Repository navigation
[Perf][Qwen4Exp] Add SM100 low-latency decode GEMM plans - #59214
Merged
gau-nernst merged 4 commits intoOct 1, 2026
Merged
Conversation
The Qwen4Exp BF16 decode GEMM selector only has plans for SM103 and SM90, so on B200 every unquantized projection falls back to cuBLAS, which picks split-K kernels plus a separate reduce at small M. Add a B200 table measured on SM100 and route capability (10, 0) to it. Each point was kept only if the CuTe skinny GEMM beats cuBLAS by at least 5% in both hot-cache and L2-flush CUDA graph replay; the winners differ from the B300 table, so it is not reused. SM103 and SM90 selection is unchanged. Signed-off-by: namgyu-youn <namgyu.dev@gmail.com>
Contributor
Author
|
cc @ZJY0516 @gau-nernst working on this module. Could you look into this when you have a change? |
namgyu-yun
marked this pull request as ready for review
September 29, 2026 12:39
namgyu-yun
requested review from
AndreasKaratzas,
WoosukKwon,
mgoin,
tlrmchlsmth,
yewentao256 and
zyongye
as code owners
September 29, 2026 12:39
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
gau-nernst
approved these changes
Sep 29, 2026
gau-nernst
left a comment
gau-nernst
left a comment
Contributor
There was a problem hiding this comment.
Some nits. Dont think some tests are needed. Also measurements should be done with flashinfer cupti cudagraph (check kernel-microbenchmark skill in vllm). U can trim the comment too (agents like verbose comment)
Signed-off-by: namgyu-youn <namgyu.dev@gmail.com>
Contributor
Author
@gau-nernst updated (1) trimmed comment, (2) drop unneeded test, (3) measurment based on your suggestion. Could you take another look? |
Contributor
|
I already approved. Will need @ZJY0516 for the stamp. |
gau-nernst
approved these changes
Oct 1, 2026
Contributor
|
/ci run |
|
✅ @namgyu-youn, CI is now available for this PR.
|
|
✅ Triggered Buildkite CI #92314 for commit |
gau-nernst
enabled auto-merge (squash)
October 1, 2026 02:34
6 tasks done
michael-gpu
pushed a commit
to practical-inference/vllm-ampere-extended
that referenced
this pull request
Oct 20, 2026
…t#59214) Signed-off-by: namgyu-youn <namgyu.dev@gmail.com> Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
6 tasks done
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Purpose
Qwen4Exp's low-latency decode GEMM table has entries for SM103 and SM90 only, so on B200 (SM100) every BF16 projection falls back to cuBLAS. This adds a B200-measured table for the TP=4 shapes and routes capability
(10, 0)to it, following #53534.Test Plan
pytest tests/kernels/test_bf16_skinny_gemm.py -k qwen4_exp # 36 passed on B200Test Result
-= keeps cuBLAS.