Add SM120 (Blackwell desktop) MXFP4 support#16975
Closed
amittell wants to merge 3 commits intosgl-project:mainfrom
Closed
Add SM120 (Blackwell desktop) MXFP4 support#16975amittell wants to merge 3 commits intosgl-project:mainfrom
amittell wants to merge 3 commits intosgl-project:mainfrom
Conversation
SM120 (RTX PRO 6000, RTX 5090) doesn't support persistent kernels the same way SM100 does. This adds SM120-specific configuration following the same pattern as vLLM PR #31089: - Use StridedLayout instead of TMA block layout - Set is_persistent=False and num_stages=1 Tested with GPT-OSS-120B on RTX PRO 6000: - 4K context: 151 tok/s - 131K context: 57 tok/s Fixes sgl-project#13061 Related: sgl-project#9707, sgl-project#12695
amittell
requested review from
AniZpZ,
BBuf,
Edwardf0t1,
FlamingoPg and
ch-wan
as code owners
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
SM120 (Blackwell desktop) doesn't support flashinfer_mxfp4 due to persistent kernel limitations. This adds: 1. server_args.py: Auto-select triton_kernel for SM120 + MXFP4 2. mxfp4.py: Use StridedLayout for SM120 with triton_kernel backend Tested on RTX PRO 6000 with GPT-OSS-120B - server starts and runs without needing to manually specify --moe-runner-backend.
This was referenced Jan 29, 2026
Collaborator
|
@amittell Could you help fix the merge conflicts? Thanks! |
This was referenced Mar 2, 2026
Collaborator
|
Hi, I add you as coauthor in #19718. Thanks~ |
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
2 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
SM120 (RTX PRO 6000, RTX 5090) doesn't support persistent kernels the same way SM100 does. This adds SM120-specific configuration following the same pattern as vLLM PR vllm-project/vllm#31089:
StridedLayoutinstead of TMA block layoutis_persistent=Falseandnum_stages=1Tested with GPT-OSS-120B on RTX PRO 6000:
Fixes #13061, related to #9707, #12695