megatron 格式转换，加载模型出错

[looput]

Describe the bug

进行模型格式转换时

CUDA_VISIBLE_DEVICES=0 swift export \
    --model /root/.cache/models/Qwen3-30B-A3B-Instruct-2507/ \
    --to_mcore true \
    --torch_dtype bfloat16 \
    --output_dir /root/.cache/models/Qwen3-30B-A3B-Instruct-2507-mcore \
    --test_convert_precision true

出现以下错误

 in _wrapper                                                                                                        
[rank0]:     return func(*args, **kwargs)                                                                           
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^                                                                           
[rank0]:   File "/root/miniconda3/envs/mcore/lib/python3.11/site-packages/transformers/modeling_utils.py", line 5176
, in from_pretrained                                                                                                
[rank0]:     ) = cls._load_pretrained_model(                                                                        
[rank0]:         ^^^^^^^^^^^^^^^^^^^^^^^^^^^                                                                        
[rank0]: TypeError: _new__load_pretrained_model() missing 1 required positional argument: 'pretrained_model_name_or_
path'                                                                                                               
[rank0]:[W917 16:53:31.821412936 ProcessGroupNCCL.cpp:1496] Warning: WARNING: destroy_process_group() was not called
 before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distribute
d.html#shutdown (function operator())         

Your hardware and system info
CUDA: 12.4
GPU: H20
Torch: 2.6.0

[Jintao-Huang]

use megatron-lm 0.13.0分支

[Jintao-Huang]

transformers版本和modelscope版本升级一下 或者直接使用镜像

[looput]

@Jintao-Huang 多谢回复~
现在版本如下，已经比较新了

[looput]

错误引起的片段如下

    @classmethod
    def _new_from_pretrained(cls, *args, **kwargs):
        if 'AutoAWQFor' in automodel_class.__name__:
            kwargs.pop('use_cache', None)
        if model_info.quant_method == 'gptq':
            cls.main_input_name = 'input_ids'
        if hasattr(cls, '_tp_plan'):  # fix tp_plan
            cls._tp_plan = cls._tp_plan or {}
        if return_dummy_model:
            origin_torch_dtype = torch.get_default_dtype()
            torch.set_default_dtype(kwargs['config'].torch_dtype)
            model = cls(copy.deepcopy(kwargs['config']))
            torch.set_default_dtype(origin_torch_dtype)
        else:
            print(args, kwargs)
            # print(cls)
            # kwargs['pretrained_model_name_or_path'] = args[0]
            # model = from_pretrained(cls, **kwargs)

            model = from_pretrained(cls, *args, **kwargs)
        return model

    PreTrainedModel.from_pretrained = _new_from_pretrained

    try:
        yield
    finally:
        PreTrainedModel.from_pretrained = classmethod(from_pretrained)

这个错误的应该是 modelopt 库与 ms-swift 的 monkey patching 机制之间的兼容性问题。

ms-swift 使用了patching 来修改 PreTrainedModel.from_pretrained 方法
modelopt 库也对 from_pretrained 方法进行了包装，创建了 _modelopt_cache
当两个库同时修改同一个方法时，参数传递出现了不匹配。

当卸载nvidia-modelopt包时，问题消失。

遗留的问题是： 如何修改该patching方式，使两者兼容了

[Jintao-Huang]

好的 感谢反馈