[WIP] FP8 support.

[Narsil]

What does this PR do?

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[flozi00]

Would it be worth to discuss about transformers engine implementation since they say it's compatible with ada lovelace and higher ?
The comment in your code says it needs Hopper+

source: https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/index.html#:~:text=Support%20for%20FP8%20on%20NVIDIA%20Hopper%20and%20NVIDIA%20Ada%20GPUs

the implementation seems to be very easy, as its already done in accelerate
https://github.com/huggingface/accelerate/blob/7aafa25673f4324cbcaeef508a13e09576f12dc4/src/accelerate/utils/transformer_engine.py#L30-L41

[Narsil]

Would it be worth to discuss about transformers engine implementation since they say it's compatible with ada lovelace and higher ? The comment in your code says it needs Hopper+

Well it works whereever FP8 works, H100 are the most common cluster ones I think.

source: https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/index.html#:~:text=Support%20for%20FP8%20on%20NVIDIA%20Hopper%20and%20NVIDIA%20Ada%20GPUs

the implementation seems to be very easy, as its already done in accelerate https://github.com/huggingface/accelerate/blob/7aafa25673f4324cbcaeef508a13e09576f12dc4/src/accelerate/utils/transformer_engine.py#L30-L41

The implementation is silently not done with dimensions are not 16 aligned, this is not acceptable in TGI.

There are 2 issues with the current PR implem.
Every call packs f16 buffer before calling gemm, that's a slowdown for everycall, + the 16 padding.

That's what makes it slow, and TransformerEngine is just as slow: https://github.com/NVIDIA/TransformerEngine/blob/main/transformer_engine/pytorch/module/linear.py#L91C13-L92

So afaik there's no win to using transformer engine over this implementation.
The real work will be removing those unecessary ops (unpacking and padding).

And since LLM mostly operate at seqlen < 16 (during decode this is pretty much batch_size), we also most likely need a customized gemv to make things faster (This limitation on size is very suprising to me)

[flozi00]

The pytorch lab implementation will have inference support later

pytorch-labs/float8_experimental#187 (comment)

I am also checking the tensorrt llm implementation of fp8