[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:dev-muse-spark-open (digest 8cfc5a72) to lmsysorg/sglang:nightly-dev-20261026-8252e3af (digest b4e4abb9).
Baseline: 2026-10-22 · lmsysorg/sglang:dev-muse-spark-open@sha256:8cfc5a72cb296771f5f8a4a4688cb6d05685610a830cbb616fa4ab0eb7bb2b0a
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 c3 PTP8x1/DTP8x0 f0db5e	3,638.57	30.25	1,380.27	2
AgentX c4 PTP8x1/DTP8x0 ec1697	4,661.62	31.84	571.75	2.08
AgentX c8 PTP4x1/DTP4x0 b47a81	14,851.49	107.36	554.12	3.21
AgentX c10 PTP8x1/DTP8x0 f5ffc4	9,024.42	66.5	1,656.93	2.26
AgentX c14 PTP8x1/DTP8x0 437c07	11,135.55	83.67	436.74	2.62
AgentX c16 PTP4x1/DTP4x0 bbd209	23,999.46	158.76	3,720.78	3.88
AgentX c20 PTP4x1/DTP4x0 cce964	32,765.88	219.9	1,657.14	5.05
AgentX c32 PTP4x1/DTP4x0 35ecb4	43,242.04	267.46	5,954.68	6.73
AgentX c40 PTP4x1/DTP4x0 7615bf	50,593.68	350.41	1,958.19	9.85

Eval	Score ↑	Samples
gsm8k/em_strict · c16	97.88%	1,319
gsm8k/em_strict · c20	97.95%	1,319
gsm8k/em_strict · c32	97.73%	1,319
gsm8k/em_strict · c40	97.42%	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 SGLang 镜像从 lmsysorg/sglang:dev-muse-spark-open（digest 8cfc5a72）更新为 lmsysorg/sglang:nightly-dev-20261026-8252e3af（digest b4e4abb9）。
**基线：**2026-10-22 · lmsysorg/sglang:dev-muse-spark-open@sha256:8cfc5a72cb296771f5f8a4a4688cb6d05685610a830cbb616fa4ab0eb7bb2b0a
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 · Queued · Run 40426236904 / attempt 1 · 2026-10-27 00:26 UTC
lmsysorg/sglang:nightly-dev-20261026-8252e3af@sha256:b4e4abb9e1dc038f7e14c85805be97e4f74abb46e541bd4edd31401d50a2e593 · 4ea8f9a3b51a · Mean latency
Change: Measure branch head 8a2db71, pinning the master image to nightly-dev-20261026-8252e3af by digest with no recipe change; Muse Spark support is upstream in SGLang 8252e3af (github.com/sgl-project/sglang/compare/bc1c2ac7...8252e3af), every recipe flag and env var still parses there, and MegaMoE no longer needs the old image's patched DeepGEMM, though the old image's exact muse-spark source commit is unlabeled (built 2026-10-14 on v0.5.21, closest branch sglang-miles-muse-spark-open).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c16 TP4 EP4 e64c8c	N/A	N/A	N/A
AgentX c8 TP4 EP4 c492ba	N/A	N/A	N/A
AgentX c14 TP8 EP8 c28d6c	N/A	N/A	N/A
AgentX c10 TP8 EP8 b7ee88	N/A	N/A	N/A
AgentX c4 TP8 EP8 eece3a	N/A	N/A	N/A
AgentX c3 TP8 EP8 1ac109	N/A	N/A	N/A
AgentX c40 TP4 EP4 HiCache e9e219	N/A	N/A	N/A
AgentX c32 TP4 EP4 HiCache e310a4	N/A	N/A	N/A
AgentX c20 TP4 EP4 HiCache 81351b	N/A	N/A	N/A

Note: All rows: pending.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: no matched baseline.

Next: Wait for the queued smoke run of all nine points and four GSM8K evals, then append the perf-changelog entry if they pass.

