fix(gpt-oss): update _patch_bridge_expert_cache_to_cpu to match Megatron-Bridge API

[aoshen02]

Summary

radixark/Megatron-Bridge@bridge added a fourth parameter hf_state_dict to ModelBridge.maybe_modify_converted_hf_weight. The monkey-patch in _patch_bridge_expert_cache_to_cpu still uses the old 3-arg signature, causing a crash during GPT-OSS colocated weight sync:

TypeError: _patch_bridge_expert_cache_to_cpu.<locals>._patched() takes 3 positional arguments but 4 were given

The fix accepts and forwards the new optional hf_state_dict argument.

Root cause

Megatron-Bridge updated ModelBridge.maybe_modify_converted_hf_weight to:

def maybe_modify_converted_hf_weight(
    self,
    task: WeightConversionTask,
    converted_weights_dict: Dict[str, torch.Tensor],
    hf_state_dict: Mapping[str, torch.Tensor],   # ← new param
) -> Dict[str, torch.Tensor]:

The patch only handled 3 args.

Fix

# before
def _patched(self, task, converted_weights_dict):
    cpu_dict = {k: v.cpu() for k, v in converted_weights_dict.items()}
    result = _orig(self, task, cpu_dict)

# after
def _patched(self, task, converted_weights_dict, hf_state_dict=None):
    cpu_dict = {k: v.cpu() for k, v in converted_weights_dict.items()}
    result = _orig(self, task, cpu_dict, hf_state_dict)

🤖 Generated with Claude Code