[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:nightly-dev-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576 to lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90.
Baseline: 2026-10-23 · lmsysorg/sglang:nightly-dev-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576
Mean latency · Sources: API 1, API 2, API 3, API 4

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 PTP4x1/DTP4x0 8a7b6a	5,174.09	37.2	742.68	3.12
AgentX c1 PTP8x1/DTP8x0 ce4f77	2,808.86	19.17	715.02	2.78
AgentX c36 PTP1x1/DTP4x6 538772	8,389.72	49.92	969.8	4.88
AgentX c48 PTP1x1/DTP4x4 c4eef4	14,033.17	91.68	1,316.09	5.17
AgentX c64 PTP1x1/DTP8x2 a39e5c	17,885.41	126.39	1,458.46	7.32
AgentX c96 PTP1x1/DTP8x1 aaa2d3	32,056.64	253.75	2,307.03	8.86
AgentX c96 PTP1x1/DTP8x2 a74110	22,692.08	175.77	2,125.03	7.9
AgentX c224 PTP1x2/DTP16x1 18fb54	37,666.39	264.42	2,586.71	9.34
AgentX c336 PTP1x3/DTP16x1 f06e2e	42,272.42	312.16	3,192.43	10.27

Eval	Score ↑	Samples
gsm8k/em_strict · c36	96.74%	1,319
gsm8k/em_strict · c48	97.19%	1,319
gsm8k/em_strict · c64	96.97%	1,319
gsm8k/em_strict · c224	97.12%	1,319
gsm8k/em_strict · c336	96.82%	1,319

Stop: Close this PR at any time to stop this candidate now; Klaud cancels its runs and deletes its branch.

中文

**目标：**将 SGLang 镜像从 lmsysorg/sglang:nightly-dev-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576 更新为 lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90。
**基线：**2026-10-23 · lmsysorg/sglang:nightly-dev-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576
平均延迟 · 来源： API 1, API 2, API 3, API 4；数值及异常说明见上表。
**停止：**随时关闭此 PR 即可立即停止该候选：Klaud 会取消其运行并删除其分支。

[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 · Failed · Run 40686805658 / attempt 1 · 2026-10-28 16:40 UTC
lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90 · 025528300e3a
Change: Both family keys move to the digest-pinned SGLang nightly 586355a8 (bundled ai-dynamo 1.7.0.dev20261027, up from 1.6.0.dev20261019), and both recipe identities record sglang 0.0.0.dev1+g586355a83 and that ai-dynamo; source diffs of SGLang ab580b2e..586355a8 show every recipe flag still parsed with unchanged EAGLE, HiCache, NIXL and modelopt_fp4 defaults.

Next: Repair 1 pins Dynamo in both family recipes to source rev 68c35a1ce41b (the direct parent of a789babbe37a; no Dynamo SGLang-component commits between it and the bundled 10-27 nightly) so /health lists instances again, then reruns the full sweep.

中文

初次尝试 · 失败 · Run 40686805658 / attempt 1 · 2026-10-28 16:40 UTC
lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90 · 025528300e3a
**变更：**两个 family 键均切换到按摘要固定的 SGLang nightly 586355a8（内置 ai-dynamo 1.7.0.dev20261027，原为 1.6.0.dev20261019），两个 recipe 的 identity 记录 sglang 0.0.0.dev1+g586355a83 及该 ai-dynamo 版本；SGLang ab580b2e..586355a8 的源码对比显示所有 recipe 参数仍可解析，EAGLE、HiCache、NIXL 与 modelopt_fp4 默认值未变。
**下一步：**修复 1 在两个 family recipe 中将 Dynamo 固定为源码 rev 68c35a1ce41b（a789babbe37a 的直接父提交；它与内置 10-27 nightly 之间没有 Dynamo SGLang 组件提交），使 /health 重新列出 instances，然后重跑完整 sweep。

[Klaud-Cold]

parked · Session 1 is waiting on 40686805658. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40686805658。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40738328317

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40738328317

[Klaud-Cold]

resumed · The session ended before finishing; session 2 continues it in a new run of its planner.

