[Perf] fuse q, k norm for Flux2Attention

[zminglei]

Motivation

fuse q, k norm for Flux2Attention

Modifications

Accuracy Tests

Benchmarking and Profiling

sglang generate --model-path /shared/public/elr-models/black-forest-labs/FLUX.2-dev --prompt "A Logo With Bold Large Text: SGL Diffusion" --save-output

Before:

[01-17 00:21:11] [DenoisingStage] average time per step: 0.4453 seconds
[01-17 00:21:11] [DenoisingStage] finished in 22.6614 seconds
[01-17 00:21:11] [DecodingStage] started...
[01-17 00:21:12] [DecodingStage] finished in 0.2300 seconds
[01-17 00:21:12] Peak GPU memory: 64.49 GB, Remaining GPU memory at peak: 75.91 GB. Components that can stay resident: ['text_encoder']
[01-17 00:21:12] Output saved to outputs/A_Logo_With_Bold_Large_Text_SGL_Diffusion_20260117-002048_b81248a0.png
[01-17 00:21:12] Pixel data generated successfully in 24.32 seconds
[01-17 00:21:12] Completed batch processing. Generated 1 outputs in 24.32 seconds
[01-17 00:21:12] Memory usage - Max peak: 66041.99 MB, Avg peak: 66041.99 MB

After:

[01-17 00:23:49] [DenoisingStage] average time per step: 0.4346 seconds
[01-17 00:23:49] [DenoisingStage] finished in 22.1209 seconds
[01-17 00:23:49] [DecodingStage] started...
[01-17 00:23:49] [DecodingStage] finished in 0.2593 seconds
[01-17 00:23:49] Peak GPU memory: 64.49 GB, Remaining GPU memory at peak: 75.91 GB. Components that can stay resident: ['text_encoder']
[01-17 00:23:50] Output saved to outputs/A_Logo_With_Bold_Large_Text_SGL_Diffusion_20260117-002326_178f3467.png
[01-17 00:23:50] Pixel data generated successfully in 23.75 seconds
[01-17 00:23:50] Completed batch processing. Generated 1 outputs in 23.76 seconds
[01-17 00:23:50] Memory usage - Max peak: 66041.99 MB, Avg peak: 66041.99 MB

24.32s -> 23.76s, around 2.5% e2e latency improvement

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 @zminglei, 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 significant optimization to the Flux2Attention mechanism by fusing the query and key normalization steps. By consolidating these operations into a single kernel, the change reduces computational overhead and improves memory access patterns, leading to a measurable performance gain. The primary goal is to enhance the inference speed of models utilizing Flux2Attention without altering the model's output accuracy.

Highlights

• Performance Optimization: Implemented a fused QK (Query-Key) normalization operation for Flux2Attention, which combines two separate RMSNorm operations into a single, more efficient kernel. This change resulted in an approximate 2.5% improvement in end-to-end latency during benchmarking.
• Tensor Handling: Enhanced the apply_qk_norm function to correctly handle 4D input tensors by ensuring they are contiguous before reshaping for the fused kernel, and then restoring their original shapes upon return.
• Code Refactoring: Replaced explicit self.norm_q(query) and self.norm_k(key) calls within the Flux2Attention module with a single call to the new apply_qk_norm utility, leveraging the fused kernel for improved efficiency.

🧠 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 introduces a fused kernel for QK normalization in Flux2Attention, which yields a nice 2.5% end-to-end latency improvement according to the benchmarks. The implementation is clean and correctly handles non-contiguous tensors for the new fused path by adding contiguity checks. This is a good robustness improvement. I've added one comment regarding applying this robustness improvement to the fallback path as well for consistency. Overall, this is a solid contribution that improves performance.

[zminglei]

/tag-and-rerun-ci retry

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[byjiang1996]

Nice!