lora微调中的LORA_CHECKPOINT参数怎么设置？

lora微调中的LORA_CHECKPOINT参数怎么设置？下图设置空文件夹（和保存目录一致），微调后将训练好的megatron模型重新转回HuggingFace格式报错
![输入图片说明](https://foruda.gitee.com/images/1713235677187089652/6ac023fd_13086124.png "屏幕截图")

```
 loading checkpoint from ./output-llama2-7b-lora at iteration 200
Traceback (most recent call last):
  File "/home/guan/temp/convert/ModelLink/tools/checkpoint/convert_ckpt.py", line 98, in <module>
Loader exited, exiting saver
    main()
  File "/home/guan/temp/convert/ModelLink/tools/checkpoint/convert_ckpt.py", line 91, in main
    loader.load_checkpoint(queue, args)
  File "/home/guan/temp/convert/ModelLink/tools/checkpoint/loader_megatron.py", line 380, in load_checkpoint
    _load_checkpoint(queue, args)
  File "/home/guan/temp/convert/ModelLink/tools/checkpoint/loader_megatron.py", line 236, in _load_checkpoint
    all_models = [get_models(tp_size, md.params_dtype)]
  File "/home/guan/temp/convert/ModelLink/tools/checkpoint/loader_megatron.py", line 161, in get_models
    load_checkpoint_mg(model_, None, None)
  File "/home/guan/temp/convert/ModelLink/megatron/checkpointing.py", line 584, in load_checkpoint
    model[0].load_state_dict(state_dict['model'], strict=strict)
  File "/home/guan/temp/convert/ModelLink/modellink/model/gpt_model.py", line 103, in load_state_dict
    self.language_model.load_state_dict(state_dict, strict=strict)
  File "/home/guan/temp/convert/ModelLink/megatron/model/language_model.py", line 608, in load_state_dict
    self.encoder.load_state_dict(state_dict_, strict=strict)
  File "/home/guan/temp/convert/ModelLink/megatron/model/transformer.py", line 1793, in load_state_dict
    super().load_state_dict(state_dict_, strict)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 2152, in load_state_dict
    raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for ParallelTransformer:
        Missing key(s) in state_dict: "layers.0.input_norm.weight", "layers.0.self_attention.query_key_value.weight", "layers.0.self_attention.dense.weight", "layers.0.post_attention_norm.weight", "layers.0.mlp.dense_h_to_4h.weight", "layers.0.mlp.dense_4h_to_h.weight", "layers.1.input_norm.weight", "layers.1.self_attention.query_key_value.weight", "layers.1.self_attention.dense.weight", "layers.1.post_attention_norm.weight", "layers.1.mlp.dense_h_to_4h.weight", "layers.1.mlp.dense_4h_to_h.weight", "layers.2.input_norm.weight", "layers.2.self_attention.query_key_value.weight", "layers.2.self_attention.dense.weight", "layers.2.post_attention_norm.weight", "layers.2.mlp.dense_h_to_4h.weight", "layers.2.mlp.dense_4h_to_h.weight", "layers.3.input_norm.weight", "layers.3.self_attention.query_key_value.weight", "layers.3.self_attention.dense.weight", "layers.3.post_attention_norm.weight", "layers.3.mlp.dense_h_to_4h.weight", "layers.3.mlp.dense_4h_to_h.weight", "layers.4.input_norm.weight", "layers.4.self_attention.query_key_value.weight", "layers.4.self_attention.dense.weight", "layers.4.post_attention_norm.weight", "layers.4.mlp.dense_h_to_4h.weight", "layers.4.mlp.dense_4h_to_h.weight", "layers.5.input_norm.weight", "layers.5.self_attention.query_key_value.weight", "layers.5.self_attention.dense.weight", "layers.5.post_attention_norm.weight", "layers.5.mlp.dense_h_to_4h.weight", "layers.5.mlp.dense_4h_to_h.weight", "layers.6.input_norm.weight", "layers.6.self_attention.query_key_value.weight", "layers.6.self_attention.dense.weight", "layers.6.post_attention_norm.weight", "layers.6.mlp.dense_h_to_4h.weight", "layers.6.mlp.dense_4h_to_h.weight", "layers.7.input_norm.weight", "layers.7.self_attention.