Repository navigation
[Klaud Cold] minimaxm31flash-fp4-gb200-vllm-agentic-dspark: day-zero MiniMax-M3.1-Flash GB200 recipe matching B200 - #4433
Conversation
…Flash GB200 recipe matching B200
|
Thanks for the contribution!
中文感谢你的贡献!
|
…gb200-dspark-b200parity # Conflicts: # inferencex-e2e/perf-changelog.yaml
There was a problem hiding this comment.
Beyond the inline findings, I compared the new GB200 recipe with the merged B200 recipe. They differ only in name, gpu_type and gpus_per_node. The fp8 drafter KV, the health-check settings and the golden-AL comments all come unchanged from B200, and the search space matches B200's entry. I found nothing GB200-specific in them to flag.
Extended reasoning...
The PR adds a GB200 srt-slurm recipe, a master-config entry and a perf-changelog block for MiniMax-M3.1-Flash on vLLM. It touches no security-sensitive code. A diff against the merged B200 sibling recipe shows only the name and GPU-topology fields changed, and the master-config search space matches B200 line for line.
Findings marked 🟡 are optional suggestions and need no follow-up push.
| description: | ||
| - "Add day-0 MiniMax-M3.1-Flash (MiniMaxAI/MiniMax-M3.1-Flash, NVFP4 experts + NVFP4 KV cache) AgentX on GB200, matching the B200 recipe (#4426) with the arm64 build of the same day-0 vLLM image (minimax-m31-flash-arm64-cu134)." | ||
| - "DSpark with the built-in drafter (probabilistic drafting, FLASH_ATTN drafter, fp8 drafter KV); TP4 c1-c32 drafting 7 tokens (golden AL 4.04); DEP4 (TP1 x DP4 + EP4 behind vllm-router) c64-c128 drafting 3 tokens (golden AL 2.88)." | ||
| pr-link: TBD |
There was a problem hiding this comment.
🔴 The changelog validator will fail this PR because the new entry's pr-link: TBD is not an accepted placeholder. The merge-prep step will also reject it. validate_perf_changelog.py:147 accepts only the PR's own URL or a PR_LINK_PLACEHOLDERS value (XXX or .../pull/XXX). Fix: set the link to https://github.com/SemiAnalysisAI/InferenceX/pull/XXX or to this PR's real URL.
Why this was flagged
The diff appends pr-link: TBD at inferencex-e2e/perf-changelog.yaml:10921. For each new PR entry, validate_perf_changelog.py:147-150 raises new PR entry must use '<expected>' or an XXX placeholder; found 'TBD', because PR_LINK_PLACEHOLDERS (validate_perf_changelog.py:22-25) holds only XXX and https://github.com/SemiAnalysisAI/InferenceX/pull/XXX. The merge-prep canonicalize_appended_links in prepare_perf_changelog_merge.py:108-111 also raises appended entry ... has unexpected pr-link 'TBD' and does not rewrite the link to the PR URL. As a result, changelog validation and the reuse-assisted merge both fail for this PR. A placeholder in the accepted form would pass and be rewritten to the real link automatically.
Verification: Trigger: the PR is merged while the appended entry still has pr-link: TBD (inferencex-e2e/perf-changelog.yaml). The PR-time CI check does not catch this. validate_added_pr_link (validate_perf_changelog.py:193) raises at line 147, because PR_LINK_PLACEHOLDERS (lines 22-25) contains only XXX and https://github.com/SemiAnalysisAI/InferenceX/pull/XXX, so the merge aborts.
| # MiniMax-M3.1-Flash AgentX on GB200: the B200 recipe and search space on the | ||
| # GB200 NVL4 tray, with the day-zero arm64 CUDA 13.4 build of the same image | ||
| # (manifest sha256:86f582dae25ea92802353727164ae4eb03b37b18baf9119b805ffaec2a7ef07c). |
There was a problem hiding this comment.
🟡 (optional) Maintainers get a header and provenance comment above the new master-config entry, which AGENTS.md forbids in every file. The three-line block at nvidia-master.yaml:5547-5549 describes the topology and says the entry reuses "the B200 recipe and search space". It also cites the B200 image's manifest digest as its source. Fix: delete the comment block above minimaxm31flash-fp4-gb200-vllm-agentic-dspark:. The B200 lineage and manifest digest are already in the perf-changelog entry and PR description.
Why this was flagged
The diff adds a comment block at inferencex-e2e/configs/nvidia-master.yaml:5547-5549 directly above the new minimaxm31flash-fp4-gb200-vllm-agentic-dspark key. AGENTS.md:60-66 ('Code comments (mandatory)') bans header comments above a recipe or section and bans comments that describe the topology or config. It also bans history comments about where code was derived from, saying git history and perf-changelog.yaml are the record. The block does all three: it describes the GB200 NVL4 tray topology, says the entry is "the B200 recipe and search space", and gives the B200 image manifest digest as provenance. The B200 sibling entry at nvidia-master.yaml:5532 has no such comment, so the new entry breaks both the rule and the nearest existing pattern.
Verification: The diff adds a three-line comment block above the new key at inferencex-e2e/configs/nvidia-master.yaml:5547-5549. AGENTS.md's "Code comments (mandatory)" section applies "to every file, including recipes, configs". The block breaks its "No header comments" and "No history" rules, but nothing at runtime is affected.
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=39942082663 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=39942082663 |
|
/use 39942082663 |
|
@Oseltamivir staged run 39942082663. This run remains available across future |
Summary
Day-zero MiniMax-M3.1-Flash AgentX on GB200 with vLLM and DSpark. The recipe and search space match the merged B200 recipe (#4426). This replaces #4424.
minimaxm31flash-fp4-gb200-vllm-agentic-dspark.minimaxm31flash/vllm/b200-fp4-mtp/agentic.yaml. Onlygpu_type(gb200) andgpus_per_node(4) differ.vllm/vllm-openai:minimax-m31-flash-arm64-cu134@sha256:5d76347…, the arm64 build of B200's image (manifestsha256:86f582da…). It already ran cleanly on GB200 in [Klaud Cold] minimaxm31flash-fp4-gb200-vllm-agentic-dspark: day-zero MiniMax-M3.1-Flash GB200 DSpark recipe #4424.Why this replaces #4424
#4424 used 3 draft tokens at TP4 (AL 2.88), and the B200 recipe uses 7 (AL 4.04). On GB200 the time per decode step was slightly lower than on B200: about 6.6 ms against 7.6 ms (TPOT × AL). But each step produced about 29% fewer tokens, so GB200 trailed B200 by 18–28% on TP4 interactivity. On DEP4, where both used 3 tokens, GB200 beat B200 by about 20% (c64) and about 10% (c128) on tokens/s/GPU.
#4424's TP4 c64 point was badly overloaded (median TTFT 131 s) and DEP4 c64 beat it on every metric, so it is dropped. DEP4 c256 is dropped to match B200.
Validation
infx generatebinds all 7 points to exactly one variant each, with the expected speculative config and golden AL.