feat: add Qwen3.5 FP8 GB200 Dynamo-SGLang MTP recipes / 新增 Qwen3.5 FP8 GB200 Dynamo-SGLang MTP 配方#2324
feat: add Qwen3.5 FP8 GB200 Dynamo-SGLang MTP recipes / 新增 Qwen3.5 FP8 GB200 Dynamo-SGLang MTP 配方#2324RohitNagraj wants to merge 2 commits into
Conversation
Add the GB200 disaggregated configuration, seven SGLang recipes, chat-formatted MTP inputs, and digest-aware Enroot import validation. 中文：新增 Qwen3.5 FP8 GB200 Dynamo-SGLang MTP 配置 添加 GB200 分离式配置、七个 SGLang 配方、MTP 聊天格式输入，以及支持摘要固定镜像的 Enroot 导入校验。
|
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 PR link for the Qwen3.5 GB200 MTP configuration. 中文：更新 #2324 的 perf-changelog 链接 记录 Qwen3.5 GB200 MTP 配置对应的 PR 链接。
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Deferring to human review — no bugs found, but this PR's launcher-script change is broader than a config addition and is worth a maintainer's eyes.
Extended reasoning...
Overview
This PR adds seven new Qwen3.5-FP8 GB200 disaggregated Dynamo-SGLang MTP benchmark recipes (config-only YAML), a corresponding nvidia-master.yaml scenario block, a perf-changelog.yaml entry, and — the part that raises the scope beyond a typical recipe PR — a rewrite of the enroot image-import logic in runners/launch_gb200-nv.sh: a new enroot_uri_for_image() helper that builds enroot URIs for digest-pinned vs. tag-only images across registries, plus a validation/atomic-replace pass (unsquashfs -l check before mv -f) in import_squash().
Security risks
No injection, auth, or data-exposure concerns — the script only builds enroot import URIs from config-controlled image strings and moves local files under a lock. No user-supplied/external input reaches this logic.
Level of scrutiny
The bug hunting system reviewed the new enroot_uri_for_image digest-splitting logic twice (double-colon-after-#, wrong split point) and found no real bug. However, import_squash is called unconditionally on every GB200 launch (import_squash "$SQUASH_FILE" "$IMAGE" / import_squash "$NGINX_SQUASH_FILE" "$NGINX_IMAGE"), so this change affects every existing GB200 recipe (dsr1, dsv4, glm5.x, kimik2.5, minimax, etc.) — not just the new Qwen3.5 recipes. A logic change to a shared, unconditionally-executed launcher path is the kind of critical-path change that benefits from a maintainer's sign-off even when no bug was found, since a subtle regression here would silently affect unrelated model sweeps.
Other factors
The seven new recipe YAMLs and master-config/changelog additions are mechanical and low-risk (new, additive, model-scoped). It's specifically the shared shell-script rewrite that pushes this past the bar for a shadow-approval.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30055331227 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30055331227 |
|
/reuse-sweep-run 30055331227 |
|
As a PR reviewer and CODEOWNER, I have reviewed this and have:
Additional detail section:
Signed: |
✅✅✅ Verdict: PASS ✅✅✅✅ Check 0 (CODEOWNER): PASS — |
3 participants
Summary
中文说明