[Misc] DP : Add ExpertTokensMetadata

[varun-sundar-rabindranath]

Purpose

This PR introduces a ExpertTokensMetadata object that could be used to communicate token-expert-routing Metadata information. With this PR, ExpertTokensMetadata is used to communicate CPU / GPU tensors that contain the number of tokens allocated to each local expert.

On main we directly send the GPU tensor containing this information. This is insufficient, as there is a use case to communicate a CPU tensor with this information as well. This information, produced by DeepEP High-Throughput kernels, could be used by DeepGemmExperts to allocate exactly the required amount of workspace. (look at #20228)

Test Plan

pytest -s tests/kernels/moe/test_modular_kernel_combinations.py from #20449
pytest -s tests/kernels/moe/test_pplx_cutlass_moe.py

Test Result

tests pass locally on H100s

(Optional) Documentation Update

[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]

Summary of Changes

Hello @varun-sundar-rabindranath, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a dedicated ExpertTokensMetadata object to standardize and enhance the communication of token-expert routing information within the MoE layers. By encapsulating both GPU and CPU tensor representations of expert token counts, it addresses a critical need for components that require CPU-side metadata, thereby improving the flexibility and potential optimizations for MoE operations.

Highlights

• New Data Structure: Introduced a new ExpertTokensMetadata dataclass in vllm/model_executor/layers/fused_moe/modular_kernel.py. This object encapsulates metadata related to token-expert routing, specifically the number of tokens allocated to each local expert, and can hold both GPU and CPU tensor representations of this information.
• API Refinement: Refactored the prepare and apply methods across various Mixture-of-Experts (MoE) layer implementations (e.g., batched_deep_gemm_moe.py, cutlass_moe.py, fused_moe.py, deepep_ht_prepare_finalize.py, etc.) to utilize the new ExpertTokensMetadata object. This replaces the direct passing of expert_num_tokens tensors, providing a more structured and flexible way to communicate this routing metadata.
• Enhanced Functionality: Enabled the communication of CPU tensors containing expert token counts, which is a new capability. This is explicitly noted as beneficial for use cases like DeepGemmExperts, allowing them to allocate workspace more efficiently based on CPU-side information produced by DeepEP High-Throughput kernels.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces a ExpertTokensMetadata object for communicating token-expert-routing Metadata information, including CPU/GPU tensors for token allocation to local experts. The code changes replace expert_num_tokens with expert_tokens_meta across multiple files to accommodate both CPU and GPU tensors. The changes also include updates to the prepare methods in several files to handle the new ExpertTokensMetadata object.

[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

[tlrmchlsmth]

Looks good to me -- I know that there are some possible approaches to reducing the worst case memory footprint that wouldn't need a CPU tensor with the number of tokens per expert. But this could be handy anyway so let's merge it for now and continue to iterate

[varun-sundar-rabindranath]

Looks good to me -- I know that there are some possible approaches to reducing the worst case memory footprint that wouldn't need a CPU tensor with the number of tokens per expert. But this could be handy anyway so let's merge it for now and continue to iterate

The deepep-high-throughput kernels output a list of expert_num_tokens on the CPU. This is specifically useful for that case atleast.