[AMD] Add MiniMax-M3-FP8 MI355X ATOM non-EAGLE3 & EAGLE3#1867
[AMD] Add MiniMax-M3-FP8 MI355X ATOM non-EAGLE3 & EAGLE3#1867seungrokj wants to merge 15 commits into
Conversation
… benchmark scripts and CI configs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
seungrokj
requested review from
1am9trash,
billishyahao,
chunfangamd and
yctseng0211
as code owners
|
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。 |
…tom-mtp Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
seungrokj
changed the title
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
…atom scripts Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27860917145 |
…-atom-mtp Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…m-mtp Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27862137852 |
…-only configs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
There was a problem hiding this comment.
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 196961a. Configure here.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27862752989 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27862873087 |
|
/merge-prs |
/merge-prs isnt an github comment fyi, it is an claude code command u type into claude code CLI https://code.claude.com/docs/en/commands @seungrokj |
…-20260622 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27996766982 |
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27999248579 |
- Add AITER_QUICK_REDUCE_QUANTIZATION=INT4, MAX_MODEL_LEN=32768, MAX_NUM_BATCHED_TOKENS=32768, MAX_NUM_SEQS=128 - Pass --max-model-len, --max-num-batched-tokens, --max-num-seqs to server - Conditional --use-chat-template based on SPEC_ARGS - Add trailing newline Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28000404764 |
2 participants
Summary
minimaxm3-fp8-mi355x-atomCI recipe: single-node ATOM benchmark for MiniMax-M3-MXFP8 on MI355Xminimaxm3-fp8-mi355x-atom-mtpCI recipe: same with EAGLE3 speculative decoding (3 draft tokens)rocm/atom-dev:MiniMax-M3-20260619image, search space mirrors the FP4 atom variants (ISL=1024,8192 OSL=1024 TP2/TP4/TP8)Test plan
mi355xrunner for both configs🤖 Generated with Claude Code
Note
Low Risk
Changes are limited to benchmark shell scripts and CI YAML; they do not alter application serving, auth, or data paths. Main operational risk is new GPU CI sweep cost or recipe drift if the pinned ATOM image or EAGLE3 draft setup fails on hardware.
Overview
Adds MI355X single-node ATOM performance CI for MiniMax-M3 MXFP8, alongside the existing vLLM and FP4 ATOM entries.
Two new config keys in
amd-master.yaml—minimaxm3-fp8-mi355x-atomandminimaxm3-fp8-mi355x-atom-mtp— both pinrocm/atom-dev:MiniMax-M3-20260622,MiniMaxAI/MiniMax-M3-MXFP8, and fixed-seq-len sweeps at ISL 1024/8192 with OSL 1024, TP4, concurrency 1–256. The MTP recipe marksspec-decoding: mtpin the search space.New benchmark drivers
minimaxm3_fp8_mi355x_atom.shandminimaxm3_fp8_mi355x_atom_mtp.shstartatom.entrypoints.openai_serverwith block size 128, 0.8 GPU memory util, and shared ROCm env (AITER_QUICK_REDUCE_QUANTIZATION=INT4, 32k context/batch limits). The MTP script adds EAGLE3 via--method eagle3, draftInferact/MiniMax-M3-EAGLE3, and 3 speculative tokens; both optionally run lm-eval whenRUN_EVAL=true.perf-changelog.yamldocuments the new recipes for changelog tracking.Reviewed by Cursor Bugbot for commit ecd0bef. Bugbot is set up for automated code reviews on this repo. Configure here.