[Klaud Cold] Update dsv4-fp4-gb200-dynamo-vllm-agentic-mtp2-agg vLLM image to v0.29.0 / 将 dsv4-fp4-gb200-dynamo-vllm-agentic-mtp2-agg 的 vLLM 镜像更新至 v0.29.0

[Klaud-Cold]

Goal: Update vLLM image from vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f to vllm/vllm-openai:v0.29.0.
Baseline: 2026-08-18 · vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f
AgentX · PTP8x1/DTP8x0 · Mean latency · Sources: API 1, API 2

Concurrency	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	N/A	N/A	N/A	N/A
4	N/A	N/A	N/A	N/A
8	N/A	N/A	N/A	N/A

Note: All rows: unavailable.

Eval: N/A

中文

**目标：**将 vLLM 镜像从 vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f 更新为 vllm/vllm-openai:v0.29.0。
**基线：**2026-08-18 · vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f
AgentX · PTP8x1/DTP8x0 · 平均延迟 · 来源： API 1, API 2；数值及异常说明见上表。

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[Klaud-Cold]

Baseline: 2026-08-18 · vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f

AgentX · PTP8x1/DTP8x0 · Mean latency

Concurrency	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	N/A	N/A	N/A	N/A
4	N/A	N/A	N/A	N/A
8	N/A	N/A	N/A	N/A

Note: All rows: unavailable.

中文

**基线：**2026-08-18 · vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f；数值及异常说明见表格。

[Klaud-Cold]

Initial attempt · Failed · Run 34662875814 / attempt 1 · 2026-09-12 01:46 UTC
vllm/vllm-openai:v0.29.0 · 127c84be90f1 · AgentX · PTP8x1/DTP8x0 · Mean latency
Change: Replace the locally built dev image (build commit unknown, 426e59f absent from upstream vLLM, built 2026-07-26, CUDA 13.0.1, FlashInfer 0.6.14) with the vLLM v0.29.0 release (commit 98dff2a8, CUDA 13.0.2, FlashInfer 0.6.18, torch 2.13.0) in the master image and both recipes with every flag kept: v0.29.0 source still accepts FLASHINFER_MLA_SPARSE_DSV4, the deprecated use_fp4_indexer_cache alias (vllm-project/vllm#52550) and the VLLM_PREFIX_CACHE_RETENTION_INTERVAL env (vllm-project/vllm#52216), MTP on the forced V2 runner, and its DP hash already ignores numa_bind so the srt-slurm v1.0.45 setup patch is skipped.

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	N/A	N/A	N/A
8	N/A	N/A	N/A

Note: c1: failed.
Note: c1: 2 request errors.
Note: All rows: Δ N/A: point unavailable.
Note: c8: cancelled.
Note: c8: request errors unavailable.

Next: Repair 1: lower gpu-memory-utilization from 0.94 to 0.90 in both recipes to restore runtime headroom for the sparse indexer workspace, then rerun the smoke.

中文

初次尝试 · 失败 · Run 34662875814 / attempt 1 · 2026-09-12 01:46 UTC
vllm/vllm-openai:v0.29.0 · 127c84be90f1 · AgentX · PTP8x1/DTP8x0 · 平均延迟
**变更：**将主配置镜像与两个配方从本地构建的开发镜像（构建提交未知，426e59f 不在上游 vLLM 中，构建于 2026-07-26，CUDA 13.0.1，FlashInfer 0.6.14）更新为 vLLM v0.29.0 正式版（提交 98dff2a8，CUDA 13.0.2，FlashInfer 0.6.18，torch 2.13.0），所有参数保持不变：v0.29.0 源码仍接受 FLASHINFER_MLA_SPARSE_DSV4、已弃用的 use_fp4_indexer_cache 别名（https://github.com/vllm-project/vllm/pull/52550）与 VLLM_PREFIX_CACHE_RETENTION_INTERVAL 环境变量（https://github.com/vllm-project/vllm/pull/52216），MTP 在强制启用的 V2 runner 上受支持，且其 DP 哈希已忽略 numa_bind，因此 srt-slurm v1.0.45 的 setup 补丁会被跳过。实测数值及异常说明见上表。
**下一步：**修复 1：将两个配方的 gpu-memory-utilization 从 0.94 降至 0.90，为稀疏索引器工作区恢复运行时显存余量，然后重新运行冒烟测试。

[Klaud-Cold]

Capacity-deferred · 2026-09-12 01:47 UTC
Initial attempt (Run 34662875814 / attempt 1) failed: c1 hit a CUDA OOM on TP rank 7 while allocating the sparse-indexer logits workspace after vLLM v0.29.0 sized the KV cache to 62.5 GiB at gpu-memory-utilization 0.94; c8 was cancelled by fail-fast.
Repair 1 (2b1f6386ae1b, gpu-memory-utilization 0.94 → 0.90 in both recipes) is pushed but was not dispatched: the gb200 capacity gate failed on two consecutive checks. Owned runs are terminal; the PR is closed and the branch released so a later autosweep can retry.

中文

容量受限延期 · 2026-09-12 01:47 UTC
初次尝试（Run 34662875814 / attempt 1）失败：vLLM v0.29.0 在 gpu-memory-utilization 0.94 下将 KV 缓存扩大到 62.5 GiB，c1 在 TP rank 7 分配稀疏索引器 logits 工作区时发生 CUDA OOM；c8 被 fail-fast 取消。
修复 1（2b1f6386ae1b，两个配方的 gpu-memory-utilization 由 0.94 降至 0.90）已推送但未派发：gb200 容量检查连续两次未通过。自有运行均已结束；PR 关闭并释放分支，以便后续自动巡检重试。

[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: 1 · Runs: 34662875814
All owned runs ended. PR closed; branch deleted for retry.

中文

capacity-deferred · 修复次数：1 · 运行：34662875814
所有自有运行均已结束。PR 已关闭；分支已删除，可重新尝试。