[Klaud Cold] Update qwen3.5-fp8-mi325x-sglang-mtp SGLang ROCm image to v0.5.20-rocm720-mi30x / 将 qwen3.5-fp8-mi325x-sglang-mtp 的 SGLang ROCm 镜像更新至 v0.5.20-rocm720-mi30x

[Klaud-Cold]

AI model disclosure: claude-opus-5-5 (Claude Opus 5.5) prepared this change as Klaud Cold.

中文

AI 模型披露：claude-opus-5-5（Claude Opus 5.5）以 Klaud Cold 身份准备了此变更。

🤖 Generated with Claude Code

[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 only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• 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：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **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]

Baseline stop · Failed · no GPU run dispatched
lmsysorg/sglang:v0.5.20-rocm720-mi30x (digest sha256:4abecde1…, SGLang v0.5.20 94602c9c, AITER 4ad99832) · 7748c32120f1 · TP8/EP1

The frozen 2026-05-18 public roster has 15 points, and the current family can produce only 5 of them. The other 10 can never have an updated-image result:

• 1k/1k c4, c8, c16, c32, c64. That scenario is no longer in qwen3.5-fp8-mi325x-sglang-mtp.
• 8k/1k c4, c8, c16, c32, c64 from the pre-SRT script recipe (point keys b28b50, 592135, 3fc78b, 31d25b, 566fdd). These identities changed when refactor: migrate fixed-sequence recipes to SRT-Slurm / 将定长配方迁移至 SRT-Slurm #3352 migrated the family to srt-slurm-recipes/qwen3.5/sglang/mi325x-fp8-mtp/8k1k.yaml.

Final validation would reject this candidate however the smoke turned out, so no benchmark was run on the old image or the new one. The change is also recorded here: the image bump plus renaming cuda-graph-max-bs to cuda-graph-max-bs-decode. SGLang v0.5.20 no longer has cuda_graph_max_bs in arg_groups/fields/exec_.py, and #3423 made the same rename for the MI355X sibling.

Next: A maintainer should re-anchor the baseline to a post-migration 8k/1k observation before this family is retried.

中文

基线停止 · 失败 · 未派发 GPU 运行
lmsysorg/sglang:v0.5.20-rocm720-mi30x（digest sha256:4abecde1…，SGLang v0.5.20 94602c9c，AITER 4ad99832）· 7748c32120f1 · TP8/EP1

冻结的 2026-05-18 公开基线共有 15 个测试点，当前配置族只能产出其中 5 个。其余 10 个点永远无法得到更新镜像的结果：

• 1k/1k c4、c8、c16、c32、c64。qwen3.5-fp8-mi325x-sglang-mtp 已不再包含该场景。
• 8k/1k c4、c8、c16、c32、c64，来自迁移前的脚本配方（点键 b28b50、592135、3fc78b、31d25b、566fdd）。refactor: migrate fixed-sequence recipes to SRT-Slurm / 将定长配方迁移至 SRT-Slurm #3352 将该配置族迁移至 srt-slurm-recipes/qwen3.5/sglang/mi325x-fp8-mtp/8k1k.yaml 后，这些标识已改变。

无论冒烟测试结果如何，最终验证都会拒绝此候选，因此旧镜像和新镜像都没有运行基准测试。此处也记录本次变更：镜像更新，并将 cuda-graph-max-bs 改名为 cuda-graph-max-bs-decode。SGLang v0.5.20 的 arg_groups/fields/exec_.py 中已没有 cuda_graph_max_bs，#3423 也为 MI355X 同类配置做了相同改名。

**下一步：**维护者需先将基线重新锚定到迁移后的 8k/1k 观测，再重试该配置族。

[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。