[Klaud Cold] Update dsv4-fp4-b200-vllm-agentic-mtp vLLM image to nightly-dev-x86_64-cu13.0.1-20ebb28 / [Klaud Cold] 将 dsv4-fp4-b200-vllm-agentic-mtp 的 vLLM 镜像更新至 nightly-dev-x86_64-cu13.0.1-20ebb28

[Klaud-Cold]

Goal: Update vLLM image from vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-591bb95 to vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f.
Baseline: 2026-09-19 · vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-591bb95
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 TP8 EP1 9c1d23	2,201.3	16.41	1,860.04	3.48
AgentX c4 TP8 EP1 f81c5a	3,874.04	27.14	764.45	5.03
AgentX c6 TP8 EP1 0d8f91	4,483.15	35.42	855.8	4.76
AgentX c10 TP8 EP1 da6c15	8,029.38	59.46	736.62	6.12
AgentX c14 TP8 EP1 68c217	9,418.59	74.89	803.57	6.91
AgentX c16 TP8 EP1 eb2826	11,719.42	78.68	932.12	8.98
AgentX c32 TP8 EP8 658e6f	17,480.3	113.95	6,331.27	13.31
AgentX c64 TP8 EP8 ac9a9a	34,561.72	261.21	7,683.01	11.9
AgentX c96 TP8 EP8 a7529d	47,863.05	385.49	9,231	13.57
AgentX c128 TP8 EP8 5f0930	54,814.79	455.61	10,673.35	14.83
AgentX c160 TP8 EP8 51422e	64,489.73	499.22	12,379.88	17.17
AgentX c192 TP8 EP8 68a170	69,323.46	522.64	14,684.94	22.13

Eval	Score ↑	Samples
gsm8k/em_strict · c192 TP8 EP8 DPA	97.04%	1,319

中文

**目标：**将 vLLM 镜像从 vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-591bb95 更新为 vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f。
**基线：**2026-09-19 · vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-591bb95
平均延迟 · 来源： API 1, API 2, API 3, API 4；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 · Passed · Run 37933966940 / attempt 1 · 2026-10-09 15:11 UTC
vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f · ddaccac3054c · Mean latency
Change: Measured ref 469be17: pin the image to nightly-dev-x86_64-cu13.0.1-20ebb28 (vllm 0.17.2rc1.dev6312+g20ebb2850, two fork commits after dev6310+g591bb95e7: an optional DSV4 dsv4_wide_eager_region attention knob whose default keeps the current narrow region on Model Runner V2, and decode CUDA-graph eligibility for one-token prompt-tail prefills); torch, FlashInfer 0.6.18.post1 and NCCL 2.30.7 are unchanged, and the recipe needs no flag changes. Source compared from the bundled wheels because these fork commits are not published in vllm-project/vllm; newer 52cb02dee drops deep_gemm_amxf4_mega_moe and VLLM_DSV4_MEGA_FP8_COMBINE, and v0.31.0/main lack them, so they are not drop-in.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP8 EP1 9c1d23	16.43 (+0.2%)	1,824.73 (-1.9%)	3.47 (-0.3%)
AgentX c32 TP8 EP8 658e6f	115.95 (+1.8%)	5,982.23 (-5.5%)	12.89 (-3.2%)

Eval	Score ↑	Samples
gsm8k/em_strict · c192	97.73% (+0.68 pp)	1,319 each
gsm8k/em_strict · c16	96.97% (N/A)	N/A/1,319 (old/new)

Note: gsm8k/em_strict · c16: Δ N/A: no matched eval baseline.

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

中文

初次尝试 · 已通过 · Run 37933966940 / attempt 1 · 2026-10-09 15:11 UTC
vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f · ddaccac3054c · 平均延迟
**变更：**测量提交 469be17：将镜像固定为 nightly-dev-x86_64-cu13.0.1-20ebb28（vllm 0.17.2rc1.dev6312+g20ebb2850，比 dev6310+g591bb95e7 多两个 fork 提交：新增可选的 DSV4 dsv4_wide_eager_region 注意力开关，默认在 Model Runner V2 上仍使用窄区域；以及对仅含一个 token 提示尾部预填充的批次启用 decode CUDA graph）；torch、FlashInfer 0.6.18.post1 与 NCCL 2.30.7 不变，配方无需修改参数。由于这些 fork 提交未发布在 vllm-project/vllm，源码对比基于镜像内置 wheel；更新的 52cb02dee 移除了 deep_gemm_amxf4_mega_moe 和 VLLM_DSV4_MEGA_FP8_COMBINE，v0.31.0/main 也没有这些选项，因此不能直接替换。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]

Final full sweep · Running · Run 37949915743 / attempt 1 · 2026-10-09 16:24 UTC
vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f · 244b82f7fe32 · Mean latency
Change: Run the complete family with the updated image; c192 DEP8 hit CUBLAS_STATUS_EXECUTION_FAILED on DP rank 4 (DSV4 indexer GEMM) during AgentX warmup, and fail-fast cancelled the other points. The c192 eval passed.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP8 EP1 9c1d23	N/A	N/A	N/A
AgentX c4 TP8 EP1 f81c5a	N/A	N/A	N/A
AgentX c6 TP8 EP1 0d8f91	N/A	N/A	N/A
AgentX c10 TP8 EP1 da6c15	N/A	N/A	N/A
AgentX c14 TP8 EP1 68c217	N/A	N/A	N/A
AgentX c16 TP8 EP1 eb2826	N/A	N/A	N/A
AgentX c32 TP8 EP8 658e6f	N/A	N/A	N/A
AgentX c64 TP8 EP8 ac9a9a	N/A	N/A	N/A
AgentX c96 TP8 EP8 a7529d	N/A	N/A	N/A
AgentX c128 TP8 EP8 5f0930	N/A	N/A	N/A
AgentX c160 TP8 EP8 51422e	N/A	N/A	N/A
AgentX c192 TP8 EP8 68a170	N/A	N/A	N/A

Note: AgentX c1 TP8 EP1 9c1d23, AgentX c4 TP8 EP1 f81c5a, AgentX c6 TP8 EP1 0d8f91, AgentX c10 TP8 EP1 da6c15, AgentX c14 TP8 EP1 68c217, AgentX c16 TP8 EP1 eb2826, AgentX c32 TP8 EP8 658e6f, AgentX c64 TP8 EP8 ac9a9a, AgentX c96 TP8 EP8 a7529d, AgentX c128 TP8 EP8 5f0930, AgentX c160 TP8 EP8 51422e: cancelled.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.
Note: AgentX c192 TP8 EP8 68a170: failed.

Eval	Score ↑	Samples
gsm8k/em_strict · c192	N/A	1,319/N/A (old/new)

Note: All rows: Δ N/A: eval score unavailable.

Next: Wait for the c16 eval, then rerun the failed and cancelled jobs on the same head to test reproducibility.

中文

最终完整 sweep · 运行中 · Run 37949915743 / attempt 1 · 2026-10-09 16:24 UTC
vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f · 244b82f7fe32 · 平均延迟
**变更：**使用更新后的镜像运行完整配置族；c192 DEP8 在 AgentX warmup 期间于 DP rank 4（DSV4 indexer GEMM）出现 CUBLAS_STATUS_EXECUTION_FAILED，fail-fast 取消了其他点位。c192 评测通过。实测数值及异常说明见上表。
**下一步：**等待 c16 评测完成，然后在同一提交上重跑失败和被取消的任务以验证是否可复现。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=37949915743

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=37949915743

[Klaud-Cold]

Repair 1/5 · Failed · Run 37949915743 / attempt 2 · 2026-10-09 17:18 UTC
vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f · 244b82f7fe32 · Mean latency
Change: Rerun unchanged at 244b82f; c192 DEP8 reproduced CUBLAS_STATUS_EXECUTION_FAILED in the DSV4 indexer GEMM (DP rank 5 on another node, about 21 minutes into AgentX warmup), and fail-fast cancelled the other points. The 20ebb28 change most likely involved, decode CUDA-graph eligibility for one-token prompt-tail prefills, has no flag, and changing CUDA-graph modes would be tuning.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP8 EP1 9c1d23	N/A	N/A	N/A
AgentX c4 TP8 EP1 f81c5a	N/A	N/A	N/A
AgentX c6 TP8 EP1 0d8f91	N/A	N/A	N/A
AgentX c10 TP8 EP1 da6c15	N/A	N/A	N/A
AgentX c14 TP8 EP1 68c217	N/A	N/A	N/A
AgentX c16 TP8 EP1 eb2826	N/A	N/A	N/A
AgentX c32 TP8 EP8 658e6f	N/A	N/A	N/A
AgentX c64 TP8 EP8 ac9a9a	N/A	N/A	N/A
AgentX c96 TP8 EP8 a7529d	N/A	N/A	N/A
AgentX c128 TP8 EP8 5f0930	N/A	N/A	N/A
AgentX c160 TP8 EP8 51422e	N/A	N/A	N/A
AgentX c192 TP8 EP8 68a170	N/A	N/A	N/A

Note: AgentX c1 TP8 EP1 9c1d23, AgentX c4 TP8 EP1 f81c5a, AgentX c6 TP8 EP1 0d8f91, AgentX c10 TP8 EP1 da6c15, AgentX c14 TP8 EP1 68c217, AgentX c16 TP8 EP1 eb2826, AgentX c32 TP8 EP8 658e6f, AgentX c64 TP8 EP8 ac9a9a, AgentX c96 TP8 EP8 a7529d, AgentX c128 TP8 EP8 5f0930, AgentX c160 TP8 EP8 51422e: cancelled.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.
Note: AgentX c192 TP8 EP8 68a170: failed.

Next: Stop: the same failure occurred twice without progress; clean up the PR so the family can be retried with a later image.

中文

修复 1/5 · 失败 · Run 37949915743 / attempt 2 · 2026-10-09 17:18 UTC
vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f · 244b82f7fe32 · 平均延迟
**变更：**在 244b82f 上原样重跑；c192 DEP8 再次在 DSV4 indexer GEMM 中出现 CUBLAS_STATUS_EXECUTION_FAILED（另一节点上的 DP rank 5，AgentX warmup 约 21 分钟），fail-fast 取消了其他点位。20ebb28 中最可能相关的改动（对一 token 提示尾部预填充启用 decode CUDA graph）没有开关，修改 CUDA graph 模式属于调优。实测数值及异常说明见上表。
**下一步：**停止：同一失败在无进展的情况下出现两次；清理 PR，以便之后使用更新的镜像重试该配置族。

[Klaud-Cold]

failed · Cleanup pending. Stop and confirm owned runs before closing.

中文

failed · 清理待完成。先停止并确认自有运行结束，再关闭 PR。

[Klaud-Cold]

failed · Repairs: 1 · Runs: 37933966940, 37949892231, 37949915743
All owned runs ended. PR closed; branch deleted for retry.

中文

failed · 修复次数：1 · 运行：37933966940, 37949892231, 37949915743
所有自有运行均已结束。PR 已关闭；分支已删除，可重新尝试。