[NV] update B300 disagg recipes

[jasonlizhengjian]

Copy of #1887 from biswapanda:pr-1787-latest--update at d08cc43b7e943a7a53c67194448d4309f338cbc4.

The original fork PR triggered run-sweep.yml, but all downstream benchmark jobs failed before checkout because fork PRs do not receive inherited repository secrets such as REPO_PAT. This same-repo copy is intended to run the sweep with the same diff against main.

Original PR: #1887

---

Note

Low Risk
Benchmark and CI YAML only; no runtime application or auth/data-path changes.

Overview
Refreshes the MiniMax-M3 MXFP8 B300 Dynamo-vLLM disaggregated benchmark sweep for 8k1k (and related search-space wiring in nvidia-master.yaml): concurrency lists, prefill/decode worker counts, and which Slurm recipe YAML each point uses.

Several new 8k1k recipes are added for TEP4 (and related) topologies—e.g. 1p2d-dep2-tep4, 1p4d-dep2-tep4, 1p4d-dep2-tep8, 4p2d-dep2-tep4—while some older combinations (e.g. conc 128-only, 3p/5p prefill, duplicate 4096 entries) are dropped or remapped. Existing recipes in that folder gain FLASHINFER + TRTLLM attention and kv-cache-dtype: fp8 on prefill and decode to align with the updated sweep.

perf-changelog.yaml documents the minimaxm3-fp8-b300-dynamo-vllm config change (image minimax-m3-0618-x86_64-cu130, TEP4/TEP8 8k1k coverage).

^{Reviewed by Cursor Bugbot for commit 6b9fe1e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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=27999202538
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27999202538

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 376f261. Configure here.}

[cursor]

Decode seq cap below concurrency

Medium Severity

The sa-bench concurrency is 4096 while decode max-num-seqs is 1024 with two decode workers. That pairing cannot honor 4096 simultaneous decode slots per worker semantics, so the 4096 sweep point may queue, fail, or report throughput that does not reflect the intended load.

Additional Locations (1)

• benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-tep4-8k1k.yaml#L69-L82

 

^{Reviewed by Cursor Bugbot for commit 376f261. Configure here.}

[cursor]

Patch apply aborts launcher

Medium Severity

The MiniMax-M3 path runs git apply for the node-IP fix with || exit 1, so if sa-submission-q2-2026 already contains that patch, checkout/setup fails and no B300 MiniMax-M3 job is submitted.

 

^{Reviewed by Cursor Bugbot for commit 376f261. Configure here.}

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28028985088
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28028985088

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28028985088
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28028985088

[Ankur-singh]

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

• 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.
• Verified that this PR passes evals.
• 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 config, no recipe PR required.

Signed: ankur-singh

[Klaud-Cold]

@Ankur-singh BLOCKED — Check 3 (recipe) fails: the sign-off's Additional detail section contains no recipe link. This gate requires a public recipe link (vllm-project/recipes or sglang cookbook); "dis-agg config, no recipe PR required" is a bare claim without a link and does not satisfy the standard.

• Check 0 — PASS. Signer owns .github/configs/nvidia-master.yaml (@ankur-singh @kedarpotdar-nv @jgangani); the multi_node recipe YAMLs fall under catch-all * @InferenceX/core, covered since the signer is a recognized CODEOWNER.
• Check 1 — PASS. Head 6b9fe1e (in PR) carries green, non-skipped multi-node 1k1k / 8k1k / eval check-runs — run 28028985088. The generic single-node */ and eval / jobs are skipped as expected for this multi-node disagg config.
• Check 2 — PASS. Real gsm8k accuracy on that run: em_strict 0.954–0.966 across all disagg points, model MiniMax-M3-MXFP8, image vllm/vllm-openai:minimax-m3-0618-x86_64-cu130 (matches this PR's config).
• Check 3 — FAIL. No recipe link present. Fix: add a link to the corresponding public recipe (or open one and link it), or confirm this gate's link requirement is waived for disagg configs.

[RohitNagraj]

/reuse-sweep-run 28028985088

[Oseltamivir]

lgtm