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

[RohitNagraj]

Summary

Adds the qwen3.5-fp8-gb300-dynamo-sglang config: Qwen3.5-397B-A17B-FP8 disaggregated multinode SGLang benchmarks via Dynamo on GB300, mirroring the gb200-fp8 recipe set.

• 6 topologies across 1k/1k and 8k/1k: 1P1D TP4 STP + wide-EP (DEP4 prefill / DEP16 decode) from 1P1D up to 8P1D, recipes under benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/
• Image: lmsysorg/sglang:nightly-dev-cu13-20260608-303757cc
• Launch script: qwen3.5 fp8 runs against srt-slurm main (the hash-pinned dynamo source install needs its cargo/maturin bootstrap; same branch the gb200-fp8 recipes use) with --no-preflight for the compute-node-local model path, and overlays a header-based NVSHMEM resolver for the DeepEP kNumMaxTopK=16 rebuild (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 DeepEP kNumMaxTopK=16 重建覆盖基于头文件的 NVSHMEM 解析脚本。

[github-actions]

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 As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

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

[github-actions]

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 As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

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 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

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

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29040316658
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29040316658

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29042476471
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29042476471

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29042476471
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=29042476471

[RohitNagraj]

/reuse-sweep-run

[Ankur-singh]

As a PR reviewer and CODEOWNER, I have reviewed this and have:

• Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
• Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
• Verified that this PR has passed PR validation. Please link to GitHub Action workflow that shows this.
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this.
• Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
• Verified that the model architecture isn't changed with benchmark hacks like using --hf-overrides to skipping indexer for every x layers on models that don't natively support this. As a general rule, we won't accept optimizations that reduces the number of model architecture FLOPs. Anything that makes that same computation run faster is fair game; FLOPs at lower precisions is fine, given that the config passes private evals. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
• If an company claims that they support vLLM/SGLang as first class LLM inference engines on their hardware, I have verified that the respective vLLM submission made using upstream https://hub.docker.com/u/vllm docker repo, upstream SGLang https://hub.docker.com/u/lmsysorg docker repo. The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet as supported by vLLM/SGLang community maintainers
• If an company claims that they support vLLM/SGLang as first class upstream in-tree LLM inference engines on their hardware, I have have verified that the respective vLLM/SGLang submission has been made before additional frameworks (TRT-LLM, ATOM, etc.). The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet.
• Verified that the single-node recipes are similar to the official vLLM recipes and/or theSGLang cookbook:
  • If they are not, I have verified that a PR has been opened in vLLM recipe repo or SGLang repo and linked it below in the additional detail section:
• If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.

Additional detail section:

• PR validation: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/29042476471 / PR 验证：https://github.com/SemiAnalysisAI/InferenceX/actions/runs/29042476471
• Evals: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/29042476471 / 评估：https://github.com/SemiAnalysisAI/InferenceX/actions/runs/29042476471
• This is a Dis-agg submission, no recipe update required. / 这是分离式（Dis-agg）提交，无需更新单节点 recipe。
• This submission is STP-only; the speculative-decoding chat-template requirement is not applicable. / 本提交仅使用单 token 预测（STP），投机解码的 chat template 要求不适用。

Signed: ankur-singh

[Klaud-Cold]

✅✅✅ Verdict: PASS ✅✅✅

✅ Check 0 (CODEOWNER): PASS — Ankur-singh is a listed owner of configs/nvidia-master.yaml; the remaining changed paths fall under the * catch-all.
✅ Check 1 (sweep on in-PR commit): PASS — head commit a4e55c6a has green, executed multi-node 1k1k/8k1k (6/6) and multi-node eval (3/3) jobs in run 29042476471; single-node *//eval / skipped only because the PR adds no single-node configs.
✅ Check 2 (evals pass): PASS — GSM8K em_strict 0.9682–0.9697 (n_eff 1319 each, 3 topologies) vs the qwen3.5 bar of 0.94 in utils/evals/thresholds.json, run on the PR's exact image lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928.
➖ Check 3 (recipe link): N/A — disaggregated/multi-node submission (benchmarks/multi_node/**, multinode: true, disagg: true, framework: dynamo-sglang); the recipe-link requirement applies to single-node recipes only.
✅ Check 4 (reuse command): PASS — /reuse-sweep-run posted by RohitNagraj (COLLABORATOR).
✅ Check 5 (latest template): PASS — all current-template items present; the one unchecked item (single-node recipe link) is explained in the additional detail section.
✅ Check 6 (upstream image / engine-first): PASS — upstream lmsysorg/sglang image on GB300; the entry's engine is upstream SGLang (Dynamo is the disagg orchestrator, not a different engine), and qwen3.5-fp4-gb300-dynamo-sglang already covers this model+SKU.
✅ Check 7 (no arch hacks): PASS — no --hf-overrides/model-config edits; the DeepEP kNumMaxTopK 8→16 rebuild raises kernel buffer capacity to fit Qwen3.5's native top-10 expert routing (no FLOPs removed).
➖ Check 8 (spec-decode chat template): N/A — all lanes are spec-decoding: "none" (STP-only); no speculative flags in the recipes.