Repository navigation
[Klaud Cold] Update Qwen3.5 FP8 MI300X SGLang AgentX MTP image to v0.5.22-rocm720-mi30x-20261024 / [Klaud Cold] 将 Qwen3.5 FP8 MI300X SGLang AgentX MTP 镜像更新至 v0.5.22-rocm720-mi30x-20261024 - #4479
Conversation
…rocm720-mi30x-20261024 Pin lmsysorg/sglang-rocm:v0.5.22-rocm720-mi30x-20261024 by digest and rename --cuda-graph-max-bs to --cuda-graph-max-bs-decode; upstream removed the deprecated alias (sgl-project/sglang#38375), which mapped to the same field. 将 Qwen3.5 MI300X FP8 SGLang AgentX MTP 镜像更新至 lmsysorg/sglang-rocm:v0.5.22-rocm720-mi30x-20261024（按 digest 固定），并将 --cuda-graph-max-bs 改为 --cuda-graph-max-bs-decode；上游已移除该弃用别名 （sgl-project/sglang#38375），两者对应同一字段。 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 40150038697 / attempt 1 · 2026-10-25 03:34 UTC
Next: Smoke benchmarks and the gsm8k eval passed with zero request errors; append the perf-changelog entry and dispatch the full-sweep-fail-fast final sweep. 中文初次尝试 · 已通过 · Run 40150038697 / attempt 1 · 2026-10-25 03:34 UTC |
|
parked · Session 1 is waiting on 40150038697. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 1 个会话正在等待 40150038697。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。 |
… MTP image update Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Klaud-Cold
added
the
full-sweep-fail-fast
|
Final full sweep · Passed · Run 40158818912 / attempt 1 · 2026-10-25 06:40 UTC
Next: Mark ready for maintainer review. 中文最终完整 sweep · 已通过 · Run 40158818912 / attempt 1 · 2026-10-25 06:40 UTC |
|
parked · Session 2 is waiting on 40158818912. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 2 个会话正在等待 40158818912。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=40158818912 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40158818912 |
|
validated · Repairs: unknown · Runs: 40150038697, 40158785046, 40158818912 中文validated · 修复次数：未知 · 运行：40150038697, 40158785046, 40158818912 |
|
Claude finished @Klaud-Cold's task in 20s —— View job LGTM - no blocking issues found Note for maintainers: the final sweep is green, but every point regressed against the 2026-08-12 baseline. Output tok/s/GPU is down 5–12% and TPOT is up 6–12%. gsm8k accuracy is unchanged. |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Goal: Update SGLang image from lmsysorg/sglang-rocm:v0.5.16-rocm720-mi30x-20260730 to lmsysorg/sglang-rocm:v0.5.22-rocm720-mi30x-20261024.
Baseline: 2026-08-12 ·
lmsysorg/sglang-rocm:v0.5.16-rocm720-mi30x-20260730Mean 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-rocm:v0.5.16-rocm720-mi30x-20260730 更新至 lmsysorg/sglang-rocm:v0.5.22-rocm720-mi30x-20261024。
**基线：**2026-08-12 ·
lmsysorg/sglang-rocm:v0.5.16-rocm720-mi30x-20260730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。