Repository navigation
[Klaud Cold] minimaxm31flash-fp4-gb200-sglang-agentic-dspark: day-zero MiniMax-M3.1-Flash GB200 SGLang recipe - #4444
Conversation
…1-Flash SGLang AgentX on GB200 Serve flags follow the SGLang cookbook's MiniMax-M3.1 cells with the day-zero dev-minimax-m31-flash image. Seven-token DSpark with synthetic acceptance from the minimaxm31flash_dspark golden curve.
|
Thanks for the contribution!
中文感谢你的贡献!
|
| - "GB200 has no cookbook cell; it runs the GB300 (4-GPU) layouts, with B200's sm_100 High-Throughput memory settings (mem-fraction-static 0.75, prefill graphs to 16384)." | ||
| - "Seven-token DSpark (the bundled draft's dspark_block_size); throughput runs use synthetic acceptance from the minimaxm31flash_dspark golden curve (AL 4.04 at 7 draft tokens, #4425)." | ||
| - "Search space: Balanced TP4 at conc 1-32; TP4 + EP4 at conc 64-128." | ||
| pr-link: TBD |
There was a problem hiding this comment.
🟡 (optional) The PR's changelog check will reject this entry, blocking CI, so the new config is never collected. pr-link: TBD is not an accepted placeholder. validate_added_pr_link only takes the PR's own URL or an XXX placeholder. Fix: set pr-link to https://github.com/SemiAnalysisAI/InferenceX/pull/XXX, or to the real PR URL.
Why this was flagged
The new tail entry at inferencex-e2e/perf-changelog.yaml:11019 sets pr-link: TBD. On PR runs, compare_entries calls validate_added_pr_link for each appended entry (inferencex-e2e/infx/workflows/validate_perf_changelog.py:192-193). With pr_number set, any link that is neither in PR_LINK_PLACEHOLDERS ({"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"}, lines 22-25) nor the expected PR URL raises ChangelogValidationError "new PR entry must use ... or an XXX placeholder; found 'TBD'" (lines 146-150). On main-branch runs the check needs CANONICAL_PR_LINK, which also rejects 'TBD' (lines 141-143). So the changelog validation fails, and result collection for minimaxm31flash-fp4-gb200-sglang-agentic-dspark cannot go ahead. No other entry in the file uses TBD. The ChangelogEntry schema types pr_link as a plain str, so the schema check alone lets it through.
Verification: nit. The trigger is merging this PR with the supported reuse-merge tool. The PR-side check-changelog job does not check pr-link, so CI is not blocked. merge_pr calls canonicalize_changelog(pr, repo) at merge_with_reuse.py:631, outside any try block. With pr_number set, 'TBD' raises ChangelogValidationError, so the reuse merge aborts with a traceback.
| # MiniMax-M3.1-Flash AgentX on GB200 with SGLang: day-zero image from the SGLang | ||
| # cookbook (https://docs.sglang.io/cookbook/autoregressive/MiniMax/MiniMax-M3.1), | ||
| # manifest digest verified 2026-10-24. Serve flags follow the cookbook's | ||
| # GB300 Balanced (TP4) and High-Throughput (TP4, EP4) cells; GB200 has no cookbook cell, so the High-Throughput arm takes B200's sm_100 memory settings (mem 0.75, prefill graphs to 16384). | ||
| # Seven-token DSpark (bundled draft); synthetic acceptance follows | ||
| # infx/golden_al_distribution/minimaxm31flash_dspark.yaml. |
There was a problem hiding this comment.
🟡 nit (optional): New section-header and provenance comments break AGENTS.md "Code comments" (no header comments above a recipe, override or section; no history or derivation). sweep:cookbook|# (TP4 Balanced|High-Throughput|Balanced:) Examples: the master header at nvidia-master.yaml:5616-5621 cites the cookbook and says "manifest digest verified 2026-10-24", a date after the PR was opened. agentic.yaml:51-52 sits above the TP4 overrides. Fix: delete these blocks and keep only short per-line technical comments. Provenance already lives in perf-changelog.yaml.
Why this was flagged
AGENTS.md lines 64-66 forbid broad comments above a recipe, override or section, and comments describing where values were derived from. The diff adds such comments at several places. One is the 6-line header above minimaxm31flash-fp4-gb200-sglang-agentic-dspark (nvidia-master.yaml:5616-5621), which cites the cookbook and a verification date of 2026-10-24, one day after the PR opened on 2026-10-23. Others are the search-space labels "# Balanced: TP4." and "# High-Throughput: TP4 + EP4 (MegaMoE)." (5634, 5636). The rest are in the recipe: "Serve flags follow the cookbook's GB300..." (agentic.yaml:24-25), and the section headers above the overrides at agentic.yaml:51-52 and 129-130. Maintainers end up with provenance text that the convention keeps in git history and perf-changelog.yaml, and the future-dated provenance claim is unreliable.
Verification: AGENTS.md:60-67 says "No header comments" above a recipe, override, or section and "No history" about where it was derived from. nvidia-master.yaml:5616-5621 adds a 6-line header with provenance and "manifest digest verified 2026-10-24", one day after the PR was opened. agentic.yaml:51-52 and 24-25 add a section header and a provenance comment. Nothing breaks at runtime.
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=39996749965 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=39996749965 |
Keep main's entries; reinsert the GB200 SGLang block and append its changelog entry.
|
/use 39996749965 |
|
@Oseltamivir staged run 39996749965. This run remains available across future |
Resolve the perf-changelog conflict with SemiAnalysisAI#4444 via prepare_perf_changelog_merge. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Summary
Day-zero MiniMax-M3.1-Flash (
MiniMaxAI/MiniMax-M3.1-Flash, NVFP4) AgentX on GB200 with SGLang, following the SGLang cookbook: https://docs.sglang.io/cookbook/autoregressive/MiniMax/MiniMax-M3.1 (config source:docs/src/snippets/configs/MiniMaxAI/minimax-m3.1.jsx).lmsysorg/sglang:dev-minimax-m31-flash, pinned by digest (verified on Docker Hub 2026-10-24).--mem-fraction-static 0.75,--cuda-graph-max-bs-prefill 16384), since B300's settings ran out of memory capturing prefill graphs on B200.--speculative-algorithm DSPARK --speculative-use-block-verification,--cuda-graph-backend-prefill breakable,minimax-m3reasoning and tool parsers,--tokenizer-worker-num 8.--speculative-dspark-block-size 7, the bundled draft'sdspark_block_size). It's set explicitly so throughput runs resolve synthetic acceptance fromminimaxm31flash_dspark.yaml([Klaud Cold] Add MiniMax-M3.1-Flash DSpark golden AL curve and SPEED-Bench collector / [Klaud Cold] 新增 MiniMax-M3.1-Flash DSpark 黄金 AL 曲线与 SPEED-Bench 采集脚本 #4425): AL 4.04 at 7 tokens, thinking on (adaptive, the default).infx generatebinds every matrix point to exactly one recipe variant, withSGLANG_SIMULATE_ACC_LEN=4.04.vLLM siblings: #4419 (GB300), #4420 (MI355X), #4421 (B300), #4422 (H100), #4423 (H200), #4426/#4431 (B200), #4433 (GB200).