Skip to content

[Klaud Cold] Update dsv4-fp8-h200-dynamo-sglang-agentic-agg SGLang image to v0.5.20-cu130 / 将 dsv4-fp8-h200-dynamo-sglang-agentic-agg 的 SGLang 镜像更新至 v0.5.20-cu130 - #3281

Closed
adibarra wants to merge 1 commit into
mainfrom
klaud/auto-0cb636efc3513aaa-b5b6c510701dbeb8
Closed

adibarra wants to merge 1 commit into
mainfrom
klaud/auto-0cb636efc3513aaa-b5b6c510701dbeb8

Conversation

@adibarra

Copy link
Copy Markdown
Collaborator

No description provided.

…ng image to v0.5.20-cu130

Update the DeepSeek-V4-Pro-0813 FP8 H200 Dynamo+SGLang AgentX aggregate
family from lmsysorg/sglang:nightly-dev-20260914-4358a161 to the
lmsysorg/sglang:v0.5.20-cu130 release. The master image and the recipe's
identity.container.image move together; the dynamo-sglang container alias,
topology, speculation, HiCache and benchmark settings are unchanged.

将 DeepSeek-V4-Pro-0813 FP8 H200 Dynamo+SGLang AgentX 聚合配置族的镜像从
lmsysorg/sglang:nightly-dev-20260914-4358a161 更新为
lmsysorg/sglang:v0.5.20-cu130 正式版。master 镜像与配方中的
identity.container.image 同步更新;dynamo-sglang 容器别名、拓扑、投机解码、
HiCache 及基准测试设置保持不变。

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor

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

Copy link
Copy Markdown
Collaborator Author

Initial attempt 0/5 · deferred (readiness-blocked) · no run dispatched · 2026-09-19T01:55Z
lmsysorg/sglang:v0.5.20-cu130 · 78aed7418ef6fe5901842eff30c1c7737760d622 · AgentX · TP8/EP1
Change: Update the master image and the recipe identity.container.image from lmsysorg/sglang:nightly-dev-20260914-4358a161 (build commit 4358a161) to the v0.5.20 release image (build commit 94602c9c, digest sha256:06e4f2ed21afde4ff513cda65070124e727ba23ccaeff7712b8c40e1097d611f, sgl-kernel 0.4.6.post1 → 0.4.7, CUDA 13.0.3 unchanged); the old commit is a direct ancestor (compare, 94 commits), server_args.py is unchanged, every recipe env var still exists, and the DSV4 changes are opt-in fp8 unified-KV paths plus HiCache/unified-radix-tree fixes. The Dynamo wheel 1.5.0.dev20260910 installs with --no-deps, so it does not alter the bundled SGLang; its compatibility with the newer SGLang internals was not verified at runtime.

No smoke or sweep was dispatched. Blockers:

  1. The public baseline could not be frozen: GET /api/v1/workflow-info?date=2026-09-16 is served content-encoding: gzip from an edge-cache hit regardless of Accept-Encoding, and the current prepare-baseline helper cannot decode it (34 retries over 60 minutes, invalid-json). Cache misses for other dates return plain JSON. PR fix(klaud): harden candidate dispatch and recovery / 加固 Klaud 候选调度与恢复 #3235 adds gzip handling to the shared reader.
  2. This session's GitHub token is not the Klaud-Cold account, so the PR author does not match and the lifecycle finish command rejects ownership; no verified receipt can be produced from this session.

Cleanup: PR closed and branch klaud/auto-0cb636efc3513aaa-b5b6c510701dbeb8 deleted so a later autosweep can retry once the cache entry expires or #3235 lands. The family claim ref was left to the lifecycle helper.

Next: Retry the candidate after the workflow-info cache entry clears or the helper decodes gzip.

中文

初始尝试 0/5 · 已推迟(readiness-blocked) · 未调度任何运行 · 2026-09-19T01:55Z
lmsysorg/sglang:v0.5.20-cu130 · 78aed7418ef6fe5901842eff30c1c7737760d622 · AgentX · TP8/EP1
**变更:**将 master 镜像及配方中的 identity.container.imagelmsysorg/sglang:nightly-dev-20260914-4358a161 更新为 v0.5.20 正式版镜像(构建提交 94602c9c,旧提交为其直接祖先,共 94 个提交);server_args.py 未变化,配方使用的环境变量均存在,DSV4 相关改动为可选的 fp8 unified-KV 路径及 HiCache/unified radix tree 修复。Dynamo wheel 以 --no-deps 安装,不会替换镜像内的 SGLang,但其与新版 SGLang 内部接口的兼容性未经运行时验证。
阻塞原因:(1) 公共 API workflow-info?date=2026-09-16 在边缘缓存命中时始终返回 gzip 编码,当前 prepare-baseline 助手无法解码(60 分钟内重试 34 次均失败),因此无法冻结基线;PR #3235 将为共享读取器加入 gzip 支持。(2) 本次会话的 GitHub 令牌并非 Klaud-Cold 账号,PR 作者不匹配,生命周期 finish 命令拒绝确认归属,无法生成已验证回执。
**清理:**已关闭 PR 并删除分支 klaud/auto-0cb636efc3513aaa-b5b6c510701dbeb8,以便后续自动扫描重试;配置族 claim 引用交由生命周期助手处理。
**下一步:**待缓存条目过期或助手支持 gzip 后重试该候选。

@adibarra adibarra closed this Sep 19, 2026
@adibarra
adibarra deleted the klaud/auto-0cb636efc3513aaa-b5b6c510701dbeb8 branch September 19, 2026 01:49
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

2 participants