Repository navigation
[Klaud Cold] minimaxm31flash-fp4-h100-vllm-agentic-dspark: day-zero MiniMax-M3.1-Flash H100 DSpark recipe - #4422
Conversation
…ash AgentX on H100/H200 Hopper TP8 recipe from recipes.vllm.ai: FP8 KV and indexer caches, NVFP4 experts on Marlin MoE, block size 128, FULL_AND_PIECEWISE graphs. No DSpark on Hopper upstream. Context kept at 1M for the uncapped AgentX corpus. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Thanks for the contribution!
中文感谢你的贡献!
|
…n DSpark Use the GB200 recipe's DSpark config (FLASH_ATTN drafter, fp8 draft KV, block rejection) and scale graph capture to four tokens per sequence. Throughput runs need the minimaxm31flash_dspark golden curve. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=39903571329 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=39903571329 |
| - "Add day-zero MiniMax-M3.1-Flash (MiniMaxAI/MiniMax-M3.1-Flash, NVFP4 experts + MXFP8 dense) AgentX on H100 and H200 with vLLM image minimax-m31-flash (https://vllm.ai/blog/2026-10-23-minimax-m31-flash)." | ||
| - "Serve flags follow the recipes.vllm.ai Hopper TP8 recipe: block size 128, FP8 KV and indexer cache, Marlin MoE backend for the NVFP4 experts, FULL_AND_PIECEWISE graphs, hybrid KV cache manager disabled, minimax_m3 reasoning and tool parsers. Context stays at 1M (upstream Hopper example uses 256K) because the AgentX corpus for this prefix is uncapped." | ||
| - "Search space: TP8 only, H100 conc 1-12 and H200 conc 1-24, no KV offloading. Upstream does not enable DSpark on Hopper." | ||
| pr-link: TBD |
There was a problem hiding this comment.
🔴 The changelog validator will reject this PR's new entry. pr-link: TBD is neither the PR URL nor an accepted placeholder. Neither config would get a valid changelog record. validate_added_pr_link accepts only XXX, https://github.com/SemiAnalysisAI/InferenceX/pull/XXX, or the canonical PR URL. Merge-time canonicalization also rewrites only those placeholders. Fix: set pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX (or the real PR number), as CONTRIBUTING.md shows.
Why this was flagged
The new entry appended at inferencex-e2e/perf-changelog.yaml:10882 has pr-link: TBD. On a PR run, validate_added_pr_link in inferencex-e2e/infx/workflows/validate_perf_changelog.py:139-150 raises ChangelogValidationError "new PR entry must use '.../pull/' or an XXX placeholder; found 'TBD'", because PR_LINK_PLACEHOLDERS (lines 22-25) holds only "XXX" and the pull/XXX URL. On a main-branch run it fails CANONICAL_PR_LINK.fullmatch with "new main-branch entry has invalid pr-link". The changelog check therefore fails for this PR. It is the only TBD pr-link in the file, and every other placeholder entry uses pull/XXX.
Verification: The appended entry at inferencex-e2e/perf-changelog.yaml:10882 still says pr-link: TBD. validate_added_pr_link (inferencex-e2e/infx/workflows/validate_perf_changelog.py:139-150) raises on it. Merge-time canonicalization fails, so the entry cannot be rewritten to the canonical link. The PR-run CI check does not raise this.
| # Serve flags follow the Hopper TP8 recipe at | ||
| # https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3.1-Flash. Hopper has no FP4 | ||
| # tensor cores, so the KV and indexer caches are FP8 and the NVFP4 experts | ||
| # run on Marlin. Upstream does not enable DSpark on Hopper. |
There was a problem hiding this comment.
🟡 nit (optional): The new recipes carry history and derivation comments, which AGENTS.md:64-66 forbids. sweep:follow the Hopper TP8 recipe|as on the MiniMax-M3 H100 arm|like the Blackwell arms|Upstream does not enable DSpark Example: h100-fp4/agentic.yaml:25-28 says where the flags came from. Fix: keep only the technical reason for each line, e.g. "Hopper has no FP4 tensor cores, so the NVFP4 experts run on Marlin". This is separate from the master-config comment finding, which does not touch the recipe files. [also at: inferencex-e2e/configs/nvidia-master.yaml:5278 - Maintainers get header and provenance comments that AGENTS.md's mandatory "Code comments" rules forbid (no header comments, no history).]
Why this was flagged
AGENTS.md:65 says comments must not say where code was derived from. AGENTS.md:64 bans comments above a recipe or section. The new recipes add "Serve flags follow the Hopper TP8 recipe at https://recipes.vllm.ai/..." at h100-fp4/agentic.yaml:25-28 and h200-fp4/agentic.yaml:25-28. They add "as on the MiniMax-M3 H100 arm" at h100-fp4/agentic.yaml:51 and "like the Blackwell arms" at line 45 of both recipes. The already-filed comment finding is anchored in nvidia-master.yaml, so fixing it leaves these recipe comments in place.