query_key_value.weight", "layers.7.self_attention.dense.weight", "layers.7.post_attention_norm.weight", "layers.7.mlp.dense_h_to_4h.weight", "layers.7.mlp.dense_4h_to_h.weight", "layers.8.input_norm.weight", "layers.8.self_attention.query_key_value.weight", "layers.8.self_attention.dense.weight", "layers.8.post_attention_norm.weight", "layers.8.mlp.dense_h_to_4h.weight", "layers.8.mlp.dense_4h_to_h.weight", "layers.9.input_norm.weight", "layers.9.self_attention.query_key_value.weight", "layers.9.self_attention.dense.weight", "layers.9.post_attention_norm.weight", "layers.9.mlp.dense_h_to_4h.weight", "layers.9.mlp.dense_4h_to_h.weight", "layers.10.input_norm.weight", "layers.10.self_attention.query_key_value.weight", "layers.10.self_attention.dense.weight", "layers.10.post_attention_norm.weight", "layers.10.mlp.dense_h_to_4h.weight", "layers.10.mlp.dense_4h_to_h.weight", "layers.11.input_norm.weight", "layers.11.self_attention.query_key_value.weight", "layers.11.self_attention.dense.weight", "layers.11.post_attention_norm.weight", "layers.11.mlp.dense_h_to_4h.weight", "layers.11.mlp.dense_4h_to_h.weight", "layers.12.input_norm.weight", "layers.12.self_attention.query_key_value.weight", "layers.12.self_attention.dense.weight", "layers.12.post_attention_norm.weight", "layers.12.mlp.dense_h_to_4h.weight", "layers.12.mlp.dense_4h_to_h.weight", "layers.13.input_norm.weight", "layers.13.self_attention.query_key_value.weight", "layers.13.self_attention.dense.weight", "layers.13.post_attention_norm.weight", "layers.13.mlp.dense_h_to_4h.weight", "layers.13.mlp.dense_4h_to_h.weight", "layers.14.input_norm.weight", "layers.14.self_attention.query_key_value.weight", "layers.14.self_attention.dense.weight", "layers.14.post_attention_norm.weight", "layers.14.mlp.dense_h_to_4h.weight", "layers.14.mlp.dense_4h_to_h.weight", "layers.15.input_norm.weight", "layers.15.self_attention.query_key_value.weight", "layers.15.self_attention.dense.weight", "layers.15.post_attention_norm.weight", "layers.15.mlp.dense_h_to_4h.weight", "layers.15.mlp.dense_4h_to_h.weight", "layers.16.input_norm.weight", "layers.16.self_attention.query_key_value.weight", "layers.16.self_attention.dense.weight", "layers.16.post_attention_norm.weight", "layers.16.mlp.dense_h_to_4h.weight", "layers.16.mlp.dense_4h_to_h.weight", "layers.17.input_norm.weight", "layers.17.self_attention.query_key_value.weight", "layers.17.self_attention.dense.weight", "layers.17.post_attention_norm.weight", "layers.17.mlp.dense_h_to_4h.weight", "layers.17.mlp.dense_4h_to_h.weight", "layers.18.input_norm.weight", "layers.18.self_attention.query_key_value.weight", "layers.18.self_attention.dense.weight", "layers.18.post_attention_norm.weight", "layers.18.mlp.dense_h_to_4h.weight", "layers.18.mlp.dense_4h_to_h.weight", "layers.19.input_norm.weight", "layers.19.self_attention.query_key_value.weight", "layers.19.self_attention.dense.weight", "layers.19.post_attention_norm.weight", "layers.19.mlp.dense_h_to_4h.weight", "layers.19.mlp.dense_4h_to_h.weight", "layers.20.input_norm.weight", "layers.20.self_attention.query_key_value.weight", "layers.20.self_attention.dense.weight", "layers.20.post_attention_norm.weight", "layers.20.mlp.dense_h_to_4h.weight", "layers.20.mlp.dense_4h_to_h.weight", "layers.21.input_norm.weight", "layers.21.self_attention.query_key_value.weight", "layers.21.self_attention.dense.weight", "layers.21.post_attention_norm.weight", "layers.21.mlp.dense_h_to_4h.weight", "layers.21.mlp.dense_4h_to_h.weight", "layers.22.input_norm.weight", "layers.22.self_attention.query_key_value