feat(minimaxm3-fp4-b200-vllm): bump vLLM nightly to 5e35a6f4, enable cutlass MSA decode / 更新 minimaxm3-fp4-b200-vllm vLLM 夜版镜像至 5e35a6f4，启用 cutlass MSA 解码后端 - #2468
Conversation
…cutlass MSA decode backend 中文：将 minimaxm3-fp4-b200-vllm vLLM 夜版镜像升级至 nightly-5e35a6f4f9bbc217c599692157ca985c894373f7，并启用 VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass 以修复固定长度推理中的 MSA 解码后端 Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
1 similar comment
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
… entry 中文：追加 minimaxm3-fp4-b200-vllm 镜像升级的 perf-changelog 条目 Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
LGTM — straightforward vLLM nightly bump for minimaxm3-fp4-b200-vllm plus a new backend env var, matching the standard image-bump pattern.
What was reviewed: the image tag bump in configs/nvidia-master.yaml, the new VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass export in the recipe script, and the perf-changelog.yaml entry (present and correctly formatted, matching the PR link and config key).
Extended reasoning...
Overview
This PR touches 3 files: it bumps the minimaxm3-fp4-b200-vllm vLLM nightly image tag in configs/nvidia-master.yaml, adds a single new env var export (VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass) to benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_b200.sh, and appends a corresponding entry to perf-changelog.yaml. No other logic in the recipe script or config is changed.
Security risks
None. This is an internal benchmark recipe pulling a pinned, content-addressed nightly image tag and setting an environment variable that only affects vLLM's internal MSA decode backend selection. No user input, auth, or secrets are involved.
Level of scrutiny
Low. This is a routine image-bump PR of the kind regularly automated in this repo (see the 'nuke' pattern for bumping single-node engine images), scoped to a single model/precision/SKU recipe. The new env var is consistent with the file's existing pattern of exporting VLLM_* tuning flags, and the changelog entry follows the established format and correctly references this PR.
Other factors
No bugs were found by the bug hunting system, and the one candidate issue raised (perf-changelog.yaml entry missing) was refuted — the entry is in fact present in the diff. The PR includes a full-sweep-fail-fast label for CI validation, which is the appropriate verification mechanism for this kind of change.
…SA decode, not env var 中文：将 VLLM_MINIMAX_M3_MSA_DECODE_BACKEND 环境变量替换为正确的 --attention_config.minimax_m3_msa_decode_backend cutlass 参数
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30821645815 |
…isable and cutlass MSA backend Disable chunked prefill and enable cutlass MSA decode backend only for conc < 64. At conc >= 64, cutlass MSA backend is skipped and chunked prefill remains enabled. 中文：仅在 conc < 64 时禁用分块预填充并启用 cutlass MSA 解码后端；conc >= 64 时跳过 cutlass MSA 后端并保留分块预填充 Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30828214188 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30841588412 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30841588412 |
|
merged vllm-project/recipes#735 |
|
/reuse-sweep-run |
1 similar comment
|
/reuse-sweep-run |
# Conflicts: # perf-changelog.yaml
2 participants
Summary
minimaxm3-fp4-b200-vllmvLLM image tonightly-5e35a6f4f9bbc217c599692157ca985c894373f7(contains fixed-len MiniMax-M3 fix)VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlassinminimaxm3_fp4_b200.shTest plan
full-sweep-fail-fastlabel runs green中文说明
minimaxm3-fp4-b200-vllmvLLM 镜像升级至nightly-5e35a6f4f9bbc217c599692157ca985c894373f7（包含固定长度 MiniMax-M3 修复）minimaxm3_fp4_b200.sh中设置VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass🤖 Generated with Claude Code