Repository navigation
[Fix] glm5.2: Align HiSparse warmup backends and move test to nightly - #44303
Merged
ispobock merged 2 commits intoOct 14, 2026
Merged
Conversation
alphabetc1
requested review from
Fridge003,
HaiShaw,
Kangyan-Zhou,
bingxche,
ispobock and
merrymercy
as code owners
October 14, 2026 14:41
Collaborator
Author
|
/rerun-test registered/e2e/models/test_dsa_glm52_hisparse.py |
Contributor
|
🚀 |
Collaborator
Author
|
/rerun-test registered/e2e/models/test_dsa_glm52_hisparse.py |
mmangkad
approved these changes
Oct 14, 2026
Contributor
|
🚀 |
This was referenced Oct 14, 2026
nvpohanh
added a commit
to YAMY1234/sglang
that referenced
this pull request
Oct 15, 2026
Picks up b677851 (sgl-project#44303), required by the CI maintenance MIN_BASE_SHA gate (sgl-project#21065): https://github.com/sgl-project/sglang/actions/runs/38550294693/job/116214243692 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
nvpohanh
added a commit
to Kh4L/sglang
that referenced
this pull request
Oct 15, 2026
Picks up the sgl-project#21065 MIN_BASE_SHA b677851 (sgl-project#44303) and recent PD-disaggregation changes on main; the relevant CI failure https://github.com/sgl-project/sglang/actions/runs/38547120923/job/115994921014 is unrelated to this PR. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
The GLM-5.2 PD + HiSparse + MTP test fails during prefill-worker startup after #44256. FlashInfer autotune runs a dummy DECODE forward on the prefill target worker, but only its prefill backend is set to
flashmla_kv. Its defaulttrtllmdecode backend rejects the packed KV cache:Expected head dim 576 for query and kv_cache, got 576 and 656.Failure: https://github.com/sgl-project/sglang/actions/runs/38575536780/job/115891733288
Modifications
flashmla_kvas well, so startup autotune consumes the same KV layout as prefill and PD transfer.extra-btonightly, using the existing 8-GPU B200 suite and retaining TP4 prefill + TP4/DP4 decode, 500-question GSM8K evaluation, speculative acceptance threshold, and both long-context cases.Accuracy Tests
git diff --checkpassed./rerun-teston this PR; GPU results are pending.Speed Tests and Profiling
Not run; this changes test configuration and scheduling.
Checklist
CI States
Latest PR Test (Base): ❌ Run #38584612690
Latest PR Test (Extra): ❌ Run #38584612406
Latest PR Test (AMD ROCm 10): ❌ Run #38584612502