[Klaud-Cold]

Goal: Update the vLLM image from vllm/vllm-openai:nightly-dev-x86_64-cu130-ac9126e58aa7 to vllm/vllm-openai:v0.32.0.
Baseline: 2026-10-22 · vllm/vllm-openai:nightly-dev-x86_64-cu130-ac9126e58aa7
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 TP2 EP1 8e063c	13,578.65	86.01	225.88	2.17
AgentX c1 TP4 EP1 b63216	7,689.42	45.97	201.09	1.68
AgentX c4 TP4 EP1 2949fa	9,834.7	68.07	149.02	1.8
AgentX c8 TP2 EP2 63b1b9	30,840.2	222.29	201.27	3.02
AgentX c8 TP4 EP1 fc3c2c	16,700.98	115.8	127.45	2.01
AgentX c16 TP2 EP2 4ec87c	56,293.65	356.33	225.53	3.5
AgentX c16 TP4 EP1 91af93	29,392.92	189.94	139.88	2.52
AgentX c24 TP2 EP2 4fbe45	89,967.93	538.79	805.18	3.86
AgentX c32 TP2 EP2 3bed93	116,349.75	673.69	1,036.83	4.48
AgentX c32 TP4 EP1 ce771f	62,554.47	368.49	187.71	3.39
AgentX c64 TP2 EP2 fa0e4a	206,155.07	1,485.23	362.45	8.6
AgentX c96 TP4 EP4 47a295	151,749.49	1,133.92	988.89	6.06
AgentX c128 TP2 EP2 597cf9	270,569.1	2,225.63	1,961.02	14.8
AgentX c128 TP4 EP1 61d9ee	166,492.11	1,319.22	413.92	9.38
AgentX c128 TP4 EP4 a2ec7a	178,100.34	1,388.18	1,153.13	7.21

Eval	Score ↑	Samples
gsm8k/em_strict · c128	97.73%	1,319
gsm8k/em_strict · c128	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-dev-x86_64-cu130-ac9126e58aa7 更新为 vllm/vllm-openai:v0.32.0。
**基线：**2026-10-22 · vllm/vllm-openai:nightly-dev-x86_64-cu130-ac9126e58aa7
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 · Queued · Run 40250403097 / attempt 1 · 2026-10-26 00:40 UTC
vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34 · 88588aaf353a
Change: Pinned vllm/vllm-openai:v0.32.0 by digest (build commit 5df63a7203cd, CUDA 13.0.2) over the ac9126e58aa7 nightly, an ancestor by 633 commits; source checks keep every recipe flag, deepseek_v41 parsers, FLASHINFER_MLA_SPARSE_DSV41, DSpark, Engram, deep_gemm_mega_moe and native KV offload, so settings are unchanged, but decoder_replay_cudagraph_capture_sizes is absent at ac9126e58aa7 yet the baseline used it, and v0.32.0 now honors it (vllm-project/vllm#59532).

Next: Wait for the five smoke throughput points and two gsm8k evals in run 40250403097, then diagnose the first server error if any.

中文

初次尝试 · 排队中 · Run 40250403097 / attempt 1 · 2026-10-26 00:40 UTC
vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34 · 88588aaf353a
**变更：**以 digest 固定 vllm/vllm-openai:v0.32.0（构建提交 5df63a7203cd，CUDA 13.0.2），替换 ac9126e58aa7 nightly（v0.32.0 之前 633 个提交的祖先）；源码检查确认所有配方参数、deepseek_v41 解析器、FLASHINFER_MLA_SPARSE_DSV41、DSpark、Engram、deep_gemm_mega_moe 与原生 KV 卸载均保留，因此设置不变；但 ac9126e58aa7 中没有 decoder_replay_cudagraph_capture_sizes，而基线使用了它，v0.32.0 现已使其生效（vllm-project/vllm#59532）。
**下一步：**等待运行 40250403097 中的五个冒烟吞吐点和两个 gsm8k 评测完成，如有失败则诊断首个服务端错误。

[Klaud-Cold]

parked · Session 1 is waiting on 40250403097. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40250403097。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]

Initial attempt · Passed · Run 40250403097 / attempt 2 · 2026-10-26 04:35 UTC
vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34 · 88588aaf353a · Mean latency
Change: Pinned vllm/vllm-openai:v0.32.0 by digest (build commit 5df63a7203cd, CUDA 13.0.2) over the ac9126e58aa7 nightly, an ancestor by 633 commits; source checks keep every recipe flag, deepseek_v41 parsers, FLASHINFER_MLA_SPARSE_DSV41, DSpark, Engram, deep_gemm_mega_moe and native KV offload, so settings are unchanged, but decoder_replay_cudagraph_capture_sizes is absent at ac9126e58aa7 yet the baseline used it, and v0.32.0 now honors it (vllm-project/vllm#59532).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP2 EP1 8e063c	86.22 (+0.2%)	210.88 (-6.6%)	2.14 (-1.4%)
AgentX c1 TP4 EP1 b63216	46.44 (+1%)	189.81 (-5.6%)	1.65 (-1.8%)
AgentX c8 TP2 EP2 63b1b9	221.7 (-0.3%)	192.79 (-4.2%)	2.97 (-1.7%)
AgentX c64 TP2 EP2 fa0e4a	1,493.84 (+0.6%)	333.6 (-8%)	8.3 (-3.5%)
AgentX c96 TP4 EP4 47a295	1,143.84 (+0.9%)	916.01 (-7.4%)	5.72 (-5.6%)

Note: AgentX c1 TP4 EP1 b63216, AgentX c64 TP2 EP2 fa0e4a: 1 request error.
Note: AgentX c8 TP2 EP2 63b1b9, AgentX c96 TP4 EP4 47a295: 2 request errors.

