gpt-oss decode performance optimization#20392
Merged
HaiShaw merged 21 commits intosgl-project:mainfrom Mar 19, 2026
Merged
Conversation
added 9 commits
March 3, 2026 05:17
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
kkHuang-amd
requested review from
AniZpZ,
BBuf,
Edwardf0t1,
FlamingoPg,
Fridge003,
HaiShaw,
b8zhong,
ch-wan,
ispobock and
merrymercy
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! |
kkHuang-amd
changed the title
kkHuang-amd
changed the title
HaiShaw
reviewed
Mar 18, 2026
Collaborator
There was a problem hiding this comment.
There are large amount of duplicated code snippets, suggest to refactor soon in follow-up PR.
@kkHuang-amd
| return concat_mla_absorb_q(q_nope, q_rope) | ||
| else: | ||
| return torch.cat([q_nope, q_rope], dim=-1) | ||
|
|
Collaborator
There was a problem hiding this comment.
Please add kernel description and comment to new functions introduced.
HaiShaw
approved these changes
Mar 18, 2026
This was referenced Mar 18, 2026
michaelzhang-ai
added a commit
that referenced
this pull request
Mar 18, 2026
Two fixes for aiter backend failures surfaced by PR #20392: 1. aiter_backend.py: Cap max_num_partitions by min(max_context_len, max_total_num_tokens). The workspace buffer was sized for the model's theoretical max context (e.g. 131K = 512 partitions = 16 GiB) when the KV cache only held 25K tokens (100 partitions = 3 GiB), causing OOM on memory-constrained CI GPUs. 2. unquant.py: Add aiter tgemm.mm fast path for unquantized linear ops, guarded by type(layer.weight.data) is torch.Tensor. Torchao-quantized weights (AffineQuantizedTensor) fail the strict type() check and fall through to F.linear, preventing NotImplementedError on gemm_a16w16.
1 task
Wangzheee
pushed a commit
to Wangzheee/sglang
that referenced
this pull request
Mar 21, 2026
Co-authored-by: wunhuang <wunhuang@amd.com>
0-693
pushed a commit
to 0-693/sglang
that referenced
this pull request
Mar 25, 2026
Co-authored-by: wunhuang <wunhuang@amd.com>
dutsc
pushed a commit
to dutsc/sglang
that referenced
this pull request
Mar 30, 2026
Co-authored-by: wunhuang <wunhuang@amd.com>
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.
Motivation
Improve the performance for gpt-oss model run
Modifications
Three parts for this PR
Accuracy Tests
Server command
SGLANG_USE_AITER=1 python3 -m sglang.launch_server --model openai/gpt-oss-120b/ --tp 8 --chunked-prefill-size 131072 --max-running-requests 128 --mem-fraction-static 0.85 --disable-radix-cache --page-size 64Client command
python3 benchmark/gsm8k/bench_sglang.py --num-questions 2000 --parallel 2000 Accuracy: 0.851 Invalid: 0.014 Latency: 47.111 s Output throughput: 9039.830 token/sBenchmarking and Profiling
Checklist
Review Process
/tag-run-ci-label,/rerun-failed-ci,/tag-and-rerun-ci