使用最新master 分支训练DeepSeek V3训练，设置DeepseekV3MoE为叶子结点报错

[Han-Huaqiao]

Reminder

• I have read the above rules and searched the existing issues.

System Info

llama-factory master分支，transformers==4.51.3

Reproduction

Put your message here.

Others

使用最新master训练时，在

if model_type in ["kimi_vl", "deepseek_v3"]:
        check_version("transformers>=4.51.1")
        from transformers.models.deepseek_v3.modeling_deepseek_v3 import DeepseekV3MoE

        _set_z3_leaf_modules(model, [DeepseekV3MoE])

代码块报错如下，这个要如何修改：

[rank0]: Traceback (most recent call last):
[rank0]:   File "/project/LLaMA-Factory/./src/train.py", line 28, in <module>
[rank0]:     main()
[rank0]:   File "/project/LLaMA-Factory/./src/train.py", line 19, in main
[rank0]:     run_exp()
[rank0]:   File "/project/LLaMA-Factory/src/llamafactory/train/tuner.py", line 107, in run_exp
[rank0]:     _training_function(config={"args": args, "callbacks": callbacks})
[rank0]:   File "/project/LLaMA-Factory/src/llamafactory/train/tuner.py", line 69, in _training_function
[rank0]:     run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
[rank0]:   File "/project/LLaMA-Factory/src/llamafactory/train/sft/workflow.py", line 52, in run_sft
[rank0]:     model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
[rank0]:   File "/project/LLaMA-Factory/src/llamafactory/model/loader.py", line 175, in load_model
[rank0]:     patch_model(model, tokenizer, model_args, is_trainable, add_valuehead)
[rank0]:   File "/project/LLaMA-Factory/src/llamafactory/model/patcher.py", line 170, in patch_model
[rank0]:     add_z3_leaf_module(model)
[rank0]:   File "/project/LLaMA-Factory/src/llamafactory/model/model_utils/moe.py", line 61, in add_z3_leaf_module
[rank0]:     _set_z3_leaf_modules(model, [DeepseekV3MoE])
[rank0]:   File "/project/LLaMA-Factory/src/llamafactory/model/model_utils/moe.py", line 33, in _set_z3_leaf_modules
[rank0]:     set_z3_leaf_modules(model, leaf_modules)
[rank0]:   File "/root/miniconda/envs/python310_torch25_cuda/lib/python3.10/site-packages/deepspeed/utils/z3_leaf_module.py", line 81, in set_z3_leaf_modules
[rank0]:     return _do_set_z3_leaf_modules(model, leaf_module_classes, True)
[rank0]:   File "/root/miniconda/envs/python310_torch25_cuda/lib/python3.10/site-packages/deepspeed/utils/z3_leaf_module.py", line 65, in _do_set_z3_leaf_modules
[rank0]:     raise ValueError(f'No modules of type {leaf_module_classes} found in model {model}')
[rank0]: ValueError: No modules of type [<class 'transformers.models.deepseek_v3.modeling_deepseek_v3.DeepseekV3MoE'>] found in model DeepseekV3ForCausalLM(
[rank0]:   (model): DeepseekV3Model(
[rank0]:     (embed_tokens): Embedding(129280, 7168)
[rank0]:     (layers): ModuleList(
[rank0]:       (0-2): 3 x DeepseekV3DecoderLayer(
[rank0]:         (self_attn): DeepseekV3FlashAttention2(
[rank0]:           (q_a_proj): Linear()
[rank0]:           (q_a_layernorm): DeepseekV3RMSNorm()
[rank0]:           (q_b_proj): Linear()
[rank0]:           (kv_a_proj_with_mqa): Linear()
[rank0]:           (kv_a_layernorm): DeepseekV3RMSNorm()
[rank0]:           (kv_b_proj): Linear()
[rank0]:           (o_proj): Linear()
[rank0]:           (rotary_emb): DeepseekV3YarnRotaryEmbedding()
[rank0]:         )
[rank0]:         (mlp): DeepseekV3MLP(
[rank0]:           (gate_proj): Linear()
[rank0]:           (up_proj): Linear()
[rank0]:           (down_proj): Linear()
[rank0]:           (act_fn): SiLU()
[rank0]:         )
[rank0]:         (input_layernorm): DeepseekV3RMSNorm()
[rank0]:         (post_attention_layernorm): DeepseekV3RMSNorm()
[rank0]:       )
[rank0]:       (3-60): 58 x DeepseekV3DecoderLayer(
[rank0]:         (self_attn): DeepseekV3FlashAttention2(
[rank0]:           (q_a_proj): Linear()
[rank0]:           (q_a_layernorm): DeepseekV3RMSNorm()
[rank0]:           (q_b_proj): Linear()
[rank0]:           (kv_a_proj_with_mqa): Linear()
[rank0]:           (kv_a_layernorm): DeepseekV3RMSNorm()
[rank0]:           (kv_b_proj): Linear()
[rank0]:           (o_proj): Linear()
[rank0]:           (rotary_emb): DeepseekV3YarnRotaryEmbedding()
[rank0]:         )
[rank0]:         (mlp): DeepseekV3MoE(
[rank0]:           (experts): ModuleList(
[rank0]:             (0-255): 256 x DeepseekV3MLP(
[rank0]:               (gate_proj): Linear()
[rank0]:               (up_proj): Linear()
[rank0]:               (down_proj): Linear()
[rank0]:               (act_fn): SiLU()
[rank0]:             )
[rank0]:           )
[rank0]:           (gate): MoEGate()
[rank0]:           (shared_experts): DeepseekV3MLP(
[rank0]:             (gate_proj): Linear()
[rank0]:             (up_proj): Linear()
[rank0]:             (down_proj): Linear()
[rank0]:             (act_fn): SiLU()
[rank0]:           )
[rank0]:         )
[rank0]:         (input_layernorm): DeepseekV3RMSNorm()
[rank0]:         (post_attention_layernorm): DeepseekV3RMSNorm()
[rank0]:       )
[rank0]:     )
[rank0]:     (norm): DeepseekV3RMSNorm()
[rank0]:   )
[rank0]:   (lm_head): Linear()
[rank0]: )

[li199959]

I've also encountered the same problem. Have you solved it?