Skip to content

[npu]add chunk gdn kernel and unify ssm state layout for ascend gdn backend - #36187

Merged
sglang-npu-bot merged 9 commits into
sgl-project:mainfrom
zhaozx-cn:add_chunk_gdn
Sep 21, 2026
Merged

sglang-npu-bot merged 9 commits into
sgl-project:mainfrom
zhaozx-cn:add_chunk_gdn

Conversation

@zhaozx-cn

@zhaozx-cn zhaozx-cn commented Aug 24, 2026 •

Copy link
Copy Markdown
Contributor

Motivation

chunk gated delta rule is now ready. with this operator adopted, the SSM states in both the prefill and MTP branches become (dv, dk), which allows us to unify the SSM state layout. this unification eliminates unnecessary transpose operations, further improving prefill performance.

Modifications

  • replace triton fla kernel with chunk gated delta rule.

  • unify ssm states layout for prefill/decode/mtp.

Accuracy Tests

qwen 3.6 35b ceval
official: 90.0
this pr: 90.1

Speed Tests and Profiling

qwen 3.6 35b
4k in 1k out NEXTN 3

Metric Before This PR Improvement
TTFT (ms) 8459 7727 -8.65%
TPOT (ms) 19.76 18.3 -7.39%

Checklist

Review and Merge Process

  1. Ping Merge Oncalls to start the process. See the PR Merge Process.
  2. Get approvals from CODEOWNERS and other reviewers.
  3. Trigger CI tests with comments or contact authorized users to do so.
    • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
  4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

CI States

Latest PR Test (Base): ❌ Run #35504083021
Latest PR Test (Extra): ❌ Run #35504082792
Latest PR Test (AMD ROCm 10): ⏳ Run #35504082956

Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
Comment thread python/sglang/srt/mem_cache/memory_pool.py
Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
@sglang-npu-bot

Copy link
Copy Markdown
Collaborator

/tag-and-rerun-ci

@github-actions github-actions Bot added the run-ci CI: run the baseline test suite on this PR label Sep 17, 2026
@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

3 similar comments
@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

2 similar comments
@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

zhaozx-cn and others added 2 commits September 20, 2026 15:57
Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
@github-actions github-actions Bot added the hicache Hierarchical Caching for SGLang label Sep 20, 2026
@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

3 similar comments
@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@zhaozx-cn

Copy link
Copy Markdown
Contributor Author

/rerun-failed-ci

@sglang-npu-bot
sglang-npu-bot merged commit 176dbcb into sgl-project:main Sep 21, 2026
497 of 564 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

hicache Hierarchical Caching for SGLang memory-pool npu run-ci CI: run the baseline test suite on this PR

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants