[FIX] model provider compatibility with Megatron-LM

[simondong1]

Issue

When using recent Megatron-LM versions (since Dec 10, 2025, NVIDIA/Megatron-LM#2608), model initialization fails with unexpected keyword argument 'pg_collection' or config because the upstream megatron get_model interface now passes these arguments explicitly. This causes convert_hf_to_torch_dist.py to fail.

from megatron

def get_model(
    model_provider_func,
    model_type=ModelType.encoder_or_decoder,
    wrap_with_ddp=True,
    config=None, # <--- NEW PARAMETER
    pg_collection=None # <--- NEW PARAMETER
):

Fix

• Custom Providers: Updated wrapped_model_provider to dynamically inspect signatures and only pass pg_collection, vp_stage, or config if the custom function accepts them.
• Standard Provider: Updated model_provider to accept pg_collection and config and forward pg_collection to GPTModel kwargs for correct distributed initialization.

Related

• Closes An error occurred while converting the Huggingface to a form that can be loaded by Megatron. #1301