[Klaud-Cold]

Goal: Update vLLM image from vllm/vllm-openai-rocm:nightly-e68efc685ab1e6f1eda004ced7157e3ca411da2f@sha256:520e098f54ec96837228aaddea0a833cbf6374dbdffd480c1d65f26ffab7701b to vllm/vllm-openai-rocm:nightly-335dd77b09161c5d63139b4e7ae51bb9928305b1@sha256:3c634cc2c51eda7eb840a236b1f09a5ecf6a4b0cb92c4b8486e2158770507291.
Baseline: 2026-10-27 · vllm/vllm-openai-rocm:nightly-e68efc685ab1e6f1eda004ced7157e3ca411da2f@sha256:520e098f54ec96837228aaddea0a833cbf6374dbdffd480c1d65f26ffab7701b
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 TP2 EP1 84969a	13,464.57	85.66	571.2	2.01
AgentX c1 TP4 EP1 c615ef	7,398.85	45.12	504.19	1.68
AgentX c2 TP2 EP1 3db9be	13,194.41	100.81	900.39	2.02
AgentX c2 TP4 EP1 b6fe6a	7,216.13	51.41	604.26	1.69
AgentX c4 TP2 EP1 4ae934	18,315.77	125.12	1,467.19	2.04
AgentX c8 TP2 EP1 2c5190	31,600.96	225.08	962.7	2.26
AgentX c16 TP2 EP1 acff22	57,244.35	364.2	899.79	3.05
AgentX c32 TP2 EP1 a12d09	110,462.16	635.85	2,884.76	4.72
AgentX c64 TP2 EP1 5d2c55	186,280.24	1,350.73	384.06	11.13

Eval	Score ↑	Samples
gsm8k/em_strict · c64	97.12%	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-e68efc685ab1e6f1eda004ced7157e3ca411da2f@sha256:520e098f54ec96837228aaddea0a833cbf6374dbdffd480c1d65f26ffab7701b 更新为 vllm/vllm-openai-rocm:nightly-335dd77b09161c5d63139b4e7ae51bb9928305b1@sha256:3c634cc2c51eda7eb840a236b1f09a5ecf6a4b0cb92c4b8486e2158770507291。
**基线：**2026-10-27 · vllm/vllm-openai-rocm:nightly-e68efc685ab1e6f1eda004ced7157e3ca411da2f@sha256:520e098f54ec96837228aaddea0a833cbf6374dbdffd480c1d65f26ffab7701b
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 41123416650 / attempt 1 · 2026-10-30 21:07 UTC
vllm/vllm-openai-rocm:nightly-335dd77b09161c5d63139b4e7ae51bb9928305b1@sha256:3c634cc2c51eda7eb840a236b1f09a5ecf6a4b0cb92c4b8486e2158770507291 · 59fbe34e18a7 · Mean latency
Change: Moved the master image to the vLLM ROCm nightly at vLLM commit 335dd77b (845 commits after e68efc68, base image digest also changed); recipe flags unchanged because a source diff of arg_utils, config/kernel.py (monokernel), config/attention.py (indexer_kv_dtype, indexer_sparse_logits now tri-state with explicit True still honored on gfx950), config/speculative.py (dspark, probabilistic drafting, block rejection) and config/engram.py (cpu_offload) found every used option still present with unchanged semantics; uncertainty remains around the new ROCm/AITER base layer and new mono decode and SWA metadata kernels (vllm-project/vllm#63885, #64101, #63689).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP2 EP1 84969a	84.19 (-1.7%)	707.7 (+23.9%)	1.8 (-10.4%)
AgentX c1 TP4 EP1 c615ef	45.04 (-0.2%)	601.51 (+19.3%)	1.51 (-10.1%)

Eval	Score ↑	Samples
gsm8k/em_strict · c64	96.89% (-0.23 pp)	1,319 each

Next: Smoke points and the c64 gsm8k eval passed on head 77a7868, so append the perf-changelog entry, merge main and dispatch the final sweep.

