Disable torch autocast context in rope forward pass

[pstjohn]

Description

We noticed issues in model convergence when we calculated our rope embeddings inside our model's forward pass, which we wrapped in a torch.autocast() context. We ended up tracing it back to the reduced precision of the rope embedding calculation, which is easily fixed by just disabling the autocast context in this function.

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:

• Change A
• Change B

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