Repository navigation
[Klaud Cold] Update GLM-5.2 FP4 B200 Dynamo-SGLang AgentX agg image to v0.5.22-cu130 / [Klaud Cold] 将 GLM-5.2 FP4 B200 Dynamo-SGLang AgentX 聚合镜像更新至 v0.5.22-cu130 - #4370
Conversation
….22-cu130 Pin lmsysorg/sglang:v0.5.22-cu130 by digest, pass the decode graph batch as cuda-graph-max-bs-decode after v0.5.22 removed the cuda-graph-max-bs alias, and pin Dynamo source b4540b87, which counts attention-DP ranks from attn_dp_size as v0.5.22 resolves --enable-dp-attention. 将 glm5.2-fp4-b200-dynamo-sglang-agentic-agg 的 SGLang 镜像更新为按摘要固定的 v0.5.22-cu130；由于 v0.5.22 移除了 cuda-graph-max-bs 别名，改用取值相同的 cuda-graph-max-bs-decode；并将 Dynamo 源码固定为 b4540b87，该版本按 attn_dp_size 统计注意力 DP rank，与 v0.5.22 对 --enable-dp-attention 的解析一致。 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 39601738601 / attempt 1 · 2026-10-21 15:31 UTC
Note: All rows: Δ N/A: dataset/scenario/statistic mismatch.
Next: Smoke throughput for all six points and the c32 GSM8K eval passed; append the changelog entry and run the full sweep with 中文初次尝试 · 已通过 · Run 39601738601 / attempt 1 · 2026-10-21 15:31 UTC |
为 GLM-5.2 B200 FP4 聚合 SGLang v0.5.22 更新添加 perf-changelog 条目。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Klaud-Cold
added
the
full-sweep-fail-fast
|
Final full sweep · Passed · Run 39626904306 / attempt 1 · 2026-10-21 17:55 UTC
Note: All rows: Δ N/A: dataset/scenario/statistic mismatch.
Next: Mark ready for maintainer review. 中文最终完整 sweep · 已通过 · Run 39626904306 / attempt 1 · 2026-10-21 17:55 UTC |
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=39626904306 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=39626904306 |
|
validated · Repairs: 0 · Runs: 39601738601, 39626809045, 39626904306 中文validated · 修复次数：0 · 运行：39601738601, 39626809045, 39626904306 |
|
/use 39626904306 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Additional findings (outside the current diff — GitHub can't attach inline comments there):
-
🔴
inferencex-e2e/configs/nvidia-master.yaml— Published results for this benchmark will still say Dynamo1.5.0.dev20260903, but the job now installs source revb4540b87. The recipe moveddynamo.sourcefromwheel: 1.5.0.dev20260903torev: b4540b8767aaa8d98582251bbe9e2ceaba1ed1b1. The masterrouter:entry at nvidia-master.yaml:7068 was not updated to match. Fix: set therouter.versionofglm5.2-fp4-b200-dynamo-sglang-agentic-aggto"b4540b8767aaa8d98582251bbe9e2ceaba1ed1b1", as the gb200 GLM-5.2 configs already do for rev71eb001e…(nvidia-master.yaml:7199, 7227).Why this was flagged
The PR changes agg-variants.yaml:9 and :18 so Dynamo installs from source commit b4540b8767aaa8d98582251bbe9e2ceaba1ed1b1 instead of the 1.5.0.dev20260903 wheel. The matching master entry still has
router: { name: dynamo-router, version: "1.5.0.dev20260903" }at nvidia-master.yaml:7068. infx/srt_slurm/generate.py:53 serializes that value into ROUTER_METADATA. infx/results/agentic/init.py:201 reads it into each result'srouterfield, so every run of this config reports the wrong Dynamo version. Before the change, the reported version matched the wheel that was installed. AGENTS.md requires the recipe and the master config to be updated together, and the repo follows that elsewhere: the gb200 GLM-5.2 recipes pinned to rev 71eb001e17fa… declare the same SHA as router.version (nvidia-master.yaml:7199, 7227). No validator compares the recipe's dynamo rev with router.version, so nothing catches the mismatch.Verification: Trigger: every run of glm5.2-fp4-b200-dynamo-sglang-agentic-agg after this PR merges. The matching master entry was not updated: nvidia-master.yaml:7068 still reads
router: { name: dynamo-router, version: "1.5.0.dev20260903" }. So every published result for this config will report a Dynamo version that the job no longer installs. Nothing in the binder catches the mismatch.
This branch has not been deployed
1 participant
Goal: Update SGLang image from
lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729tolmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae.Baseline: 2026-10-19 ·
lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729Mean latency · Sources: API 1, API 2, API 3
Stop: Add the
klaud-abortlabel at any time to stop this candidate now; Klaud cancels its runs, closes this PR and does not retry it. Remove the label to let the planner pick it up again.中文
**目标：**将 SGLang 镜像从
lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729更新至lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae。**基线：**2026-10-19 ·
lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729平均延迟 · 来源： API 1, API 2, API 3；数值及异常说明见上表。
**停止：**随时添加
klaud-abort标签即可立即停止该候选：Klaud 会取消其运行、关闭此 PR 且不再重试。移除该标签后，规划器会在下次扫描时重新考虑它。