[model] fix: support MCore FSDP factory wrappers

[cuichenx]

Summary

• keep conversion unwrap_model() compatible with both the Bridge-pinned legacy MCore FSDP class API and the newer MCore factory API
• unwrap both concrete FullyShardedDataParallelV1 and FullyShardedDataParallelV2 wrappers without passing the factory function to isinstance()
• add regression coverage for both API generations, including the legacy case where the V1/V2 symbols do not exist

Root cause

Newer MCore versions expose mcore_fsdp_adapter.FullyShardedDataParallel as a factory function and provide V1/V2 as the concrete wrapper classes. Bridge conversion code included the factory in an isinstance() tuple, which raises before MFSDP conversion reaches weight loading.

This is separate from the TP=2 vocabulary-scatter defect fixed by #5353. That fix is present on current main; this PR removes the newer MCore compatibility blocker that prevented QA from re-verifying the three MFSDP conversion examples.

Compatibility

• legacy API: use FullyShardedDataParallel when it is a class; V1/V2 symbols may be absent
• factory API: use the explicit, independent V1 and V2 wrapper classes
• no training-path, dependency, CI, public API, or MCore submodule changes

Validation

• Red on clean origin/main c62be0886a:
  • uv run python -m pytest tests/unit_tests/models/test_conversion_utils.py -k unwrap_model -vv
  • factory case failed with TypeError: isinstance() arg 2 must be a type, a tuple of types, or a union; legacy case passed
• Green on ea19acb7f:
  • uv run python -m pytest tests/unit_tests/models/test_conversion_utils.py -vv — 7 passed
  • uv run pre-commit run --all-files — passed

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[cuichenx]

/ok to test 62097ac

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyu-33]

/ok to test ea19acb

[claude]

LGTM — the change correctly handles the Megatron-Core FullyShardedDataParallel transition from a class to a factory function.

The isinstance(megatron_fsdp, type) guard is right: a factory function is not a type, so it falls through to the V1/V2 branch, while a real class is used directly. Both paths are covered by the two new unit tests (test_unwrap_model_supports_mcore_fsdp_factory and test_unwrap_model_supports_legacy_mcore_fsdp_class). Test coverage for the changed logic is good.

Minor note (non-blocking): in the factory branch, FullyShardedDataParallelV1/V2 are assumed to exist. If a future MCore ships a factory FullyShardedDataParallel without those two attributes, this raises AttributeError rather than a clear message — acceptable for now, since the factory contract implies they exist.

Suggested test cases

• No perf tests impacted.

[yaoyu-33]

Superseded by #5431, which has now merged the MCore FSDP factory-wrapper compatibility fix into main together with focused coverage in tests/unit_tests/models/test_conversion_unwrap_utils.py. Rebasing this PR now conflicts on the same implementation and would only duplicate the code/tests, so closing this PR rather than resolving it as a duplicate.