[NV] minimaxm3-fp4-b200-dynamo-vllm-mtp - #2431
Conversation
Port the B300 MTP EAGLE3 recipe set (PR #2182) to B200: - Six srt-slurm recipe YAMLs under b200-fp4/8k1k/mtp/ (gpu_type: b200) - minimaxm3-fp4-b200-dynamo-vllm-mtp master-config entry (runner: b200-multinode) - perf-changelog entry to trigger the sweep 中文：将 B300 MTP EAGLE3 配方集（PR #2182）移植到 B200： - 在 b200-fp4/8k1k/mtp/ 下新增六个 srt-slurm 配方 YAML（gpu_type: b200） - 新增 minimaxm3-fp4-b200-dynamo-vllm-mtp 主配置条目（runner: b200-multinode） - 新增 perf-changelog 条目以触发扫描 Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
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 |
…-mtp 中文：补全 minimaxm3-fp4-b200-dynamo-vllm-mtp 的 perf-changelog pr-link Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
xinli-sw
changed the title
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
LGTM — this is a mechanical B300→B200 port of the MiniMax M3 NVFP4 MTP/EAGLE3 recipe set, following the established day-zero porting pattern used for prior SKU ports (e.g. Kimi K2.6, Qwen3.5).
What was reviewed: verified all 6 new recipe YAMLs differ from their B300 counterparts only in gpu_type/container/name plus a consistently-applied VLLM_FLASHINFER_ALLREDUCE_BACKEND env var; confirmed the new nvidia-master.yaml entry's CONFIG_FILE paths and conc-list values line up 1:1 with the added recipe files' concurrency lists; confirmed no launcher changes were needed per the existing cp -rT tree-copy behavior. The bug-hunting system flagged that the B200 port standardizes on the image/env pairing used by only one of the six B300 topologies (the -legacy-dep4 variant) rather than mixing in the newer image used by the other five — verifiers ruled this out as a real bug, and each new recipe is internally self-consistent, so I'm treating it as an intentional day-zero choice that CI's full-sweep-fail-fast run will validate.
Extended reasoning...
Overview
This PR ports the existing MiniMax M3 NVFP4 B300 Dynamo-vLLM disaggregated EAGLE3/MTP recipe set to B200: 6 new srt-slurm recipe YAMLs under benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/, one new minimaxm3-fp4-b200-dynamo-vllm-mtp entry in configs/nvidia-master.yaml, and a perf-changelog.yaml entry. No launcher code changes. This is a pure-config/benchmark-recipe PR with no application or inference-server code touched.
Security risks
None. These are YAML benchmark/config files consumed by the internal sweep launcher; there is no user input, auth, or executable code path introduced.
Level of scrutiny
Low-to-moderate. This is a config-only day-zero hardware port following an established, previously-merged pattern (same shape as PR #2360 Kimi K2.6 B200 port and others). Functional correctness of the recipe (does it actually boot/serve on B200) is gated by the full-sweep-fail-fast label already applied, so CI itself is the real validation gate, not static review.
Other factors
I cross-checked the new nvidia-master.yaml search-space entries against the 6 new recipe files: the CONFIG_FILE paths and conc-list/concurrencies values match 1:1, and each recipe only differs from its B300 source by gpu_type, container, name, and a consistently-added VLLM_FLASHINFER_ALLREDUCE_BACKEND env var. I also independently examined the candidate issue about the B200 port using a single image/env pairing (matching only one of the six B300 topologies) across all six new topologies — this is real (confirmed by diffing the B300 source files directly), but it is applied uniformly and self-consistently across all 6 new files rather than being a mismatched or partial port, and the bug-hunting system's verifiers already concluded it isn't a functional bug. No outstanding review comments exist on this PR to address.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30578251019 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30578251019 |
中文：将 B200 fp4 MTP 镜像更新为 vllm/vllm-openai:v0.26.0 Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30590756683 |
…2380053bff4de 中文：将 B200 fp4 MTP 镜像更新为 nightly-6f91edf96d3f3272945809c04702380053bff4de Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
15db419 to
8591db2
Compare
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30603194145 |
8591db2 to
9aa51d9
Compare
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30635925754 |
…00-dynamo-vllm-mtp-dayzero
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30685134115 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30685850055 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30699032832 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30699373728 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30729399693 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30729630905 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30730440421 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30731060643 |
…00-dynamo-vllm-mtp-dayzero
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30733028640 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30733231070 |
1 participant
No description provided.