中文

resumed · 会话未完成即已结束；第 2 个会话将在其规划运行的新运行中继续。

[Klaud-Cold]

Repair 1/5 · Failed · Run 40728600329 / attempt 1 · 2026-10-28 17:45 UTC
lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90 · fcb03bebf516
Change: Both family recipes now install Dynamo from source rev 68c35a1ce41b (identity ai-dynamo 1.6.0), the direct parent of a789babbe37a (#16454), which made frontend /health omit the instances list srtctl readiness needs; Dynamo's components/src/dynamo/sglang has no commits between that rev and the bundled 1.7.0.dev20261027, and SGLang stays at 586355a8 with no engine patches; main was merged with the perf-changelog entry kept at the tail.

Next: Repair 2 sets Dynamo's documented DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT, _ENGINE_REQUEST_WAITING_LIMIT and _REQUEST_QUEUE_LIMIT to the gate's pre-engine-report defaults (10000/10000/20000) on every worker role, restoring the baseline's pass-through queueing into SGLang, then reruns the full sweep.

中文

修复 1/5 · 失败 · Run 40728600329 / attempt 1 · 2026-10-28 17:45 UTC
lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90 · fcb03bebf516
**变更：**两个 family recipe 现从源码 rev 68c35a1ce41b 安装 Dynamo（identity 记为 ai-dynamo 1.6.0），即 a789babbe37a（#16454）的直接父提交；该提交使前端 /health 不再返回 srtctl 就绪检查所需的 instances 列表；Dynamo 的 components/src/dynamo/sglang 在该 rev 与内置 1.7.0.dev20261027 之间没有提交，SGLang 仍为 586355a8 且无引擎补丁；已合并 main，perf-changelog 条目保持在末尾。
**下一步：**修复 2 在所有 worker 角色上将 Dynamo 文档化的 DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT、_ENGINE_REQUEST_WAITING_LIMIT 与 _REQUEST_QUEUE_LIMIT 设为该准入门在引擎上报前的默认值（10000/10000/20000），恢复基线中请求直接排入 SGLang 的行为，然后重跑完整 sweep。

[Klaud-Cold]

parked · Session 2 is waiting on 40728600329. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40728600329。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。

[Klaud-Cold]

resumed · The session ended before finishing; session 3 continues it in a new run of its planner.

中文

resumed · 会话未完成即已结束；第 3 个会话将在其规划运行的新运行中继续。

[Klaud-Cold]

Repair 2/5 · Queued · Run 40738328317 / attempt 1 · 2026-10-28 17:45 UTC
lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90 · b9ec45672690
Change: Every worker role in both family recipes sets DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT=10000, _ENGINE_REQUEST_WAITING_LIMIT=10000 and _REQUEST_QUEUE_LIMIT=20000, the fallback limits Dynamo's admission gate (lib/llm/src/admission_gate.rs at 68c35a1ce41b) uses before the engine reports max_num_seqs, so bursts queue in SGLang as with the baseline's pre-gate Dynamo; Dynamo stays at 68c35a1ce41b and SGLang at 586355a8 with no engine patches.

Next: The full-sweep-fail-fast final sweep on this head is queued with 9 benchmark and 9 eval jobs after check-final passed; the session parks until they finish.

中文

修复 2/5 · 排队中 · Run 40738328317 / attempt 1 · 2026-10-28 17:45 UTC
lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90 · b9ec45672690
**变更：**两个 family recipe 的所有 worker 角色设置 DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT=10000、_ENGINE_REQUEST_WAITING_LIMIT=10000 与 _REQUEST_QUEUE_LIMIT=20000，即 Dynamo 准入门（68c35a1ce41b 的 lib/llm/src/admission_gate.rs）在引擎上报 max_num_seqs 前使用的回退限制，使突发请求像基线中无准入门的 Dynamo 一样在 SGLang 内排队；Dynamo 仍为 68c35a1ce41b，SGLang 仍为 586355a8，且无引擎补丁。
**下一步：**check-final 通过后，该 head 上的 full-sweep-fail-fast 最终 sweep 已排队，共 9 个基准与 9 个评测任务；会话将暂停，直到它们完成。

[Klaud-Cold]

parked · Session 3 is waiting on 40738328317. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 3 个会话正在等待 40738328317。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。