Integration with Hugging Face transformers library

[SunMarc]

Hi neuralmagic team !

Very nice work with AutoFP8 ! We were thinking of integrating AutoFP8 in transformers, so that users can run your checkpoints directly with transformers ! We would simply replace the linear layers by its quantized version. Hence, we would only support the inference. Let us know if you agree with this ! The goal would be to explose the quantized linear layer class in this repo (I see that you have several quantized linear) and import it in transformers.

I will be leading the integration, so any help is appreciated ! Also, are there any big blockers that I might not have seen ?

Thanks in advance !

[robertgshaw2-neuralmagic]

Hey @SunMarc - we are planning to push most of our development into llm-compressor and compressed-tensors which are the successors to this mini-repo that we are already working on integrating it into transformers (https://github.com/huggingface/transformers/pull/31704)

This supports:

• mixed precision w4a16 / w8a16
• w8a8 int8 (activation quantization)
• w8a8 fp8 (float point quantization)

We also support the following algorithms which can be applied to both fp8 and int8 and int4 models:

• ptq
• gptq
• smoothquant
• sparsegpt

We would prefer to put efforts related to transformers behind this framework (including doing a surge on fp8 and int8 compute with our cutlass kernels that we use in vllm)

[robertgshaw2-neuralmagic]

Couple other notes for fp8 on various compute capabilities:

• For pre-ampere, we could emulate (since the conversion is quick)
• For ampere GPUs, we can add support for our fp8 marlin kernels (mixed precision)
• For lovelace/hopper GPUs we can use torch.scaled_mm to make it easy

[robertgshaw2-neuralmagic]

For MoEs:

• we could consider adding the triton kernels we have in vllm as well