Verification: AGENTS.md "Code comments (mandatory)" applies "to every file, including recipes, configs". It says "No history. Do not write comments that describe ... where it was derived from". h100-fp4/agentic.yaml lines 25-28 and h200-fp4/agentic.yaml lines 25-28 contain "# Serve flags follow the Hopper TP8 recipe", which is a derivation comment. Nothing breaks at runtime.
| # 2026-10-23 (amd64 is minimax-m31-flash-x86_64-cu134). Serve flags follow the | ||
| # Hopper TP8 recipe: FP8 KV and indexer caches, NVFP4 experts on Marlin MoE. | ||
| # Upstream does not enable DSpark on Hopper. | ||
| minimaxm31flash-fp4-h100-vllm-agentic: |
There was a problem hiding this comment.
🟡 (optional) Operators who run full-sweep --model-prefix minimaxm3, such as the EVALS.md MiniMax-M3 smoke runbook, will now also get MiniMax-M3.1-Flash H100/H200 jobs. The filter at generate.py:1169 uses key.startswith(p), and minimaxm31flash-… keys start with minimaxm3. Fix: match only the whole prefix (key == p or key.startswith(f"{p}-"), the rule ci_priority.py:43 already uses), so minimaxm3 stops selecting minimaxm31flash. The same change stops dsv4 from selecting dsv41flash, which already happens on the base branch.
Why this was flagged
expand_full_sweep at inferencex-e2e/infx/matrix/generate.py:1169 keeps a master config when any(key.startswith(p) for p in options.model_prefix). This PR adds the keys minimaxm31flash-fp4-h100-vllm-agentic (nvidia-master.yaml:5279) and minimaxm31flash-fp4-h200-vllm-agentic (:5301). They are the first keys whose name starts with minimaxm3 without being MiniMax-M3. inferencex-e2e/infx/evals/EVALS.md:100-106 documents full-sweep ... --model-prefix minimaxm3 as the command that builds the MiniMax-M3 deployment-smoke matrix. On the base branch that command matches only minimaxm3 configs. After merge it also emits the 7 H100 and 8 H200 Flash points. Those take whole 8-GPU h100-dsxe and h200-dgxc nodes and mix a different model into the MiniMax-M3 smoke results. Nothing guards against this, because the --model-prefix help text at generate.py:1600 promises a model-prefix filter but the code compares config-key prefixes.
Verification: Trigger: an operator runs full-sweep --model-prefix minimaxm3, the documented smoke command in inferencex-e2e/infx/evals/EVALS.md:101-106. inferencex-e2e/infx/matrix/generate.py:1169 keeps a config when any(key.startswith(p) for p in options.model_prefix), with no delimiter. At the base commit every matching key is minimaxm3-.... At HEAD the two new minimaxm31flash-... keys (nvidia-master.yaml:5279, :5301) also match.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…to match the golden curve Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
/use 39903571329 |
|
@Oseltamivir staged run 39903571329. This run remains available across future |
Summary
Day-zero MiniMax-M3.1-Flash DSpark (
MiniMaxAI/MiniMax-M3.1-Flash) AgentX on H100 with vLLM, from https://vllm.ai/blog/2026-10-23-minimax-m31-flash. This follows the Hopper section of https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3.1-Flash.vllm/vllm-openai:minimax-m31-flash, pinned to the multiarch digestsha256:86f582da…(amd64 =minimax-m31-flash-x86_64-cu134). This is the same pin as the B200/B300 PR.--block-size 128,--kv-cache-dtype fp8,indexer_kv_dtype: fp8,--moe-backend marlin(NVFP4 experts; Hopper has no FP4 tensor cores),FULL_AND_PIECEWISEgraphs,--disable-hybrid-kv-cache-manager, and theminimax_m3reasoning and tool parsers.max-model-lenis 1M rather than the 256K in the upstream Hopper example. The AgentX corpus for this prefix is uncapped, and the Blackwell arms also run at 1M.FLASH_ATTNdrafter, fp8 draft KV, block rejection). Upstream doesn't enable DSpark on Hopper, so whether the drafter runs on Hopper is still unverified. Throughput runs use synthetic acceptance fromminimaxm31flash_dspark.yaml, which isn't committed yet (an AL curve run is in progress). The sweep is on hold until it lands.h100-dsxe), no KV offload. Scaled from the MiniMax-M3 H100 HBM cliff (c5–c6), using the smaller 250 GB checkpoint.Siblings: #4423 (H200), #4419 (GB200/GB300), #4420 (MI355X), #4421 (B200/B300).
🤖 Generated with Claude Code