[Klaud-Cold]

Goal: Update vLLM image from vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded to vllm/vllm-openai:nightly-edad8b03c5285a1c12e7e4f83c1af709b319a01b@sha256:2e735ffeef26457f8a3ff44a62ae182870b10af37adc06ecbb08baca152365c6.
Baseline: 2026-10-23 · vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded
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	2,350.11	17.27	1,203.5	3.09
2	2,488.81	20.48	872.26	3.28
4	3,809.25	26.66	907.4	3.76
6	4,747.99	36.66	829.98	4.7
8	6,406.3	47.77	849.68	4.65
10	7,788.12	58.04	882.56	5.13
12	7,852.88	61.02	1,027.48	5.86
14	8,925.9	72.64	1,342.91	7.49

Eval	Score ↑	Samples
gsm8k/em_strict · c14	97.04%	1,319
minimax_m3_full/em_strict · c1	98%	102
minimax_m3_full/em_strict · c2	99%	102
minimax_m3_full/em_strict · c4	99%	102
minimax_m3_full/em_strict · c6	99%	102
minimax_m3_full/em_strict · c8	99%	102
minimax_m3_full/em_strict · c10	99%	102
minimax_m3_full/em_strict · c12	99%	102
minimax_m3_full/em_strict · c14	99%	102

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-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded 更新至 vllm/vllm-openai:nightly-edad8b03c5285a1c12e7e4f83c1af709b319a01b@sha256:2e735ffeef26457f8a3ff44a62ae182870b10af37adc06ecbb08baca152365c6。
**基线：**2026-10-23 · vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded
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 40686348183 / attempt 1 · 2026-10-28 13:45 UTC
vllm/vllm-openai:nightly-edad8b03c5285a1c12e7e4f83c1af709b319a01b@sha256:2e735ffeef26457f8a3ff44a62ae182870b10af37adc06ecbb08baca152365c6 · 6cbbb96e7ad6 · AgentX · TP8/EP1 · Mean latency
Change: Measured ref e07adbe moves the master image to the 2026-10-28 vLLM nightly edad8b03 (CUDA 13.0.2, image label commit verified) with no recipe changes; the source diff from 83c078d9 adds a Hopper query-tiled MiniMax-M3 sparse prefill kernel (vllm#57420), auto NUMA binding by default, a Marlin MoE post-GEMM2 hook, NCCL 2.31.2 and flashinfer 0.7.2.post1, while every used flag, parser and env var is unchanged.

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	17.24 (N/A)	1,215.7 (N/A)	3.09 (N/A)

Note: All rows: Δ N/A: no matched baseline.

Eval	Score ↑	Samples
gsm8k/em_strict · c14	97.42% (+0.38 pp)	1,319 each
minimax_m3_full/em_strict · c1	99% (+1.00 pp)	102 each

Next: Append the perf-changelog entry, merge main, check the final matrix and dispatch the full AgentX final sweep with full-sweep-fail-fast.

中文

初次尝试 · 已通过 · Run 40686348183 / attempt 1 · 2026-10-28 13:45 UTC
vllm/vllm-openai:nightly-edad8b03c5285a1c12e7e4f83c1af709b319a01b@sha256:2e735ffeef26457f8a3ff44a62ae182870b10af37adc06ecbb08baca152365c6 · 6cbbb96e7ad6 · AgentX · TP8/EP1 · 平均延迟
**变更：**测量提交 e07adbe 将主配置镜像更新为 2026-10-28 的 vLLM nightly edad8b03（CUDA 13.0.2，已核实镜像标签中的提交），配方不变；相对 83c078d9 的源码差异新增 Hopper 上 MiniMax-M3 的 query-tiled 稀疏 prefill 内核（vllm#57420）、默认自动 NUMA 绑定、Marlin MoE 的 GEMM2 后置钩子、NCCL 2.31.2 与 flashinfer 0.7.2.post1，所用的参数、解析器和环境变量均未变化。实测数值及异常说明见上表。
**下一步：**追加 perf-changelog 条目、合并 main、检查最终矩阵，并以 full-sweep-fail-fast 调度完整的 AgentX 最终 sweep。

[Klaud-Cold]

parked · Session 1 is waiting on 40686348183. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40686348183。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 40699607468 / attempt 1 · 2026-10-28 16:27 UTC
vllm/vllm-openai:nightly-edad8b03c5285a1c12e7e4f83c1af709b319a01b@sha256:2e735ffeef26457f8a3ff44a62ae182870b10af37adc06ecbb08baca152365c6 · 99be0a93d154 · 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	17.27 (+0%)	1,203.37 (-0%)	3.08 (-0.3%)
2	20.6 (+0.6%)	915.35 (+4.9%)	3.33 (+1.5%)
4	25.62 (-3.9%)	897.77 (-1.1%)	3.56 (-5.3%)
6	36.56 (-0.3%)	819.94 (-1.2%)	4.73 (+0.6%)
8	47.67 (-0.2%)	840.93 (-1%)	4.53 (-2.6%)
10	58.15 (+0.2%)	893.95 (+1.3%)	5.03 (-1.9%)
12	60.54 (-0.8%)	1,013.37 (-1.4%)	5.82 (-0.7%)
14	73.19 (+0.8%)	1,221.79 (-9%)	7.19 (-4%)

Note: c6, c12: 1 request error.

Eval	Score ↑	Samples
minimax_m3_full/em_strict · c1	98% (+0.00 pp)	102 each
minimax_m3_full/em_strict · c2	99% (+0.00 pp)	102 each
minimax_m3_full/em_strict · c4	99% (+0.00 pp)	102 each
minimax_m3_full/em_strict · c6	99% (+0.00 pp)	102 each
minimax_m3_full/em_strict · c8	99% (+0.00 pp)	102 each
minimax_m3_full/em_strict · c10	99% (+0.00 pp)	102 each
minimax_m3_full/em_strict · c12	99% (+0.00 pp)	102 each
gsm8k/em_strict · c14	97.35% (+0.30 pp)	1,319 each
minimax_m3_full/em_strict · c14	99% (+0.00 pp)	102 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40699607468 / attempt 1 · 2026-10-28 16:27 UTC
vllm/vllm-openai:nightly-edad8b03c5285a1c12e7e4f83c1af709b319a01b@sha256:2e735ffeef26457f8a3ff44a62ae182870b10af37adc06ecbb08baca152365c6 · 99be0a93d154 · 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 40699607468. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40699607468。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=40699607468

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40699607468

[Klaud-Cold]

validated · Repairs: unknown · Runs: 40686348183, 40699523338, 40699607468
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数：未知 · 运行：40686348183, 40699523338, 40699607468
所有自有运行均已结束。完整 sweep 已验证；已就绪，等待审查。

[claude]

Code review found no issues

No high-confidence issues detected in this change.