Add Qwen3.5 FP8 GB300 disaggregated multinode SGLang benchmarks via Dynamo / 新增 Qwen3.5 FP8 GB300 分离式多节点 SGLang（Dynamo）基准测试 - #2137
Conversation
…ynamo 6 topologies across 1k/1k and 8k/1k: 1P1D TP4 STP plus wide-EP (DEP4 prefill / DEP16 decode) from 1P1D up to 8P1D, mirroring the gb200-fp8 recipe set. Runs qwen3.5 fp8 against srt-slurm main (hash-pinned dynamo source install needs its cargo/maturin bootstrap) with --no-preflight for the compute-node-local model path, and overlays a header-based NVSHMEM resolver for the DeepEP kNumMaxTopK=16 rebuild. 中文：新增 Qwen3.5 FP8 GB300 分离式多节点 SGLang（Dynamo）基准测试。 覆盖 1k/1k 与 8k/1k 共 6 种拓扑：1P1D TP4 STP 及 1P1D 至 8P1D 的宽 EP （预填充 DEP4 / 解码 DEP16），与 gb200-fp8 配方集一致。fp8 使用 srt-slurm main 分支（哈希固定的 dynamo 源码安装需要其 cargo/maturin 引导），对计算节点本地模型路径使用 --no-preflight，并为 DeepEP kNumMaxTopK=16 重建覆盖基于头文件的 NVSHMEM 解析脚本。
|
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 |
中文：更新 perf-changelog 中 #2137 的 pr-link。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29040316658 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Update the pinned container image from nightly-dev-cu13-20260608-303757cc to nightly-dev-cu13-20260709-074bb928 across the master config entry, the six gb300-fp8 recipes, and the perf-changelog entry. 中文：将 qwen3.5-fp8-gb300-dynamo-sglang 的容器镜像从 nightly-dev-cu13-20260608-303757cc 更新为 nightly-dev-cu13-20260709-074bb928，同步更新主配置条目、六个 gb300-fp8 配方及 perf-changelog 条目。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29042476471 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29042476471 |
|
As a PR reviewer and CODEOWNER, I have reviewed this and have:
Additional detail section:
Signed: |
✅✅✅ Verdict: PASS ✅✅✅✅ Check 0 (CODEOWNER): PASS — Ankur-singh is a listed owner of |
…pk=10 The DeepEP snapshot baked into the GB cu13 sglang images caps the low-latency dispatch/combine paths at kNumMaxTopK=11, which already accommodates Qwen3.5's top-10 expert routing. The only lower cap (kNumMaxTopk=9) is in the opt-in overlapped-combine kernel, which these recipes do not enable. Drop setup_script: rebuild-deepep.sh from the wide-EP recipes, the vendored rebuild script, and its launch-script overlay. No other sglang recipe carries a setup script. 中文：移除 DeepEP 重建步骤——GB cu13 sglang 镜像内置的 DeepEP 低延迟 dispatch/combine 路径默认 kNumMaxTopK=11，已满足 Qwen3.5 的 top-10 专家 路由；唯一更低的上限（kNumMaxTopk=9）位于这些配方未启用的 overlapped- combine 内核中。删除宽 EP 配方中的 setup_script、内置重建脚本及启动 脚本中的覆盖逻辑。
…amo-sglang # Conflicts: # perf-changelog.yaml
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29133711332 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29133711332 |
Pin the source-install path to the latest published srt-slurm release instead of the moving main branch. 中文：将 Qwen3.5 FP8 的 srt-slurm 源码安装路径从持续变化的 main 分支固定到最新发布标签 v1.0.25。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29207681289 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29207681289 |
|
/reuse-sweep-run |
|
/reuse-sweep-run |
# Conflicts: # perf-changelog.yaml
|
As a PR reviewer and CODEOWNER, I have reviewed this and have:
Additional detail section:
Signed: |
✅✅✅ Verdict: PASS ✅✅✅✅ Check 0 (CODEOWNER): PASS — Ankur-singh is a listed owner of |
Resolve conflicts from #2137 (Qwen3.5-FP8 GB300 SGLang): - configs/nvidia-master.yaml: keep our dsv4 agentic keys and main's new qwen3.5-fp8-gb300-dynamo-sglang key (both at EOF). - perf-changelog.yaml: keep main's qwen3.5-fp8 entry and our dsv4 entry (appended at EOF); drop the "Prune ... (decode-side MTP only)" description line. - launch_gb300-nv.sh auto-merged (our v1.0.10 block + main's qwen3.5-fp8 routing).
5 participants
Summary
Adds the
qwen3.5-fp8-gb300-dynamo-sglangconfig: Qwen3.5-397B-A17B-FP8 disaggregated multinode SGLang benchmarks via Dynamo on GB300, mirroring the gb200-fp8 recipe set.benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/lmsysorg/sglang:nightly-dev-cu13-20260608-303757ccmain(the hash-pinned dynamo source install needs its cargo/maturin bootstrap; same branch the gb200-fp8 recipes use) with--no-preflightfor the compute-node-local model path, and overlays a header-based NVSHMEM resolver for the DeepEPkNumMaxTopK=16rebuild (Qwen3.5 routes top-10 experts).中文：新增
qwen3.5-fp8-gb300-dynamo-sglang配置——Qwen3.5-397B-A17B-FP8 在 GB300 上的分离式多节点 SGLang（Dynamo）基准测试，与 gb200-fp8 配方集一致。覆盖 1k/1k 与 8k/1k 共 6 种拓扑（1P1D TP4 STP 及 1P1D 至 8P1D 宽 EP）。fp8 使用 srt-slurm main 分支，对计算节点本地模型路径使用--no-preflight，并为 DeepEPkNumMaxTopK=16重建覆盖基于头文件的 NVSHMEM 解析脚本。