[Klaud Cold] Update dsr1-fp8-mi325x-sglang-mtp SGLang image to v0.5.21-rocm720-mi30x / [Klaud Cold] 将 dsr1-fp8-mi325x-sglang-mtp SGLang 镜像更新至 v0.5.21-rocm720-mi30x

[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:v0.5.12-rocm700-mi30x to lmsysorg/sglang:v0.5.21-rocm720-mi30x@sha256:0534eb382626ef55ace8be24f7d81cc6eb7451acac8a5727848397a5e9309d91.
Baseline: 2026-05-18 · lmsysorg/sglang:v0.5.12-rocm700-mi30x
Mean latency · Sources: API 1, API 2

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1k/1k c4 TP8 EP1 5785cc	66.26	32.97	385.33	14.31
1k/1k c8 TP8 EP1 d615a8	116.87	58.65	281.8	16.4
1k/1k c16 TP8 EP1 408635	200.08	99.51	313.87	19.34
1k/1k c32 TP8 EP1 9663a6	314.63	157.56	377.76	24.4
1k/1k c64 TP8 EP1 f4bd94	439.65	219.77	465.6	35.05
8k/1k c4 TP8 EP1 c3cd4f	N/A	N/A	N/A	N/A
8k/1k c4 TP8 EP1 f2d179	224.8	25.01	965.6	18.52
8k/1k c8 TP8 EP1 374ce0	N/A	N/A	N/A	N/A
8k/1k c8 TP8 EP1 9dfb76	381.95	42.94	897.03	21.93
8k/1k c16 TP8 EP1 287862	621.2	68.83	920.66	27.24
8k/1k c16 TP8 EP1 c40a77	N/A	N/A	N/A	N/A
8k/1k c32 TP8 EP1 00c54c	N/A	N/A	N/A	N/A
8k/1k c32 TP8 EP1 bebf91	842.61	94.23	1,428.05	39.98
8k/1k c64 TP8 EP1 7b183b	N/A	N/A	N/A	N/A
8k/1k c64 TP8 EP1 a8e734	1,027.76	114.02	1,962.01	67.03

Note: 8k/1k c4 TP8 EP1 c3cd4f, 8k/1k c8 TP8 EP1 374ce0, 8k/1k c16 TP8 EP1 c40a77, 8k/1k c32 TP8 EP1 00c54c, 8k/1k c64 TP8 EP1 7b183b: unavailable.

Eval: N/A

中文

**目标：**将 SGLang 镜像从 lmsysorg/sglang:v0.5.12-rocm700-mi30x 更新为 lmsysorg/sglang:v0.5.21-rocm720-mi30x@sha256:0534eb382626ef55ace8be24f7d81cc6eb7451acac8a5727848397a5e9309d91。
**基线：**2026-05-18 · lmsysorg/sglang:v0.5.12-rocm700-mi30x
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]

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 · Reason: baseline-point-mismatch · Repairs: 0 · Runs: —
All owned runs ended. PR closed; branch deleted for retry.

中文

failed · 原因：baseline-point-mismatch · 修复次数：0 · 运行：—
所有自有运行均已结束。PR 已关闭；分支已删除，可重新尝试。