TransformerEngine icon indicating copy to clipboard operation
TransformerEngine copied to clipboard

Add primary weighs fp8 support for mxfp8

Open kunlunl opened this issue 3 months ago • 1 comments

Description

This PR mainly adds the partial cast feature for mxfp8 primary weights. In FSDP, since each forward and backward pass requires gathering params, it's better to only gather the fp8 weights (row-wise for forward, column-wise for backward)

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)
  • [x] 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:

  • Add partial cast for mxfp8 primary weights

Checklist:

  • [x] I have read and followed the contributing guidelines
  • [x] The functionality is complete
  • [x] I have commented my code, particularly in hard-to-understand areas
  • [ ] I have made corresponding changes to the documentation
  • [x] My changes generate no new warnings
  • [x] I have added tests that prove my fix is effective or that my feature works
  • [x] New and existing unit tests pass locally with my changes

kunlunl avatar Aug 11 '25 16:08 kunlunl