[Klaud-Cold]

Goal: Update vLLM image from vllm/vllm-openai-rocm:nightly-e9757321527ca1ecd514c07c1418dd2c53da3d19 to vllm/vllm-openai-rocm:nightly-5d35ce3934a7804b298559d2d99eec33792e1eb9.
Baseline: 2026-09-24 · vllm/vllm-openai-rocm:nightly-e9757321527ca1ecd514c07c1418dd2c53da3d19
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 8ccc66	1,988.43	14.94	1,144.79	4.38
AgentX c4 TP8 EP1 d833aa	3,329.2	24.8	775.34	5.12
AgentX c8 TP8 EP1 2b3c37	5,870.41	44.55	701.82	6.13
AgentX c16 TP8 EP1 0a4d00	10,422.63	74.78	1,093.06	8.96
AgentX c32 TP8 EP1 2b9041	18,197.64	118.41	1,509.48	15.69
AgentX c40 TP8 EP1 23d2c9	18,874.99	140.15	1,489.52	20.19
AgentX c48 TP8 EP1 c9ac7c	19,113.15	134.72	2,672.62	26.11
AgentX c56 TP8 EP1 518617	15,699.21	120.42	5,652.87	39.09
AgentX c64 TP8 EP1 206f05	20,458.87	167.87	2,460.68	38.61
AgentX c64 TP8 EP1 db46f3	10,325.15	72.91	24,374.65	65.21
AgentX c128 TP8 EP1 e2dfca	22,960.83	195.05	4,559.45	86.96
AgentX c192 TP8 EP1 4dbcb1	26,109.86	196.48	6,175.66	125.92

Eval	Score ↑	Samples
gsm8k/em_strict · c192	97.65%	1,319

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-rocm:nightly-e9757321527ca1ecd514c07c1418dd2c53da3d19 更新为 vllm/vllm-openai-rocm:nightly-5d35ce3934a7804b298559d2d99eec33792e1eb9。
**基线：**2026-09-24 · vllm/vllm-openai-rocm:nightly-e9757321527ca1ecd514c07c1418dd2c53da3d19
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 40503448191 / attempt 1 · 2026-10-27 14:56 UTC
vllm/vllm-openai-rocm:nightly-5d35ce3934a7804b298559d2d99eec33792e1eb9@sha256:0c030ba9e2e5fd2f45f2a048a1c8bb239d9674b63297166b79e258b4ce65c6cc · ccd4fcd15796 · AgentX · TP8/EP1 · Mean latency
Change: At measured ref 7095711, pin nightly-5d35ce39 by digest (ROCm SDK 10.0.0, AITER v0.1.26, Triton 3.8) over nightly-e9757321 (ROCm 7.2.3, AITER v0.1.22.post1); the vLLM e9757321...5d35ce39 source diff keeps every recipe flag and env valid, so no recipe change; VLLM_PREFIX_CACHE_RETENTION_INTERVAL is unread by both builds, and the old tag is gone from Docker Hub, so its stack is inferred from the vLLM Dockerfile at e9757321.

Concurrency	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1	15.87 (N/A)	947.2 (N/A)	4.03 (N/A)
64	170.64 (N/A)	2,418.66 (N/A)	38.43 (N/A)

Note: All rows: Δ N/A: no matched baseline.
Note: c64: 1 request error.

Eval	Score ↑	Samples
gsm8k/em_strict · c192 TP8 DPA	97.35% (-0.30 pp)	1,319 each
gsm8k/em_strict · c64 TP8	97.27% (N/A)	1,319

Note: gsm8k/em_strict · c64 TP8: Δ N/A: no matched eval baseline.

Next: Smoke startup, both throughput points and both gsm8k evals passed (one InvalidInferenceResultError in 5,135 c64 requests; hipBLASLt fell back to hipBLAS with warnings only), so append the perf-changelog entry and run the full final sweep.

中文

初次尝试 · 已通过 · Run 40503448191 / attempt 1 · 2026-10-27 14:56 UTC
vllm/vllm-openai-rocm:nightly-5d35ce3934a7804b298559d2d99eec33792e1eb9@sha256:0c030ba9e2e5fd2f45f2a048a1c8bb239d9674b63297166b79e258b4ce65c6cc · ccd4fcd15796 · AgentX · TP8/EP1 · 平均延迟
**变更：**在实测 ref 7095711 上，按 digest 固定 nightly-5d35ce39（ROCm SDK 10.0.0、AITER v0.1.26、Triton 3.8），替换 nightly-e9757321（ROCm 7.2.3、AITER v0.1.22.post1）；vLLM e9757321...5d35ce39 源码对比显示配方的所有参数与环境变量仍然有效，因此配方不变；两个版本都不读取 VLLM_PREFIX_CACHE_RETENTION_INTERVAL，旧标签已不在 Docker Hub 上，其软件栈依据 e9757321 的 vLLM Dockerfile 推断。实测数值及异常说明见上表。
**下一步：**smoke 的服务启动、两个吞吐量点和两个 gsm8k 评测均已通过（c64 的 5,135 个请求中有 1 个 InvalidInferenceResultError；hipBLASLt 仅以警告形式回退到 hipBLAS），因此追加 perf-changelog 条目并运行完整的最终扫描。

