[AMD] Add MiniMax-M3-FP4 MI355X ATOM EAGLE3 only by seungrokj · Pull Request #1866 · SemiAnalysisAI/InferenceX

seungrokj · 2026-06-20T04:28:28Z

Summary

Add minimaxm3-fp4-mi355x-atom-mtp CI recipe: single-node ATOM benchmark for MiniMax-M3-MXFP4 on MI355X with EAGLE3 speculative decoding (3 draft tokens)
Add benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh with TP/EP/DP-attention support
Search space mirrors minimaxm3-fp4-mi355x-atom: ISL=1024 and ISL=8192, TP2/TP4/TP8

Test plan

Server launches correctly with EAGLE3 spec args
TP/EP/DP-attention combinations pass through correctly
CI sweep triggers on mi355x runner

🤖 Generated with Claude Code

Note

Low Risk
Changes are limited to CI config, benchmark shell scripts, and changelog; they only affect automated MI355X sweeps, not application runtime code.

Overview
Adds minimaxm3-fp4-mi355x-atom-mtp, a new AMD master CI recipe for single-node ATOM serving of MiniMax-M3-MXFP4 on MI355X with EAGLE3 speculative decoding (spec-decoding: mtp), pinned to rocm/atom-dev:MiniMax-M3-20260622.

Introduces minimaxm3_fp4_mi355x_atom_mtp.sh, which starts atom.entrypoints.openai_server with the Inferact/MiniMax-M3-EAGLE3 draft (3 speculative tokens), optional expert parallel and DP attention, then runs the standard fixed-seq-len throughput path (and optional lm-eval). The YAML search space covers ISL 1024/8192, OSL 1024, TP2/TP4, and concurrency ranges aligned with the non-MTP minimaxm3-fp4-mi355x-atom recipe (without the base recipe’s TP8 row on 8k).

Documents the new config key in perf-changelog.yaml.

^{Reviewed by Cursor Bugbot for commit 65a9d7d. Bugbot is set up for automated code reviews on this repo. Configure here.}

Add single-node ATOM MTP (EAGLE3) benchmark for MiniMax-M3-MXFP4 on MI355X with TP/EP/DP-attention support and matching amd-master.yaml entry. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

github-actions · 2026-06-20T04:28:35Z

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

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

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

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

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 · 2026-06-20T04:28:35Z