[PyTorch] Store module extra state in tensor

[timmoon10]

Description

TE modules implement get_extra_state/set_extra_state in order to include FP8 state in checkpoints. We currently pickle the FP8 state and store in a io.BtyesIO object, but this is problematic because PyTorch makes no guarantees if the extra state is not a torch.Tensor. This has resulted in problems with ONNX export and Hugging Face Transformers.

https://github.com/NVIDIA/TransformerEngine/pull/363 changed from storing the extra state in a torch.Tensor to io.BytesIO in order to reduce the overhead from GPU-CPU memory transfers. This PR restores the original torch.Tensor format, but performs the memory transfers asynchronously to reduce overhead. It's similar to the approach used in the operation-based API (https://github.com/NVIDIA/TransformerEngine/pull/1063). It should be backward compatible and I've been able to load existing checkpoints. The attention docs mention extra state, but I don't think this PR affects it.

Fixes https://github.com/NVIDIA/TransformerEngine/issues/1317.

Type of change

• [ ] Documentation change (change only to the documentation, either a fix or a new content)
• [x] 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 refractor

Changes

• Store module extra state in tensor instead of io.BytesIO

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

[timmoon10]

/te-ci pytorch L0 L1