[AMD] Add MiniMax-M3-FP8 MI355X ATOMESH

[seungrokj]

Summary

• Add minimaxm3-fp8-mi355x-atom-disagg CI recipe: multi-node disaggregated prefill-decode on MI355X via ATOM for MiniMax-M3-MXFP8
• Align server settings with slurm reference script: MEM_FRAC_STATIC=0.8, MAX_NUM_SEQS=128, BLOCK_SIZE=128, MAX_MODEL_LEN=32768, KV_CACHE_DTYPE=auto
• server_atom.sh: fix _MAX_CONC assignment before cudagraph size check; gate ATOM_MOE_GU_ITLV and AITER_BF16_FP8_MOE_BOUND on DeepSeek-V4-Pro only; use ${KV_CACHE_DTYPE:-fp8} default
• Search space: ISL=8192 and ISL=1024, 1P1D TP4, conc 1–512

Test plan

• CI sweep on mi355x-disagg runner triggers correctly
• --kv-cache-dtype is not passed when KV_CACHE_DTYPE=auto
• Decode node cudagraph sizes scale with max concurrency

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches shared multi-node ATOM launch paths (server_atom.sh, env_atom.sh, job.slurm) used by other disagg recipes; behavior changes are mostly gated by model name but could affect non–DSv4 atom-disagg runs.

Overview
Adds minimaxm3-fp8-mi355x-atom-disagg to AMD CI: multi-node 1P1D prefill–decode on MI355X with ATOM + mooncake, TP4, 1k/1k and 8k/1k, concurrency 1–512, via new launcher minimaxm3_fp8_mi355x_atom-disagg.sh (reference tuning: MEM_FRAC_STATIC=0.8, block size 128, MAX_MODEL_LEN=32768, KV_CACHE_DTYPE=auto, no MTP).

server_atom.sh is generalized for non–DeepSeek-V4 models: MEM_FRACTION → MEM_FRAC_STATIC, optional MAX_MODEL_LEN / MAX_NUM_BATCHED_TOKENS, MTP/spec args, KV dtype omitted when auto, model-specific parallel flags (DSv4 TBO/HF overrides vs AITER_QUICK_REDUCE_QUANTIZATION=INT4 elsewhere), and eval-built server commands. env_atom.sh applies ATOM_MOE_GU_ITLV / AITER_BF16_FP8_MOE_BOUND only for DeepSeek-V4-Pro. job.slurm passes BENCH_REQUEST_RATE and the new atom-disagg env vars; models_atom.yaml registers MiniMax-M3 MXFP4/MXFP8. bench.sh uses --dsv4 for MTP on DeepSeek-V4-Pro. perf-changelog.yaml documents the recipe.

^{Reviewed by Cursor Bugbot for commit c363d91. 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]

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]

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。

[functionstackx]

will review it in an bit but seems like we need to merge vllm disagg first?

i thought we chatted about this before about sglang/vllm native engine first back in april 17 & your thumbs up means u acholwdege the guidelines

#1043 (comment)

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27859909769
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27859909769

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27861799455
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27861799455

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27862212112
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27862212112

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27862212112
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27862212112

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

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

^{Reviewed by Cursor Bugbot for commit cb168e4. Configure here.}

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27874092441
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27874092441

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27874309402
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27874309402

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27874309402
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27874309402

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27874309402
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27874309402

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27894796811
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27894796811

[functionstackx]

hi @seungrokj @andyluo7 @chunfangamd

vLLM FP8 Disagg M3 has merged so we can merge this PR once yall has followed the PR review process and properly filled in the PR review form so that the signoff CI verification can trigger https://github.com/SemiAnalysisAI/InferenceX/blob/main/.github/workflows/codeowner-signoff-verify.yml

can u please use this PR CHECKLIST template for reviewing it https://github.com/SemiAnalysisAI/InferenceX/blob/main/docs/PR_REVIEW_CHECKLIST.md

here is an example of how to do it #1891 (comment)

If you don't follow the PR CHECKLIST template including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have in ur approval comment, it wouldn't trigger our codeowner signoff CI verifcation https://github.com/SemiAnalysisAI/InferenceX/blob/main/.github/workflows/codeowner-signoff-verify.yml

[Klaud-Cold]

@functionstackx BLOCKED — not a CODEOWNER for the AMD config changed here, no executed sweep/eval on any in-PR commit, and no recipe link. The signed comment is also a request to the AMD reviewers to fill in the checklist, not a completed sign-off.

