[Klaud-Cold]

Goal: Update vLLM image from vllm/vllm-openai:nightly-dev-x86_64-cu130-ac9126e58aa7 to vllm/vllm-openai:nightly-fe4df3df0eb703862c686b256e9fe44fa94c8945@sha256:8e5e3042302f29e63b1722ceb1db8edf96beafc5dfc19f5eebb5a668470b71b4.
Baseline: 2026-10-02 · vllm/vllm-openai:nightly-dev-x86_64-cu130-ac9126e58aa7
Mean latency · Sources: API 1, API 2, API 3

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 f41bbd	7,537.41	45.46	316.92	1.69
AgentX c4 TP4 EP1 a60967	10,268.87	70.09	211.63	1.82
AgentX c8 TP2 EP2 c75e8c	30,793.29	222.51	213.18	2.92
AgentX c8 TP4 EP1 74519e	16,585.72	115.03	175.2	2.04
AgentX c16 TP2 EP2 c95c57	56,827.42	358.48	200.7	3.38
AgentX c16 TP4 EP1 5f4875	29,331.81	190.22	187.76	2.56
AgentX c32 TP2 EP2 2e8842	119,446.55	687.84	221.67	4.64
AgentX c64 TP2 EP2 7b01d6	214,245.91	1,525.81	268.59	7.81
AgentX c128 TP2 EP2 0e11a9	299,060.19	2,406.26	389.92	12.95
AgentX c192 TP2 EP2 b6fb07	406,373.87	3,059.33	513.06	18.05

Eval	Score ↑	Samples
gsm8k/em_strict · c16	97.35%	1,319
gsm8k/em_strict · c192	97.27%	1,319

中文

**目标：**将 vLLM 镜像从 vllm/vllm-openai:nightly-dev-x86_64-cu130-ac9126e58aa7 更新为 vllm/vllm-openai:nightly-fe4df3df0eb703862c686b256e9fe44fa94c8945@sha256:8e5e3042302f29e63b1722ceb1db8edf96beafc5dfc19f5eebb5a668470b71b4。
**基线：**2026-10-02 · vllm/vllm-openai:nightly-dev-x86_64-cu130-ac9126e58aa7
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 · Failed · Run 38330982761 / attempt 1 · 2026-10-13 01:25 UTC
vllm/vllm-openai:nightly-fe4df3df0eb703862c686b256e9fe44fa94c8945@sha256:8e5e3042302f29e63b1722ceb1db8edf96beafc5dfc19f5eebb5a668470b71b4 · 801ae81adcba · Mean latency
Change: Pin the official vLLM nightly fe4df3df (CUDA 13.0, 555 main commits after ac9126e5) by digest with no recipe flag changes; every used option still parses in the new source, e.g. decoder_replay_cudagraph_capture_sizes from #59532 and deep_gemm_mega_moe after #58053 (compare); the old dev image is labelled ac9126e5 yet already accepted #59532 options, so its exact source is uncertain.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 95d8c2	N/A	N/A	N/A
AgentX c8 TP2 EP2 8b11f9	N/A	N/A	N/A

Note: All rows: failed.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: no matched baseline.

Next: Retry the same head once the b300-dsxe partition accepts jobs.

中文

初次尝试 · 失败 · Run 38330982761 / attempt 1 · 2026-10-13 01:25 UTC
vllm/vllm-openai:nightly-fe4df3df0eb703862c686b256e9fe44fa94c8945@sha256:8e5e3042302f29e63b1722ceb1db8edf96beafc5dfc19f5eebb5a668470b71b4 · 801ae81adcba · 平均延迟
**变更：**按摘要固定官方 vLLM nightly fe4df3df（CUDA 13.0，比 ac9126e5 多 555 个 main 提交），配方参数不变；所用选项在新源码中均可解析，例如 #59532 引入的 decoder_replay_cudagraph_capture_sizes 和 #58053 之后的 deep_gemm_mega_moe（对比）；旧 dev 镜像标注为 ac9126e5 却已接受 #59532 的选项，其确切源码存在不确定性。实测数值及异常说明见上表。
**下一步：**待 b300-dsxe 分区恢复接收作业后，在同一提交上重试。

[Klaud-Cold]

Infrastructure retry 0 · Passed · Run 38330982761 / attempt 2 · 2026-10-13 03:17 UTC
vllm/vllm-openai:nightly-fe4df3df0eb703862c686b256e9fe44fa94c8945@sha256:8e5e3042302f29e63b1722ceb1db8edf96beafc5dfc19f5eebb5a668470b71b4 · 801ae81adcba · Mean latency
Change: Rerun the unchanged head after Slurm rejected every b300-dsxe submission with an inactive or drained partition.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 f41bbd	45.54 (N/A)	303.99 (N/A)	1.67 (N/A)
AgentX c8 TP2 EP2 c75e8c	221.85 (N/A)	155.98 (N/A)	2.91 (N/A)

Note: All rows: Δ N/A: dataset/scenario/statistic mismatch.
Note: AgentX c8 TP2 EP2 c75e8c: 1 request error.

Eval	Score ↑	Samples
gsm8k/em_strict · c192	97.57% (+0.30 pp)	1,319 each
gsm8k/em_strict · c16	97.5% (+0.15 pp)	1,319 each

Next: Append the changelog entry and run the final full sweep.

中文

基础设施重试 0 · 已通过 · Run 38330982761 / attempt 2 · 2026-10-13 03:17 UTC
vllm/vllm-openai:nightly-fe4df3df0eb703862c686b256e9fe44fa94c8945@sha256:8e5e3042302f29e63b1722ceb1db8edf96beafc5dfc19f5eebb5a668470b71b4 · 801ae81adcba · 平均延迟
**变更：**由于 b300-dsxe 分区不可用（inactive 或 drain），Slurm 拒绝了全部提交，现以未修改的提交重新运行。实测数值及异常说明见上表。
**下一步：**追加 changelog 条目并运行最终完整 sweep。

[Klaud-Cold]

capacity-deferred · Cleanup pending. Stop and confirm owned runs before closing.

中文

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

[Klaud-Cold]

capacity-deferred · Repairs: 0 · Runs: 38330982761, 38341119182
All owned runs ended. PR closed; branch deleted for retry.

中文

capacity-deferred · 修复次数：0 · 运行：38330982761, 38341119182
所有自有运行均已结束。PR 已关闭；分支已删除，可重新尝试。