Add Qwen3.5 B300 Dynamo-SGLang AgentX recipes / 添加 Qwen3.5 B300 Dynamo-SGLang AgentX 配方 - #2978
Add Qwen3.5 B300 Dynamo-SGLang AgentX recipes / 添加 Qwen3.5 B300 Dynamo-SGLang AgentX 配方#2978RohitNagraj wants to merge 4 commits into
Conversation
添加 Qwen3.5 NVFP4 B300 Dynamo-SGLang AgentX 配置和配方，并校验合成接受长度与已提交的黄金曲线一致。
|
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 |
在性能变更日志中补充 Qwen3.5 B300 PR #2978 链接。
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.
Beyond what's already flagged inline, I checked whether the launch script's job-name sed override (sed -i "s/^name:.*/.../") would silently no-op on the two new recipes that wrap everything under a top-level base: key (where name: is indented two spaces) — that base:/zip_override_running_cap: matrix format predates this PR (already used by the glm5/glm5.2 recipes), so it's pre-existing behavior rather than a regression introduced here.
Extended reasoning...
This run reported a confirmed finding (perf-changelog.yaml's new entry still has a placeholder pull/XXX pr-link) that will be posted inline, plus additional verified findings that are not being surfaced individually in this review body per the dedup/cap policy. Since verified findings exist, approval is not appropriate. I independently re-examined the candidate concern that two of the four new recipe YAMLs indent name: under a top-level base: key, which would make the launch script's unconditional sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" fail to match and silently leave the recipe's default job name in place. I confirmed via git grep against the pre-PR commit that this base: + zip_override_running_cap: matrix wrapping format already exists in the repository (glm5 and glm5.2 recipes use it), so the same sed limitation already applied before this change — it is not a new regression introduced by this diff, just an existing gap this PR's new recipes happen to also hit. Given the outstanding verified-but-unposted findings, a human should still review before merge.
Additional findings (no inline location):
-
🟡
perf-changelog.yaml— The new changelog entry's pr-link is still the placeholder.../pull/XXXinstead of the real PR URL, unlike every other historical entry which links a real PR. Fix: replace with the actual pull request URL before merge, per the append-only changelog convention that requires a real, verifiable PR link for traceability.Extended reasoning...
perf-changelog.yaml is append-only and byte-sensitive; the new entry at the physical tail (config-keys: qwen3.5-fp4-b300-dynamo-sglang-agentic-pareto) ends with
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX. All prior entries (2925, 2954, 2955, 2969, ...) use real PR numbers. If merged as-is, anyone auditing this recipe/config change later via the changelog gets a dead/placeholder link instead of the actual PR, breaking the traceability the append-only convention exists to guarantee.Verification: nit. The changelog entry added by this diff at line 7149 of perf-changelog.yaml ends with
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX(a placeholder), for config-keysqwen3.5-fp4-b300-dynamo-sglang-agentic-pareto, while adjacent entries such as #2925 (line 7143) use real PR URLs. The finding is factually real. Caveat to the candidate's wording: it claims every other… | nit.…
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34515330805 |
延长 Qwen3.5 AgentX 配方的 Slurm 时间限制。
合并 origin/main，并将本分支的性能变更日志条目重新追加到文件末尾。
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34548026742 |
1 participant
Description
lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729image.draft tokens = steps + 1invariant.中文说明
lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729。Related Issue
N/A
Type of Change
Checklist
perf-changelog.yamland have not edited historical entriesOWNER/MEMBER/COLLABORATOR) has commented/reuse-sweep-runon this PRNote
Medium Risk
Changes benchmark launch paths and fail-closed synthetic acceptance for Qwen3.5, which can block misconfigured runs but also alters throughput vs eval injection behavior on B300 DSXE.
Overview
Adds Qwen3.5 NVFP4 B300 Dynamo-SGLang AgentX coverage via a new
qwen3.5-fp4-b300-dynamo-sglang-agentic-paretoentry innvidia-master.yamland four srt-slurm agentic recipes (TP2+HiCache+MTP with zip batch caps, TP2 c44, TP2EP2 high-throughput, TP4 low-lat), all pinned tolmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729and wired toagentic_srt.shwith synthetic acceptance lengths per concurrency point.Synthetic acceptance is tightened for Qwen3.5:
inject_synthetic_acceptance.pynow resolves AL fromgolden_al_distribution/qwen3.5_mtp.yamlfrom the recipe’sspeculative-num-steps, rejects mismatchedSYNTHETIC_ACCEPTANCE_LENGTH/NUM_SPEC_TOKENS, and the SGLang injector enforcesspeculative-num-draft-tokens = steps + 1with correct nested env injection.launch_b300-dsxe.shalways runs the injector (not onlyEVAL_ONLY) and mapsqwen3.5-fp4toQwen3.5-397B-A17B-NVFP4(V2 kept as a separate alias). Docs,perf-changelog.yaml, and slurm injector tests cover the pareto points and fail-closed behavior.Reviewed by Cursor Bugbot for commit 6b7dcd0. Bugbot is set up for automated code reviews on this repo. Configure here.