[PyTorch] Avoid registering FP8 scale update in ops without backward pass

[timmoon10]

Description

We keep track of the first TE module within an fp8_autocast context so that we can perform an FP8 scale update during its backward pass. However, the op fuser does not consider the possibility that it might not perform a backward pass. This was causing divergence problems in some finetuning workflows where the first layer does not require any grads.

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

• Avoid registering FP8 scale update in ops without backward pass

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

[timmoon10]

/te-ci pytorch