Optimize GDN decode for Qwen3 Next

[samuellees]

Motivation

As MoE/FullAttn/GEMM has been optimized, the GDN kernel becomes a bottleneck for Q3N. For example, decode linear attention takes about 15~23% with cc=64/256 with ISL=1k in NVFP4 inference on B200.

This PR apply an optimize on GDN decode kernel, which brings significant benefits e2e.

Ref: vllm-project/vllm#31722

cc @yizhang2077

Modifications

Accuracy Tests and Benchmarking

TEP4, MTP OFF, B200
It gives about 10% e2e improvements on gsm8k dataset.

# Before
Accuracy: 0.945
Invalid: 0.000
Latency: 74.870 s
Output throughput: 2967.035 token/s

# After
Accuracy: 0.943
Invalid: 0.000
Latency: 67.961 s
Output throughput: 3266.447 token/s

Benchmarking and Profiling

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.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello @samuellees, 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 critical optimization to the Gated Delta Rule (GDN) decoding kernel, specifically targeting the Qwen3 Next model. By adjusting a key parameter in the fused recurrent attention mechanism, the PR significantly improves overall inference performance, leading to higher output throughput and reduced latency, addressing a previously identified bottleneck.

Highlights

• GDN Kernel Optimization: The BV block size in the fused_recurrent_gated_delta_rule_fwd function has been increased from 8 to 32. This change is aimed at improving the performance of the Gated Delta Rule (GDN) kernel, which was identified as a bottleneck for Qwen3 Next decoding.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

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
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 pull request 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 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 optimizes the Gated Delta Rule (GDR) kernel by increasing the block size BV from 8 to 32 in fused_recurrent_gated_delta_rule_fwd. This is a good performance tuning that, as the benchmarks show, significantly improves throughput.

However, this optimization seems to be applied inconsistently across the codebase. I've found a few other places where BV is still capped at 8. Given the PR's goal is to optimize decoding, these should probably be updated as well to ensure consistent performance improvements across all relevant kernels.

Specifically, please consider applying the same change to:

• python/sglang/srt/layers/attention/fla/fused_recurrent.py at line 543 in fused_recurrent_gated_delta_rule_update_fwd.
• python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py at line 184 in fused_sigmoid_gating_delta_rule_update.
• python/sglang/srt/layers/attention/fla/kda.py at line 56 in fused_recurrent_kda_fwd.

Applying this optimization consistently would likely yield further performance benefits, especially in decoding scenarios which seems to be the main goal of this PR.

[Fridge003]

/rerun-stage unit-test-backend-4-gpu

[github-actions]

✅ Triggered unit-test-backend-4-gpu to run independently (skipping dependencies).

[github-actions]

🔗 View workflow run

[Fridge003]

/rerun-stage unit-test-backend-4-gpu

[github-actions]

✅ Triggered unit-test-backend-4-gpu to run independently (skipping dependencies).

[github-actions]

🔗 View workflow run

[Fridge003]

/tag-and-rerun-ci

[Fridge003]

/rerun-stage unit-test-backend-4-gpu

[github-actions]

✅ Triggered unit-test-backend-4-gpu to run independently (skipping dependencies).

[github-actions]

🔗 View workflow run