Update MiniMax-M3 GB200 aggregate recipes / 更新 MiniMax-M3 GB200 聚合配方 - #2925
Conversation
更新 MiniMax-M3 GB200 聚合配方，固定服务运行时，并加入 TP4 常驻、TP4 延迟 SimpleCPU KV 卸载和双节点 TP8 常驻配置。
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
将 MiniMax-M3 GB200 变更日志条目关联到对应的拉取请求。
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
I reviewed this PR and didn't find any bugs. Because it changes multi-node cluster launch behavior (srt-slurm checkout strategy) and a benchmark script's readiness-wait gating shared across ~100+ recipes, a human look would still be worthwhile.
What was reviewed:
- configs/nvidia-master.yaml search-space changes cross-checked against the three new recipe YAMLs (image/container, router version, topology fields all match).
- Confirmed
model.containerequals the pinned image in all three new nightly-native recipe YAMLs. - Checked the EVAL_ONLY guard removal in agentic_srt.sh and its updated test — the test now asserts the readiness wait runs unconditionally, matching the code.
- perf-changelog.yaml entry appended at file tail without altering existing bytes; still carries a placeholder
pull/XXXlink.
Extended reasoning...
Overview
This PR reworks the MiniMax-M3 GB200 FP4 aggregate agentic-coding benchmark: it trims the search space in configs/nvidia-master.yaml to three topology arms (TP4 resident, TP4 with vllm-simple DRAM KV offload, and two-node TP8 resident), adds three new 'nightly-native' recipe YAMLs, pins a new vLLM nightly image and Dynamo router version, changes the srt-slurm checkout strategy in runners/launch_gb200-nv.sh from a shared branch-pinned shallow clone to an independent full-clone-then-checkout of a specific commit, removes an EVAL_ONLY conditional gating a readiness wait in benchmarks/multi_node/agentic_srt.sh, and updates/renames a corresponding unit test.
Security risks
None identified. This is benchmark configuration and CI/orchestration tooling; no user input, auth, or crypto paths are touched. The git clone/checkout change fetches from a fixed public GitHub URL and pins by commit SHA verified via git rev-parse HEAD, which is arguably safer than the prior branch-based pin.
Level of scrutiny
The bug-hunting system ran to a dry streak (i.e., exhausted its search without finding further issues) and reported no bugs, with two candidate concerns (the always-on readiness wait's blast radius across shared recipes, and the full-clone vs. scoped-clone divergence from sibling model paths) investigated and explicitly ruled out. The changes are mechanical and pattern-following (config/recipe pins, changelog append, test rename), but they touch a shared script (agentic_srt.sh) used by many other models' recipes and a cluster launcher script, which is why I lean toward a human co-reviewing rather than a confident stand-alone approval, even absent concrete findings.
Other factors
The test coverage for the EVAL_ONLY removal was updated appropriately and follows the repo's test-quality guidance (no frozen image/pin literals baked into new assertions beyond what's already an existing pattern for this test file). The perf-changelog entry's pull/XXX placeholder is a known pre-merge TODO already flagged as a convention item, not a functional bug.
This review covers commit 0cf6543, which is no longer the latest commit on this pull request; later commits are not covered by it.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34384638294 |
|
/reuse-sweep-run 34384638294 |
|
As a PR reviewer and CODEOWNER, I have reviewed this and have:
Additional detail section:
Signed: |
✅✅✅ Verdict: PASS ✅✅✅✅ Check 0 (CODEOWNER): PASS — |
Resolve conflicts in configs/nvidia-master.yaml and perf-changelog.yaml. - nvidia-master: keep this branch's rewrite of minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp (nightly image, TP4 resident / TP4 lazy SimpleCPU offload / two-node TP8 resident, the -nightly-native recipe files), which supersedes main's PR #2807 tuning of the same entry. Main's disagg-mtp rewrite and the new Qwen3.5 GB300 Pareto entries are retained unchanged. - perf-changelog: keep both sides; main's #2812/#2807/#2623 entries stay in place and this branch's #2925 entry is appended at the end. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using default effort 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 6aa4605. Configure here.
合并 main 并解决变更日志冲突。
恢复仅在评测模式下等待 AgentX 端点。
4 participants
Description / 描述
Update the MiniMax-M3 GB200 aggregate configuration with single-node TP4 resident, single-node TP4 lazy SimpleCPU KV offload, and two-node TP8 resident variants.
Pin the serving stack to
vllm/vllm-openai:nightly-9ea8f3ffc354901b740f0b31988900897b7221d7, Dynamo1.5.0.dev20260908, and NVIDIA/srt-slurmd50ee7280c33d469df8708e363e23be2456e94fb.Use worker-native
minimax_m3tool and reasoning parsing, EAGLE3 with FlashInfer, andFULL_AND_PIECEWISECUDA graphs.Wait for the OpenAI chat route before AgentX trace replay and explicitly enable chat templating in all three speculative-decoding recipes.
更新 MiniMax-M3 GB200 聚合配置，加入单节点 TP4 常驻、单节点 TP4 延迟 SimpleCPU KV 卸载和双节点 TP8 常驻三种部署方式。
将服务栈固定为
vllm/vllm-openai:nightly-9ea8f3ffc354901b740f0b31988900897b7221d7、Dynamo1.5.0.dev20260908和 NVIDIA/srt-slurmd50ee7280c33d469df8708e363e23be2456e94fb。使用 worker 原生的
minimax_m3工具与推理解析器、基于 FlashInfer 的 EAGLE3，以及FULL_AND_PIECEWISECUDA 图模式。在 AgentX 轨迹回放前等待 OpenAI chat 路由就绪，并在三个推测解码配方中显式启用聊天模板。
Validation / 验证
srtctl dry-runat the pinned revision / 三个配方均在固定版本上通过srtctl dry-runType of change / 变更类型
Note
Low Risk
Benchmark and CI configuration only—pinned images, recipes, and launcher clone logic—with no application auth or production serving code changes.
Overview
Refreshes MiniMax-M3 GB200 aggregate AgentX by adding three new srt-slurm recipes (single-node TP4 GPU-resident, single-node TP4 with lazy SimpleCPU KV offload, two-node TP8 resident) and rewiring
minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtpinnvidia-master.yamlto use them.The stack moves from vLLM v0.27.1 to a pinned vLLM nightly, Dynamo 1.5.0.dev20260908, and srt-slurm commit
d50ee728…(GB200 launcher checks out that SHA instead of v1.0.50). Recipes enable Dynamo KV routing, workerminimax_m3reasoning/tool parsers, EAGLE3 speculative decoding, FULL_AND_PIECEWISE CUDA graphs,use_chat_template: true, and synthetic acceptance env vars; the simple-offload variant addskv-transfer-config/VLLM_USE_SIMPLE_KV_OFFLOAD.Search space now includes no-offload TP4 and two-node TP8 points, adjusts dram-utilization and concurrency for the vLLM-simple offload arm, and points
CONFIG_FILEat the new YAML paths.perf-changelog.yamldocuments the change.Reviewed by Cursor Bugbot for commit 2b0b640. Bugbot is set up for automated code reviews on this repo. Configure here.