Add H20-3e fused MoE kernel tuning configs for DeepSeek-R1/V3

[Xu-Wenqing]

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

Purpose

H20-3e is H20 GPU with 141G memory. Add Moe tuning configs for DeepSeek-R1/V3

Test Plan

Tuning command:

python /mnt/workspace/vllm/benchmarks/kernels/benchmark_moe.py --model /mnt/data/DeepSeek-R1-0528 -tp 8 --dtype fp8_w8a8 --tune 

Deploy DeepSeek-R1-0528:

vllm serve /mnt/data/DeepSeek-R1-0528 --host 0.0.0.0 --port 8000 --root '/' --no-enable-prefix-caching --trust-remote-code --tensor-parallel-size 8 --served-model-name DeepSeek-R1-0528

Benchmark:

python3 -m sglang.bench_serving --tokenizer /mnt/data/DeepSeek-R1-0528 --host 0.0.0.0 --port 8000 --backend vllm --dataset-name random --random-input 1024 --random-output 512 --max-concurrency 8 --num-prompt 200

Test Result

Results without Moe tuning:

============ Serving Benchmark Result ============
Backend:                                 vllm      
Traffic request rate:                    inf       
Max request concurrency:                 8         
Successful requests:                     200       
Benchmark duration (s):                  276.03    
Total input tokens:                      103005    
Total generated tokens:                  53590     
Total generated tokens (retokenized):    53336     
Request throughput (req/s):              0.72      
Input token throughput (tok/s):          373.17    
Output token throughput (tok/s):         194.15    
Total token throughput (tok/s):          567.31    
Concurrency:                             7.80      
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   10761.65  
Median E2E Latency (ms):                 10697.00  
---------------Time to First Token----------------
Mean TTFT (ms):                          162.99    
Median TTFT (ms):                        154.33    
P99 TTFT (ms):                           470.98    
---------------Inter-Token Latency----------------
Mean ITL (ms):                           39.88     
Median ITL (ms):                         37.86     
P95 ITL (ms):                            38.65     
P99 ITL (ms):                            123.22    
Max ITL (ms):                            1000.79   
==================================================

Results with Moe tuning:

============ Serving Benchmark Result ============
Backend:                                 vllm      
Traffic request rate:                    inf       
Max request concurrency:                 8         
Successful requests:                     200       
Benchmark duration (s):                  238.14    
Total input tokens:                      103005    
Total generated tokens:                  53590     
Total generated tokens (retokenized):    53449     
Request throughput (req/s):              0.84      
Input token throughput (tok/s):          432.55    
Output token throughput (tok/s):         225.04    
Total token throughput (tok/s):          657.58    
Concurrency:                             7.77      
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   9255.99   
Median E2E Latency (ms):                 9082.86   
---------------Time to First Token----------------
Mean TTFT (ms):                          157.92    
Median TTFT (ms):                        149.74    
P99 TTFT (ms):                           463.50    
---------------Inter-Token Latency----------------
Mean ITL (ms):                           34.17     
Median ITL (ms):                         32.04     
P95 ITL (ms):                            32.67     
P99 ITL (ms):                            125.23    
Max ITL (ms):                            978.02    
==================================================

[gemini-code-assist]

Hello @Xu-Wenqing, 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!

Summary of Changes

Hello! Gemini here, providing a summary of this pull request to help everyone quickly understand the changes and context.

This PR, authored by Xu-Wenqing, aims to improve the performance of the fused MoE kernel specifically for the DeepSeek-R1/V3 model when running on NVIDIA H20-3e GPUs using fp8_w8a8 precision. The core change involves adding a new configuration file containing optimized kernel tuning parameters for various batch sizes. The PR description includes benchmark results demonstrating the effectiveness of these tuning configurations, showing improvements in request and token throughput, as well as reduced end-to-end and inter-token latency compared to running without these specific tunings.

Highlights

• MoE Kernel Tuning: Adds specific tuning configurations for the fused Mixture-of-Experts (MoE) kernel.
• Hardware Specific: The tuning configurations are tailored for the NVIDIA H20-3e GPU.
• Model Specific: The tuning is intended for the DeepSeek-R1/V3 model.
• Performance Improvement: Benchmark results show performance gains (throughput and latency reduction) when using these tuning configs.
• Data Type: The configurations are for the fp8_w8a8 data type.

Changelog

• vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json
  • Added a new JSON file containing fused MoE kernel tuning parameters.
  • Includes optimized BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K, GROUP_SIZE_M, num_warps, and num_stages for various batch sizes (1 to 4096).
  • These parameters are specific to the NVIDIA H20-3e GPU, fp8_w8a8 dtype, and MoE configuration with E=256, N=256, block_shape=[128,128].

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.

---

Kernels run so fast,
Tuning makes the speed increase,
GPU hums along.

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. ↩

[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 introduces new fused MoE kernel tuning configurations for the NVIDIA H20-3e GPU with fp8_w8a8 data type, specifically for E=256,N=256 and block_shape=[128,128]. The changes are well-documented in the PR description, including a clear purpose, test plan, and benchmark results that demonstrate a notable performance improvement.

The added JSON configuration file is well-structured, and its filename adheres to the established conventions, which is great for maintainability. The tuned parameters appear consistent with what one might expect from a kernel tuning process.

Overall, this is a valuable addition that enhances performance for the specified hardware and model configuration. Excellent work!

Summary of Findings

• No issues meeting severity criteria: No issues of medium, high, or critical severity were found in the submitted JSON configuration file. The configurations appear to be the result of a successful tuning process, validated by the provided benchmark improvements.

Merge Readiness

The pull request is well-prepared and the added configurations are backed by positive benchmark results. The code (the new JSON configuration file) is clear and follows existing patterns. Based on this review, the changes appear to be in good shape for merging. As always, ensure any other necessary reviews or CI checks are completed. I am unable to approve the pull request myself, so please ensure it is reviewed and approved by authorized maintainers.

[Xu-Wenqing]

@mgoin @DarkLight1337

[mgoin]

LGTM, thank you for the contribution!