[feat] added the optimized config for Qwen3-30B-A3B Fp8

[sara4dev]

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

#22294

Test Plan

Test Result

Performance without tuned config - TP1/FP8

(pid=58866) INFO 08-10 23:45:45 [__init__.py:241] Automatically detected platform cuda.
(BenchmarkWorker pid=58866) WARNING 08-10 23:45:46 [fused_moe.py:701] Using default MoE config. Performance might be sub-optimal! Config file not found at /root/vllm/vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json
Batch size: 1, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 44.82 us
Batch size: 2, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 59.71 us
Batch size: 4, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 90.60 us
Batch size: 8, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 129.78 us
Batch size: 16, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 185.30 us
Batch size: 24, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 216.45 us
Batch size: 32, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 238.53 us
Batch size: 48, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 257.88 us
Batch size: 64, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 263.15 us
Batch size: 96, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 271.29 us
Batch size: 128, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1}
Kernel time: 274.56 us
Batch size: 256, config: {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}
Kernel time: 358.08 us
Batch size: 512, config: {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}
Kernel time: 393.85 us
Batch size: 1024, config: {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}
Kernel time: 537.17 us
Batch size: 1536, config: {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}
Kernel time: 728.29 us
Batch size: 2048, config: {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}
Kernel time: 912.21 us
Batch size: 3072, config: {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}
Kernel time: 1279.56 us
Batch size: 4096, config: {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}
Kernel time: 1647.03 us

Performance with tuned config- TP1/FP8

(pid=66907) INFO 08-11 04:28:27 [__init__.py:241] Automatically detected platform cuda.
(BenchmarkWorker pid=66907) INFO 08-11 04:28:28 [fused_moe.py:694] Using configuration from /root/vllm/vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json for MoE layer.
Batch size: 1, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 1, 'num_warps': 4, 'num_stages': 4}
Kernel time: 41.61 us
Batch size: 2, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 4, 'num_stages': 3}
Kernel time: 55.53 us
Batch size: 4, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 4, 'num_stages': 3}
Kernel time: 77.27 us
Batch size: 8, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 256, 'GROUP_SIZE_M': 16, 'num_warps': 4, 'num_stages': 4}
Kernel time: 117.03 us
Batch size: 16, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 4, 'num_stages': 3}
Kernel time: 162.91 us
Batch size: 24, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 4, 'num_stages': 4}
Kernel time: 192.34 us
Batch size: 32, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 4, 'num_stages': 3}
Kernel time: 208.46 us
Batch size: 48, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 256, 'GROUP_SIZE_M': 1, 'num_warps': 8, 'num_stages': 2}
Kernel time: 227.08 us
Batch size: 64, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 256, 'GROUP_SIZE_M': 1, 'num_warps': 4, 'num_stages': 4}
Kernel time: 234.61 us
Batch size: 96, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 256, 'GROUP_SIZE_M': 1, 'num_warps': 8, 'num_stages': 2}
Kernel time: 240.62 us
Batch size: 128, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 16, 'num_warps': 4, 'num_stages': 4}
Kernel time: 246.60 us
Batch size: 256, config: {'BLOCK_SIZE_M': 16, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 4, 'num_stages': 3}
Kernel time: 261.87 us
Batch size: 512, config: {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 8, 'num_stages': 3}
Kernel time: 286.08 us
Batch size: 1024, config: {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 4, 'num_stages': 3}
Kernel time: 339.66 us
Batch size: 1536, config: {'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 16, 'num_warps': 8, 'num_stages': 3}
Kernel time: 391.48 us
Batch size: 2048, config: {'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 256, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 8, 'num_stages': 4}
Kernel time: 456.74 us
Batch size: 3072, config: {'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 256, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 1, 'num_warps': 8, 'num_stages': 4}
Kernel time: 584.17 us
Batch size: 4096, config: {'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 256, 'BLOCK_SIZE_K': 128, 'GROUP_SIZE_M': 16, 'num_warps': 8, 'num_stages': 4}
Kernel time: 717.16 us

(Optional) Documentation Update

[gemini-code-assist]

Code Review

This pull request adds an optimized configuration for Fused MoE for what appears to be the Qwen3-30B-A3B model. However, the pull request description is empty and does not include any benchmark results to demonstrate the performance improvements from this new configuration. To properly evaluate this change, it's crucial to provide context and data that justify adding this specialized configuration. Please update the description with details about the test plan and the performance results (e.g., latency improvements) as per the PR template.

[gemini-code-assist]

The configurations for batch sizes "2", "4", "16", "32", and "256" are identical. Similarly, the configurations for "48" and "96" are identical, and for "2048" and "3072" as well.

While it's possible for different batch sizes to share an optimal configuration, having this many duplicates is unusual and may indicate that the tuning process was not exhaustive or that there's a potential copy-paste error. This could mean the configurations are not fully optimized for each batch size, which would defeat the purpose of this file.

Could you please double-check the tuning results and confirm these configurations are correct? If some are redundant, they could be removed to simplify the file, since the runtime will select the configuration for the closest batch size.

[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.

🚀

[jeejeelee]

Thank you very much for your contribution. If you could provide a description like #21700, that would be perfect

[sara4dev]

Sure, I'll get it in that format.