[Klaud Cold] minimaxm3-fp4-mi355x-atom-agentic-mtp: day-zero MiniMax-M3 MXFP4 ATOM AgentX recipe on MI355X / MI355X 上 MiniMax-M3 MXFP4 ATOM AgentX 首发配方

[functionstackx]

Summary / 摘要

Day-zero MiniMax-M3 MXFP4 AgentX recipe on MI355X served by ATOM with EAGLE3 speculative decoding. minimaxm3 currently has no ATOM coverage in AgentX — the vLLM MI355X arm is the only MI355X entry — so this adds the ATOM arm.

在 MI355X 上新增由 ATOM 服务的 MiniMax-M3 MXFP4 AgentX 配方，采用 EAGLE3 投机解码。minimaxm3 目前在 AgentX 中没有 ATOM 覆盖（MI355X 上仅有 vLLM 分支），本 PR 补上 ATOM 分支。

Config key / 配置项

minimaxm3-fp4-mi355x-atom-agentic-mtp — amd/MiniMax-M3-MXFP4, rocm/atom-dev:nightly_202608251555 (latest ATOM nightly, verified on Docker Hub), cluster:mi355x-amds, TP4, conc [1, 4, 8, 12, 16], no KV offloading.

Recipe decisions / 配方要点

• TP4. MiniMax-M3 has four KV heads, so TP4 gives exactly one KV head per rank and keeps the AITER sparse-attention fast path — the same reason the vLLM MI355X sibling runs TP4.

• EAGLE3, 3 draft tokens, drafter Inferact/MiniMax-M3-EAGLE3 (the PoR draft the retired ATOM MI355X recipe used). Throughput runs pin --spec-decode-acceptance-length 2.83 from golden_al_distribution/minimaxm3_eagle3.yaml (thinking_on[3]); eval-only runs drop the pin and use real target verification.

• Prefix caching enabled. The retired single-turn 8k1k ATOM recipe passed --no-enable_prefix_caching; that is deliberately not carried over — trace replay is exactly the workload prefix caching pays for.

• No --hf-overrides index-cache block. The vLLM arm sets {"text_config": {"use_index_cache": true, "index_topk_freq": 4}}; this ATOM recipe deliberately does not.

• No launcher change needed: runners/launch_mi355x-amds.sh already resolves <prefix>_<precision>_mi355x_<framework><_mtp>.sh under agentic/, and already routes amd/MiniMax-M3* weights to the NFS HF cache.

• TP4：MiniMax-M3 有 4 个 KV head，TP4 使每个 rank 恰好一个 KV head，可走 AITER 稀疏注意力快速路径，与 vLLM MI355X 同类配方一致。

• EAGLE3、3 个草稿 token，草稿模型 Inferact/MiniMax-M3-EAGLE3。吞吐运行按 golden_al_distribution/minimaxm3_eagle3.yaml（thinking_on[3]）锁定 --spec-decode-acceptance-length 2.83；仅评测运行取消锁定，使用真实目标验证。

• 启用前缀缓存：已退役的单轮 8k1k ATOM 配方使用 --no-enable_prefix_caching，此处刻意不沿用——轨迹回放正是前缀缓存最有价值的负载。

• 不添加 --hf-overrides 索引缓存参数：vLLM 分支设置了 {"text_config": {"use_index_cache": true, "index_topk_freq": 4}}，本 ATOM 配方刻意不设置。

• 无需改动 launcher：runners/launch_mi355x-amds.sh 已能解析 agentic/ 下的脚本名，并已将 amd/MiniMax-M3* 权重指向 NFS HF 缓存。

Files / 改动文件

• benchmarks/single_node/agentic/minimaxm3_fp4_mi355x_atom_mtp.sh (new)
• configs/amd-master.yaml — new entry after the vLLM MI355X sibling
• perf-changelog.yaml — appended entry

Note / 说明

A sibling PR builds the same config key from the upstream ROCm/ATOM Qwen3.5 recipe shape instead. The two are mutually exclusive candidates — merge whichever sweeps green, not both.

另有一个同类 PR 按上游 ROCm/ATOM Qwen3.5 recipe 的形态实现同一配置项。两者互斥，只应合并全量 sweep 通过的那一个。

🤖 Generated with Claude Code

---

Note

Low Risk
Benchmark and YAML config only; no changes to application runtime, auth, or shared orchestration logic beyond registering a new sweep recipe.

Overview
Adds day-zero AgentX coverage for amd/MiniMax-M3-MXFP4 on MI355X served by ATOM (alongside the existing vLLM arm), via new config key minimaxm3-fp4-mi355x-atom-agentic-mtp and benchmark script minimaxm3_fp4_mi355x_atom_mtp.sh.

The recipe runs TP4 / EP1 / no DP-attention / no KV offload, sweeps concurrency 1, 4, 8, 12, 16, and uses EAGLE3 with drafter Inferact/MiniMax-M3-EAGLE3 and three draft tokens. Throughput runs pin --spec-decode-acceptance-length to 2.83 from the committed golden curve; EVAL_ONLY drops the pin for real target verification. The ATOM server enables prefix caching, PTPC FP8 online quant, Triton sparse attention (ATOM_FORCE_ATTN_TRITON), and required atom: Prometheus metrics for AIPerf.

perf-changelog.yaml documents the new matrix entry.

^{Reviewed by Cursor Bugbot for commit 28edf09. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

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。参见 GitHub 关于重新运行失败任务的文档

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=32919182029
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=32919182029