.weight", "layers.22.self_attention.dense.weight", "layers.22.post_attention_norm.weight", "layers.22.mlp.dense_h_to_4h.weight", "layers.22.mlp.dense_4h_to_h.weight", "layers.23.input_norm.weight", "layers.23.self_attention.query_key_value.weight", "layers.23.self_attention.dense.weight", "layers.23.post_attention_norm.weight", "layers.23.mlp.dense_h_to_4h.weight", "layers.23.mlp.dense_4h_to_h.weight", "layers.24.input_norm.weight", "layers.24.self_attention.query_key_value.weight", "layers.24.self_attention.dense.weight", "layers.24.post_attention_norm.weight", "layers.24.mlp.dense_h_to_4h.weight", "layers.24.mlp.dense_4h_to_h.weight", "layers.25.input_norm.weight", "layers.25.self_attention.query_key_value.weight", "layers.25.self_attention.dense.weight", "layers.25.post_attention_norm.weight", "layers.25.mlp.dense_h_to_4h.weight", "layers.25.mlp.dense_4h_to_h.weight", "layers.26.input_norm.weight", "layers.26.self_attention.query_key_value.weight", "layers.26.self_attention.dense.weight", "layers.26.post_attention_norm.weight", "layers.26.mlp.dense_h_to_4h.weight", "layers.26.mlp.dense_4h_to_h.weight", "layers.27.input_norm.weight", "layers.27.self_attention.query_key_value.weight", "layers.27.self_attention.dense.weight", "layers.27.post_attention_norm.weight", "layers.27.mlp.dense_h_to_4h.weight", "layers.27.mlp.dense_4h_to_h.weight", "layers.28.input_norm.weight", "layers.28.self_attention.query_key_value.weight", "layers.28.self_attention.dense.weight", "layers.28.post_attention_norm.weight", "layers.28.mlp.dense_h_to_4h.weight", "layers.28.mlp.dense_4h_to_h.weight", "layers.29.input_norm.weight", "layers.29.self_attention.query_key_value.weight", "layers.29.self_attention.dense.weight", "layers.29.post_attention_norm.weight", "layers.29.mlp.dense_h_to_4h.weight", "layers.29.mlp.dense_4h_to_h.weight", "layers.30.input_norm.weight", "layers.30.self_attention.query_key_value.weight", "layers.30.self_attention.dense.weight", "layers.30.post_attention_norm.weight", "layers.30.mlp.dense_h_to_4h.weight", "layers.30.mlp.dense_4h_to_h.weight", "layers.31.input_norm.weight", "layers.31.self_attention.query_key_value.weight", "layers.31.self_attention.dense.weight", "layers.31.post_attention_norm.weight", "layers.31.mlp.dense_h_to_4h.weight", "layers.31.mlp.dense_4h_to_h.weight", "final_norm.weight".
        Unexpected key(s) in state_dict: "layers.0.self_attention.query_key_value.lora_A.default.weight", "layers.0.self_attention.query_key_value.lora_B.default.weight", "layers.0.self_attention.dense.lora_A.default.weight", "layers.0.self_attention.dense.lora_B.default.weight", "layers.0.mlp.dense_h_to_4h.lora_A.default.weight", "layers.0.mlp.dense_h_to_4h.lora_B.default.weight", "layers.0.mlp.dense_4h_to_h.lora_A.default.weight", "layers.0.mlp.dense_4h_to_h.lora_B.default.weight", "layers.1.self_attention.query_key_value.lora_A.default.weight", "layers.1.self_attention.query_key_value.lora_B.default.weight", "layers.1.self_attention.dense.lora_A.default.weight", "layers.1.self_attention.dense.lora_B.default.weight", "layers.1.mlp.dense_h_to_4h.lora_A.default.weight", "layers.1.mlp.dense_h_to_4h.lora_B.default.weight", "layers.1.mlp.dense_4h_to_h.lora_A.default.weight", "layers.1.mlp.dense_4h_to_h.lora_B.default.weight", "layers.2.self_attention.query_key_value.lora_A.default.weight", "layers.2.self_attention.query_key_value.lora_B.default.weight", "layers.2.self_attention.dense.lora_A.default.weight", "
```

Ascend/MindSpeed-LLM
暂停

内容风险标识

评论 (3)

Ascend/MindSpeed-LLM暂停 .gitee-modal { width: 500px !important; }

内容风险标识