[Klaud-Cold]

parked · Session 1 is waiting on 40503448191. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40503448191。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 40526180403 / attempt 1 · 2026-10-27 21:53 UTC
vllm/vllm-openai-rocm:nightly-5d35ce3934a7804b298559d2d99eec33792e1eb9@sha256:0c030ba9e2e5fd2f45f2a048a1c8bb239d9674b63297166b79e258b4ce65c6cc · 29655ec51977 · Mean latency
Change: Validate the complete updated-image family at this head.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP8 EP1 8ccc66	15.86 (+6.2%)	950.3 (-17%)	3.97 (-9.4%)
AgentX c4 TP8 EP1 d833aa	24.72 (-0.3%)	729.14 (-6%)	4.78 (-6.6%)
AgentX c8 TP8 EP1 2b3c37	45.31 (+1.7%)	659.5 (-6%)	5.44 (-11.3%)
AgentX c16 TP8 EP1 0a4d00	79.22 (+5.9%)	825.51 (-24.5%)	7.51 (-16.2%)
AgentX c32 TP8 EP1 2b9041	135 (+14%)	1,152.7 (-23.6%)	11.64 (-25.8%)
AgentX c40 TP8 EP1 23d2c9	164.55 (+17.4%)	1,286.59 (-13.6%)	13.32 (-34%)
AgentX c48 TP8 EP1 c9ac7c	178.81 (+32.7%)	1,599.49 (-40.2%)	16.46 (-37%)
AgentX c56 TP8 EP1 518617	201.68 (+67.5%)	1,749.03 (-69.1%)	20.93 (-46.5%)
AgentX c64 TP8 EP1 206f05	175.43 (+4.5%)	2,390.6 (-2.8%)	36.27 (-6.1%)
AgentX c64 TP8 EP1 db46f3	199.37 (+173.5%)	3,087.74 (-87.3%)	27.57 (-57.7%)
AgentX c128 TP8 EP1 e2dfca	215.03 (+10.2%)	4,042.16 (-11.3%)	80.76 (-7.1%)
AgentX c192 TP8 EP1 4dbcb1	232.63 (+18.4%)	5,054.97 (-18.1%)	106.5 (-15.4%)

Note: AgentX c32 TP8 EP1 2b9041: 1 request error.
Note: AgentX c48 TP8 EP1 c9ac7c, AgentX c56 TP8 EP1 518617, AgentX c64 TP8 EP1 db46f3: 2 request errors.

Eval	Score ↑	Samples
gsm8k/em_strict · c64	97.04% (N/A)	1,319
gsm8k/em_strict · c192	97.04% (-0.61 pp)	1,319 each

Note: gsm8k/em_strict · c64: Δ N/A: no matched eval baseline.

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40526180403 / attempt 1 · 2026-10-27 21:53 UTC
vllm/vllm-openai-rocm:nightly-5d35ce3934a7804b298559d2d99eec33792e1eb9@sha256:0c030ba9e2e5fd2f45f2a048a1c8bb239d9674b63297166b79e258b4ce65c6cc · 29655ec51977 · 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 40526180403. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40526180403。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=40526180403

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40526180403

[Klaud-Cold]

validated · Repairs: unknown · Runs: 40503448191, 40526089554, 40526180403
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数：未知 · 运行：40503448191, 40526089554, 40526180403
所有自有运行均已结束。完整 sweep 已验证；已就绪，等待审查。

[claude]

I reviewed this image bump and didn't find any bugs. amd-master.yaml has designated CODEOWNERS, though, so one of them should still sign off before merge.

What was reviewed:

• The image on dsv4-fp4-mi355x-vllm-agentic-mtp now uses a tag plus a sha256 digest, the same form other entries in amd-master.yaml already use.
• The changelog change adds one block at the end of the file and leaves every existing byte alone. It uses kebab-case fields, lists the same config key whose image changed, sets scenario-type agentic-coding, and has an English-only description.

Extended reasoning...

The change edits one image line in inferencex-e2e/configs/amd-master.yaml and adds one 8-line block to the end of inferencex-e2e/perf-changelog.yaml. It has no security-sensitive surface: it only swaps a container image reference for a benchmark config. The change is mechanical and correct. I deferred only because .github/CODEOWNERS gives amd-master.yaml to specific AMD owners, so a human owner needs to sign off.