[AMD] Add MiniMax-M3-FP8 MI355X ATOM EAGLE3 only by seungrokj · Pull Request #1867 · SemiAnalysisAI/InferenceX

seungrokj · 2026-06-20T04:51:44Z

Summary

Add minimaxm3-fp8-mi355x-atom CI recipe: single-node ATOM benchmark for MiniMax-M3-MXFP8 on MI355X
Add minimaxm3-fp8-mi355x-atom-mtp CI recipe: same with EAGLE3 speculative decoding (3 draft tokens)
Both use rocm/atom-dev:MiniMax-M3-20260619 image, search space mirrors the FP4 atom variants (ISL=1024,8192 OSL=1024 TP2/TP4/TP8)

Test plan

Server launches correctly for FP8 checkpoint
EAGLE3 spec args pass through correctly for mtp variant
CI sweep triggers on mi355x runner for both configs

🤖 Generated with Claude Code

Note

Low Risk
Benchmark and CI config only; no changes to auth, data paths, or production serving logic.

Overview
Adds minimaxm3-fp8-mi355x-atom-mtp to AMD CI so MiniMax-M3 MXFP8 on MI355X can be swept under ATOM with EAGLE3 (spec-decoding: mtp, 3 draft tokens). The recipe pins rocm/atom-dev:MiniMax-M3-20260622, MiniMaxAI/MiniMax-M3-MXFP8, fixed-seq-len at ISL 1024/8192 and OSL 1024, and TP4 concurrency 1–256.

Introduces benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi355x_atom_mtp.sh, which starts atom.entrypoints.openai_server with EAGLE3 (Inferact/MiniMax-M3-EAGLE3), block size 128, and the usual serving limits, then runs the standard throughput benchmark (optional lm-eval when RUN_EVAL=true).

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

^{Reviewed by Cursor Bugbot for commit 3ac8eeb. Bugbot is set up for automated code reviews on this repo. Configure here.}

… benchmark scripts and CI configs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

github-actions · 2026-06-20T04:51:52Z

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>

…atom scripts Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

github-actions · 2026-06-20T05:40:22Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27860917145
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?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>

github-actions · 2026-06-20T06:21:09Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27862137852
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27862137852

…-only configs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

github-actions · 2026-06-20T06:28:34Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27862752989
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27862752989

github-actions · 2026-06-20T07:53:52Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27862873087
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27862873087

seungrokj · 2026-06-20T14:23:31Z

/merge-prs

functionstackx · 2026-06-20T23:17:56Z

/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>

github-actions · 2026-06-23T03:01:15Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27996766982
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27996766982

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

github-actions · 2026-06-23T03:13:04Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27999248579
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?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>

github-actions · 2026-06-23T04:19:33Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28000404764
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28000404764

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

github-actions · 2026-06-23T05:56:10Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28004827939
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28004827939

github-actions · 2026-06-23T08:10:24Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28004827939
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28004827939

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…log entry Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

github-actions · 2026-06-23T08:16:02Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28011998988
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28011998988

cursor · 2026-06-23T08:16:13Z

+    "${SPEC_ARGS[@]}" \
+    --block-size 128 \
+    --gpu-memory-utilization $MEM_FRAC_STATIC \
+    --max-model-len $MAX_MODEL_LEN \


Hardcoded max model length

High Severity

The benchmark script sets MAX_MODEL_LEN to 32768 and passes it to the atom server, overwriting the matrix value (isl + osl + 256, e.g. 9472 for 8k/1k). Sibling atom recipes use the supplied scenario length. Extra KV reservation can prevent the server from starting or waste GPU time under high concurrency with EAGLE3.

^{Reviewed by Cursor Bugbot for commit 0d1ed51. Configure here.}

github-actions · 2026-06-23T09:43:32Z

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28012203541
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28012203541

seungrokj · 2026-06-23T10:02:19Z

/reuse-sweep-run

chunfangamd

[x] Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
[x] Verified that this PR has passed PR validation.
[x] Verified that this PR passes evals.
[x] Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
[x] minimaxm3-fp8-mi355x-vllm-mtp already merged into the InferenceX repo

cursor

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

There are 2 total unresolved issues (including 1 from previous review).

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

^{Reviewed by Cursor Bugbot for commit 3ac8eeb. Configure here.}

cursor · 2026-06-23T13:20:29Z

      - { tp: 8, conc-start: 1, conc-end: 2 }

+minimaxm3-fp8-mi355x-atom-mtp:
+  image: rocm/atom-dev:MiniMax-M3-20260622


Container image tag mismatch

Medium Severity

The new minimaxm3-fp8-mi355x-atom-mtp entry pins rocm/atom-dev:MiniMax-M3-20260622, while the perf-changelog line added in the same change documents MiniMax-M3-20260619. CI pulls the yaml image tag, so one of these is wrong and can fail the job or run an unintended build.