中文

初次尝试 · 已通过 · Run 41123416650 / attempt 1 · 2026-10-30 21:07 UTC
vllm/vllm-openai-rocm:nightly-335dd77b09161c5d63139b4e7ae51bb9928305b1@sha256:3c634cc2c51eda7eb840a236b1f09a5ecf6a4b0cb92c4b8486e2158770507291 · 59fbe34e18a7 · 平均延迟
**变更：**将 master 镜像切换到 vLLM commit 335dd77b 的 vLLM ROCm nightly（比 e68efc68 新 845 个提交，基础镜像 digest 也已变化）；对 arg_utils、config/kernel.py（monokernel）、config/attention.py（indexer_kv_dtype，indexer_sparse_logits 改为三态但显式 True 在 gfx950 上仍生效）、config/speculative.py（dspark、probabilistic drafting、block rejection）和 config/engram.py（cpu_offload）的源码对比显示所用选项均保留且语义不变，因此 recipe 参数未改；新的 ROCm/AITER 基础层以及新的 mono decode 和 SWA 元数据 kernel（vllm-project/vllm#63885、#64101、#63689）仍存在不确定性。实测数值及异常说明见上表。
**下一步：**smoke 点和 c64 gsm8k 评测已在 head 77a7868 上通过，接下来追加 perf-changelog 条目、合并 main 并调度最终 sweep。

[Klaud-Cold]

parked · Session 1 is waiting on 41123416650. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 41123416650。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 41128818347 / attempt 1 · 2026-10-30 22:49 UTC
vllm/vllm-openai-rocm:nightly-335dd77b09161c5d63139b4e7ae51bb9928305b1@sha256:3c634cc2c51eda7eb840a236b1f09a5ecf6a4b0cb92c4b8486e2158770507291 · 0bd34344dc94 · Mean latency
Change: Validate the complete updated-image family at this head.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP2 EP1 84969a	88.59 (+3.4%)	446.98 (-21.7%)	1.8 (-10.4%)
AgentX c1 TP4 EP1 c615ef	46.84 (+3.8%)	381.06 (-24.4%)	1.51 (-10.1%)
AgentX c2 TP2 EP1 3db9be	105.94 (+5.1%)	745.09 (-17.2%)	1.81 (-10.4%)
AgentX c2 TP4 EP1 b6fe6a	56.26 (+9.4%)	645.62 (+6.8%)	1.52 (-10.1%)
AgentX c4 TP2 EP1 4ae934	126.25 (+0.9%)	1,245.77 (-15.1%)	1.82 (-10.8%)
AgentX c8 TP2 EP1 2c5190	227.49 (+1.1%)	842.75 (-12.5%)	2.17 (-4%)
AgentX c16 TP2 EP1 acff22	367.36 (+0.9%)	713.66 (-20.7%)	2.91 (-4.6%)
AgentX c32 TP2 EP1 a12d09	638.47 (+0.4%)	2,762.82 (-4.2%)	4.66 (-1.3%)
AgentX c64 TP2 EP1 5d2c55	1,386.2 (+2.6%)	309.01 (-19.5%)	10.44 (-6.2%)

Note: AgentX c1 TP4 EP1 c615ef, AgentX c32 TP2 EP1 a12d09: 1 request error.
Note: AgentX c64 TP2 EP1 5d2c55: 3 request errors.

Eval	Score ↑	Samples
gsm8k/em_strict · c64	97.42% (+0.30 pp)	1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 41128818347 / attempt 1 · 2026-10-30 22:49 UTC
vllm/vllm-openai-rocm:nightly-335dd77b09161c5d63139b4e7ae51bb9928305b1@sha256:3c634cc2c51eda7eb840a236b1f09a5ecf6a4b0cb92c4b8486e2158770507291 · 0bd34344dc94 · 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 41128818347. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 41128818347。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=41128818347

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=41128818347

[Klaud-Cold]

/use 41128818347

[Klaud-Cold]

validated · Repairs: unknown · Runs: 41123416650, 41128768985, 41128818347
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数：未知 · 运行：41123416650, 41128768985, 41128818347
所有自有运行均已结束。完整 sweep 已验证；已就绪，等待审查。

[github-actions]

@Klaud-Cold staged run 41128818347.

• DeepSeek-V4.1-Flash · agentic-traces · FP4 · mi355x_vllm · 2026-10-30

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

[claude]

Claude finished @Klaud-Cold's task in 12s —— View job

---

LGTM - no blocking issues found

[claude]

Code review found no issues

No high-confidence issues detected in this change.