[AMD] Fp8 prefill integration with radix cache path for dpsk models by 1am9trash · Pull Request #20187 · sgl-project/sglang

1am9trash · 2026-03-09T11:17:45Z

Motivation

Previously, fp8 prefill attention on dpsk models did not cover the radix-cache path.

Modifications

This PR enables fp8 prefill attention for the radix-cache path.
To reduce extra element-wise casts, it also uses fused_gemm_afp4wfp4_split_cat, following the same design principle as the existing fp8 prefill path.

(Not cover fp8 fused gemm yet.)

Accuracy Tests

Tested with radix-cache on/off and fp8/bf16 prefill.

Benchmarking and Profiling

Compared fp8 vs. bf16 prefill on dspk-r1-mxfp4 with 70k/200 and radix-cache on.

server cmd:

SGLANG_AITER_FP8_PREFILL_ATTN=1 python3 -m sglang.launch_server --model-path DeepSeek-R1-MXFP4-Preview/ --tensor-parallel-size 8 --trust-remote-code --chunked-prefill-size 131072 --host 0.0.0.0 --port 9000 --log-requests --mem-fraction-static 0.8 --max-running-requests 64 --kv-cache-dtype fp8_e4m3 --attention-backend aiter --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

client cmd:

python3 -m sglang.bench_serving --host localhost --port 9000 --model DeepSeek-R1-MXFP4-Preview/ --dataset-name random --random-input 70000 --random-output 200 --random-range-ratio 1.0 --num-prompt [cc * 8] --max-concurrency [cc]

Speed on MI355 (cc=1-16):

total tok/s: 1.17×-1.26×
ttft: 1.07×-1.29×

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

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

…p4_split_cat

…ix_cache_path

gemini-code-assist · 2026-03-09T11:17:58Z

Summary of Changes

Hello, 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 extends the functionality of FP8 prefill attention to include the radix-cache path for dpsk models, addressing a previous limitation. By integrating this capability and optimizing the K/V generation process with a fused GEMM operation, the changes aim to enhance the overall efficiency and performance of the attention mechanism, leading to faster inference speeds.

Highlights

FP8 Prefill Attention for Radix Cache: Enabled FP8 prefill attention for the radix-cache path, which was previously unsupported in dpsk models.
Optimized K/V Generation: Implemented fused_gemm_afp4wfp4_split_cat to reduce extra element-wise casts during K/V generation, aligning with existing FP8 prefill optimizations.
Performance Improvements: Demonstrated significant speedups on MI355, with total tokens/second increasing by 1.17x-1.26x and time-to-first-token by 1.07x-1.29x compared to BF16 prefill.

🧠 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.

Changelog

python/sglang/srt/layers/attention/aiter_backend.py
- Added a new method mla_fp8_prefill_attn to encapsulate the logic for FP8 prefill attention, including type casting and kernel calls.
- Updated the init_forward_metadata method to correctly retrieve and utilize kv_indptr for metadata generation.
- Modified the call to make_mla_prefill_ps_meta_data to use kv_indptr instead of qo_indptr for accurate metadata setup.
- Adjusted the calculation of total_s to directly use forward_batch.seq_lens_sum for efficiency.
- Refactored the forward_extend method to delegate FP8 prefill attention computation to the newly introduced mla_fp8_prefill_attn method.
- Implemented conditional logic in forward_extend to use fused_gemm_afp4wfp4_split_cat for optimized K/V generation when FP8 prefill is active and weights are in torch.uint8 format.

Activity

No specific activity (comments, reviews, or progress updates) has been recorded for this pull request yet.

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.

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 integrates Fp8 prefill with the radix cache path for dpsk models, which is a valuable performance enhancement. The changes are well-structured, primarily involving the addition of a new mla_fp8_prefill_attn function by refactoring existing code and extending it to handle the radix cache path. A new optimization for MXFP4 weights using a fused kernel is also introduced, which is a nice touch. The logic appears sound. I have one minor suggestion to improve code conciseness.

