[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:nightly-dev-20260916-c9a8fba9 to lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469.
Baseline: 2026-09-22 · lmsysorg/sglang:nightly-dev-20260916-c9a8fba9
Mean latency · Sources: API 1, API 2, API 3

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c64 PTP8x1/DTP8x1 f7c942	25,134.99	181.41	2,358.22	7.32
AgentX c128 PTP8x1/DTP8x1 55b477	38,023.26	305.47	4,445.61	8.47
AgentX c256 PTP8x2/DTP8x1 b5a0b5	47,447.77	345.79	6,016.56	11.22

Eval	Score ↑	Samples
gsm8k/em_strict · c64	97.65%	1,319
gsm8k/em_strict · c128	97.57%	1,319
gsm8k/em_strict · c256	97.57%	1,319

中文

**目标：**将 SGLang 镜像从 lmsysorg/sglang:nightly-dev-20260916-c9a8fba9 更新至 lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469。
**基线：**2026-09-22 · lmsysorg/sglang:nightly-dev-20260916-c9a8fba9
平均延迟 · 来源： API 1, API 2, API 3；数值及异常说明见上表。

[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 38721756264 / attempt 1 · 2026-10-15 15:38 UTC
lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 · b301e68225f8 · Mean latency
Change: Image-only bump at measured head 11db101 to SGLang v0.5.21 (release commit e00930c5, a descendant of c9a8fba9 per git merge-base) pinned by manifest digest; every recipe server arg still exists in sglang arg_groups at v0.5.21, the DeepSeek-V4.1-only DSpark PD check does not apply because DeepSeek-V4-Pro-0813 reports model_type deepseek_v4, megamoe now explicitly accepts SM100, and SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT is no longer read because sglang#40313 moved SWA leaf capping into the unified radix cache by default; v0.5.21 predates sglang#41818, so the pinned Dynamo 1.5.0.dev20260914 dp-attention path is unchanged.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c64 PTP8x1/DTP8x1 f7c942	179.5 (N/A)	2,430.61 (N/A)	7.19 (N/A)
AgentX c128 PTP8x1/DTP8x1 55b477	290.78 (N/A)	6,567.03 (N/A)	8.48 (N/A)
AgentX c256 PTP8x2/DTP8x1 b5a0b5	346.85 (N/A)	5,960.28 (N/A)	11.12 (N/A)

Note: All rows: Δ N/A: dataset/scenario/statistic mismatch.

Eval	Score ↑	Samples
gsm8k/em_strict · c64	97.5% (-0.15 pp)	1,319 each
gsm8k/em_strict · c128	97.57% (+0.00 pp)	1,319 each
gsm8k/em_strict · c256	97.35% (-0.23 pp)	1,319 each

Next: Final run 38756295960: the c128 gsm8k eval failed before scoring because the eval client hit a Hugging Face 429 API rate limit loading openai/gsm8k, and fail-fast cancelled the c64 eval; after the benchmark jobs finish, rerun the failed jobs on the same head as an infrastructure retry.

中文

初次尝试 · 已通过 · Run 38721756264 / attempt 1 · 2026-10-15 15:38 UTC
lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 · b301e68225f8 · 平均延迟
**变更：**在实测提交 11db101 上仅更新镜像至按 manifest 摘要固定的 SGLang v0.5.21（发布提交 e00930c5，经 git merge-base 确认为 c9a8fba9 的后代）；配方的所有服务器参数在 v0.5.21 的 sglang arg_groups 中仍然存在，仅针对 DeepSeek-V4.1 的 DSpark PD 校验不适用于 model_type 为 deepseek_v4 的 DeepSeek-V4-Pro-0813，megamoe 现已显式支持 SM100，且 sglang#40313 已将 SWA 叶子截断默认并入统一 radix cache，因此不再读取 SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT；v0.5.21 早于 sglang#41818，固定的 Dynamo 1.5.0.dev20260914 的 dp-attention 路径不受影响。实测数值及异常说明见上表。
**下一步：**最终运行 38756295960：c128 gsm8k 评测在评分前失败，原因是评测客户端加载 openai/gsm8k 时触发 Hugging Face 429 API 限流，fail-fast 随后取消了 c64 评测；待基准任务结束后，在同一提交上将失败任务作为基础设施重试重新运行。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=38756295960

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38756295960

[Klaud-Cold]

unexpected-error · Cleanup pending. Stop and confirm owned runs before closing.

中文

unexpected-error · 清理待完成。先停止并确认自有运行结束，再关闭 PR。

[Klaud-Cold]

unexpected-error · Repairs: unknown · Runs: 38721756264, 38756216808, 38756295960
All owned runs ended. PR closed; branch deleted for retry.

中文

unexpected-error · 修复次数：unknown · 运行：38721756264, 38756216808, 38756295960
所有自有运行均已结束。PR 已关闭；分支已删除，可重新尝试。