Use single mma warp group for short q_len in FA to optimize decoding performance#18985
Use single mma warp group for short q_len in FA to optimize decoding performance#18985Fridge003 merged 4 commits intosgl-project:mainfrom
Conversation
Qiaolin-Yu
requested review from
BBuf,
FlamingoPg,
HaiShaw,
ispobock,
merrymercy,
yizhang2077 and
zhyncs
as code owners
Summary of ChangesHello @Qiaolin-Yu, 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 focuses on optimizing Flash Attention for short query lengths by updating the Highlights
🧠 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
Activity
Using Gemini Code AssistThe 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
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 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
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request updates the flash-attention dependency and related build configurations to enable using a single MMA for short query lengths, which shows significant performance improvements in the provided benchmarks. The changes include updating a git submodule, adding new compiler flags for SM90, and modifying file globbing patterns in CMake. My review confirms the changes are logical, and I've provided suggestions to make the globbing patterns more robust to avoid potential issues with incorrect file matching.
|
|
||
| file(GLOB FA3_BF16_GEN_SRCS | ||
| "${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdimall_bf16*_sm90.cu") | ||
| "${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdim[0-9]*_bf16*_sm90.cu") |
There was a problem hiding this comment.
The glob pattern [0-9]* can match an empty string, which might lead to unintended files being included if, for example, a file named flash_fwd_hdim_bf16_sm90.cu exists. To ensure that you only match files with at least one digit in the head dimension, it's safer to use [0-9][0-9]*.
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdim[0-9][0-9]*_bf16*_sm90.cu")
| # FP16 source files - use individual hdim files instead of hdimall to avoid ptxas crash | ||
| file(GLOB FA3_FP16_GEN_SRCS | ||
| "${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdimall_fp16*_sm90.cu") | ||
| "${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdim[0-9]*_fp16*_sm90.cu") |
There was a problem hiding this comment.
For consistency and to prevent accidentally matching files without a head dimension, it's better to use [0-9][0-9]* instead of [0-9]*. This ensures at least one digit is present.
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdim[0-9][0-9]*_fp16*_sm90.cu")
| # FP8 source files | ||
| file(GLOB FA3_FP8_GEN_SRCS | ||
| "${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdimall_e4m3*_sm90.cu") | ||
| "${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdim[0-9]*_e4m3*_sm90.cu") |
There was a problem hiding this comment.
For consistency and to prevent accidentally matching files without a head dimension, it's better to use [0-9][0-9]* instead of [0-9]*. This ensures at least one digit is present.
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdim[0-9][0-9]*_e4m3*_sm90.cu")
|
/tag-and-rerun-ci |
Qiaolin-Yu
changed the title
|
/rerun-stage stage-c-test-8-gpu-h200 |
|
✅ Triggered |
|
/rerun-stage stage-c-test-4-gpu-b200 |
|
✅ Triggered |
…o xverse_moe * 'xverse_moe' of https://github.com/xiaobaicxy/sglang: (275 commits) fix: add missing blank line after docstring in serving_transcription.py (sgl-project#19206) Whisper model support & `/v1/audio/transcriptions` endpoint & benchmark (sgl-project#16983) fix: patch docker image fixes (sgl-project#19100) [PD-Disagg] Unify prefill info data transition flow, all with `PrefillServerInfo` (sgl-project#19195) [CI] Tiny enhance the dp attention load blance benchmark (sgl-project#19194) add new ci user (sgl-project#19133) [CI] fix the teardown output of disaggregation test (sgl-project#19193) [PD-Disagg] Support query dp rank from bootstrap server. (sgl-project#19168) [Kernel Slimming] Migrate AWQ marlin repack kernel to JIT (sgl-project#18949) [Diffusion] Match rotary_embedding module name style (sgl-project#19179) [Refactor] Split rotary_embedding.py into a modular package (sgl-project#19144) [NPU] bump sgl-kernel-npu to 2026.02.01.post2 (sgl-project#19178) Use single mma warp group for short q_len in FA to optimize decoding performance (sgl-project#18985) Reorganize topk logic to clean up code and expose logical experts (sgl-project#16945) [ROCm] Use unreg path for custom all-reduce during CUDA graph capture (sgl-project#19162) [diffusion] feat: detect Flux2 custom VAE path from component_paths (sgl-project#19170) [AMD] ENV flags tuning and cleanup (sgl-project#19176) Fix bench_one_batch_server by moving the print statements (sgl-project#19175) Update rocm7.2 Dockerfile to install amdsmi for QuickReduce Initialization (sgl-project#19091) Revert "Refactor graph input buffers (sgl-project#18991)" (sgl-project#19173) ...
5 participants
sgl-project/sgl-flash-attn#34
accuracy:
[{'eval_name': 'gpqa', 'model_name': 'dummy-medium_temp1.0_20260218_210226', 'metric': 0.7051767676767676}]
Benchmarking
input len 32000, output len 2000
bs1:
bs16:
bs32:
bs64
profiling
bs 64 before
profiling after (also enables pdl with #18756)