Modularize fused experts and integrate PPLX kernels

[bnellnm]

This PR defines a set of base classes used to make MoE kernels more modular. The goal is to be able to utilize different communication mechanisms with any fused MoE kernel without needing to have combinatoric implementations.

The fused moe kernels are broken down into the following components:

[Router] → [Quantize-Dispatch] → [Permute-Experts-Unpermute] → [Combine]                                                                                                       

Each component will be independent of the others except for [Quantize-Dispatch] and `[Combine] (see below). The components can then be mixed and matched with so that DP+EP can be supported easily for multiple MoE kernel implementations.

The following main classes are defined:

• FusedMoEQuantizeDispatchCombine - an abstract base class for quantization, dispatching and combing. The dispatch method takes care of any needed quantization and the combine method applies weights and does the final reduction of the output.
• FusedMoEPermuteExpertsUnpermute - an abstract base class for the main fused MoE operation. One important feature to note is that this class does not apply topk weights or reduce the final output.
• FusedMoEModularKernel - an interface class that combines a FusedMoEQuantizeDispatchCombine and a
• FusedMoEPermuteExpertsUnpermute to provide the standard fused MoE kernel interface.
• StandardDispatchCombine - a concrete class that can be used for serial Triton, DeepGemm and CUTLASS moe implementations.

The implementations for the DeepGemm and CUTLASS moe functions have been replaced with the modularized versions. There's also a modularized version for the Triton kernels but it will not be enabled by default.

[Quantize-Dispatch] and [Combine] functionality are bundled into a single class FusedMoEQuantizeDispatchCombine since they could use collective communication mechanisms that need to be consistent.

cc @ElizaWszola , @varun-sundar-rabindranath

[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.

🚀

[tlrmchlsmth]

For readers: We're doing this to support the pplx-kernel integration. We can use this structure for DeepEP as well.

Right now our fused MoE is implemented as something very very roughly like:

[Router] → [Quantize] → [Experts + topk_weight scaling + reduction]

This is a problem as the topk_weight scaling and reduction now need to happen during combine. We need to fit dispatch in there as well.

This PR defines a set of base classes used to make MoE kernels more modular. The goal is to be able to utilize different communication mechanisms with any fused MoE kernel without needing to have combinatoric implementations.

The fused moe kernels are broken down into the following components:

[Router] → [Quantize-Dispatch] → [Permute-Experts-Unpermute] → [Combine]

☝️ This is what @bnellnm and I agreed on. However:

The other option we originally considered was:

[Router] → [Quantize-Dispatch-Permute] → [Experts] → [Unpermute-Combine]

Right now I am thinking that permute/unpermute will (unfortunately) depend both on the implementation of both dispatch/combine and experts, so we should consider breaking that out.

[tlrmchlsmth]

nice and clean

[abcdabcd987]

One thing I forgot to put in our examples is -- Please call the destructor! ata.destroy()

[bnellnm]

One thing I forgot to put in our examples is -- Please call the destructor! ata.destroy()

Do all references to an AllToAll need to be destroyed? The current plan is to have a cache (that will be in a different PR) manage all the AllToAll instances and the PplxDispatchCombine would hold on to a reference of one of the cached objects.

[abcdabcd987]

Do all references to an AllToAll need to be destroyed?

No. Call destroy() only when you are shutting down the engine (or removing the model from GPU, etc...)

You are right to cache the object. It is supposed to be reused across layers and across runs.

[mergify]

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

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