[PyTorch] Skip test_nvfp4_partial_cast_matches_full test when NVFP4 is not available.

[ksivaman]

Description

#2691 introduced this test but the NVFP4 condition wasn't checked properly.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Skip test_nvfp4_partial_cast_matches_full test when NVFP4 is not available.

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[ksivaman]

/te-ci pytorch L1

[greptile-apps]

Greptile Summary

This PR fixes a misplaced pytest.skip() call introduced in #2691. The NVFP4 availability guard was placed inside _test_nvfp4_partial_cast_matches_full, a worker function that runs inside a torch.distributed.run subprocess — completely outside the pytest process. Calling pytest.skip() there raised a BaseException that was uncaught, causing the subprocess to exit non-zero and subprocess.run(..., check=True) to throw subprocess.CalledProcessError, making the test fail on machines without NVFP4 instead of skipping it.

The fix moves the guard to test_nvfp4_partial_cast_matches_full, which runs directly under pytest, so pytest.skip() is handled correctly and the subprocess is never launched on unsupported hardware.

• The root cause is correctly identified: pytest.skip() must only be called from code running under the pytest process.
• The worker function _test_nvfp4_partial_cast_matches_full is only reachable via the subprocess launched by the now-guarded test function, so removing the check from the worker is safe.
• No other code paths call _test_nvfp4_partial_cast_matches_full directly.
• Change is minimal and targeted with no side-effects on other tests.

Confidence Score: 5/5

• This PR is safe to merge; it is a small, well-scoped bug fix with no risk of regressions.
• The change is a single-file, four-line relocation of a pytest.skip() call from a subprocess worker to the correct pytest-context test function. The logic is sound, the fix is the minimal correct solution, and there are no new dependencies or side-effects introduced.
• No files require special attention.

Important Files Changed

Filename	Overview
tests/pytorch/distributed/test_cast_master_weights_to_fp8.py	Moves is_nvfp4_available guard from subprocess worker _test_nvfp4_partial_cast_matches_full to the pytest-facing test_nvfp4_partial_cast_matches_full, correctly preventing a spurious subprocess.CalledProcessError when NVFP4 is unavailable.

Sequence Diagram

sequenceDiagram
    participant pytest as pytest runner
    participant T as test_nvfp4_partial_cast_matches_full
    participant SP as subprocess (torch.distributed.run)
    participant W as _test_nvfp4_partial_cast_matches_full (worker)

    pytest->>T: invoke test

    alt NVFP4 not available (NEW behaviour)
        T->>T: is_nvfp4_available() → False
        T->>pytest: pytest.skip(reason)
        Note over SP,W: subprocess never launched
    else NVFP4 available
        T->>SP: subprocess.run(..., check=True)
        SP->>W: run_parallel_nvfp4_partial_cast_test()
        W->>W: distributed quantisation assertions
        W-->>SP: success / AssertionError
        SP-->>T: exit 0 / non-zero
        T-->>pytest: pass / CalledProcessError
    end

Loading

_{Last reviewed commit: c1fb5d7}

[KshitijLakhani]

LGTM !
Good to merge once CI passes fully !
Thanks !