[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:nightly-dev-cu13-20260913-fa663e72 to lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae.
Baseline: 2026-10-09 · lmsysorg/sglang:nightly-dev-cu13-20260913-fa663e72
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 c100 PTP1x1/DTP4x1 86e6a6	N/A	N/A	N/A	N/A
AgentX c200 PTP1x1/DTP4x1 353efc	85,309.99	818.1	8,871.06	10.13
AgentX c600 PTP1x3/DTP4x2 3e58dc	N/A	N/A	N/A	N/A
AgentX c1000 PTP1x5/DTP4x3 c65354	101,367.28	972.04	8,217.9	13.46

Note: AgentX c100 PTP1x1/DTP4x1 86e6a6, AgentX c600 PTP1x3/DTP4x2 3e58dc: unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c200	97.57%	1,319
gsm8k/em_strict · c1000	97.73%	1,319

Stop: Add the klaud-abort label at any time to stop this candidate now; Klaud cancels its runs, closes this PR and does not retry it. Remove the label to let the planner pick it up again.

中文

**目标：**将 SGLang 镜像从 lmsysorg/sglang:nightly-dev-cu13-20260913-fa663e72 更新为 lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae。
**基线：**2026-10-09 · lmsysorg/sglang:nightly-dev-cu13-20260913-fa663e72
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：**随时添加 klaud-abort 标签即可立即停止该候选：Klaud 会取消其运行、关闭此 PR 且不再重试。移除该标签后，规划器会在下次扫描时重新考虑它。

[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 · Failed · Run 40071590738 / attempt 1 · 2026-10-24 08:30 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 601e531e49a8 · Mean latency
Change: Only the family master image changes, pinned by digest, because the v0.5.22 source keeps every recipe flag and value choice and already contains the old nightly commit fa663e72; v0.5.22 drops the deprecated SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK alias, so the TP memory-balance check is re-enabled, but the shared disagg-variants.yaml recipe is left unchanged (sources: sgl-project/sglang fa663e72 and v0.5.22 environ.py, distributed/bootstrap.py, arg_groups).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1000 PTP1x5/DTP4x3 c65354	N/A	N/A	N/A
AgentX c600 PTP1x3/DTP4x2 3e58dc	N/A	N/A	N/A
AgentX c200 PTP1x1/DTP4x1 353efc	N/A	N/A	N/A
AgentX c100 PTP1x1/DTP4x1 86e6a6	N/A	N/A	N/A

Note: AgentX c1000 PTP1x5/DTP4x3 c65354, AgentX c600 PTP1x3/DTP4x2 3e58dc, AgentX c200 PTP1x1/DTP4x1 353efc: cancelled.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.
Note: AgentX c100 PTP1x1/DTP4x1 86e6a6: failed.

Eval	Score ↑	Samples
gsm8k/em_strict · c100	97.73% (N/A)	1,319

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

Next: Classified as infrastructure: c100 and the c200 eval both ran on batch3 runner gb300-nv_18 (r03 nodes), whose host exports DYN_TCP_RPC_HOST=enP22s22f0np0, an interface name the family's pinned Dynamo ad6f7d54 runtime 1.4.2 rejects with Invalid TCP bind address, while the c100 eval on gb300-nv_03 served and scored gsm8k normally and baseline run 37833971361 had no such variable; rerun the failed and cancelled jobs on the same head as infrastructure retry 1.

中文

初次尝试 · 失败 · Run 40071590738 / attempt 1 · 2026-10-24 08:30 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 601e531e49a8 · 平均延迟
**变更：**仅更新该配置族的 master 镜像并按摘要固定，因为 v0.5.22 源码保留了配方使用的所有参数及取值，且已包含旧 nightly 提交 fa663e72；v0.5.22 移除了已弃用的 SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK 别名，TP 内存均衡检查因此重新启用，但共享的 disagg-variants.yaml 配方保持不变（来源：sgl-project/sglang fa663e72 与 v0.5.22 的 environ.py、distributed/bootstrap.py、arg_groups）。实测数值及异常说明见上表。
**下一步：**归类为基础设施问题：c100 与 c200 评测都运行在 batch3 runner gb300-nv_18（r03 节点）上，其主机导出 DYN_TCP_RPC_HOST=enP22s22f0np0，这是网卡名，该配置族固定的 Dynamo ad6f7d54 runtime 1.4.2 会以 Invalid TCP bind address 拒绝；而在 gb300-nv_03 上的 c100 评测正常服务并完成 gsm8k 评分，基线运行 37833971361 也没有该变量；在同一 head 上以基础设施重试 1 重跑失败和被取消的作业。

