[PyT] Fix PyTorch BF16 reduction numerics tests

[jberchtold-nvidia]

Description

Fixes PyTorch tests to match tolerances on L40 that were changed in this PR #3440 . Now BF16 partial results reductions are disabled

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

• Guard tests with PyTorch flag to disable bf16 partial result reductions in reference

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

[greptile-apps]

Greptile Summary

The PR aligns plain-PyTorch numerical references with Transformer Engine’s higher-precision BF16 reduction behavior.

• Adds a context manager that temporarily disables reduced-precision BF16 reductions and reliably restores the prior setting.
• Applies the precision setting automatically to every test in the PyTorch numerics module.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
tests/pytorch/test_numerics.py	Adds a scoped autouse fixture that disables PyTorch BF16 partial-result reductions during each numerical test and restores the original backend setting afterward.

_{Reviews (3): Last reviewed commit: "Apply BF16 precision fixture to all nume..." | Re-trigger Greptile}

[jberchtold-nvidia]

/te-ci L0 pytorch

[jberchtold-nvidia]

/te-ci L0 pytorch