[Klaud Cold] Update dsv4-fp4-gb300-dynamo-sglang-agentic-disagg SGLang image to v0.5.21-cu130 / [Klaud Cold] 将 dsv4-fp4-gb300-dynamo-sglang-agentic-disagg 的 SGLang 镜像更新至 v0.5.21-cu130

[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:nightly-dev-cu13-20260916-c9a8fba9 to lmsysorg/sglang:v0.5.21-cu130.
Baseline: 2026-10-08 · lmsysorg/sglang:nightly-dev-cu13-20260916-c9a8fba9
Mean latency · Sources: API 1, API 2

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c120 PTP8x1/DTP16x1 544619	26,304.26	205.34	3,806.9	6.63
AgentX c480 PTP8x1/DTP16x1 ddf5ab	85,649.42	657.39	9,050.56	9.41
AgentX c960 PTP8x2/DTP16x1 7034ec	115,825.03	865.59	9,019.85	13.99
AgentX c1440 PTP8x3/DTP16x1 98c007	123,251.46	922.41	9,516.72	18.41
AgentX c2400 PTP8x4/DTP16x1 e959ca	127,047.36	965.45	15,838.86	28.38

Eval: N/A

中文

**目标：**将 SGLang 镜像从 lmsysorg/sglang:nightly-dev-cu13-20260916-c9a8fba9 更新至 lmsysorg/sglang:v0.5.21-cu130。
**基线：**2026-10-08 · lmsysorg/sglang:nightly-dev-cu13-20260916-c9a8fba9
平均延迟 · 来源： API 1, API 2；数值及异常说明见上表。

[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 run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• 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：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **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。

[Klaud-Cold]

Initial attempt · Queued · Run 38697092913 / attempt 1 · 2026-10-15 11:17 UTC
lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 · b301e68225f8 · Mean latency
Change: Image-only bump: v0.5.21 descends from c9a8fba9 (github.com/sgl-project/sglang/compare/c9a8fba9...v0.5.21); all recipe engine flags, torch/sglang-kernel/flashinfer/DeepGEMM/Mooncake 0.3.13/CUDA 13.0.3 pins and the Dynamo wheel's sglang imports are unchanged, and SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT was already inert before its removal (sglang PR 40313); upstream defaults now enable deferred decode KV release and the rust radix core, and the old nightly tag is no longer listed on Docker Hub, so its exact bundled commit is inferred from the tag name.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c120 PTP8x1/DTP16x1 544619	N/A	N/A	N/A
AgentX c960 PTP8x2/DTP16x1 7034ec	N/A	N/A	N/A
AgentX c1440 PTP8x3/DTP16x1 98c007	N/A	N/A	N/A
AgentX c2400 PTP8x4/DTP16x1 e959ca	N/A	N/A	N/A

Note: All rows: pending.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.

Next: All eight smoke jobs on run 38697092913 were still queued for gb300-nv after four hours at background priority, with cluster utilization now at or above 80 percent; keep the queued work and wait.

中文

初次尝试 · 排队中 · Run 38697092913 / attempt 1 · 2026-10-15 11:17 UTC
lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 · b301e68225f8 · 平均延迟
**变更：**仅更新镜像：v0.5.21 是 c9a8fba9 的后代（github.com/sgl-project/sglang/compare/c9a8fba9...v0.5.21）；配方引擎参数、torch/sglang-kernel/flashinfer/DeepGEMM/Mooncake 0.3.13/CUDA 13.0.3 版本以及 Dynamo wheel 的 sglang 导入均未变化，SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT 在被移除前已无效（sglang PR 40313）；上游默认值现已启用 deferred decode KV release 与 rust 基数树核心，旧 nightly 标签已不在 Docker Hub 上列出，其打包提交仅根据标签名推断。实测数值及异常说明见上表。
**下一步：**运行 38697092913 的八个冒烟任务在后台优先级下排队等待 gb300-nv 已超过四小时，目前集群利用率已达到或超过 80%；保留排队任务并继续等待。

[Klaud-Cold]

unexpected-error · Cleanup pending. Stop and confirm owned runs before closing.

中文

unexpected-error · 清理待完成。先停止并确认自有运行结束，再关闭 PR。

[Klaud-Cold]

unexpected-error · Repairs: unknown · Runs: 38697092913
All owned runs ended. PR closed; branch deleted for retry.

中文

unexpected-error · 修复次数：unknown · 运行：38697092913
所有自有运行均已结束。PR 已关闭；分支已删除，可重新尝试。