Add fp8-fused gemm kernel

[sfc-gh-reyazda]

This PR adds the new fused kernel for the Dense GeMM using fp8-quantized weight.

[jeffra]

One thing that needs to be resolved before merging this, this kernel requires triton==2.3.0. This should be checked at runtime and communicated to users somehow.

[HeyangQin]

Hi @jeffra. To clarify, does this kernel require exactly triton==2.3 or triton>=2.3?

[jeffra]

Hi @jeffra. To clarify, does this kernel require exactly triton==2.3 or triton>=2.3?

@sfc-gh-reyazda would know better, I am not sure if we've tested with newer triton than 2.3.0. I have not personally tested this at least.

[sfc-gh-reyazda]

Hi @jeffra. To clarify, does this kernel require exactly triton==2.3 or triton>=2.3?

It needs that specific version, unfortunately triton keeps changing/improving and their APIs change too so it is hard to track it properly. That's also another motivation to move to cutlass soon and have a more solid implementation to work independent of other libraries. On the other hand, Triton gives the flexibility to run on various hardwares. So, it is always a tradeoff. I think we need to have some more discussions on such dependencies later in a different discussion.
Best,
Reza