Additional Locations (1)

perf-changelog.yaml#L4007-L4008

^{Reviewed by Cursor Bugbot for commit 3ac8eeb. Configure here.}

feat: add minimaxm3-fp8-mi355x-atom and minimaxm3-fp8-mi355x-atom-mtp…

4d484fb

… benchmark scripts and CI configs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

seungrokj requested a review from a team June 20, 2026 04:51

seungrokj requested review from 1am9trash, billishyahao, chunfangamd and yctseng0211 as code owners June 20, 2026 04:51

github-project-automation Bot added this to InferenceMAX Board Jun 20, 2026

chore: add perf-changelog entries for minimaxm3-fp8-mi355x-atom and a…

45b360c

…tom-mtp Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

seungrokj changed the title ~~feat: add minimaxm3-fp8-mi355x-atom and atom-mtp single-node ATOM benchmarks~~ [AMD] Add MiniMax-M3-FP8 MI355X ATOM non-EAGLE3 & EAGLE3 Jun 20, 2026

cursor Bot reviewed Jun 20, 2026

View reviewed changes

Comment thread benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi355x_atom_mtp.sh Outdated

Comment thread benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi355x_atom.sh Outdated

functionstackx requested changes Jun 20, 2026

View reviewed changes

Comment thread benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi355x_atom_mtp.sh Outdated

fix: remove stray GitHub review comment artifacts from minimaxm3_fp8 …

4700bd4

…atom scripts Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

seungrokj added AMD full-sweep-enabled labels Jun 20, 2026

fix: use MiniMaxAI/MiniMax-M3-MXFP8 model id for minimaxm3-fp8-mi355x…

e17cbaa

…-atom-mtp Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cursor Bot reviewed Jun 20, 2026

View reviewed changes

Comment thread .github/configs/amd-master.yaml

seungrokj and others added 2 commits June 20, 2026 14:52

Merge branch 'main' into amd/m3_atom_fp8

f48400e

fix: correct model id in amd-master.yaml for minimaxm3-fp8-mi355x-ato…

a068fde

…m-mtp Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

fix: trim minimaxm3-fp8-mi355x-atom and atom-mtp search spaces to TP4…

196961a

…-only configs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cursor Bot reviewed Jun 20, 2026

View reviewed changes

Comment thread .github/configs/amd-master.yaml

fix: correct minimaxm3-fp8-mi355x-atom-mtp ISL=1024 search space to tp4

e016169

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Merge branch 'main' into amd/m3_atom_fp8

3913e9a

Merge branch 'main' into amd/m3_atom_fp8

f3def1c

seungrokj removed the full-sweep-enabled label Jun 23, 2026

seungrokj and others added 2 commits June 23, 2026 10:51

Merge branch 'main' into amd/m3_atom_fp8

9f9c9a7

fix: bump minimaxm3-fp8-mi355x-atom and atom-mtp images to MiniMax-M3…

5c44abf

…-20260622 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

seungrokj added the full-sweep-enabled label Jun 23, 2026

fix: disable prefix caching for minimaxm3-fp8-mi355x-atom and atom-mtp

d2cd74e

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

seungrokj and others added 2 commits June 23, 2026 12:36

Merge branch 'main' into amd/m3_atom_fp8

ecd0bef

fix: bump MAX_NUM_SEQS to 256 for minimaxm3-fp8-mi355x-atom scripts

45dfb13

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cursor Bot reviewed Jun 23, 2026

View reviewed changes

Comment thread benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi355x_atom.sh Outdated

Merge branch 'main' into amd/m3_atom_fp8

60deebe

fix: cap minimaxm3-fp8-mi355x-atom search-space conc-end to 128

7557212

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

seungrokj changed the title ~~[AMD] Add MiniMax-M3-FP8 MI355X ATOM non-EAGLE3 & EAGLE3~~ [AMD] Add MiniMax-M3-FP8 MI355X ATOM EAGLE3 only Jun 23, 2026

fix: remove minimaxm3-fp8-mi355x-atom recipe, script, and perf-change…

0d1ed51

…log entry Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

cursor Bot reviewed Jun 23, 2026

View reviewed changes

chunfangamd approved these changes Jun 23, 2026

View reviewed changes

Merge branch 'main' into amd/m3_atom_fp8

3ac8eeb

cursor Bot reviewed Jun 23, 2026

View reviewed changes

Conversation

seungrokj commented Jun 20, 2026 • edited by cursor Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

github-actions Bot commented Jun 20, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Jun 20, 2026

Uh oh!

Uh oh!

github-actions Bot commented Jun 20, 2026

Uh oh!

Uh oh!

github-actions Bot commented Jun 20, 2026

Uh oh!

github-actions Bot commented Jun 20, 2026

Uh oh!

seungrokj commented Jun 20, 2026

Uh oh!

functionstackx commented Jun 20, 2026

Uh oh!

github-actions Bot commented Jun 23, 2026

Uh oh!

github-actions Bot commented Jun 23, 2026

Uh oh!

github-actions Bot commented Jun 23, 2026

Uh oh!

Uh oh!

github-actions Bot commented Jun 23, 2026

Uh oh!

github-actions Bot commented Jun 23, 2026

Uh oh!

github-actions Bot commented Jun 23, 2026

Uh oh!

cursor Bot Jun 23, 2026

Choose a reason for hiding this comment

Hardcoded max model length

Uh oh!

github-actions Bot commented Jun 23, 2026

Uh oh!

seungrokj commented Jun 23, 2026

Uh oh!

chunfangamd left a comment

Choose a reason for hiding this comment

Uh oh!

cursor Bot left a comment

Choose a reason for hiding this comment

Uh oh!

cursor Bot Jun 23, 2026

Choose a reason for hiding this comment

Container image tag mismatch

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

seungrokj commented Jun 20, 2026 •

edited by cursor Bot

Loading