[NV] initial submission for minimax-m3 fp8 GB200 disagg receipes

[richardhuo-nv]

[NV] MiniMax-M3 FP8 GB200 disagg recipes

Summary

• Add 11 recipe files under benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/ covering disaggregated prefill/decode topologies on GB200 NVL72 (4 GPU/node):

  • 1k1k (5 shapes): 1p1d-dep4-dep8, 1p1d-dep4-tep8, 1p2d-dep4-dep8, 1p2d-dep4-tep8, 1p4d-dep4-tp4-marlin
  • 8k1k (6 shapes): 1p2d-dep4-dep8, 1p2d-dep4-tep4, 1p2d-dep4-tep8, 2p2d-dep4-dep8, 3p2d-dep4-dep8, 5p2d-dep4-dep8
  • All prefill workers use DEP4 (TP1 DP4 EP, 4 GPU/worker). Decode modes: DEP8, TEP8, TEP4, TP4+Marlin.
  • Image: vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223

• Add configs/minimax-m3-gb200-vllm-fixes.sh — setup script that patches the vLLM FlashInfer MNNVL all-reduce for one-shot support and makes the MiniMax-M3 MSA prefill top-k slice contiguous before CSR construction.

• Update runners/launch_gb200-nv.sh:

  • Add minimaxm3/fp8 model path case (/mnt/lustre01/models/MiniMax-M3-MXFP8)
  • Introduce uses_watchtower_shared_fs() helper, replacing hard-coded minimaxm2.5 || kimik2.5 checks throughout — minimaxm3 is now included in the watchtower shared-FS code paths
  • Add atomic import_squash() with flock serialization and temp-file swap, replacing bare enroot import calls to prevent race conditions from concurrent matrix jobs
  • Add minimaxm3/fp8 srt-slurm clone + recipe overlay branch, with SRTCTL_SETUP_SCRIPT wired to minimax-m3-gb200-vllm-fixes.sh
  • Refactor srtctl apply into a SRTCTL_APPLY_ARGS array so --setup-script is conditionally appended

• Add minimaxm3-fp8-gb200-dynamo-vllm to nvidia-master.yaml with the full 1k1k and 8k1k search spaces referencing the new recipe files

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

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. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

---

感谢你的贡献！对于 vLLM 与 SGLang，请确保你的 recipe 与官方 vLLM recipes 和/或 SGLang cookbook 保持一致

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

如果不一致，请先创建一个 PR，之后我们才能将你的单节点 PR 合并到 master 分支。让我们确保文档保持一流水准，使整个 ML 社区都能从你的辛勤工作中受益！谢谢

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

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。如果选择重新运行失败的任务，PR 作者有责任确保其最终通过。参见 GitHub 关于重新运行失败任务的文档：https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

一般而言，PR 作者应先向相应公司的 CODEOWNERS 请求审阅并获得 PR 批准，然后再请求核心维护者审阅。

如需更多帮助，PR 作者可通过 Slack 联系核心维护者。

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

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. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

---

感谢你的贡献！对于 vLLM 与 SGLang，请确保你的 recipe 与官方 vLLM recipes 和/或 SGLang cookbook 保持一致

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

如果不一致，请先创建一个 PR，之后我们才能将你的单节点 PR 合并到 master 分支。让我们确保文档保持一流水准，使整个 ML 社区都能从你的辛勤工作中受益！谢谢

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

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。如果选择重新运行失败的任务，PR 作者有责任确保其最终通过。参见 GitHub 关于重新运行失败任务的文档：https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

一般而言，PR 作者应先向相应公司的 CODEOWNERS 请求审阅并获得 PR 批准，然后再请求核心维护者审阅。

如需更多帮助，PR 作者可通过 Slack 联系核心维护者。

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28205303704
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28205303704

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28205303704
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28205303704

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28205303704
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28205303704

[Ankur-singh]

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. Link
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. Link
• Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
• If an company claims that they support vLLM/SGLang as first class 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 NVL72, 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 the single-node recipes are similar to the official vLLM recipes and/or theSGLang cookbook:
  • If they are not, I have verified that a PR has been opened in vLLM recipe repo or SGLang repo and linked it 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:

• This is a Dis-agg submission, no recipe updated required.

Signed: ankur-singh

[Klaud-Cold]

@Ankur-singh BLOCKED — Check 3 fails: the sign-off contains no recipe link. The "Dis-agg submission, no recipe updated required" note in the Additional detail section does not satisfy the link-required standard. Add a link to the corresponding recipe PR/page (vllm-project/recipes, sgl-project/sglang cookbook, recipes.vllm.ai, or docs.sglang.io/cookbook) — or, if disagg recipes are genuinely out of scope, confirm that exemption explicitly.

• Check 0 PASS — @ankur-singh is a listed owner of .github/configs/nvidia-master.yaml; remaining paths fall under catch-all * @InferenceX/core and are covered by a recognized CODEOWNER.
• Check 1 PASS — green executed sweep+eval on in-PR head 752562ae: multi-node 1k1k / (5), multi-node 8k1k / (6), multi-node eval / (6) all success. https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28205303704
• Check 2 PASS — 5 gsm8k em_strict scores 0.9530–0.9598 (above bar), model MiniMax-M3-MXFP8 on image vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223, matching the PR's config.
• Check 3 FAIL — no recipe link in the Additional detail section (link required).

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28391386827
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28391386827

[adibarra]

/reuse-sweep-run