[jberchtold-nvidia]

Description

Checks CuDNN version and if below the fixed version does unfused norm and MXFP8 quantize to prevent producing incorrect results.

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

• Check CuDNN version and perform unfused layernorm in JAX if it is below the fixed version
• If below the fixed CuDNN version, lessen the strictness of the tolerances for MXFP8 norm tests as unfused layernorm loses some precision with the intermediate cast into the input dtype if that is less than fp32

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

[jberchtold-nvidia]

/te-ci L0

[jberchtold-nvidia]

/te-ci L0

[jberchtold-nvidia]

/te-ci L0

[jberchtold-nvidia]

Pushed branch to main repo instead of fork by mistake. Closing to replace with #1700