[Klaud Cold] Update dsr1-fp4-b200-sglang SGLang image to v0.5.21-cu130 / [Klaud Cold] 将 dsr1-fp4-b200-sglang 的 SGLang 镜像更新至 v0.5.21-cu130

[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:v0.5.19-cu130 to lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469.
Baseline: 2026-09-16 · lmsysorg/sglang:v0.5.19-cu130
Mean latency · Sources: API 1, API 2

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
8k/1k c1 TP4 EP1 3bfc5c	407.4	45.51	208.32	5.27
8k/1k c1 TP4 EP1 620784	N/A	N/A	N/A	N/A
8k/1k c2 TP4 EP1 2a3d39	N/A	N/A	N/A	N/A
8k/1k c2 TP4 EP1 bab69f	704.42	79.05	257.26	6.04
8k/1k c4 TP4 EP1 684a05	N/A	N/A	N/A	N/A
8k/1k c4 TP4 EP1 a8bdd9	1,179.46	131.21	288.36	7.15
8k/1k c8 TP4 EP1 06004a	N/A	N/A	N/A	N/A
8k/1k c8 TP4 EP1 425d06	1,828.31	205.53	379.49	9.17
8k/1k c16 TP4 EP1 83585a	N/A	N/A	N/A	N/A
8k/1k c16 TP4 EP1 af369a	2,584.54	286.38	650.3	12.92
8k/1k c32 TP4 EP1 bea97b	3,487.5	390	1,034.79	18.97
8k/1k c32 TP4 EP1 e5e256	N/A	N/A	N/A	N/A
8k/1k c64 TP4 EP4 502b5d	N/A	N/A	N/A	N/A
8k/1k c64 TP4 EP4 a9e5b1	4,918.77	545.69	1,319.72	27.24
8k/1k c128 TP4 EP4 c77207	6,637.64	735.22	1,834.97	40.75
8k/1k c128 TP4 EP4 e8c737	N/A	N/A	N/A	N/A
8k/1k c256 TP4 EP4 368d10	8,365.57	929.78	6,928.52	60.21
8k/1k c256 TP4 EP4 e506d2	N/A	N/A	N/A	N/A

Note: 8k/1k c1 TP4 EP1 620784, 8k/1k c2 TP4 EP1 2a3d39, 8k/1k c4 TP4 EP1 684a05, 8k/1k c8 TP4 EP1 06004a, 8k/1k c16 TP4 EP1 83585a, 8k/1k c32 TP4 EP1 e5e256, 8k/1k c64 TP4 EP4 502b5d, 8k/1k c128 TP4 EP4 e8c737, 8k/1k c256 TP4 EP4 e506d2: unavailable.

Eval: N/A

中文

**目标：**将 SGLang 镜像从 lmsysorg/sglang:v0.5.19-cu130 更新为 lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469。
**基线：**2026-09-16 · lmsysorg/sglang:v0.5.19-cu130
平均延迟 · 来源： API 1, API 2；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 37009882020 / attempt 1 · 2026-10-02 13:30 UTC
lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 · 15b01bb526bd · Mean latency
Change: Pin SGLang v0.5.21-cu130 by digest and replace the removed --cuda-graph-max-bs and --disable-piecewise-cuda-graph aliases with their v0.5.19 targets --cuda-graph-max-bs-decode and --cuda-graph-backend-prefill=disabled (v0.5.19 aliases, v0.5.21 fields); the bundled image source is commit e00930c5 and the old flags fail argparse there; measured ref d090548.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
8k/1k c1 TP4 EP1 620784	44.44 (N/A)	224.01 (N/A)	5.39 (N/A)
8k/1k c64 TP4 EP4 502b5d	547.16 (N/A)	1,315.66 (N/A)	27.16 (N/A)

Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c32	95.45% (N/A)	N/A/1,319 (old/new)
gsm8k/em_strict · c128	95.38% (N/A)	N/A/1,319 (old/new)

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

Next: Stop before the final sweep: check-final rejects the matrix because the 9 published baseline points (c1-c256, run 34557177019 at 25f8f98) predate the #3352 srt-recipe migration and cannot be regenerated.

中文

初次尝试 · 已通过 · Run 37009882020 / attempt 1 · 2026-10-02 13:30 UTC
lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 · 15b01bb526bd · 平均延迟
**变更：**按 digest 固定 SGLang v0.5.21-cu130，并将已移除的 --cuda-graph-max-bs 与 --disable-piecewise-cuda-graph 别名替换为其在 v0.5.19 中对应的 --cuda-graph-max-bs-decode 与 --cuda-graph-backend-prefill=disabled；镜像内源码为 e00930c5，旧参数在其中无法通过 argparse；测量提交为 d090548。实测数值及异常说明见上表。
**下一步：**在最终 sweep 前停止：check-final 拒绝该矩阵，因为 9 个已发布基线点（c1-c256，run 34557177019，提交 25f8f98）早于 #3352 的 srt-recipe 迁移，无法重新生成。

[Klaud-Cold]

failed · Cleanup pending. Stop and confirm owned runs before closing.

中文

failed · 清理待完成。先停止并确认自有运行结束，再关闭 PR。

[Klaud-Cold]

failed · Repairs: 0 · Runs: 37009882020, 37013298309
All owned runs ended. PR closed; branch deleted for retry.

中文

failed · 修复次数：0 · 运行：37009882020, 37013298309
所有自有运行均已结束。PR 已关闭；分支已删除，可重新尝试。