训练时使用的QLoRA 4rank，进行cuda模型合并导出时出现，KeyError: 'base_model.model.model.model.layers.14.mlp.down_proj'

[xiaoheiyue]

System Info

File "/home/mukuro/projects/LLaMA-Factory/src/llamafactory/model/adapter.py", line 299, in init_adapter
model = _setup_lora_tuning(
^^^^^^^^^^^^^^^^^^^
File "/home/mukuro/projects/LLaMA-Factory/src/llamafactory/model/adapter.py", line 181, in _setup_lora_tuning
model: "LoraModel" = PeftModel.from_pretrained(model, adapter, **init_kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/mukuro/softwares/miniconda3/envs/qwen2.5/lib/python3.11/site-packages/peft/peft_model.py", line 545, in from_pretrained
model.load_adapter(
File "/home/mukuro/softwares/miniconda3/envs/qwen2.5/lib/python3.11/site-packages/peft/peft_model.py", line 1151, in load_adapter
self._update_offload(offload_index, adapters_weights)
File "/home/mukuro/softwares/miniconda3/envs/qwen2.5/lib/python3.11/site-packages/peft/peft_model.py", line 1028, in _update_offload
safe_module = dict(self.named_modules())[extended_prefix]
~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
KeyError: 'base_model.model.model.model.layers.14.mlp.down_proj'

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder
• My own task or dataset (give details below)

Reproduction

llamafactory 使用 peft 0.12.0

Expected behavior

我希望能解决这个问题正常合并

[JINO-ROHIT]

looks like some mismatch, can you ensure using

print(dict(model.named_modules()).keys())

[xiaoheiyue]

looks like some mismatch, can you ensure using

print(dict(model.named_modules()).keys())

是查看原模型的吗？训练的 LoRA adapter 和模型层数是一样的，每个层 att qkv 还有 mlp 的 那些也都有。

[xiaoheiyue]

少了一个 .model 的前缀。

[xiaoheiyue]

      if index:
            print("index", index)
            print('model.layers.14.mlp.gate_proj.weight' in index)				
            offload_index = {
                p: {
                    "safetensors_file": index[p]["safetensors_file"],
                    "weight_name": p,
                    "dtype": str(weight_map[p].dtype).replace("torch.", ""),
                }
                for p in weight_map.keys()
                if p in disk_modules
            }
            kwargs["offload_index"] = offload_index

    if (getattr(model, "hf_device_map", None) is not None) and len(
        set(model.hf_device_map.values()).intersection({"cpu", "disk"})
    ) > 0:
        remove_hook_from_submodules(model)

这是不是意味着原始模型少了 14 层