[Klaud Cold] Update qwen3.5-fp8-gb300-dynamo-sglang-mtp SGLang image to v0.5.19-cu130 / 将 qwen3.5-fp8-gb300-dynamo-sglang-mtp 的 SGLang 镜像更新至 v0.5.19-cu130

[Klaud-Cold]

Goal: Update SGLang image from lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3 to lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9.
Baseline: 2026-07-30 · lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3
Mean latency · Sources: API 1, API 2, API 3

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
8k/1k c1 PTP4x1/DTP4x1 66c1c7	N/A	N/A	N/A	N/A
8k/1k c2 PTP4x1/DTP4x1 e015a0	N/A	N/A	N/A	N/A
8k/1k c8 PTP4x1/DTP4x1 7bc1d9	N/A	N/A	N/A	N/A
8k/1k c32 PTP8x1/DTP8x1 e10f13	N/A	N/A	N/A	N/A
8k/1k c48 PTP8x1/DTP8x1 435bbd	N/A	N/A	N/A	N/A
8k/1k c80 PTP8x1/DTP8x1 7bd8f7	N/A	N/A	N/A	N/A
8k/1k c480 PTP4x3/DTP16x1 6e9d7c	N/A	N/A	N/A	N/A
8k/1k c768 PTP4x4/DTP16x1 42ea07	N/A	N/A	N/A	N/A
8k/1k c1280 PTP4x6/DTP16x1 7b76a3	N/A	N/A	N/A	N/A
8k/1k c1344 PTP4x7/DTP16x1 437e88	N/A	N/A	N/A	N/A
8k/1k c1920 PTP4x8/DTP16x1 115634	N/A	N/A	N/A	N/A
8k/1k c2304 PTP4x8/DTP16x1 b971bf	N/A	N/A	N/A	N/A

Note: All rows: unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c80	97.04%	1,319
gsm8k/em_strict · c480	96.82%	1,319
gsm8k/em_strict · c768	97.12%	1,319
gsm8k/em_strict · c1280	96.66%	1,319
gsm8k/em_strict · c1344	96.74%	1,319
gsm8k/em_strict · c2304	97.19%	1,319

中文

**目标：**将 SGLang 镜像从 lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3 更新为 lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9。
**基线：**2026-07-30 · lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3
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! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[Klaud-Cold]

Baseline: 2026-07-30 · lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3

Mean latency

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
8k/1k c1 PTP4x1/DTP4x1 66c1c7	N/A	N/A	N/A	N/A
8k/1k c2 PTP4x1/DTP4x1 e015a0	N/A	N/A	N/A	N/A
8k/1k c8 PTP4x1/DTP4x1 7bc1d9	N/A	N/A	N/A	N/A
8k/1k c32 PTP8x1/DTP8x1 e10f13	N/A	N/A	N/A	N/A
8k/1k c48 PTP8x1/DTP8x1 435bbd	N/A	N/A	N/A	N/A
8k/1k c80 PTP8x1/DTP8x1 7bd8f7	N/A	N/A	N/A	N/A
8k/1k c480 PTP4x3/DTP16x1 6e9d7c	N/A	N/A	N/A	N/A
8k/1k c768 PTP4x4/DTP16x1 42ea07	N/A	N/A	N/A	N/A
8k/1k c1280 PTP4x6/DTP16x1 7b76a3	N/A	N/A	N/A	N/A
8k/1k c1344 PTP4x7/DTP16x1 437e88	N/A	N/A	N/A	N/A
8k/1k c1920 PTP4x8/DTP16x1 115634	N/A	N/A	N/A	N/A
8k/1k c2304 PTP4x8/DTP16x1 b971bf	N/A	N/A	N/A	N/A

Note: All rows: unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c80	97.04%	1,319
gsm8k/em_strict · c480	96.82%	1,319
gsm8k/em_strict · c768	97.12%	1,319
gsm8k/em_strict · c1280	96.66%	1,319
gsm8k/em_strict · c1344	96.74%	1,319
gsm8k/em_strict · c2304	97.19%	1,319

中文

**基线：**2026-07-30 · lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3；数值及异常说明见表格。

[Klaud-Cold]

Compatibility check · incompatible · no GPU dispatch · 2026-09-13T00:51:17Z
lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9 · 920cfe158f0600119bcebede61f95dcf01711c4b · 8k/1k · 7 srt-slurm recipes · Dynamo 46520ca59afe992fb5ef61b3197b2316f8df9b2b
Change: Bumped the master image and the seven referenced recipe model.container values from v0.5.14-cu130 to the digest-pinned v0.5.19-cu130 manifest list; its arm64 image is labelled build commit 0bcd822 from sglang actions run 33912440803 (created 2026-09-04), and the annotated v0.5.19 tag dereferences to the same commit.

