Repository navigation
[Klaud Cold] minimaxm31flash-fp4-mi355x-sglang-agentic-dspark: disable sparse index top-k sharing (SGLANG_MINIMAX_M3_INDEX_TOPK_FREQ=1) - #4469
Conversation
Oseltamivir
added
the
full-sweep-fail-fast
|
Thanks for the contribution!
中文感谢你的贡献！
|
Oseltamivir
added
non-canary-full-sweep-enabled
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| - minimaxm31flash-fp4-mi355x-sglang-agentic-dspark | ||
| description: | ||
| - "Set SGLANG_MINIMAX_M3_INDEX_TOPK_FREQ=1 on MI355X SGLang MiniMax-M3.1-Flash AgentX: every sparse layer computes its own index top-k instead of SGLang's ROCm default of sharing one across each pair of layers (sgl-project/sglang#36527)." | ||
| pr-link: TBD |
There was a problem hiding this comment.
🔴 The new changelog entry will fail the perf-changelog validator on this PR's CI run, which blocks the benchmark from being scheduled and merged. pr-link: TBD is neither the canonical PR URL nor an accepted placeholder. validate_perf_changelog.py:22-25 only accepts "XXX" or "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX". Fix: set pr-link: XXX, or use the real https://github.com/SemiAnalysisAI/InferenceX/pull/<n>. The merge tooling then rewrites the XXX placeholder to the canonical link.
Why this was flagged
The PR appends a block ending in pr-link: TBD at perf-changelog.yaml:11108. On a PR run, validate_added_pr_link in infx/workflows/validate_perf_changelog.py:139-150 checks every added entry. It raises ChangelogValidationError("new PR entry must use '...pull/' or an XXX placeholder; found 'TBD'") unless the link is in PR_LINK_PLACEHOLDERS {"XXX", ".../pull/XXX"} or equals the expected PR URL. On a main-branch run the CANONICAL_PR_LINK regex rejects it as well. The placeholder canonicalizer in prepare_perf_changelog_merge.py also only handles the XXX placeholders (its tests use "XXX"), so nothing rewrites TBD on merge. The base branch has no such entry. CONTRIBUTING.md:206 documents the .../pull/XXX form, and none of the other 11,000+ lines of entries uses TBD.
Verification: Trigger: merging through the supported reuse path (merge_with_reuse) while the entry still has pr-link: TBD. TBD passes PR CI and the sweep is still scheduled. compare_entries(base, head, pr_number) (prepare_perf_changelog_merge.py:88) calls validate_added_pr_link (validate_perf_changelog.py:193). That raises because 'TBD' is not in PR_LINK_PLACEHOLDERS (lines 22-25).
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40110172586 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40110172586 |
|
/use 40110172586 |
|
@Oseltamivir staged run 40110172586. This run remains available across future |
Follow-up to #4445. Sets
SGLANG_MINIMAX_M3_INDEX_TOPK_FREQ=1on the MI355X SGLang MiniMax-M3.1-Flash AgentX recipe.SGLang defaults this to 2 on ROCm (sgl-project/sglang#36527, environ.py#L2033-L2039): each pair of sparse-attention layers shares one index top-k selection. NVIDIA always runs with 1. Setting 1 here makes every layer compute its own selection, matching NVIDIA and removing the stale-selection approximation, so the throughput and eval comparison against #4445 shows what sharing buys.
Only the env var changes; search space, image and serve flags are unchanged.