feat(minimaxm3-fp4-b300-dynamo-vllm-mtp): bump vLLM nightly to 5e35a6f4, enable cutlass MSA decode / 更新 minimaxm3-fp4-b300-dynamo-vllm-mtp vLLM 镜像至 5e35a6f4，启用 cutlass MSA 解码后端

[xinli-sw]

Summary

• Bump minimaxm3-fp4-b300-dynamo-vllm-mtp vLLM image to nightly-5e35a6f4f9bbc217c599692157ca985c894373f7 (contains fixed-len MiniMax-M3 fix)
• Update 5 B300 FP4 8k1k MTP recipe YAML files: update model.container and add VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass to prefill and decode environments
• 2p1d-dep2-dep4-eagle3-8k1k.yaml intentionally NOT updated (used by minimaxm3-fp4-b300-dynamo-vllm-mtp-legacy-dep4 which pins a different image)

Test plan

• Full sweep with full-sweep-fail-fast label runs green

中文说明

• 将 minimaxm3-fp4-b300-dynamo-vllm-mtp vLLM 镜像升级至 nightly-5e35a6f4f9bbc217c599692157ca985c894373f7
• 更新 5 个 B300 FP4 8k1k MTP recipe YAML 文件：同步 model.container，并在预填充与解码环境中添加 VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass
• 2p1d-dep2-dep4-eagle3-8k1k.yaml 有意保留不变（由 minimaxm3-fp4-b300-dynamo-vllm-mtp-legacy-dep4 使用，该配置固定使用不同镜像）

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[claude]

LGTM — straightforward vLLM nightly bump and MSA decode backend flag for the minimax-m3 B300 FP4 MTP recipes.

What was reviewed: verified the container tag bump (4080263b → 5e35a6f4) and new VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass env var are applied consistently across all 5 recipe YAMLs, master config, and changelog; confirmed the intentionally-excluded 2p1d-dep2-dep4 recipe (pinned to the legacy-dep4 image) was correctly left untouched.

Extended reasoning...

Overview

This PR bumps the vLLM nightly image tag for the minimaxm3-fp4-b300-dynamo-vllm-mtp config from nightly-4080263b to nightly-5e35a6f4 and adds VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass to the prefill/decode environments across 5 B300 FP4 8k1k MTP recipe YAMLs, plus the corresponding configs/nvidia-master.yaml entry and a perf-changelog.yaml record. This follows the repo's standard nightly-image-bump pattern used for other model configs.

Security risks

None. These are benchmark recipe configs (container tags and env vars) with no code, auth, or credential changes.

Level of scrutiny

Low — this is a mechanical, repeated edit (image tag + one env var) applied identically across a known set of files, consistent with prior nightly-bump PRs in this repo. The PR correctly excludes the 2p1d-dep2-dep4 recipe, which pins a different (legacy) image, matching the stated intent in the description.

Other factors

No bugs were found by the bug hunting system. I cross-checked that all 5 modified recipe files, the master config, and the changelog are internally consistent, and that the excluded legacy recipe was correctly left alone. The PR is gated on a full benchmark sweep (full-sweep-fail-fast label) before merge, which will catch any functional regressions from the image bump itself.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30821722120
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30821722120

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30824007160
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30824007160

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30828244671
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30828244671

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30838377466
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30838377466

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30838377466
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30838377466

[xinli-sw]

/reuse-sweep-run

[kedarpotdar-nv]

As a PR reviewer and CODEOWNER, I have reviewed this and have:

• Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
• Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
• Verified that this PR has passed PR validation. Please link to GitHub Action workflow that shows this.
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this.
• Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
• For agentic workloads: verified that speculative-decoding configs (EAGLE / MTP / draft models) run with simulated synthetic acceptance, with the acceptance-length value taken from the committed golden AL curve in golden_al_distribution/ for that model, thinking mode, and draft length. A submission may choose any supported draft length, but it may not substitute a different acceptance target.
• Verified against the current MODELS.md that this PR does not submit a deprecated model, scenario, or model-scenario combination.
• Verified that the model architecture isn't changed with benchmark hacks like using --hf-overrides to skipping indexer for every x layers on models that don't natively support this. As a general rule, we won't accept optimizations that reduces the number of model architecture FLOPs. Anything that makes that same computation run faster is fair game; FLOPs at lower precisions is fine, given that the config passes private evals. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
• If an company claims that they support vLLM/SGLang as first class LLM inference engines on their hardware, I have verified that the respective vLLM submission made using upstream https://hub.docker.com/u/vllm docker repo, upstream SGLang https://hub.docker.com/u/lmsysorg docker repo. The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet as supported by vLLM/SGLang community maintainers
• If an company claims that they support vLLM/SGLang as first class upstream in-tree LLM inference engines on their hardware, I have have verified that the respective vLLM/SGLang submission has been made before additional frameworks (TRT-LLM, ATOM, etc.). The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet.
• Verified that every single-node vLLM/SGLang recipe in this PR is documented in the official vLLM recipes and/or the SGLang cookbook:
  • I linked the corresponding upstream PR in the vLLM recipe repo or SGLang repo and verified that it is MERGED before this InferenceX PR merges. An opened, draft, or closed-without-merge upstream PR does not satisfy this requirement. If the matching recipe was already published, I linked the published recipe/cookbook page in the additional detail section below.
• Verified that this PR does not patch the inference engine or serving stack — the pinned image must run as shipped. This covers .patch files / git apply / patch, inline patches embedded in benchmark scripts (e.g. a python3/sed heredoc that rewrites installed engine sources before serving), in-place edits of site-packages, monkey-patching, overwriting container files, and installing forked/rebuilt engine wheels on top of the pinned image. The only exception is a patch covered by a filled-out waiver at docs/waiver/<PR_NUMBER>.md — named after the PR that introduces the patch and filed in that same PR, stating what is patched, why the unmodified upstream image cannot run this benchmark, the upstream PR/issue link, and the removal plan — which I have linked below in the additional detail section.
• If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.

Additional detail section:

• Validation and evals: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/30838377466
• Speculative decoding uses chat templates in all five updated multi-node recipes.
• Upstream recipe check is N/A because this PR changes only disaggregated multi-node recipes.
• Uses the upstream vllm/vllm-openai image; no serving-stack patches or model-architecture overrides are introduced.
• Authorized sweep reuse: #2471 (comment)

Signed: kedarpotdar-nv

[Klaud-Cold]

❌❌❌ REJECTED ❌❌❌

@kedarpotdar-nv — blocking: every benchmark config this PR updates is the MiniMax-M3 Single-turn 8k1k scenario, which MODELS.md deprecates after 2026-08-03; the review date is 2026-08-04. A fresh sign-off cannot fix this — the submission targets a deprecated model-scenario combination.

✅ Check 0 (CODEOWNER): PASS — kedarpotdar-nv is a listed owner of configs/nvidia-master.yaml; the remaining changed paths carry only the catch-all owner.
✅ Check 1 (passing sweep on in-PR commit): PASS — head 4b53f2c has all five multi-node 8k1k / and five multi-node eval / check-runs green (executed, not skipped) in run 30838377466.
✅ Check 2 (evals pass): PASS — GSM8K em_strict 0.948–0.958 (n_eff 1319) across all 5 lane configs, run on vllm/vllm-openai:nightly-5e35a6f4…, the same image this PR pins.
➖ Check 3 (recipe link): N/A — disaggregated/multi-node submission (dynamo-vllm, multinode: true, disagg: true); the recipe-link requirement applies to single-node recipes only.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by xinli-sw (COLLABORATOR).
✅ Check 5 (latest checklist template): PASS — every current-template item is present and checked.
✅ Check 6 (upstream image / engine-first): PASS — upstream vllm/vllm-openai image on B300; the PR adds no new framework entries.
❌ Check 7 (deprecated models/scenarios): FAIL — MiniMax-M3 (minimaxm3) Single-turn 8k1k is deprecated after 2026-08-03 per MODELS.md ("Monday, August 3, 2026 is the last day…"; matrix row: "Single-turn 8k1k (until 2026-08-03)"), and this PR changes only minimaxm3 8k1k configs (minimaxm3-fp4-b300-dynamo-vllm-mtp plus its 5 recipe YAMLs) with review date 2026-08-04. The sign-off's checked MODELS.md item is incorrect.
✅ Check 8 (no architecture hacks): PASS — changes are an image bump, kv-cache-dtype: fp8 (lower precision, evals pass) and minimax_m3_msa_decode_backend: cutlass (kernel selection); no FLOPs-reducing overrides.
✅ Check 9 (spec-decode chat templates): PASS — all 5 EAGLE3 recipes benchmark via sa-bench with use_chat_template: true.
✅ Check 10 (no engine patches): PASS — YAML config changes only; the pinned image runs as shipped.
➖ Check 11 (agentic golden AL): N/A — no agentic spec-decode changes; the non-agentic EAGLE3 configs correctly use real acceptance (no synthetic knobs).