[Klaud Cold] Update DeepSeek-V4.1-Flash FP4 H100 vLLM AgentX image to nightly-49cb2bbf / [Klaud Cold] 将 DeepSeek-V4.1-Flash FP4 H100 vLLM AgentX 镜像更新至 nightly-49cb2bbf

[Klaud-Cold]

Goal: Update vLLM image from vllm/vllm-openai:nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3 to vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:a8d84c552462d1c60da4d7fc97f8f8f48a8fbfbf6820f73477fa0efe2571bf9d.
Baseline: 2026-09-18 · vllm/vllm-openai:nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3
AgentX · TP8/EP1 · Mean latency · Sources: API 1, API 2, API 3

Concurrency	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	2,685.27	18.82	695.58	3.05
2	2,290.78	19.7	1,115.38	4.95
4	4,055.62	27.96	547.38	3.67
8	7,092.9	51.59	546.41	4.88
16	11,241.98	78.75	994.18	8.27
20	12,673.19	89.81	2,051.87	10.98
24	10,339.29	76.84	5,959.07	19.41
28	2,815.93	19.64	45,478.73	95.27

Eval	Score ↑	Samples
gsm8k/em_strict · c28	97.57%	1,319

中文

**目标：**将 vLLM 镜像从 vllm/vllm-openai:nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3 更新为 vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:a8d84c552462d1c60da4d7fc97f8f8f48a8fbfbf6820f73477fa0efe2571bf9d。
**基线：**2026-09-18 · vllm/vllm-openai:nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3
AgentX · TP8/EP1 · 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 · Passed · Run 38933896017 / attempt 1 · 2026-10-16 19:19 UTC
vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:a8d84c552462d1c60da4d7fc97f8f8f48a8fbfbf6820f73477fa0efe2571bf9d · ea0db03844a4 · AgentX · TP8/EP1 · Mean latency
Change: Update only the master image to the digest-pinned 2026-10-16 vLLM nightly 49cb2bbf; the source diff from cd10ed6f (vllm-project/vllm@cd10ed6...49cb2bb) keeps every used DSpark, Engram, parser, Rust-frontend and V2-runner option, adds SM90 changes (Humming over Marlin, TP-split DSv4.1 fused_wqa_wkv, no DSpark draft-slot budget reservation), and the baseline GSM8K row used runner label h100-dgxc so its delta may be N/A.

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	18.92 (N/A)	558.24 (N/A)	3.03 (N/A)

Note: All rows: Δ N/A: dataset/scenario/statistic mismatch.

Eval	Score ↑	Samples
gsm8k/em_strict · c28	97.35% (N/A)	N/A/1,319 (old/new)

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

Next: Append the perf-changelog entry, then run the full fail-fast sweep over all eight baseline points.

中文

初次尝试 · 已通过 · Run 38933896017 / attempt 1 · 2026-10-16 19:19 UTC
vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:a8d84c552462d1c60da4d7fc97f8f8f48a8fbfbf6820f73477fa0efe2571bf9d · ea0db03844a4 · AgentX · TP8/EP1 · 平均延迟
**变更：**仅将主配置镜像更新为按 digest 固定的 2026-10-16 vLLM nightly 49cb2bbf；与 cd10ed6f 的源码对比（vllm-project/vllm@cd10ed6...49cb2bbf71cf）保留了所用的全部 DSpark、Engram、解析器、Rust 前端与 V2 runner 选项，新增 SM90 改动（Humming 优先于 Marlin、DSv4.1 fused_wqa_wkv 按 TP 切分、DSpark 草稿槽不再占用 token 预算），且基线 GSM8K 行使用 runner 标签 h100-dgxc，因此其差值可能为 N/A。实测数值及异常说明见上表。
**下一步：**追加 perf-changelog 条目，然后对全部八个基线点运行完整的 fail-fast 扫描。

[Klaud-Cold]

Initial attempt · Running · Run 38954171661 / attempt 1 · 2026-10-16 20:19 UTC
vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:a8d84c552462d1c60da4d7fc97f8f8f48a8fbfbf6820f73477fa0efe2571bf9d · 254639cf7eb9
Change: Add the perf-changelog entry for the passed smoke head and start the full-sweep-fail-fast sweep over all eight AgentX points plus the c28 GSM8K eval.

Next: c4 failed at startup on dsxe-sa-h100-prd0-gpu-00 after Lustre checkpoint reads stalled and the shared filesystem returned EIO, which cancelled c2, c8, c16, c24 and c28; after c1 and c20 finish, rerun the failed and cancelled jobs on this head as an infrastructure retry.

中文

初次尝试 · 运行中 · Run 38954171661 / attempt 1 · 2026-10-16 20:19 UTC
vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:a8d84c552462d1c60da4d7fc97f8f8f48a8fbfbf6820f73477fa0efe2571bf9d · 254639cf7eb9
**变更：**为已通过冒烟测试的提交添加 perf-changelog 条目，并对全部八个 AgentX 点及 c28 GSM8K 评测启动 full-sweep-fail-fast 扫描。实测数值及异常说明见上表。
**下一步：**c4 在 dsxe-sa-h100-prd0-gpu-00 上启动失败：Lustre 检查点读取停滞，共享文件系统返回 EIO，导致 c2、c8、c16、c24 和 c28 被取消；待 c1 与 c20 完成后，在此提交上将失败与取消的任务作为基础设施重试重新运行。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=38954171661

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38954171661

[Klaud-Cold]

Final full sweep · Passed · Run 38954171661 / attempt 2 · 2026-10-17 04:17 UTC
vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:a8d84c552462d1c60da4d7fc97f8f8f48a8fbfbf6820f73477fa0efe2571bf9d · 254639cf7eb9 · 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	18.89 (N/A)	526.05 (N/A)	3.03 (N/A)
2	22.39 (N/A)	444.67 (N/A)	3.33 (N/A)
4	27.98 (N/A)	411.45 (N/A)	3.58 (N/A)
8	51.81 (N/A)	376.51 (N/A)	4.55 (N/A)
16	85.14 (N/A)	359.32 (N/A)	5.69 (N/A)
20	108.51 (N/A)	395.9 (N/A)	6.74 (N/A)
24	121.92 (N/A)	485 (N/A)	7.46 (N/A)
28	129.17 (N/A)	1,070.1 (N/A)	10.06 (N/A)

Note: All rows: Δ N/A: dataset/scenario/statistic mismatch.
Note: c16, c24, c28: 2 request errors.
Note: c20: 3 request errors.

Eval	Score ↑	Samples
gsm8k/em_strict · c28	97.5% (N/A)	N/A/1,319 (old/new)

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

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 38954171661 / attempt 2 · 2026-10-17 04:17 UTC
vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:a8d84c552462d1c60da4d7fc97f8f8f48a8fbfbf6820f73477fa0efe2571bf9d · 254639cf7eb9 · 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]

validated · Repairs: unknown · Runs: 38933896017, 38954124360, 38954171661
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数：unknown · 运行：38933896017, 38954124360, 38954171661
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.