[PyTorch] Fix MXFP8 master weight cast when a rank owns no shard.

[rapatel]

Description

Fix MXFP8 master weight cast when a rank owns no shard.

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

Please list the changes introduced in this PR:

• Fix MXFP8 master weight cast when a rank owns no shard.
• Adds tests/pytorch/mxfp8/test_mxfp8_master_weight_empty_shard.py.

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

This PR fixes MXFP8 master-weight casting when a rank owns no shard by deriving the fallback amax-buffer dtype from the quantized model weight. It also adds regression coverage for empty and populated shards and safely tears down a process group created by the test fixture.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/tensor/utils.py	Initializes the master-weight dtype from the model weight so empty-shard ranks participate in amax reduction with the same dtype as populated ranks.
tests/pytorch/mxfp8/test_mxfp8_master_weight_empty_shard.py	Adds empty-shard regression coverage and conditionally destroys only the distributed process group created by the fixture.

_{Reviews (2): Last reviewed commit: "[PyTorch] Destroy the process group crea..." | Re-trigger Greptile}