Update flashinfer to 0.6.1

[Fridge003]

Motivation

flashinfer -> 0.6.1
flashinfer-cubin -> 0.6.1

PRs dependent on this upgrade:
#15546
#15422
#15514
#15347
#14668
#16232
#16279
#16892
#16534
#12787
...

Modifications

Accuracy Tests

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.
• Work with maintainers to merge your PR. See the PR Merge Process

[gemini-code-assist]

Summary of Changes

Hello @Fridge003, 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 updates the FlashInfer library to version 0.6.0 across the project's configuration and dependency management files. This essential upgrade ensures that the project is aligned with the latest FlashInfer release, which is a prerequisite for several other pending pull requests and helps maintain compatibility and leverage new optimizations from the library.

Highlights

• FlashInfer Version Update: The FLASHINFER_VERSION argument in the Dockerfile has been updated from 0.5.3 to 0.6.0.
• Python Dependency Alignment: The flashinfer_python and flashinfer_cubin dependencies in pyproject.toml have been upgraded to version 0.6.0 to match the new FlashInfer release.
• Engine Version Assertion: The version assertion for flashinfer_python in engine.py has been updated to 0.6.0 to ensure compatibility checks are current.
• CI Script Update: The FLASHINFER_VERSION variable in the ci_install_dependency.sh script has been updated to 0.6.0.

🧠 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. ↩

[Fridge003]

cc @DarkSharpness

[gemini-code-assist]

Code Review

This pull request updates the flashinfer dependency to version 0.6.0. The changes are consistent across the Dockerfile, Python project dependencies, runtime version assertions, and CI installation scripts. My review includes one suggestion to improve maintainability by centralizing the version number.

[Fridge003]

/tag-and-rerun-ci again

[elvischenv]

The Flashinfer MoE API breakage changes(remove tile_tokens_dim) should come together with this PR, otherwise the tests using Flashinfer MoE will fail.

[Fridge003]

We are currently blocked at the update of FA4 #15182
Seems the FA kernels need to be upgraded first

[Fridge003]

Just checking this failing case: https://github.com/sgl-project/sglang/actions/runs/21020274512/job/60599048636?pr=15551（Seems to be the only non flaky test)

Seems the behavior of BatchDecodeWithPagedKVCacheWrapper in flashinfer backend has changed when upgrading from 0.5.3 to 0.6.1:

In 0.5.3, it will be dispatched to BatchPrefillWithPagedKVCacheKernel in cuda graph

While in 0.6.1, it will be dispatched to PrefillWithKVCacheKernel instead. But it's like 2x slower on the same workload

As a workaround, I will set fa3 as the default backend for Mixtral model to unblock CI.

My commands for profiling:

python3 -m sglang.launch_server --model-path mistralai/Mixtral-8x7B-Instruct-v0.1 --tp 2 --device cuda --host 127.0.0.1 --port 21000

python -m sglang.bench_serving --backend sglang --model mistralai/Mixtral-8x7B-Instruct-v0.1 --num-prompts 10 --sharegpt-output-len 10 --profile --port 21000

[Fridge003]

/rerun-stage performance-test-2-gpu

[github-actions]

✅ Triggered performance-test-2-gpu to run independently (skipping dependencies).

[github-actions]

🔗 View workflow run

[Fridge003]

/rerun-stage unit-test-backend-4-gpu-gb200

[github-actions]

✅ Triggered unit-test-backend-4-gpu-gb200 to run independently (skipping dependencies).

[github-actions]

🔗 View workflow run

[Swipe4057]

Just checking this failing case: https://github.com/sgl-project/sglang/actions/runs/21020274512/job/60599048636?pr=15551（Seems to be the only non flaky test)

Seems the behavior of BatchDecodeWithPagedKVCacheWrapper in flashinfer backend has changed when upgrading from 0.5.3 to 0.6.1:

In 0.5.3, it will be dispatched to BatchPrefillWithPagedKVCacheKernel in cuda graph

While in 0.6.1, it will be dispatched to `PrefillWithKVCacheKernel` instead. But it's like 2x slower on the same workload

As a workaround, I will set fa3 as the default backend for Mixtral model to unblock CI.

My commands for profiling:

python3 -m sglang.launch_server --model-path mistralai/Mixtral-8x7B-Instruct-v0.1 --tp 2 --device cuda --host 127.0.0.1 --port 21000

python -m sglang.bench_serving --backend sglang --model mistralai/Mixtral-8x7B-Instruct-v0.1 --num-prompts 10 --sharegpt-output-len 10 --profile --port 21000

Not only mixtral, Qwen 3 235B bandwidth collapsed on flashinfer

[Edenzzzz]

@Fridge003 PrefillWithKVCacheKernel is the FA3 paged attention in FlashInfer. I have tested a long time ago that it's not faster than FA2(flashinfer-ai/flashinfer#1340), and @zihaoye is working on optimizing FA3 (flashinfer-ai/flashinfer#2192).
Though I'm not sure why upgrading leads to calling FA3, here it's fixed to FA2.

sglang/python/sglang/srt/layers/attention/flashinfer_backend.py

Line 267 in 3908985

backend="fa2",