[Klaud Cold] Update dsv41flash-fp4-mi325x-vllm-agentic-dspark vLLM ROCm image to nightly-dee37d89 / 将 dsv41flash-fp4-mi325x-vllm-agentic-dspark 的 vLLM ROCm 镜像更新至 nightly-dee37d89

[adibarra]

Goal: Update vLLM ROCm image from vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657 to vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910.
Baseline: 2026-09-18 · vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657
AgentX · TP8/EP1 · Mean latency · Sources: API 1, API 2

Concurrency	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	2,028.48	15.39	1,143.13	4.71
2	2,196.82	18.06	890.28	5.11
4	3,234.51	24.34	912.28	5.65
8	5,766.62	43.96	836.1	7.18
16	9,495.44	69.76	1,250.1	12.97
32	15,287.59	103.09	1,460.34	23.4

Note: 1: 11 request errors.
Note: 2: 22 request errors.
Note: 4: 44 request errors.
Note: 8: 87 request errors.
Note: 16: 177 request errors.
Note: 32: 354 request errors.

Eval	Score ↑	Samples
gsm8k/em_strict · c32	97.04%	1,319

中文

**目标：**将 vLLM ROCm 镜像从 vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657 更新为 vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910。
**基线：**2026-09-18 · vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657
AgentX · TP8/EP1 · 平均延迟 · 来源： API 1, API 2；数值及异常说明见上表。

AI model disclosure

Claude Fable 5.1 (claude-fable-5-1, the identifier exposed by the runtime), running autonomously as Klaud Cold, prepared this PR end to end: engine source comparison, the image edit, baseline freeze, benchmark dispatch, monitoring and reporting. No other model or delegated agent contributed.

中文

本 PR 由 Claude Fable 5.1（运行时暴露的模型标识为 claude-fable-5-1）以 Klaud Cold 身份自主完成：引擎源码对比、镜像修改、基线冻结、基准测试调度、监控与报告。没有其他模型或委派代理参与。

🤖 Generated with Claude Code

[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 only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• 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：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **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。

[adibarra]

Initial attempt · Passed · Run 35346379012 / attempt 1 · 2026-09-18 13:57 UTC
vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910 · d84aefcdc172 · AgentX · TP8/EP1 · Mean latency
Change: Update from the 2026-09-12 to the 2026-09-18 upstream ROCm nightly (compare eed1f3d0...dee37d89, same base-nightly layer) because v0.29.0 predates DeepSeek V4.1; the recipe's flags, deepseek_v41 tokenizer/parsers, DSpark keys and aiter MoE backend are unchanged upstream, and ROCm fixes #57229, #57132, #55966 are in range.

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	15.65 (+1.7%)	1,115.27 (-2.4%)	4.52 (-4%)

Eval	Score ↑	Samples
gsm8k/em_strict · c32	97.27% (+0.23 pp)	1,319 each

Next: Append the perf-changelog entry, verify the full matrix and label the final full sweep.

中文

初次尝试 · 已通过 · Run 35346379012 / attempt 1 · 2026-09-18 13:57 UTC
vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910 · d84aefcdc172 · AgentX · TP8/EP1 · 平均延迟
**变更：**从 2026-09-12 的上游 ROCm nightly 更新到 2026-09-18 的 nightly（eed1f3d0 至 dee37d89，base-nightly 层相同），因为 v0.29.0 早于 DeepSeek V4.1 支持；配方参数、deepseek_v41 分词器/解析器、DSpark 配置键与 aiter MoE 后端在上游未变，并包含 ROCm 修复 #57229、#57132、#55966。实测数值及异常说明见上表。
**下一步：**追加 perf-changelog 条目，校验完整矩阵并为最终完整 sweep 添加标签。

[adibarra]

Final full sweep · Passed · Run 35353314220 / attempt 1 · 2026-09-18 15:34 UTC
vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910 · 99b06b5ce09f · AgentX · TP8/EP1 · Mean latency
Change: Run the complete family (AgentX c1-c32, TP8, gsm8k eval at c32) on the updated image with the perf-changelog entry appended at this head.

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	15.72 (+2.2%)	1,075.34 (-5.9%)	4.53 (-3.8%)
2	18.62 (+3.1%)	923.26 (+3.7%)	4.95 (-3.1%)
4	24.18 (-0.7%)	833.58 (-8.6%)	5.24 (-7.3%)
8	44.53 (+1.3%)	891.27 (+6.6%)	6.77 (-5.7%)
16	70.38 (+0.9%)	1,167.13 (-6.6%)	12.21 (-5.9%)
32	104.51 (+1.4%)	1,539.16 (+5.4%)	23.01 (-1.7%)

Note: c32: 2 request errors.

Eval	Score ↑	Samples
gsm8k/em_strict · c32	97.42% (+0.38 pp)	1,319 each

Next: Maintainer review: the lifecycle helper cannot mark this PR ready because the session token is not the Klaud-Cold account, so readiness and the reuse command are left to maintainers.

中文

最终完整 sweep · 已通过 · Run 35353314220 / attempt 1 · 2026-09-18 15:34 UTC
vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910 · 99b06b5ce09f · AgentX · TP8/EP1 · 平均延迟
**变更：**在此提交上以更新后的镜像运行完整配置族（AgentX c1-c32、TP8、c32 gsm8k eval），perf-changelog 条目已追加。实测数值及异常说明见上表。
**下一步：**维护者审查：由于会话令牌并非 Klaud-Cold 账号，生命周期助手无法将此 PR 标记为就绪，就绪与复用命令交由维护者处理。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=35353314220

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=35353314220

[adibarra]

uncertain · Cleanup by the agent (the finish helper could not run) · 2026-09-18 15:36 UTC

The updated image passed the smoke run (35346379012) and the complete final sweep on head 99b06b5c (35353314220): all six AgentX points (c1–c32, TP8) and the c32 gsm8k eval, matching every frozen baseline point by identity. Results are in the attempt comments above.

The lifecycle helper cannot certify or mark this PR ready: the session's AGENT_PAT resolves to a user account rather than Klaud-Cold, so infx.klaud treats the PR and its runs as unowned, and run-sweep.yml skipped the klaud-sweep-manifest upload for the same reason. This is a credential/configuration mismatch, not image incompatibility. Owned runs are terminal; the sweep label is removed, the PR is closed as a draft and the branch is retained for maintainer review. A maintainer may reopen and take over this PR; run 35353314220 is a complete green sweep on the current head.

中文

uncertain · 由代理完成清理（finish 助手无法运行）· 2026-09-18 15:36 UTC

更新后的镜像通过了 smoke 运行（35346379012）以及提交 99b06b5c 上的完整最终 sweep（35353314220）：全部六个 AgentX 点（c1–c32，TP8）与 c32 gsm8k eval，并按标识覆盖了所有冻结基线点，结果见上方评论。

生命周期助手无法认证或将此 PR 标记为就绪：本会话的 AGENT_PAT 对应用户账号而非 Klaud-Cold，因此 infx.klaud 将该 PR 及其运行视为非自有，run-sweep.yml 也因此跳过了 klaud-sweep-manifest 上传。这是凭据/配置不匹配，而非镜像不兼容。自有运行均已结束；已移除 sweep 标签，PR 以草稿状态关闭并保留分支供维护者审查。维护者可重新打开并接管此 PR；运行 35353314220 是当前提交上的完整绿色 sweep。