[wip] optimize memory-bound perf for machete

[czhu-cohere]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

For memory-bound workloads we observed Marlin has much lower latency than Machete. Add different kernel/schedule config for small batch sizes (M <= 16); specifically, change the kernel schedule from TmaWarpSpecializedCooperative -> TmaWarpSpecialized and use the compatible Default scheduler instead of StreamK (which is not supported by non-coop/pingpong schedule).

I think the main intuition is that for Cooperative schedule, we have 1 producer warp group (responsible for loading) and 2 consumer warp groups (responsible for math)

The perf difference w/wo the change is summarized below

todo

Some other notes

• Keeping the epilogue to use Cooperative schedule is crucial
• Comparing different benchmark methods (benchmark_machete.py vs triton do_bench vs ncu vs vllm e2e profiling)
  • benchmark_machete.py seems to match the actual kernel execution time in the forward pass (from torch profiler used for e2e vllm profiling best)
  • do_bench is consistently a bit slower
  • the duration reported by ncu may not be accurate due to the profiling/metrics collection overhead.

Other things I tried that did not work

• Different combinations of TmaWarpSpecialized{Pingpong/cooperative} for kernel/epilogue schedule
• Sweeping tile/cluster shapes
• Eliminate shared bank conflicts reported in original kernel chosen for bs=1
• Shrink the tile shape from (128, 16, 64) -> (128, 8, 64) which shrinks the mma atom from 64x16x16 -> 64x8x16 (in the transposed problem, K corresponds to the activation batch dimension). The motivation is that e.g. for bs=1, majority of 64x16x16 is wasted, and 64x8x16 is the smallest atom. However, the latency of 64x16 is the same as 64x8, and also we cannot use the widest CopyR2S instruction in the epilogue so the result is not any faster.

The following are some issues I will try to address in a future PR (I'm not super familiar with CUTLASS yet).

• ncu report some bank conflicts for loading A from smem -> rmem. Eliminating these may improve by ~10% and approach parity with Marlin.
• Because we changed the scheduler, there are some wave quantization effects on the M dimension of the transposed problem (equivalently, the number of columns of the original weight matrix W). For larger values of M, this is less of an issue.

Test Plan

1. test_machete_mm.py for correctness
2. lm-eval with generation task, force small batch size so it dispatches to correct kernel
3. benchmark_machete.py for kernel latency
4. e2e forward pass profiling to see the TPOT difference

Test Result

TODO

(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 @czhu-cohere, 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 aims to optimize the memory-bound performance of machete quantization kernels. My changes involve a significant refactoring of how these kernels are generated, allowing for more dynamic selection of kernel schedules based on input dimensions. This enables specific performance optimizations for certain matrix shapes and includes the necessary configuration for a new model.

Highlights

• Machete Kernel Generation Refactoring: I've refactored the machete kernel generation process to dynamically select the KernelSchedule based on the ScheduleConfig. This moves away from hardcoded kernel types, allowing for more flexible and optimized kernel instantiation.
• Performance Optimization for Small M: Specifically, for matrix multiplication cases where M (batch size or sequence length) is small (1-16), I've enabled the TmaWarpSpecialized kernel schedule. This is intended to improve performance for memory-bound scenarios in these specific dimensions.
• New Model Weight Shape Configuration: I've added specific weight shape configurations for the CohereLabs/c4ai-command-a-03-2025 model within the benchmarks/kernels/weight_shapes.py file. This provides necessary dimensions for gate_up_proj, down_proj, qkv_proj, and o_proj for this model.

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 optimizes memory-bound performance for the machete kernel by making the kernel schedule configurable based on the problem size. For small M dimensions, a different kernel schedule and tile scheduler are now used. The changes are implemented by modifying the C++ code generator (generate.py) to support schedule overrides and updating the machete_mm_kernel.cuh to use this new configuration. A new model has also been added to the benchmarks. The changes appear logical and well-implemented. I have one minor suggestion to improve code consistency in the benchmark configuration file.