Repository navigation
[Klaud Cold] Update Qwen3.5 FP8 B300 SGLang AgentX MTP image to v0.5.22-cu130 / [Klaud Cold] 将 Qwen3.5 FP8 B300 SGLang AgentX MTP 镜像更新至 v0.5.22-cu130 - #4482
Conversation
Pin lmsysorg/sglang:v0.5.22-cu130 by digest and rename the removed --cuda-graph-max-bs alias to --cuda-graph-max-bs-decode. 将 Qwen3.5 FP8 B300 SGLang AgentX MTP 镜像更新至 v0.5.22-cu130（按 digest 固定）， 并将已移除的 --cuda-graph-max-bs 别名改为 --cuda-graph-max-bs-decode。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Thanks for the contribution!
中文感谢你的贡献！
|
|
Initial attempt · Passed · Run 40159476208 / attempt 1 · 2026-10-25 08:19 UTC
Next: Smoke and gsm8k passed with zero request errors; append the perf-changelog entry and run the full-sweep-fail-fast final sweep. 中文初次尝试 · 已通过 · Run 40159476208 / attempt 1 · 2026-10-25 08:19 UTC |
|
parked · Session 1 is waiting on 40159476208. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 1 个会话正在等待 40159476208。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。 |
|
resumed · The session ended before finishing; session 2 continues it in a new run of its planner. 中文resumed · 会话未完成即已结束；第 2 个会话将在其规划运行的新运行中继续。 |
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Klaud-Cold
added
the
full-sweep-fail-fast
|
Final full sweep · Running · Run 40176215702 / attempt 1 · 2026-10-25 08:21 UTC Next: Wait for the full-sweep-fail-fast final sweep covering all 34 baseline AgentX points and the c72 gsm8k eval. 中文最终完整 sweep · 运行中 · Run 40176215702 / attempt 1 · 2026-10-25 08:21 UTC |
|
parked · Session 2 is waiting on 40176215702. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 2 个会话正在等待 40176215702。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。 |
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40176215702 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40176215702 |
|
resumed · The session ended before finishing; session 3 continues it in a new run of its planner. 中文resumed · 会话未完成即已结束；第 3 个会话将在其规划运行的新运行中继续。 |
|
Final full sweep · Failed · Run 40176215702 / attempt 2 · 2026-10-25 20:15 UTC Next: Rerun the failed and cancelled jobs on the same head as infrastructure retry 1, reusing the 26 passed benchmarks and the gsm8k eval. 中文最终完整 sweep · 失败 · Run 40176215702 / attempt 2 · 2026-10-25 20:15 UTC |
|
Infrastructure retry 1/2 · Queued · Run 40176215702 / attempt 3 · 2026-10-25 20:15 UTC Next: Wait for the 8 rerun jobs; a passing rerun completes all 34 baseline AgentX points and the gsm8k eval for finish. 中文基础设施重试 1/2 · 排队中 · Run 40176215702 / attempt 3 · 2026-10-25 20:15 UTC |
|
parked · Session 3 is waiting on 40176215702. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 3 个会话正在等待 40176215702。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。 |
github-actions
Bot
added
the
klaud-parked
|
Mean latency
|
|
Mean latency
|
|
Final full sweep · Passed · Run 40176215702 / attempt 4 · 2026-10-27 14:26 UTC Next: Mark ready for maintainer review. Results continue in numbered report comments. 中文最终完整 sweep · 已通过 · Run 40176215702 / attempt 4 · 2026-10-27 14:26 UTC |
|
validated · Repairs: unknown · Runs: 40159476208, 40176195205, 40176215702 中文validated · 修复次数：未知 · 运行：40159476208, 40176195205, 40176215702 |
github-actions
Bot
added
klaud-finished
![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. Both changed config paths have NVIDIA CODEOWNERS, so their sign-off is still needed. One more check for a human: the MTP draft precision under v0.5.22 can only be confirmed by running the image.
What was reviewed:
- Master config: the image moves to the same digest-pinned
v0.5.22-cu130tag that other entries in nvidia-master.yaml already use. - Recipe: all 34
cuda-graph-max-bskeys were renamed tocuda-graph-max-bs-decodewith the same values, and no old keys are left. The new key name matches existing SGLang recipes. - perf-changelog.yaml: the new entry is only appended at the end, and it covers the affected config key.
Extended reasoning...
The change touches a single NVIDIA master-config entry, the matching single-node B300 FP8 MTP agentic recipe and the perf changelog. It touches nothing security-sensitive. The edits are mechanical: an image bump to a digest already used in that file, a key rename with values unchanged, and an entry appended to the changelog. The deciding factor is ownership: CODEOWNERS assigns nvidia-master.yaml and b300-* srt-slurm recipes to NVIDIA owners. Also, the repo conventions require a reviewer to check the effective MTP draft precision whenever the image is bumped.
Goal: Update SGLang image from lmsysorg/sglang:v0.5.16-cu130 to lmsysorg/sglang:v0.5.22-cu130.
Baseline: 2026-08-07 ·
lmsysorg/sglang:v0.5.16-cu130Mean latency · Sources: API 1, API 2, API 3, API 4
Stop: Close this PR at any time to stop this candidate now; Klaud cancels its runs and deletes its branch.
中文
**目标：**将 SGLang 镜像从 lmsysorg/sglang:v0.5.16-cu130 更新至 lmsysorg/sglang:v0.5.22-cu130。
**基线：**2026-08-07 ·
lmsysorg/sglang:v0.5.16-cu130平均延迟 · 来源： API 1, API 2, API 3, API 4；数值及异常说明见上表。
**停止：**随时关闭此 PR 即可立即停止该候选：Klaud 会取消其运行并删除其分支。