中文

初次尝试 · 排队中 · Run 40426236904 / attempt 1 · 2026-10-27 00:26 UTC
lmsysorg/sglang:nightly-dev-20261026-8252e3af@sha256:b4e4abb9e1dc038f7e14c85805be97e4f74abb46e541bd4edd31401d50a2e593 · 4ea8f9a3b51a · 平均延迟
**变更：**测量分支 head 8a2db71，以 digest 将主镜像固定为 nightly-dev-20261026-8252e3af，recipe 不变；Muse Spark 支持已合入 SGLang 8252e3af（github.com/sgl-project/sglang/compare/bc1c2ac7...8252e3af），recipe 使用的全部参数和环境变量在该版本中仍可解析，MegaMoE 也不再依赖旧镜像中打过补丁的 DeepGEMM；但旧镜像的确切 muse-spark 源码提交未标注（2026-10-14 基于 v0.5.21 构建，最接近的分支为 sglang-miles-muse-spark-open）。实测数值及异常说明见上表。
**下一步：**等待排队中的 smoke 运行（九个点和四个 GSM8K 评测）完成，若通过则追加 perf-changelog 条目。

[Klaud-Cold]

parked · Session 1 is waiting on 40426236904. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40426236904。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]

Initial attempt · Passed · Run 40426236904 / attempt 3 · 2026-10-27 07:48 UTC
lmsysorg/sglang:nightly-dev-20261026-8252e3af@sha256:b4e4abb9e1dc038f7e14c85805be97e4f74abb46e541bd4edd31401d50a2e593 · 4ea8f9a3b51a · Mean latency
Change: Measure branch head 8a2db71, pinning the master image to nightly-dev-20261026-8252e3af by digest with no recipe change; Muse Spark support is upstream in SGLang 8252e3af (github.com/sgl-project/sglang/compare/bc1c2ac7...8252e3af), every recipe flag and env var still parses there, and MegaMoE no longer needs the old image's patched DeepGEMM, though the old image's exact muse-spark source commit is unlabeled (built 2026-10-14 on v0.5.21, closest branch sglang-miles-muse-spark-open).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c16 TP4 EP4 e64c8c	157.52 (N/A)	3,934.87 (N/A)	3.87 (N/A)
AgentX c8 TP4 EP4 c492ba	107.69 (N/A)	540.88 (N/A)	3.24 (N/A)
AgentX c14 TP8 EP8 c28d6c	83.48 (N/A)	460.6 (N/A)	2.64 (N/A)
AgentX c10 TP8 EP8 b7ee88	67.08 (N/A)	1,547.5 (N/A)	2.29 (N/A)
AgentX c4 TP8 EP8 eece3a	32.11 (N/A)	502.77 (N/A)	2.05 (N/A)
AgentX c3 TP8 EP8 1ac109	29.66 (N/A)	1,296.96 (N/A)	1.98 (N/A)
AgentX c40 TP4 EP4 HiCache e9e219	353.5 (N/A)	1,799.55 (N/A)	9.48 (N/A)
AgentX c32 TP4 EP4 HiCache e310a4	263.63 (N/A)	6,081.72 (N/A)	6.79 (N/A)
AgentX c20 TP4 EP4 HiCache 81351b	218.83 (N/A)	1,686.7 (N/A)	5.12 (N/A)

Note: All rows: Δ N/A: no matched baseline.
Note: AgentX c40 TP4 EP4 HiCache e9e219: 2 request errors.
Note: AgentX c32 TP4 EP4 HiCache e310a4: 1 request error.

Eval	Score ↑	Samples
gsm8k/em_strict · gsm8k/em_strict c16	97.8% (-0.08 pp)	1,319 each
gsm8k/em_strict · gsm8k/em_strict c20	97.8% (-0.15 pp)	1,319 each
gsm8k/em_strict · gsm8k/em_strict c32	97.65% (-0.08 pp)	1,319 each
gsm8k/em_strict · gsm8k/em_strict c40	97.5% (+0.08 pp)	1,319 each

Next: Smoke points and GSM8K evals passed on run attempt 3 after two scheduler preemptions, so append the perf-changelog entry and dispatch the full-sweep-fail-fast final sweep.