Eval	Score ↑	Samples
gsm8k/em_strict · c128	97.27% (-0.45 pp)	1,319 each
gsm8k/em_strict · c128	97.12% (-0.38 pp)	1,319 each

Next: Smoke attempt 1 was preempted and its rerun passed all five points and both gsm8k evals (0-2 InvalidInferenceResultError requests per point, baseline count unpublished); append the perf-changelog entry and dispatch the full-sweep-fail-fast final sweep.

中文

初次尝试 · 已通过 · Run 40250403097 / attempt 2 · 2026-10-26 04:35 UTC
vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34 · 88588aaf353a · 平均延迟
**变更：**以 digest 固定 vllm/vllm-openai:v0.32.0（构建提交 5df63a7203cd，CUDA 13.0.2），替换 ac9126e58aa7 nightly（v0.32.0 之前 633 个提交的祖先）；源码检查确认所有配方参数、deepseek_v41 解析器、FLASHINFER_MLA_SPARSE_DSV41、DSpark、Engram、deep_gemm_mega_moe 与原生 KV 卸载均保留，因此设置不变；但 ac9126e58aa7 中没有 decoder_replay_cudagraph_capture_sizes，而基线使用了它，v0.32.0 现已使其生效（vllm-project/vllm#59532）。实测数值及异常说明见上表。
**下一步：**冒烟运行第 1 次尝试被抢占，重跑后五个吞吐点和两个 gsm8k 评测全部通过（每个点 0-2 个 InvalidInferenceResultError 请求，基线未公布该计数）；下一步追加 perf-changelog 条目并以 full-sweep-fail-fast 发起最终扫描。

[Klaud-Cold]

Final full sweep · Passed · Run 40269923234 / attempt 1 · 2026-10-26 08:18 UTC
vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34 · 7c13e420c340 · Mean latency
Change: Validate the complete updated-image family at this head.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP2 EP1 8e063c	86.01 (+0%)	212.02 (-6.1%)	2.15 (-0.9%)
AgentX c1 TP4 EP1 b63216	46.48 (+1.1%)	189.45 (-5.8%)	1.65 (-1.8%)
AgentX c4 TP4 EP1 2949fa	68.74 (+1%)	139.91 (-6.1%)	1.79 (-0.6%)
AgentX c8 TP2 EP2 63b1b9	221.91 (-0.2%)	194.91 (-3.2%)	2.99 (-1%)
AgentX c8 TP4 EP1 fc3c2c	116.69 (+0.8%)	123.34 (-3.2%)	1.96 (-2.5%)
AgentX c16 TP2 EP2 4ec87c	359.47 (+0.9%)	222.61 (-1.3%)	3.59 (+2.6%)
AgentX c16 TP4 EP1 91af93	192 (+1.1%)	130.48 (-6.7%)	2.3 (-8.7%)
AgentX c24 TP2 EP2 4fbe45	536.38 (-0.4%)	771.97 (-4.1%)	3.83 (-0.8%)
AgentX c32 TP2 EP2 3bed93	677.51 (+0.6%)	912.75 (-12%)	4.32 (-3.6%)
AgentX c32 TP4 EP1 ce771f	370.3 (+0.5%)	166.41 (-11.3%)	3.23 (-4.7%)
AgentX c64 TP2 EP2 fa0e4a	1,481.61 (-0.2%)	343.36 (-5.3%)	8.4 (-2.3%)
AgentX c96 TP4 EP4 47a295	1,144.85 (+1%)	906.44 (-8.3%)	5.74 (-5.3%)
AgentX c128 TP2 EP2 597cf9	2,213.44 (-0.5%)	2,576.9 (+31.4%)	14.45 (-2.4%)
AgentX c128 TP4 EP1 61d9ee	1,350.58 (+2.4%)	316.77 (-23.5%)	8.7 (-7.2%)
AgentX c128 TP4 EP4 a2ec7a	1,406.97 (+1.4%)	1,048.45 (-9.1%)	6.79 (-5.8%)

Note: AgentX c1 TP4 EP1 b63216, AgentX c8 TP2 EP2 63b1b9, AgentX c24 TP2 EP2 4fbe45, AgentX c32 TP2 EP2 3bed93: 1 request error.
Note: AgentX c16 TP4 EP1 91af93: 2 request errors.
Note: AgentX c32 TP4 EP1 ce771f, AgentX c64 TP2 EP2 fa0e4a, AgentX c96 TP4 EP4 47a295: 3 request errors.
Note: AgentX c128 TP2 EP2 597cf9: 4 request errors.
Note: AgentX c128 TP4 EP1 61d9ee: 7 request errors.

Eval	Score ↑	Samples
gsm8k/em_strict · c128	97.5% (-0.23 pp)	1,319 each
gsm8k/em_strict · c128	97.27% (-0.23 pp)	1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40269923234 / attempt 1 · 2026-10-26 08:18 UTC
vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34 · 7c13e420c340 · 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 40269923234. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40269923234。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=40269923234

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40269923234

[Klaud-Cold]

validated · Repairs: unknown · Runs: 40250403097, 40269875283, 40269923234
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数：未知 · 运行：40250403097, 40269875283, 40269923234
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 21s —— View job

---

LGTM - no blocking issues found

I checked three things:

• The image is public and pinned by digest. 101 other image entries in the master configs are pinned by digest the same way.
• The new perf-changelog.yaml entry is at the end of the file and covers dsv41flash-fp4-b200-vllm-agentic-dspark.
• The final full sweep (run 40269923234) passed.

[claude]

Code review found no issues

No high-confidence issues detected in this change.