[Klaud Cold] Update qwen3.5-fp8-b300-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260907-30705c00 and move to cluster:b300-dsxe - #2881
Conversation
…to nightly-dev-cu13-20260907-30705c00 and move to cluster:b300-dsxe Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
de9c9ce to
847e9dd
Compare
|
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 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
I reviewed this PR and found no bugs. Because the runner change means the whole AgentX curve reruns on the newly added cluster:b300-dsxe fleet rather than an append-only bump, a human look at the results would still be worthwhile.
- Confirmed cluster:b300-nv's labels were already dropped and cluster:b300-dsxe exists in configs/runners.yaml (#2826), matching the retirement claim.
- perf-changelog.yaml appends correctly at the tail with the right config-keys/scenario-type shape.
- pr-link is still the literal placeholder string, not yet the real PR URL the changelog gate requires.
- Could not confirm from this branch that the sibling B200/H200/H100 Qwen3.5 SGLang AgentX-mtp recipes are already on the claimed nightly-dev-cu13-20260907-30705c00 tag; they still show v0.5.16-cu130 (or an older nightly for the H200 hicache variant) in nvidia-master.yaml here.
Extended reasoning...
Overview
The diff touches exactly two files: configs/nvidia-master.yaml (bumps the qwen3.5-fp8-b300-sglang-agentic-mtp recipe's SGLang image tag and moves its runner from the retired cluster:b300-nv to cluster:b300-dsxe) and perf-changelog.yaml (a single tail-appended entry documenting both changes). No source code, workflow logic, or security-sensitive paths are touched — this is a pure recipe/config edit.
Security risks
None. No secrets, auth, or code-execution paths are touched; this is a YAML data change consumed by the existing sweep/validation tooling.
Level of scrutiny
Low-to-moderate. The change is small and mechanical (image tag + runner label + changelog entry), matching the repo's well-established "Klaud Cold" image-bump pattern. However, two things push this past a rubber-stamp approval: (1) the runner change is explicitly non-append-only per the changelog text itself — the entire curve reruns on cluster:b300-dsxe, a fleet added only days ago (commit ee8f457, #2826) — so there is real behavioral risk beyond a typical append-only point addition; (2) the changelog's pr-link is still a literal placeholder rather than a real PR URL, which is expected to be self-corrected by the standard Klaud Cold post-PR-create amend step but is nonetheless an open item at the time of this review.
Other factors
I independently verified: cluster:b300-dsxe exists in configs/runners.yaml and cluster:b300-nv's runner labels were already removed in #2826 (12 other recipes in nvidia-master.yaml are still pinned to the now-labelless cluster:b300-nv, which commit ee8f457's own message acknowledges as a known, separately-tracked issue — not something this PR needs to fix). I also spot-checked the changelog's claim that sibling B200/H200/H100 Qwen3.5 SGLang AgentX-mtp recipes already moved to nightly-dev-cu13-20260907-30705c00: in this branch's snapshot of nvidia-master.yaml, qwen3.5-fp8-b200-sglang-agentic-mtp, qwen3.5-fp4-b200-sglang-agentic-mtp, qwen3.5-fp8-h200-sglang-agentic-mtp, and qwen3.5-fp8-h100-sglang-agentic-mtp are still on lmsysorg/sglang:v0.5.16-cu130, and qwen3.5-fp8-h200-sglang-agentic-hicache-mtp is on a different, older nightly (nightly-dev-cu13-20260815-a5ba081f). This could simply reflect that PRs #2861/#2862/#2868/#2869 haven't landed in this local checkout yet, but I could not confirm that from the available history, so I'm flagging it as something worth a human's attention rather than treating it as a confirmed defect (it echoes a candidate issue an earlier investigation round already looked at and ruled out). No bugs were reported by the automated hunt, and the change is small enough that I lean toward defer rather than a blocking finding, but the non-append-only runner move plus the open pr-link placeholder are enough that I don't have the high confidence required to approve outright.
This review covers commit de9c9ce, which is no longer the latest commit on this pull request; later commits are not covered by it.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34177489662 |
1 similar comment
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34177489662 |
1 participant
Summary
Update the SGLang image for
qwen3.5-fp8-b300-sglang-agentic-mtpfromlmsysorg/sglang:v0.5.16-cu130tolmsysorg/sglang:nightly-dev-cu13-20260907-30705c00, and move the recipe from the retiredcluster:b300-nvrunner tocluster:b300-dsxe.sha256:19b8fa1223cc339c1eae7a5b703f1a8c2543b5b119155bf3d7efaef18f77f007; same tag as the other Qwen3.5 SGLang AgentX bumps ([Klaud Cold] Update qwen3.5-fp8-b200-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260907-30705c00 #2861, [Klaud Cold] Update qwen3.5-fp4-b200-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260907-30705c00 #2862, [Klaud Cold] Update qwen3.5-fp8-h200-sglang-agentic-mtp (+hicache-mtp) SGLang image to nightly-dev-cu13-20260907-30705c00 #2868, [Klaud Cold] Update qwen3.5-fp8-h100-sglang-agentic-mtp SGLang image to nightly-dev-cu13-20260907-30705c00 #2869).qwen3.5-fp8-b300-sglang-agentic-power-ab, a controlled FP8/FP4 power matrix pinned to one SGLang build across both precisions.cluster:b300-nvwas retired in feat(runners): add the B300 DSXE cluster and retire the B300 NV launcher / 添加 B300 DSXE 集群并下线 B300 NV 启动脚本 #2826 (launcher and runner labels removed), so a sweep on it can never be scheduled. Repointed tocluster:b300-dsxe, the same move [Klaud Cold] glm5.2-fp4-b300-sglang-agentic-mtp: extend the AgentX curve to concurrency 48 and move to cluster:b300-dsxe / 将 GLM-5.2 B300 SGLang AgentX 曲线扩展至并发度 48 并迁移至 cluster:b300-dsxe #2829 makes for the GLM-5.2 FP4 B300 sibling. Because the runner changed this is not an append-only bump; the whole curve reruns on DSXE.Recipes touched:
qwen3.5-fp8-b300-sglang-agentic-mtpTest plan
🤖 Generated with Claude Code
Note
Low Risk
Benchmark recipe and changelog only; no application runtime or security paths. Main operational note is non-comparable perf history across the fleet move.
Overview
Updates
qwen3.5-fp8-b300-sglang-agentic-mtpinnvidia-master.yaml: SGLang moves fromv0.5.16-cu130tonightly-dev-cu13-20260907-30705c00(aligned with other Qwen3.5 SGLang AgentX image bumps), and the runner switches from retiredcluster:b300-nvtocluster:b300-dsxeso sweeps can schedule again.Adds a matching
perf-changelog.yamlentry noting unchanged TP/EP grids and script settings, thatqwen3.5-fp8-b300-sglang-agentic-power-abstays pinned for cross-precision parity, and that the runner change forces a full curve rerun on DSXE (not append-only).Reviewed by Cursor Bugbot for commit 847e9dd. Bugbot is set up for automated code reviews on this repo. Configure here.