[2/2] Support MHA prefill with FlashAttention 4.

[lifuhuang]

(co-author @hyhieu)

Updates: the kernel change has been checked in separately in: #10940

Motivation

Add support for FA4 MHA prefill, changes mostly based on: #9428

Modifications

Accuracy Tests

lm_eval \
  --model local-chat-completions \
  --model_args model=gpt-oss,base_url=http://127.0.0.1:30010/v1/chat/completions,num_concurrent=128,timeout=999999,max_gen_toks=2048 \
  --tasks gsm8k \
  --batch_size 128 \
  --apply_chat_template \
  --num_fewshot 8

FA4:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	8	exact_match	↑	0.8931	±	0.0085
		strict-match	8	exact_match	↑	0.3419	±	0.0131

Baseline (TRTLLM-MHA):

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	8	exact_match	↑	0.8840	±	0.0088
		strict-match	8	exact_match	↑	0.3207	±	0.0129

Benchmarking and Profiling

For the model I am benchmarking: openai/gpt-oss-20b. I am not seeing significant difference between fa4 and trtllm-mha, both are significantly faster than triton, expectedly.

FA4:

TRTLLM-MHA

Triton

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 @lifuhuang, 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 significantly enhances the system's ability to leverage FlashAttention v4 for prefill operations, aiming to improve performance and efficiency. It achieves this by removing previous hardcoded limitations on fa4 usage and by refactoring the attention backend selection process into a dedicated utility, making the system more robust and easier to configure for different attention mechanisms.

Highlights

• FlashAttention v4 Support: FlashAttention v4 (fa4) is now fully supported for Multi-Head Attention (MHA) prefill operations, removing previous restrictions that limited its use to specific model architectures or prevented its application during prefill.
• Backend Selection Refactoring: The logic for determining the appropriate prefill and decode attention backends has been centralized into a new utility function, determine_attention_backends, improving code modularity and maintainability.
• Expanded Backend Options: The list of supported attention backends for GptOssForCausalLM models has been updated to include 'fa4', allowing for more flexible and potentially faster attention computations.

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 adds support for MHA prefill with FlashAttention 4 by removing assertions that previously restricted its use. The changes also include a refactoring of the attention backend determination logic into a new utility function.

My main feedback is to further improve the refactoring by moving the new determine_attention_backends function into the ServerArgs class as a method. This would improve code structure by eliminating a dependency from the low-level server_args.py module to the large utils.py module, which is better for maintainability. I've left specific suggestions on how to achieve this.

[lifuhuang]

TODO: Need to check in first and bump up sgl-kernel.

[lifuhuang]

waiting for #10940 to merge first.

[Swipe4057]

The H100-H200 series is not supported by FA4, right?

[cicirori]

The H100-H200 series is not supported by FA4, right?

It was restricted to sm100 because of we only tested this on blackwell primary optimization object.
and FA4 do support sm90 when page_table is None.
It would be nice if you can tell if we can just enable sm90 in MHA prefill.

[zhyncs]

@lifuhuang can u upgrade sgl-kernel v0.3.15 in this pr

[lifuhuang]

@lifuhuang can u upgrade sgl-kernel v0.3.15 in this pr

done