[QUESTION]How to convert a model from torchdist to hugginface？

[Listen-WLS]

Your question
Ask a clear and concise question about Megatron-LM. Tag the @megatron-oncall
to get oncall's attention to this issue.
When i run the convert pipline this err will come：File "Megatron-LM-core_v0.15.0/tools/checkpoint/loader_base.py", line 208, in get_models_for_pipeline_stage
this_model = model_provider(pre_process=pre_process,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: model_provider() missing 1 required positional argument: 'model_builder'

[sbhavani]

@Listen-WLS Thanks for reporting this, you've likely encountered a bug in the checkpoint converter code. The model_provider function signature was updated.

That said, for converting Megatron checkpoints to Hugging Face format, I'd recommend checking out Megatron Bridge.

Megatron Bridge provides bidirectional converters that support:
• Hugging Face → Megatron (for continued pretraining)
• Megatron → Hugging Face (for exporting trained models back to HF format)

This is the recommended path for checkpoint format conversion between Megatron and Hugging Face ecosystems.