config(minimaxm3): refresh full GB200 Dynamo-vLLM disaggregated curve / 刷新完整 GB200 Dynamo-vLLM 解耦曲线 - #2807
Conversation
|
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 |
1 similar comment
|
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 |
添加 MiniMax-M3 GB200 Dynamo-vLLM 解耦配方，配置 Mooncake 主机 KV 存储、NIXL UCX 传输和已提交的 thinking-on EAGLE3-GQA 接受长度。
379ca95 to
15cb539
Compare
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33660947747 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.
Looks good — the placeholder pull/XXX link flagged inline is a minor, easily-fixed nit that doesn't affect functional correctness.
Extended reasoning...
Overview
This PR adds three new SLURM benchmark recipe YAMLs for MiniMax-M3 on GB200-FP4 (disaggregated prefill/decode with TP4/TP4, TP4/TP8, and TP4/TP4x2 decode configurations), updates configs/nvidia-master.yaml to reference these new recipes (switching from a pinned image tag to a sha256 digest, bumping the Dynamo router version, and adding three new agentic-coding search-space entries with Mooncake DRAM KV-offloading), and appends a corresponding entry to perf-changelog.yaml.
Security risks
None. This is purely benchmark/config data — no code execution paths, auth, or secrets are touched. The image reference change from a tag to a sha256 digest is actually a security improvement (immutable, verifiable reference).
Level of scrutiny
Low-to-moderate. The changes are declarative YAML config/benchmark recipes with no runtime code logic, and they follow the same structural patterns as existing entries in the same files. The one confirmed issue (a literal pull/XXX placeholder in the new perf-changelog entry, unlike every other entry which has a concrete PR link) is a trivial oversight that is already flagged as an inline comment and easily fixed before merge — it doesn't affect functional behavior.
Other factors
I additionally checked the previously-flagged candidate about max-num-seqs: 1 in the TP8 decode config of disagg-1p1d-tp4-tp8-c1-agentic.yaml. It corresponds to conc-list: [1] in nvidia-master.yaml for that same scenario (this recipe is specifically the concurrency=1 benchmark point), so the hardcoded batch size of 1 is intentional and consistent with the scenario's purpose, not a bug. The other two new recipe files (c24, c8/c16) correctly omit an explicit max-num-seqs override since they target higher concurrency. Given the single confirmed finding is minor and cosmetic, and no other issues were found, I'm approving.
将 MiniMax-M3 GB200 分离式 AgentX 配置与仓库中已提交的 EAGLE3-GQA thinking-on 接受长度保持一致。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33663459461 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33663459461 |
恢复 MiniMax-M3 GB200 解耦 AgentX 的完整配置矩阵，使完整扫描覆盖现有及新增的所有拓扑点。
RohitNagraj
changed the title
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33663459461 |
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 38157cd. Configure here.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33686851838 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=33686851838 |
1 participant
Description
Refresh the complete MiniMax-M3 NVFP4 GB200 Dynamo-vLLM disaggregated AgentX configuration so the full sweep covers every configured topology point. Add three Mooncake host-KV recipes while retaining the DEP8 prefill and DEP4 decode topology on the shared pinned image and Dynamo version. Configure NIXL over UCX, explicit thinking mode, and the committed MiniMax-M3 EAGLE3-GQA acceptance target.
中文说明
刷新完整的 MiniMax-M3 NVFP4 GB200 Dynamo-vLLM 解耦 AgentX 配置，使完整扫描覆盖所有已配置的拓扑点。新增三个 Mooncake 主机 KV 配方，同时在统一固定的镜像和 Dynamo 版本上保留 DEP8 prefill 与 DEP4 decode 拓扑。配置基于 UCX 的 NIXL 传输、显式 thinking 模式，以及仓库中已提交的 MiniMax-M3 EAGLE3-GQA 接受长度目标。
Related Issue
N/A
Type of Change
Checklist
perf-changelog.yamland have not edited historical entriesOWNER/MEMBER/COLLABORATOR) has commented/reuse-sweep-runon this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the sweep label will no longer automatically kick off new sweeps. Remove and re-add the label to force one.