中文

初次尝试 · 已通过 · Run 40426236904 / attempt 3 · 2026-10-27 07:48 UTC
lmsysorg/sglang:nightly-dev-20261026-8252e3af@sha256:b4e4abb9e1dc038f7e14c85805be97e4f74abb46e541bd4edd31401d50a2e593 · 4ea8f9a3b51a · 平均延迟
**变更：**测量分支 head 8a2db71，以 digest 将主镜像固定为 nightly-dev-20261026-8252e3af，recipe 不变；Muse Spark 支持已合入 SGLang 8252e3af（github.com/sgl-project/sglang/compare/bc1c2ac7...8252e3af），recipe 使用的全部参数和环境变量在该版本中仍可解析，MegaMoE 也不再依赖旧镜像中打过补丁的 DeepGEMM；但旧镜像的确切 muse-spark 源码提交未标注（2026-10-14 基于 v0.5.21 构建，最接近的分支为 sglang-miles-muse-spark-open）。实测数值及异常说明见上表。
**下一步：**在两次调度抢占后，run attempt 3 上的 smoke 点和 GSM8K 评测均已通过，因此追加 perf-changelog 条目并启动 full-sweep-fail-fast 最终扫描。

[Klaud-Cold]

Final full sweep · Passed · Run 40469520772 / attempt 1 · 2026-10-27 09:35 UTC
lmsysorg/sglang:nightly-dev-20261026-8252e3af@sha256:b4e4abb9e1dc038f7e14c85805be97e4f74abb46e541bd4edd31401d50a2e593 · 212b89552c8a · Mean latency
Change: Validate the complete updated-image family at this head.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c3 PTP8x1/DTP8x0 f0db5e	29.75 (-1.6%)	1,184.54 (-14.2%)	1.99 (-0.5%)
AgentX c4 PTP8x1/DTP8x0 ec1697	32.45 (+1.9%)	667.87 (+16.8%)	2.06 (-1%)
AgentX c8 PTP4x1/DTP4x0 b47a81	107.37 (+0%)	545.08 (-1.6%)	3.24 (+0.9%)
AgentX c10 PTP8x1/DTP8x0 f5ffc4	66.91 (+0.6%)	1,601.17 (-3.4%)	2.28 (+0.9%)
AgentX c14 PTP8x1/DTP8x0 437c07	83.44 (-0.3%)	454.86 (+4.1%)	2.61 (-0.4%)
AgentX c16 PTP4x1/DTP4x0 bbd209	158.7 (-0%)	3,735.43 (+0.4%)	3.84 (-1%)
AgentX c20 PTP4x1/DTP4x0 cce964	219.39 (-0.2%)	1,672.64 (+0.9%)	5.17 (+2.4%)
AgentX c32 PTP4x1/DTP4x0 35ecb4	268.39 (+0.3%)	5,804.64 (-2.5%)	6.6 (-1.9%)
AgentX c40 PTP4x1/DTP4x0 7615bf	352.42 (+0.6%)	1,863.55 (-4.8%)	9.53 (-3.2%)

Note: AgentX c14 PTP8x1/DTP8x0 437c07, AgentX c32 PTP4x1/DTP4x0 35ecb4: 1 request error.
Note: AgentX c40 PTP4x1/DTP4x0 7615bf: 2 request errors.

Eval	Score ↑	Samples
gsm8k/em_strict · c16	97.5% (-0.38 pp)	1,319 each
gsm8k/em_strict · c20	97.65% (-0.30 pp)	1,319 each
gsm8k/em_strict · c32	97.65% (-0.08 pp)	1,319 each
gsm8k/em_strict · c40	97.95% (+0.53 pp)	1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40469520772 / attempt 1 · 2026-10-27 09:35 UTC
lmsysorg/sglang:nightly-dev-20261026-8252e3af@sha256:b4e4abb9e1dc038f7e14c85805be97e4f74abb46e541bd4edd31401d50a2e593 · 212b89552c8a · 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 40469520772. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40469520772。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=40469520772

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40469520772

[Klaud-Cold]

validated · Repairs: unknown · Runs: 40426236904, 40469445594, 40469520772
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数：未知 · 运行：40426236904, 40469445594, 40469520772
所有自有运行均已结束。完整 sweep 已验证；已就绪，等待审查。

[claude]

Claude finished @Klaud-Cold's task in 18s —— View job

---

LGTM - no blocking issues found

The image is a public Docker Hub reference pinned by digest. The perf-changelog entry is appended at the end of the file and covers the changed config key.

[claude]

Code review found no issues

No high-confidence issues detected in this change.