[Klaud-Cold]

Infrastructure retry 1/2 · Running · Run 40071590738 / attempt 2 · 2026-10-24 12:27 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 601e531e49a8 · Mean latency
Change: No change: the same head and image are rerun because the first failure came from the batch3 runner host environment, not the recipe or image.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1000 PTP1x5/DTP4x3 c65354	N/A	N/A	N/A
AgentX c600 PTP1x3/DTP4x2 3e58dc	1,022.05 (N/A)	3,700.34 (N/A)	14.27 (N/A)
AgentX c200 PTP1x1/DTP4x1 353efc	902.56 (+10.3%)	3,724.3 (-58%)	11.14 (+10%)
AgentX c100 PTP1x1/DTP4x1 86e6a6	532.67 (N/A)	1,635 (N/A)	8.65 (N/A)

Note: AgentX c1000 PTP1x5/DTP4x3 c65354: pending.
Note: AgentX c1000 PTP1x5/DTP4x3 c65354: request errors unavailable.
Note: AgentX c1000 PTP1x5/DTP4x3 c65354, AgentX c600 PTP1x3/DTP4x2 3e58dc, AgentX c100 PTP1x1/DTP4x1 86e6a6: Δ N/A: point unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c100	97.73% (N/A)	1,319
gsm8k/em_strict · c200	97.73% (N/A)	1,319
gsm8k/em_strict · c600	97.5% (N/A)	1,319

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

Next: The c1000 eval failed again only on runner gb300-nv_18 (r03 nodes) with the same Dynamo Invalid TCP bind address host-environment error while every job on other runners passed; wait for the running c1000 benchmark, then rerun the failed c1000 eval on the same head as infrastructure retry 2.

中文

基础设施重试 1/2 · 运行中 · Run 40071590738 / attempt 2 · 2026-10-24 12:27 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 601e531e49a8 · 平均延迟
**变更：**无改动：在同一 head 和镜像上重跑，因为首次失败来自 batch3 runner 的主机环境，而非配方或镜像。实测数值及异常说明见上表。
**下一步：**c1000 评测再次仅在 runner gb300-nv_18（r03 节点）上以相同的 Dynamo Invalid TCP bind address 主机环境错误失败，而其他 runner 上的所有作业均通过；等待正在运行的 c1000 基准测试完成后，在同一 head 上以基础设施重试 2 重跑失败的 c1000 评测。

[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]

Infrastructure retry 2/2 · Running · Run 40071590738 / attempt 3 · 2026-10-24 12:58 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 601e531e49a8 · Mean latency
Change: No change: the same head and image are rerun because the c1000 eval failure came from the gb300-nv_18 runner host environment, not the recipe or image.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1000 PTP1x5/DTP4x3 c65354	1,026.94 (+5.6%)	3,796.35 (-53.8%)	15.63 (+16.1%)
AgentX c600 PTP1x3/DTP4x2 3e58dc	1,022.05 (N/A)	3,700.34 (N/A)	14.27 (N/A)
AgentX c200 PTP1x1/DTP4x1 353efc	902.56 (+10.3%)	3,724.3 (-58%)	11.14 (+10%)
AgentX c100 PTP1x1/DTP4x1 86e6a6	532.67 (N/A)	1,635 (N/A)	8.65 (N/A)

Note: AgentX c600 PTP1x3/DTP4x2 3e58dc, AgentX c100 PTP1x1/DTP4x1 86e6a6: Δ N/A: point unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c100	97.73% (N/A)	1,319
gsm8k/em_strict · c200	97.73% (N/A)	1,319
gsm8k/em_strict · c600	97.5% (N/A)	1,319

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

Next: Wait for the rerun c1000 eval; the four passed smoke benchmarks and the c100, c200 and c600 evals from earlier run attempts are reused.

中文

基础设施重试 2/2 · 运行中 · Run 40071590738 / attempt 3 · 2026-10-24 12:58 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 601e531e49a8 · 平均延迟
**变更：**无改动：在同一 head 和镜像上重跑，因为 c1000 评测失败来自 gb300-nv_18 runner 的主机环境，而非配方或镜像。实测数值及异常说明见上表。
**下一步：**等待重跑的 c1000 评测完成；复用之前运行尝试中已通过的四个冒烟基准测试以及 c100、c200、c600 评测。