[Klaud-Cold]

Goal: Update vLLM image from vllm/vllm-openai:nightly-e4a5b5537344b103ed10de7764378ef09604ef3d@sha256:6b90105076216e861cd30fae8828efef9263a0399137b894289f3b6a685bf4ff to vllm/vllm-openai:nightly-00b6c4c565a8732c131e1f452d4a17803bb0842b@sha256:76a9614ed4ed20694914c87e82a0344bd73e1c6e081459a2467175ab903657e4.
Baseline: 2026-10-26 · vllm/vllm-openai:nightly-e4a5b5537344b103ed10de7764378ef09604ef3d@sha256:6b90105076216e861cd30fae8828efef9263a0399137b894289f3b6a685bf4ff
AgentX · TP8/EP1 · Mean latency · Sources: API 1, API 2, API 3, API 4

Concurrency	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	2,840.02	19.37	515.04	2.85
2	2,926.98	23.05	451.56	3.11
4	4,163.15	28.54	400.72	3.25
8	7,299.08	52.99	369.77	4.01
16	13,530.31	85.7	357.31	5.28
20	16,791.03	109.75	388.37	6.27
24	19,779.81	122.8	577.8	7.33
28	19,743.59	128.9	1,261.34	9.83

Eval	Score ↑	Samples
gsm8k/em_strict · c28	97.5%	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 vLLM 镜像从 vllm/vllm-openai:nightly-e4a5b5537344b103ed10de7764378ef09604ef3d@sha256:6b90105076216e861cd30fae8828efef9263a0399137b894289f3b6a685bf4ff 更新为 vllm/vllm-openai:nightly-00b6c4c565a8732c131e1f452d4a17803bb0842b@sha256:76a9614ed4ed20694914c87e82a0344bd73e1c6e081459a2467175ab903657e4。
**基线：**2026-10-26 · vllm/vllm-openai:nightly-e4a5b5537344b103ed10de7764378ef09604ef3d@sha256:6b90105076216e861cd30fae8828efef9263a0399137b894289f3b6a685bf4ff
AgentX · TP8/EP1 · 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 · Passed · Run 40448923378 / attempt 1 · 2026-10-27 05:51 UTC
vllm/vllm-openai:nightly-00b6c4c565a8732c131e1f452d4a17803bb0842b@sha256:76a9614ed4ed20694914c87e82a0344bd73e1c6e081459a2467175ab903657e4 · e1356ab6e506 · AgentX · TP8/EP1 · Mean latency
Change: Image-only bump from vLLM nightly e4a5b55 to nightly 00b6c4c (39 upstream commits, same CUDA 13.0.2 base, no dependency changes); recipe flags unchanged because the diff (vllm-project/vllm@e4a5b55...00b6c4c) renames no used option and only touches the Engram offload meta-init path, speculator max-model-len sync, priority-preemption scheduling and Rust-frontend structured-output backend selection (default xgrammar kept).

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	19.29 (-0.4%)	531.15 (+3.1%)	2.86 (+0.4%)

Note: All rows: 1 request error.

Eval	Score ↑	Samples
gsm8k/em_strict · c28	97.42% (-0.08 pp)	1,319 each

Next: Smoke c1 benchmark and c28 gsm8k eval passed on measured commit 49ff151; append the perf-changelog entry and dispatch the full-sweep-fail-fast final sweep.

中文

初次尝试 · 已通过 · Run 40448923378 / attempt 1 · 2026-10-27 05:51 UTC
vllm/vllm-openai:nightly-00b6c4c565a8732c131e1f452d4a17803bb0842b@sha256:76a9614ed4ed20694914c87e82a0344bd73e1c6e081459a2467175ab903657e4 · e1356ab6e506 · AgentX · TP8/EP1 · 平均延迟
**变更：**仅更新镜像：vLLM nightly e4a5b55 升级到 nightly 00b6c4c（上游 39 个提交，CUDA 13.0.2 基础不变，无依赖变化）；recipe 参数保持不变，因为该差异（vllm-project/vllm@e4a5b55...00b6c4c565a8732c131e1f452d4a17803bb0842b）未重命名任何所用选项，只涉及 Engram offload 的 meta 初始化、speculator 的 max-model-len 同步、priority 抢占调度以及 Rust 前端 structured outputs 后端选择（默认仍为 xgrammar）。实测数值及异常说明见上表。
**下一步：**在实测提交 49ff151 上，smoke c1 基准与 c28 gsm8k 评测均已通过；接下来追加 perf-changelog 条目并通过 full-sweep-fail-fast 调度最终 sweep。

[Klaud-Cold]

parked · Session 1 is waiting on 40448923378. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40448923378。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 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]

Final full sweep · Passed · Run 40457810253 / attempt 1 · 2026-10-27 10:10 UTC
vllm/vllm-openai:nightly-00b6c4c565a8732c131e1f452d4a17803bb0842b@sha256:76a9614ed4ed20694914c87e82a0344bd73e1c6e081459a2467175ab903657e4 · 5ac6f3a18b70 · AgentX · TP8/EP1 · Mean latency
Change: Validate the complete updated-image family at this head.

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	19.35 (-0.1%)	529.79 (+2.9%)	2.85 (+0%)
2	23 (-0.2%)	425.11 (-5.9%)	3.13 (+0.6%)
4	28.37 (-0.6%)	399.4 (-0.3%)	3.29 (+1.2%)
8	52.92 (-0.1%)	376.85 (+1.9%)	4.05 (+1%)
16	85.74 (+0.1%)	372.73 (+4.3%)	5.25 (-0.6%)
20	109.79 (+0%)	403.71 (+3.9%)	6.23 (-0.6%)
24	122.69 (-0.1%)	543.17 (-6%)	7.31 (-0.3%)
28	129.58 (+0.5%)	1,231.11 (-2.4%)	9.82 (-0.1%)

Note: c16: 5 request errors.
Note: c20, c24: 2 request errors.
Note: c28: 3 request errors.

Eval	Score ↑	Samples
gsm8k/em_strict · c28	97.35% (-0.15 pp)	1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40457810253 / attempt 1 · 2026-10-27 10:10 UTC
vllm/vllm-openai:nightly-00b6c4c565a8732c131e1f452d4a17803bb0842b@sha256:76a9614ed4ed20694914c87e82a0344bd73e1c6e081459a2467175ab903657e4 · 5ac6f3a18b70 · AgentX · TP8/EP1 · 平均延迟
**变更：**验证此提交更新镜像后的完整配置族。实测数值及异常说明见上表。
**下一步：**标记为就绪，等待维护者审查。

[Klaud-Cold]

parked · Session 2 is waiting on 40457810253. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40457810253。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=40457810253

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40457810253

[Klaud-Cold]

validated · Repairs: unknown · Runs: 40448923378, 40457782702, 40457810253
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数：未知 · 运行：40448923378, 40457782702, 40457810253
所有自有运行均已结束。完整 sweep 已验证；已就绪，等待审查。

[claude]

Claude finished @Klaud-Cold's task in 12s —— View job

---

LGTM, no blocking issues found.

The image bump in nvidia-master.yaml points to a public Docker Hub vllm/vllm-openai nightly, pinned by digest. The matching perf-changelog.yaml entry for dsv41flash-fp4-h100-vllm-agentic-dspark is at the end of the file, and the final full sweep passed.

[claude]

Code review found no issues

No high-confidence issues detected in this change.