Add DSV4 B200 disaggregated Dynamo SGLang MTP configuration / 新增 DSV4 B200 分离式 Dynamo SGLang MTP 配置 - #2252
Add DSV4 B200 disaggregated Dynamo SGLang MTP configuration / 新增 DSV4 B200 分离式 Dynamo SGLang MTP 配置#2252RohitNagraj wants to merge 9 commits into
Conversation
Add five recipe-backed 8k/1k MTP topologies and config-scoped B200 launcher handling. 中文：新增 DSV4 FP4 B200 分离式 Dynamo SGLang MTP 配置，加入五个由 recipe 驱动的 8k/1k MTP 拓扑，并将 B200 启动器处理限定到该配置。
|
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 |
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 |
Record the public pull request URL in the final changelog entry. 中文：在性能变更日志的最后一个条目中记录公开拉取请求链接。
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
LGTM — straightforward addition of a new DSV4 NVFP4 B200 dynamo-sglang MTP disaggregated benchmark recipe, following existing patterns in this file.
Extended reasoning...
Overview
This PR adds a new benchmark recipe: DSV4 NVFP4 on B200 using dynamo-sglang with MTP (speculative decoding), disaggregated prefill/decode. It adds five new srt-slurm recipe YAML files (1p1d, 1p6d, and three mid-curve DP/EP topologies), a corresponding entry in configs/nvidia-master.yaml, a perf-changelog.yaml entry, and extends runners/launch_b200-dgxc.sh to support this new framework/model/precision/spec-decoding combination.
Security risks
None. All changes are benchmark configuration (YAML) and CI/runner shell script logic for launching internal SLURM benchmark jobs. No user input handling, auth, or credential logic is touched.
Level of scrutiny
Low-to-moderate. The five new YAML recipe files are pure data (env vars, sglang server flags, resource topology) with no control flow, so they carry no correctness risk beyond being internally consistent, which they are (prefill/decode node counts match the resources block, deepep configs are used consistently with enable-dp-attention, etc.). The one piece of actual logic is the runners/launch_b200-dgxc.sh change, which extends an existing if/elif chain to special-case the new dynamo-sglang plus dsv4 plus fp4 plus mtp combination (model path resolution, a new no-preflight flag, framework validation, and repo checkout/copy). This follows the same structural pattern already used for the adjacent dsr1/dynamo-sglang and dynamo-vllm/dsv4 branches, so it's a low-risk, mechanical extension rather than a new design.
Other factors
The bug-hunting pass found no issues, and my own read of the launcher script diff didn't surface any additional problems — the new PREFLIGHT_ARGS array is initialized unconditionally before use, and the updated multinode-framework-validation conditional correctly permits both the pre-existing dynamo-vllm path and the new dynamo-sglang plus mtp path for dsv4. This matches the shape of many prior benchmark-recipe-addition PRs in this repo that have been approved without human review.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29540123128 |
2f1eeb4 to
8e6ce84
Compare
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29656832531 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29656873340 |
5 similar comments
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29656873340 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29656873340 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29656873340 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29656873340 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29656873340 |
中文：配置 DSV4 B200 SGLang 模型加载、RDMA 传输和固定版本的 uv 引导程序。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30593374709 |
Remove synthetic checkpoint metadata and the loader-isolation override, select the checkpoint-compatible MoE backend, and update the staged checkpoint path while preserving the existing precision, topology, and transport settings.\n\n中文：移除合成的检查点元数据和仅用于隔离加载问题的设置，选择与检查点兼容的 MoE 后端，并更新已暂存的检查点路径，同时保留现有精度、拓扑和传输设置。
RohitNagraj
changed the title
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30663023108 |
Restore the Mooncake registered-memory pool and MNNVL selection for all B200 MTP recipes, remove the forced HCA list, and use the SGLang image containing DSV4 disaggregated MTP state transfer. 中文：为所有 B200 MTP recipe 恢复 Mooncake 注册内存池和 MNNVL 传输选择，移除强制 HCA 列表，并使用包含 DSV4 分离式 MTP 状态传输支持的 SGLang 镜像。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30712963208 |
Replace the unavailable SGLang image tag across the DSV4 B200 master configuration, five recipes, and changelog entry. Synchronize the branch with current main and keep the PR changelog entry at the end. 中文：将 DSV4 B200 主配置、五个 recipe 和变更日志条目中的不可用 SGLang 镜像标签替换为已发布版本。同步当前 main，并将本 PR 的变更日志条目保留在文件末尾。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30949937760 |
中文：将 DSV4 B200 SGLang 配置与最新 main 对齐，并保留已验证的镜像和启动修复。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31068141306 |
Replace the Mooncake MNNVL-specific memory pool settings with the B200 NIXL and UCX transport across all five disaggregated recipes. Refresh the branch from main and keep the PR changelog entry appended at the end. 中文：将五个分离式配方中的 Mooncake MNNVL 专用内存池配置替换为 B200 NIXL 和 UCX 传输。同步最新 main，并保持该 PR 的变更日志条目位于文件末尾。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31211885397 |
2 similar comments
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31211885397 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=31211885397 |
2 participants
Summary
Validation
中文说明
验证