Delete CUTLASS FP8 blockwise for SM90 and SM100, move SM120 to JIT and add SwapAB - #30438
Merged
Conversation
…mm files Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Dedup getSMVersion/alloc_workspace_tensor into shared utils.cuh, delete dead fp8_blockwise benchmark-only entry points and force_noswap branch, reuse the canonical is_sm120_supported() instead of local reimpls, fix CI runner_config to 1-gpu-small and drop the stale disabled= reason, and trim/remove explanatory comments in the new fp8 blockwise files. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
b8zhong
changed the title
added 2 commits
July 7, 2026 23:07
Collaborator
Author
|
/rerun-test test/registered/jit/test_fp8_blockwise_gemm.py |
Contributor
|
Results for 🚀 |
b8zhong
marked this pull request as ready for review
July 7, 2026 23:14
b8zhong
requested review from
AniZpZ,
BBuf,
DarkSharpness,
Edwardf0t1,
FlamingoPg,
HaiShaw,
HydraQYH,
celve,
ch-wan,
ispobock,
merrymercy,
yizhang2077 and
yuan-luo
as code owners
July 7, 2026 23:14
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
Collaborator
Author
|
The following failure is fixed on main : https://github.com/sgl-project/sglang/actions/runs/29057323200/job/86430764736?pr=30438 |
Collaborator
Author
|
/rerun-failed-ci |
Collaborator
Author
|
The multimodal tests are fixed in #30716 |
Collaborator
Author
|
/rerun-failed-ci |
Contributor
|
Tested PR head In the post-page-mask 120-step C1 trace:
Dense FP8 is the largest measured post-fix decode kernel family. These numbers |
# Conflicts: # test/registered/quant/test_fp8_blockwise_row_padding.py
sgl-kernel/tests/test_per_token_group_quant_8bit.py and the matching benchmark script still imported from sglang.srt.layers.quantization.fp8_kernel, which was moved to sglang.kernels.ops.quantization.fp8_kernel on main, breaking CI collection after the merge.
…8-blockwise # Conflicts: # python/sglang/jit_kernel/csrc/gemm/nvfp4/nvfp4_expert_quant.cuh # python/sglang/jit_kernel/csrc/gemm/nvfp4/nvfp4_quant_kernels.cuh # python/sglang/jit_kernel/csrc/gemm/nvfp4/nvfp4_scaled_mm_common.cuh # python/sglang/jit_kernel/csrc/moe/nvfp4_blockwise_moe.cuh # sgl-kernel/benchmark/bench_fp8_blockwise_gemm.py
Collaborator
|
Nice! |
5 tasks
At1a8
added a commit
to At1a8/sglang
that referenced
this pull request
Jul 14, 2026
…o JIT and add SwapAB (sgl-project#30438)" This reverts commit 7431f35.
At1a8
added a commit
to At1a8/sglang
that referenced
this pull request
Jul 15, 2026
… SM120 to JIT and add SwapAB (sgl-project#30438)"" This reverts commit ac2470d.
This was referenced Jul 15, 2026
Chronostasys
pushed a commit
to MindLab-Research/sglang
that referenced
this pull request
Aug 24, 2026
…d add SwapAB (sgl-project#30438) Co-authored-by: Brayden Zhong <brayden.zhong@radixark.ai> Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> Co-authored-by: root <root@sgl-b300-inference.datacrunch.io> Co-authored-by: Brayden Zhong <brayden@radixark.ai>
jakki-amd
pushed a commit
to jakki-amd/sglang
that referenced
this pull request
Sep 9, 2026
…d add SwapAB (sgl-project#30438) Co-authored-by: Brayden Zhong <brayden.zhong@radixark.ai> Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> Co-authored-by: root <root@sgl-b300-inference.datacrunch.io> Co-authored-by: Brayden Zhong <brayden@radixark.ai>
This was referenced Sep 10, 2026
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
4 participants
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.
Recommended migration plan
Only SM120 needs this for now. Also, there has been no development on this kernel for 1 year+.
For future development (Rubin, etc), we'd use Cute-DSL. Ideally, some kernel experts can also help us migrate this one to Cute-DSL (as long as the perf is the same)
For M = 1 on
Qwen/Qwen3.6-27B-FP8, speed increase from 42 to 46.76 TPS.And here is the difference of using StreamK scheduler on M >= 64 (when below this threshold, we currently use SwapAB to transpose the A and B operand for better MMA utilization)
CI States
Latest PR Test (Base): 🚫 Run #29298468640
Latest PR Test (Extra): ❌ Run #29298468570