• Check 0 (CODEOWNER): FAIL — PR changes .github/configs/amd-master.yaml, owned by @billishyahao @chunfangamd @seungrokj @yctseng0211 @1am9trash. functionstackx is not in that list and is not listed anywhere in CODEOWNERS. Sign-off must come from an AMD owner.
• Check 1 (green executed sweep+eval on an in-PR commit): FAIL — No in-PR commit has green, executed multi-node */ + eval / check-runs. Head c0a813b has them all skipped (reuse run, executed nothing on this commit: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28130510662); no other of the 33 PR commits has any benchmark/eval check-run. Fix: run (or re-anchor via /reuse-sweep-run) a passing full sweep on a commit currently in this PR.
• Check 2 (eval accuracy): FAIL — no passing eval run exists on any in-PR commit, so accuracy cannot be verified.
• Check 3 (recipe linked + complete): FAIL — the sign-off comment contains no recipe link (no vllm-project/recipes, sgl-project/sglang, recipes.vllm.ai, or docs.sglang.io). It also lacks the required checklist/As a PR reviewer and CODEOWNER, I have reviewed this sign-off text.

[billishyahao]

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

• Verified that this is the latest version of PR_REVIEW_CHECKLIST.md
• Verified that the general code quality meets the InferenceX standard and does not make it worse.
• Verified that this PR has passed PR validation: https://inferencex.semianalysis.com/inference?unofficialRun=27996171132
• Verified that this PR passes evals: [](https://inferencex.semianalysis.com/evaluation?unofficialRun=27996171132)
• vLLM/SGLang submitted before ATOM: vLLM FP8 disagg M3 minimaxm3-fp8-mi355x-vllm-disagg #1762 merged 2026-06-24.
• Single-node recipes similar to official vLLM/SGLang recipes:
• If any criteria cannot be satisfied, reasoning provided below.

Additional detail section:

This is a multi-node FP8 disaggregated ATOMesh config (week-1 MI355X support); no single-node recipe PR required. vLLM equivalent merged in #1762.

Signed: billishyahao

[seungrokj]

@functionstackx can you please approve this ?

[Klaud-Cold]

@billishyahao Not ready to merge: no passing sweep/eval exists on any commit in this PR, and the sign-off has no public recipe link.

• Check 0 PASS — billishyahao is a named owner of .github/configs/amd-master.yaml; remaining paths fall under the * @InferenceX/core catch-all and are covered.
• Check 1 FAIL — No passing sweep/eval on any commit currently in this PR. Every benchmark/eval check-run (single-node */, multi-node */, eval /) is skipped on all 34 PR commits (head a2e7439); only the reuse-sweep-gate gate is green. The signed-off run 27996171132 executed on fa89765aa9c838c6143e1b2d16dd71bccac40cbc, which is not in GET /pulls/1865/commits (rebased out) — validate_reusable_run will reject it. Fix: run or /reuse-sweep-run a full green sweep on a commit currently in the PR.
• Check 2 BLOCKED — No valid in-PR eval run to verify accuracy against; depends on Check 1.
• Check 3 FAIL — The Additional detail section contains no link to a public recipe (vllm-project/recipes or sgl-project/sglang). It only references InferenceX PR minimaxm3-fp8-mi355x-vllm-disagg #1762, which is an internal PR, not a published/public recipe. A recipe link is required even for a multi-node config.