[Klaud Cold] Update qwen3.5-fp4-gb300-dynamo-sglang-agentic-pp-pareto SGLang image to v0.5.21-cu130 / [Klaud Cold] 将 qwen3.5-fp4-gb300-dynamo-sglang-agentic-pp-pareto 的 SGLang 镜像更新至 v0.5.21-cu130

[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:nightly-dev-cu13-20260831-bb5e6198 to lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469.
Baseline: 2026-09-09 · lmsysorg/sglang:nightly-dev-cu13-20260831-bb5e6198
Mean latency · Sources: API 1, API 2, API 3

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c565 PTP1x3/DTP4x4 5ac19c	87,274.51	836.15	4,201.16	6.16
AgentX c704 PTP1x3/DTP4x2 f8ec59	131,367.59	1,266.3	5,057.49	11.01

Eval	Score ↑	Samples
gsm8k/em_strict · c565	97.35%	1,319
gsm8k/em_strict · c704	97.73%	1,319

中文

**目标：**将 SGLang 镜像从 lmsysorg/sglang:nightly-dev-cu13-20260831-bb5e6198 更新为 lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469。
**基线：**2026-09-09 · lmsysorg/sglang:nightly-dev-cu13-20260831-bb5e6198
平均延迟 · 来源： API 1, API 2, API 3；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 38663648843 / attempt 1 · 2026-10-15 06:45 UTC
lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 · c2aa0c36b1af
Change: SGLang bb5e6198 → v0.5.21 (e00930c5, 1591 ahead/0 behind; FlashInfer 0.6.18, sglang-kernel 0.4.7) removed deprecated aliases, so decode uses --cuda-graph-max-bs-decode and --mamba-radix-cache-strategy, both roles use SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK=0, and prefill drops PYTHONOPTIMIZE=1 because check_pipeline_parallel_compat now allows PP + MTP on disaggregated prefill; Dynamo 61c37bc is kept though upstream only certified 0.5.21 in 5c8b22b3.

Next: All four smoke jobs remain queued on gb300-nv after more than five hours behind higher-priority work; keep waiting without cancelling, then diagnose the first server error if any job fails.

中文

初次尝试 · 排队中 · Run 38663648843 / attempt 1 · 2026-10-15 06:45 UTC
lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 · c2aa0c36b1af
**变更：**SGLang bb5e6198 → v0.5.21（e00930c5，领先 1591、落后 0；FlashInfer 0.6.18，sglang-kernel 0.4.7）移除了弃用别名，因此 decode 改用 --cuda-graph-max-bs-decode 与 --mamba-radix-cache-strategy，两个角色改用 SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK=0；由于 check_pipeline_parallel_compat 已允许分离式 prefill 上的 PP + MTP，prefill 移除 PYTHONOPTIMIZE=1；Dynamo 保持 61c37bc，但上游直到 5c8b22b3 才适配 0.5.21。实测数值及异常说明见上表。
**下一步：**四个冒烟任务在 gb300-nv 上排在更高优先级任务之后已超过五小时；继续等待且不取消，若有任务失败则诊断第一个服务端错误。

[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: 38663648843
All owned runs ended. PR closed; branch deleted for retry.

中文

unexpected-error · 修复次数：unknown · 运行：38663648843
所有自有运行均已结束。PR 已关闭；分支已删除，可重新尝试。