[Kernel] Refactor Cutlass c3x

[varun-sundar-rabindranath]

Refactor Cutlass c3x kernels for better maintainability and easier experimentation.

• Break scaled_mm_c3x.cu into,
  • scaled_mm_c3x.cuh : All the base cutlass c3x code (cutlass_3x_gemm and cutlass_gemm_caller).
  • scaled_mm_c3x_sm90_fp8_dispatch.cuh : All fp8 kernels along with the gemm shape based dispatch function.
  • scaled_mm_c3x_sm90_int8_dispatch.cuh : All int8 kernels along with the gemm shape based dispatch function.
  • scaled_mm_c3x.cu : interfaces expected by scaled_mm_entry.cu

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[varun-sundar-rabindranath]

@tlrmchlsmth @ProExpertProg @LucasWilkinson PTAL. Thanks!

[LucasWilkinson]

LGTM (just FYI may conflict with #9855)

[tlrmchlsmth]

LGTM

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @varun-sundar-rabindranath.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[varun-sundar-rabindranath]

@tlrmchlsmth re-requesting review as the PR is now rebased.

[tlrmchlsmth]

Looks good! (Likely wait for #10995)

[tlrmchlsmth]

Actually do you need to turn off clang-format here? The reason for turning it off is that CUTLASS headers need to be included in a specific order but it looks like that's not the case following the refactor

[varun-sundar-rabindranath]

Using clang-format in this block, turns,

#include <cudaTypedefs.h>

#if defined CUDA_VERSION && CUDA_VERSION >= 12000

#include "scaled_mm_c3x_sm90_fp8_dispatch.cuh"
#include "scaled_mm_c3x_sm90_int8_dispatch.cuh"

#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"

using namespace vllm;

into

#include <cudaTypedefs.h>

#if defined CUDA_VERSION && CUDA_VERSION >= 12000

    #include "scaled_mm_c3x_sm90_fp8_dispatch.cuh"
    #include "scaled_mm_c3x_sm90_int8_dispatch.cuh"

    #include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"

using namespace vllm;

the #if seems trigger inconsistent indenting. I switched off clang-format in this block to avoid that.

I moved the original clang-format toggle to scaled_mm_c3x.cuh.

vllm/csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cuh

Line 3 in dad8e47

// clang-format will break include orders

[varun-sundar-rabindranath]

Had second thoughts about this and removed the clang-format block. It is probably better to stick to the convention.