HaiShaw · 2026-03-10T06:17:41Z

/tag-and-rerun-ci

HaiShaw · 2026-03-10T09:48:32Z

@1am9trash Let's extend fp8 prefill attention to (in follow-up PRs):

other MHA models:
gfx942

…gl-project#20187)

1am9trash and others added 6 commits March 9, 2026 15:45

Use kv_indptr for radix cache case

1999481

Use correct total seq lens including prefix

e1090ae

Add mla_fp8_prefill_attn() for both no prefix and radix-cache case

835861d

Use mla_fp8_prefill_attn() in no prefix case

750193c

Use mla_fp8_prefill_attn() in radix-cache case with fused_gemm_afp4wf…

a606f35

…p4_split_cat

Merge branch 'sgl-project:main' into fp8_prefill_integration_with_rad…

28d68ac

…ix_cache_path

1am9trash requested review from Fridge003, HaiShaw, Qiaolin-Yu, hebiao064, ispobock and merrymercy as code owners March 9, 2026 11:17

gemini-code-assist Bot reviewed Mar 9, 2026

View reviewed changes

Comment thread python/sglang/srt/layers/attention/aiter_backend.py

1am9trash changed the title ~~[AMD] Fp8 prefill integration with radix cache path for dpsk models~~ [DO NOT MERGE][AMD] Fp8 prefill integration with radix cache path for dpsk models Mar 9, 2026

1am9trash changed the title ~~[DO NOT MERGE][AMD] Fp8 prefill integration with radix cache path for dpsk models~~ [AMD] Fp8 prefill integration with radix cache path for dpsk models Mar 9, 2026

1am9trash and others added 2 commits March 10, 2026 11:20

Merge branch 'main' into fp8_prefill_integration_with_radix_cache_path

ab6e9ed

Merge branch 'main' into fp8_prefill_integration_with_radix_cache_path

e73a305

github-actions Bot added the run-ci label Mar 10, 2026

HaiShaw approved these changes Mar 10, 2026

View reviewed changes

HaiShaw merged commit 6407891 into sgl-project:main Mar 10, 2026
143 of 164 checks passed

liubiyongge pushed a commit to liubiyongge/sglang that referenced this pull request Mar 13, 2026

[AMD] Fp8 prefill integration with radix cache path for dpsk models (s…

25c09d2

…gl-project#20187)

Wangzheee pushed a commit to Wangzheee/sglang that referenced this pull request Mar 21, 2026

[AMD] Fp8 prefill integration with radix cache path for dpsk models (s…

2d5c4bb

…gl-project#20187)

JustinTong0323 pushed a commit to JustinTong0323/sglang that referenced this pull request Apr 7, 2026

[AMD] Fp8 prefill integration with radix cache path for dpsk models (s…

0082ab3

…gl-project#20187)

yhyang201 pushed a commit to yhyang201/sglang that referenced this pull request Apr 22, 2026

[AMD] Fp8 prefill integration with radix cache path for dpsk models (s…

b6a4e04

…gl-project#20187)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[AMD] Fp8 prefill integration with radix cache path for dpsk models#20187

[AMD] Fp8 prefill integration with radix cache path for dpsk models#20187
HaiShaw merged 8 commits intosgl-project:mainfrom
1am9trash:fp8_prefill_integration_with_radix_cache_path

1am9trash commented Mar 9, 2026 •

edited

Loading

Uh oh!

gemini-code-assist Bot commented Mar 9, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

HaiShaw commented Mar 10, 2026

Uh oh!

HaiShaw commented Mar 10, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

1am9trash commented Mar 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

Review Process

Uh oh!

gemini-code-assist Bot commented Mar 9, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

HaiShaw commented Mar 10, 2026

Uh oh!

HaiShaw commented Mar 10, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

1am9trash commented Mar 9, 2026 •

edited

Loading