[PyTorch] FP8 AllToAll

[yaox12]

Description

Add a FP8AllToAll layer, which conducts cast_to_fp8 -> all_to_all in fp8 -> cast_from_fp8. We're getting about 5% end to end performance gain in Mixtral 8x7B and 8x22B training with the parallelism configs where alltoall happens on inter-node connections such as IB/RoCE.

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)

Changes

Please list the changes introduced in this PR:

• Add a FP8AllToAll layer

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
• [x] 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

[yaox12]

The implementation looks reasonable, although this functionality seems somewhat Mixtral-specific. Perhaps it would be better to have this class live inside NeMo?

This implementation is used to improve the performance of AllToAllTokenDispatcher in MCore, which is used not only in Mixtral but also other MoE models, such as GPT-MoE, DBRX, Grok and so on. But you suggestion is reasonable. It should be fine to live in MCore. I will discuss with other folks in MoE development.