[Refactor] Remove Unused Naive Moe Kernels

[yewentao256]

Purpose

Fixed #21124

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

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 either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Code Review

This pull request is a refactoring effort to remove unused "naive" Mixture-of-Experts (MoE) kernels. The changes are clean and involve the deletion of C++ kernels, Python wrappers, benchmarks, and tests related to the old moe_permute and moe_unpermute operations. This cleanup improves the maintainability of the codebase. The changes appear to be correct and self-contained. I have reviewed the removals and found no issues.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @yewentao256.

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

[yewentao256]

Seems stagnant, could you take a look? Thanks! @WoosukKwon

[tlrmchlsmth]

I think we're going to integrate them here https://github.com/vllm-project/vllm/pull/17934/files

[yewentao256]

@tlrmchlsmth Thanks for letting me know.
Seems that the pr is same with the #20903 @varun-sundar-rabindranath ?
Do you know which one is faster?

[varun-sundar-rabindranath]

Hi @yewentao256 - I believe there is value is having the cuda permute/unpermute kernels.

I experimented with both the triton kernels (introduced in #20903) and the cuda kernels (#20982) - and decided to merge the triton kernels due to slightly better TTFT.
Arguments for having the cuda kernel:

• I think there will be regimes where the cuda kernels will be faster
• In the context of DeepGemmExperts, the cuda kernels are slower because they do some unnecessary extra work - eliminating that will make them faster.
• The cuda kernels would probably be useful in Cutlass grouped gemm data-prep (permute/unpermute)

The goal is to have a PermuteUnpermute abstract class and have both Triton and CUDA implementations - that way we can have a dispatching logic based on the number of tokens to decide which implementation to use.

[yewentao256]

@varun-sundar-rabindranath
Make sense, thanks for letting me know the context!

[yewentao256]

Close this pr as not needed