feat: add fused moe config for Qwen3-Next-80B-A3B-Instruct on B200

[zixuanzhang226]

Motivation

Add fused MoE config for Qwen3-Next-80B-A3B-Instruct on NVIDIA B200

Accuracy Tests

With config:

Without config:

Benchmarking

With config:

============ Serving Benchmark Result ============
Successful requests:                     512
Maximum request concurrency:             256
Request rate configured (RPS):           12.00
Benchmark duration (s):                  108.26
Total input tokens:                      1789330
Total generated tokens:                  768000
Request throughput (req/s):              4.73
Request goodput (req/s):                 4.73
Output token throughput (tok/s):         7094.27
Total Token throughput (tok/s):          23622.92
---------------Time to First Token----------------
Mean TTFT (ms):                          683.71
Median TTFT (ms):                        497.53
P50 TTFT (ms):                           497.53
P90 TTFT (ms):                           1422.95
P95 TTFT (ms):                           1652.33
P99 TTFT (ms):                           2197.77
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          31.87
Median TPOT (ms):                        33.19
P50 TPOT (ms):                           33.19
P90 TPOT (ms):                           35.73
P95 TPOT (ms):                           36.00
P99 TPOT (ms):                           36.41
---------------Inter-token Latency----------------
Mean ITL (ms):                           31.87
Median ITL (ms):                         24.51
P50 ITL (ms):                            24.51
P90 ITL (ms):                            25.25
P95 ITL (ms):                            28.83
P99 ITL (ms):                            161.44
==================================================

Without config:

============ Serving Benchmark Result ============
Successful requests:                     512
Maximum request concurrency:             256
Request rate configured (RPS):           12.00
Benchmark duration (s):                  114.89
Total input tokens:                      1789330
Total generated tokens:                  768000
Request throughput (req/s):              4.46
Request goodput (req/s):                 4.46
Output token throughput (tok/s):         6684.48
Total Token throughput (tok/s):          22258.37
---------------Time to First Token----------------
Mean TTFT (ms):                          850.26
Median TTFT (ms):                        592.93
P50 TTFT (ms):                           592.93
P90 TTFT (ms):                           1909.28
P95 TTFT (ms):                           2278.52
P99 TTFT (ms):                           2565.70
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          33.98
Median TPOT (ms):                        35.52
P50 TPOT (ms):                           35.52
P90 TPOT (ms):                           37.65
P95 TPOT (ms):                           37.89
P99 TPOT (ms):                           38.23
---------------Inter-token Latency----------------
Mean ITL (ms):                           33.98
Median ITL (ms):                         26.17
P50 ITL (ms):                            26.17
P90 ITL (ms):                            26.69
P95 ITL (ms):                            26.93
P99 ITL (ms):                            150.05
==================================================

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.

[gemini-code-assist]

Summary of Changes

Hello @zixuanzhang226, 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 an optimized configuration for the Qwen3-Next-80B-A3B-Instruct large language model, leveraging a fused Mixture-of-Experts architecture on NVIDIA B200 hardware. The primary goal is to boost inference performance, which has been successfully validated through comprehensive benchmarking, showing improved efficiency and responsiveness.

Highlights

• New Fused MoE Configuration: A new fused Mixture-of-Experts (MoE) configuration has been added specifically for the Qwen3-Next-80B-A3B-Instruct model, optimized for deployment on NVIDIA B200 GPUs.
• Performance Improvement: Benchmarking results demonstrate significant performance enhancements with the new configuration, including increased request throughput (from 4.46 req/s to 4.73 req/s), higher output token throughput (from 6684.48 tok/s to 7094.27 tok/s), and reduced Time to First Token (TTFT) and Time per Output Token (TPOT) latencies.

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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description

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 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

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 new fused MoE configuration for the Qwen3-Next-80B-A3B-Instruct model running on NVIDIA B200 GPUs. The change consists of adding a single JSON file with tuned Triton kernel parameters for various batch sizes. The benchmark results provided in the description clearly demonstrate a performance improvement with this new configuration. The implementation is straightforward and correct. This is a valuable addition for optimizing performance on newer hardware.

[zhyncs]

Did you tune this on the sglang b200 using a script, or did you copy and paste it? Please paste the gsm8k result.

[zixuanzhang226]

Did you tune this on the sglang b200 using a script, or did you copy and paste it? Please paste the gsm8k result.

I tuned it using benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py, and the gsm8k result screenshot was included in the PR description already. Thanks!