Skip to content

[CI] Fix AMD Extra small timeout and test hygiene - #40721

Open
hdt98 wants to merge 14 commits into
sgl-project:mainfrom
hdt98:ci/test-hygiene-fixes
Open

hdt98 wants to merge 14 commits into
sgl-project:mainfrom
hdt98:ci/test-hygiene-fixes

Conversation

@hdt98

@hdt98 hdt98 commented Sep 22, 2026 •

Copy link
Copy Markdown
Contributor

Motivation

Split the CI/test-hygiene fixes out of #39273 so the Flash PR stays focused on GLM-5.3 Flash enablement.

Modifications

  • Raise AMD Extra small CI timeout from 45 to 60 minutes
  • Use UCX automatic device selection on MI35X fabric runners
  • Retarget stale MiniMax subblock sparse mocks
  • Set tp_size in the hash top-k EP override, which main's topology check rejects without it
  • Clarify the MLX shutdown-loop test comment
  • Fix Kimi-K3 KDA decode FlyDSL imports for current AITER, which removed aiter.ops.flydsl.utils and aiter.ops.flydsl.kernels.vector (co-authored with @long10024070)

The MiniMax and hash top-k changes are identical to #41002, so either PR can merge first without conflicts.

Accuracy Tests

Not applicable; no model-output path changes.

Speed Tests and Profiling

Not applicable; no inference-speed path changes.

Checklist

  • Format with pre-commit
  • Add or update focused tests where applicable
  • Follow SGLang code style guidance

CI States

Latest PR Test (Base): ❌ Run #36758661833
Latest PR Test (Extra): ❌ Run #36758661019
Latest PR Test (AMD ROCm 10): ❌ Run #36758661457

hdt98 added a commit to hdt98/sglang that referenced this pull request Sep 25, 2026
Per review: the AMD Extra timeout, NIXL/UCX autoselection, MiniMax mock
retarget, and MLX comment are unrelated to GLM-5.3-Flash and are split into
their own PR.
@hdt98
hdt98 force-pushed the ci/test-hygiene-fixes branch from 0d569a0 to c0ccd62 Compare September 25, 2026 15:24
@kevin-mii

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 25, 2026
hdt98 and others added 3 commits September 26, 2026 16:33
The topology check from sgl-project#40340 rejects moe_ep_size=4 with tp_size=1.
Same change as sgl-project#41002.

Co-authored-by: xinguozhu-2026 <xinguo.zhu@intel.com>
AITER no longer ships aiter.ops.flydsl.utils or aiter.ops.flydsl.kernels.vector,
so the MI35X KDA decode test failed at import. Use the MLIR vector dialect
already imported as mlir_vector and drop the removed availability helper.

Co-authored-by: long10024070 <long10024070@users.noreply.github.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

amd apple-silicon jit-kernel 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