_stop`.
+            mask_id (`int`, *optional*, defaults to 156895):
+                The token ID used as a placeholder for tokens that are yet to be generated. This is central to the
+                iterative refinement algorithm.
+
+        Return:
+            `torch.Tensor`: A string containing the generated token IDs, starting
+            after the prompt and stopping at the first `eos_id` or `gen_length`.
+        """
+        steps = min(steps, gen_length // minimal_topk)
+        input_ids = inputs.to(self.device)
+
+        prompt_length = input_ids.shape[1]
+        num_blocks = (prompt_length + gen_length + block_length - 1) // block_length
+        total_length = num_blocks * block_length
+
+        block_mask = torch.tril(torch.ones(num_blocks, num_blocks, device=self.device))
+        block_diffusion_attention_mask = (
+            block_mask.repeat_interleave(block_length, dim=0)
+            .repeat_interleave(block_length, dim=1)
+            .unsqueeze(0)
+            .unsqueeze(0)
+        ).bool()
+        block_diffusion_attention_mask = torch.where(
+            block_diffusion_attention_mask, 0.0, float("-inf")
+        ).to(torch.bfloat16)
+
+        position_ids = torch.arange(total_length, device=self.device).unsqueeze(0)
+        x = torch.full((1, total_length), mask_id, dtype=torch.long, device=self.device)
+        x[:, :prompt_length] = input_ids.clone()
+
+        prompt_index_full = torch.zeros_like(x, dtype=torch.bool)
+        prompt_index_full[:, :prompt_length] = True
+
+        prefill_blocks = prompt_length // block_length
+
+        denoising_steps_per_block = steps
+        num_transfer_tokens_schedule = self._get_num_transfer_tokens(
+            block_length, denoising_steps_per_block
+        )
+        for num_block in range(prefill_blocks, num_blocks):
+            current_window_end = (num_block + 1) * block_length
+            cur_x = x[:, :current_window_end]
+            cur_attn_mask = block_diffusion_attention_mask[
+                :, :, :current_window_end, :current_window_end
+            ]
+            cur_position_ids = position_ids[:, :current_window_end]
+
+            for step in range(denoising_steps_per_block):
+                active_block_mask = cur_x[:, -block_length:] == mask_id
+                if active_block_mask.sum() == 0:
+                    break
+
+                logits = self.forward(
+                    cur_x,
+                    attention_mask=cur_attn_mask,
+                    position_ids=cur_position_ids,
+                ).logits
+
+                active_logits = logits[:, -block_length:, :]
+                x0, x0_p = self._sample_with_temperature_topk_topp(
+                    active_logits, temperature=temperature, top_k=top_k, top_p=top_p
+                )
+
+                num_to_transfer = num_transfer_tokens_schedule[step].item()
+                transfer_index = torch.zeros_like(x0, dtype=torch.bool)
+
+                confidence = torch.where(active_block_mask, x0_p, -torch.inf)
+                high_conf_mask = confidence[0] > threshold
+                num_high_confidence = high_conf_mask.sum().item()
+
+                if num_high_confidence >= num_to_transfer:
+                    transfer_index[0] = high_conf_mask
+                else:
+                    _, idx = torch.topk(
+                        confidence[0],
+                        k=min(num_to_transfer, active_block_mask.sum().item()),
+                    )
+                    transfer_index[0, idx] = True
+
+                if transfer_index.any():
+                    cur_x[:, -block_length:][transfer_index] = x0[transfer_index]
+                if eos_early_stop and (x0[transfer_index] == eos_id).any():
+                    eos_pos_in_x = (cur_x[0] == eos_id).nonzero(as_tuple=True)
+                    if len(eos_pos_in_x[0]) > 0:
+                        eos_pos = eos_pos_in_x[0][0].item()
+                        if (cur_x[0, prompt_length:eos_pos] != mask_id).all():
+                            final_x = x[:, :total_length][:, : eos_pos + 1]
+                            return final_x
+
+            x[:, :current_window_end] = cur_x
+            if (
+                eos_id is not None
+                and (x[0, prompt_length:current_window_end] == eos_id).any()
+            ):
+                break
+
+        generated_answer = x[:, : prompt_length + gen_length]
+
+        mask_positions = (generated_answer[0][input_ids.shape[1] :] == eos_id).nonzero(
+            as_tuple=True
+        )[0]
+        if len(mask_positions) > 0:
+            first_mask_position = mask_positions[0].item()
+        else:
+            first_mask_position = gen_length
+        return generated_answer[
+            :, input_ids.shape[1] : input_ids.shape[1] + first_mask_position + 1
+        ]

special_tokens_map.json

@@ -0,0 +1,8 @@
+{
+  "bos_token": "<|startoftext|>",
+  "cls_token": "[CLS]",
+  "eos_token": "<|endoftext|>",
+  "gmask_token": "[gMASK]",
+  "pad_token": "<|endoftext|>",
+  "mask_token": "<|mask|>"
+}

tokenizer_config.json

@@ -0,0 +1,18 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "bos_token": "<|startoftext|>",
+  "chat_template": "{% set thinking_option = 'off' %}\n{{- '<role>SYSTEM</role>' }}\n{%- if messages[0].role == 'system' %}\n    {{- messages[0].content + '\\n' }}\n{%- endif %}\n{%- if tools %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call>\\n\" }}\n{%- endif %}\n{{- 'detailed thinking ' + thinking_option + '<|role_end|>' }}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if message.content is string %}\n        {%- set content = message.content %}\n    {%- else %}\n        {%- set content = '' %}\n    {%- endif %}\n    {%- if message.role == \"user\" %}\n        {{- '<role>HUMAN</role>' + message.content + '<|role_end|>' }}\n    {%- elif message.role == \"system\" and not loop.first %}\n        {{- '<role>SYSTEM</role>' + message.content + '<|role_end|>' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is string %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in content %}\n                {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n                {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- if loop.index0 > ns.last_query_index %}\n            {%- if reasoning_content %}\n                {{- '<role>ASSISTANT</role>' + '\\n<think>\\n' + reasoning_content.strip('\\n') + '\\n</think>\\n\\n' + content.lstrip('\\n') }}\n            {%- else %}\n                {{- '<role>ASSISTANT</role>' + content }}\n            {%- endif %}\n        {%- else %}\n            {{- '<role>ASSISTANT</role>' + content }}\n        {%- endif %}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- if tool_call.function %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>\\n{\"name\": \"' }}\n                {{- tool_call.name }}\n                {{- '\", \"arguments\": ' }}\n                {%- if tool_call.arguments is string %}\n                    {{- tool_call.arguments }}\n                {%- else %}\n                    {{- tool_call.arguments | tojson }}\n                {%- endif %}\n                {{- '}\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|role_end|>' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<role>OBSERVATION</role>' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|role_end|>' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<role>ASSISTANT</role>' }}\n{%- endif %}",
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "[CLS]",
+  "eos_token": "<|endoftext|>",
+  "mask_token": "<|mask|>",
+  "fast_tokenizer": true,
+  "gmask_token": "[gMASK]",
+  "merges_file": null,
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "trust_remote_code": true,
+  "vocab_file": null
+}