[Klaud Cold] Update glm5.2-fp8-h200-dynamo-sglang-agentic-mtp-agg SGLang image to v0.5.20-cu130 / 将 glm5.2-fp8-h200-dynamo-sglang-agentic-mtp-agg 的 SGLang 镜像更新至 v0.5.20-cu130

[adibarra]

No description provided.

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[adibarra]

Initial update 0/5 · deferred · no run dispatched · 2026-09-18 UTC
lmsysorg/sglang:v0.5.20-cu130@sha256:06e4f2ed21afde4ff513cda65070124e727ba23ccaeff7712b8c40e1097d611f · 315c3c89cdfa68fc9e989c260e548a4ee830b60d · AgentX · TP8/EP1
Change: Replace the 2026-09-14 nightly with the digest-pinned v0.5.20 release; the v0.5.20 tag (94602c9c) is 94 commits ahead of and 0 behind the nightly commit 4358a161, while v0.5.19 predates it by 633 commits and was not used. Every recipe flag and SGLANG_* variable used by the five aggregate recipes is present at both refs with unchanged environ.py defaults, and the HiCache get_device_accessible_ptr fallback (sgl-project/sglang#39516) is included. The amd64 image layers contain ai-dynamo and ai-dynamo-runtime 1.5.0.dev20260917, sglang 0.5.20 and sglang-kernel 0.4.7 (the Dockerfile installs the latest ai-dynamo pre-release at build time), so dynamo.install: false is kept and identity.frameworks was updated to match. Master router.version stays 1.5.0.dev20260913 because it is part of the generated point identity that the frozen baseline must match; the bundled router is 1.5.0.dev20260917.

Blocker: prepare-baseline cannot freeze the 2026-09-18 public baseline for this aggregated multinode family. The five published rows from run 35177419428 carry decode_num_workers: 1 and num_decode_gpu: 8 (ingestion derives them from DECODE_NUM_WORKERS), whereas infx/klaud/reporting.py::public_point projects the generator's aggregated decode.num-worker: 0, so every row is discarded as a different topology and the helper raises "No verified public baseline points for the selected family". A second, separate issue: the CDN serves workflow-info?date=2026-09-18 gzip-encoded even for identity requests, which the helper's urllib reader rejects as invalid JSON; the topology mismatch was reproduced after decompressing that response locally. No GPU work was dispatched and no result was fabricated. Fixing the helper projection for aggregated multinode families is a shared-code change outside Klaud scope.

Next: Maintainer review of the baseline helper; this exact candidate is retained.

中文

初始更新 0/5 · 已推迟 · 未派发任何运行 · 2026-09-18 UTC
lmsysorg/sglang:v0.5.20-cu130@sha256:06e4f2ed21afde4ff513cda65070124e727ba23ccaeff7712b8c40e1097d611f · 315c3c89cdfa68fc9e989c260e548a4ee830b60d · AgentX · TP8/EP1
**变更：**用按摘要固定的 v0.5.20 发布版替换 2026-09-14 nightly；v0.5.20 标签（94602c9c）领先 nightly 提交 4358a161 94 个提交且不落后，而 v0.5.19 落后 633 个提交，故未采用。五个聚合配方使用的全部参数与 SGLANG_* 环境变量在两个版本中均存在且默认值不变，并包含 HiCache 回退修复（sgl-project/sglang#39516）。镜像层内含 ai-dynamo 1.5.0.dev20260917、sglang 0.5.20、sglang-kernel 0.4.7，因此保留 dynamo.install: false 并同步更新 identity.frameworks。主配置 router.version 保持 1.5.0.dev20260913，因为它属于基线必须匹配的生成点标识；镜像内置的路由器为 1.5.0.dev20260917。

阻塞：prepare-baseline 无法为该聚合多节点系列冻结 2026-09-18 的公开基线。运行 35177419428 发布的五行数据带有 decode_num_workers: 1 与 num_decode_gpu: 8，而 infx/klaud/reporting.py::public_point 按生成器的聚合 decode.num-worker: 0 投影，导致所有行被当作不同拓扑丢弃并报 "No verified public baseline points for the selected family"。另一独立问题：CDN 对 workflow-info?date=2026-09-18 即便在 identity 请求下也返回 gzip 编码，助手的 urllib 读取器将其判为无效 JSON；本地解压后复现了上述拓扑不匹配。未派发任何 GPU 工作，也未伪造任何结果。修复聚合多节点系列的助手投影属于共享代码改动，超出 Klaud 范围。

**下一步：**由维护者审查基线助手；保留该候选。

[adibarra]

uncertain · Repairs: 0 · Runs: —
No GPU work was dispatched. The finish lifecycle command could not verify this session because the PR author is not the Klaud-Cold account in this run (candidate ownership mismatch), so cleanup was performed directly: the draft PR is closed and the exact-candidate branch is retained pending maintainer review of the baseline-helper defect described above.

中文

uncertain · 修复次数：0 · 运行：—
未派发任何 GPU 工作。本次运行中 PR 作者并非 Klaud-Cold 账户，finish 生命周期命令无法验证本会话（候选归属不匹配），因此直接完成清理：草稿 PR 已关闭，保留精确候选分支，等待维护者审查上文所述的基线助手缺陷。