Finding

• SGLang #36681 (commit 94183a8, 2026-08-27) moved python/sglang/srt/server_args_config_parser.py to python/sglang/srt/utils/server_args_config_parser.py. v0.5.14 ships only the old path and v0.5.19 only the new one.
• All seven recipes pin Dynamo 46520ca (2026-05-19) with install: true. On this family's GB300 dcgm-power lane, runners/launch_gb300-nv.sh clones the pinned srt-slurm producer 6fc1bed0, whose srtctl _hash_cached_source_install builds ai-dynamo-runtime from source at that hash and pip install -e the Dynamo tree into the SGLang container. At that commit components/src/dynamo/sglang/args.py line 18 imports ConfigArgumentMerger from the removed path and main.py imports args, so every python -m dynamo.sglang prefill and decode worker exits with ModuleNotFoundError before loading weights.
• The same failure was observed at runtime on gb300-nv with this image in PR #2909 (run 34303559979: ModuleNotFoundError: No module named 'sglang.srt.server_args_config_parser') and cleared there only by moving the Dynamo pin; the identical GB200 recipes reached the same conclusion in PR #3061.
• Upstream Dynamo fixed the import in #14054 (commit f3d0ed0, 2026-09-02), which is not reachable from 46520ca.
• Moving the Dynamo pin changes dynamo.hash in all seven recipes and the family's router.version in the master entry, which is outside the image-only scope, and patching the installed Dynamo is prohibited. No in-scope repair exists, so the smoke run was not dispatched.
• Everything else checked out: every sglang_config flag and SGLANG_* key the recipes use is still defined in v0.5.19. --mamba-scheduler-strategy and --cuda-graph-max-bs were already deprecated aliases of --mamba-radix-cache-strategy and --cuda-graph-max-bs-decode at v0.5.14 with an unchanged mapping, and no_buffer, trtllm_mha, flashinfer_trtllm, fp8_e4m3 and mooncake remain accepted values. Bundled dependencies move: sgl-kernel 0.4.4 → 0.4.6.post1, DeepGEMM 0.1.3 → 0.1.7, FlashInfer 0.6.12 → 0.6.18, Mooncake 0.3.11.post1 → 0.3.13, torch 2.11.0 → 2.13.0, transformers 5.8.1 → 5.12.1, CUDA 13.0.1 → 13.0.3, DeepEP source build → sgl-deep-ep==0.1.2, and Python now runs from an /opt/sglang venv (Dockerfile v0.5.14, Dockerfile v0.5.19). These are read from the labelled build commit, not from the image layers, and their runtime effects remain unmeasured.

Next: A maintainer retargets the family's Dynamo pin (dynamo.hash in the seven recipes and router.version in the master entry) to a commit or wheel that includes f3d0ed0 and re-releases this candidate with release-candidate.

中文

兼容性检查 · 不兼容 · 未调度 GPU · 2026-09-13T00:51:17Z
lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9 · 920cfe158f0600119bcebede61f95dcf01711c4b · 8k/1k · 7 个 srt-slurm 配方 · Dynamo 46520ca59afe992fb5ef61b3197b2316f8df9b2b
**变更：**将主镜像及其引用的七个配方 model.container 从 v0.5.14-cu130 更新为按摘要固定的 v0.5.19-cu130 清单列表；其 arm64 镜像标签指向 sglang 构建提交 0bcd822，与带注解的 v0.5.19 标签解析结果一致。

**发现：**SGLang #36681（提交 94183a8，2026-08-27）将 server_args_config_parser.py 移至 sglang/srt/utils/，v0.5.19 不再提供旧路径。七个配方固定的 Dynamo 46520ca（2026-05-19，install: true）由本配置族 GB300 dcgm-power 通道所用的 srt-slurm 生产者 6fc1bed0 从源码安装到 SGLang 容器中；该提交的 components/src/dynamo/sglang/args.py 第 18 行在模块导入时从旧路径导入 ConfigArgumentMerger，因此每个 python -m dynamo.sglang prefill/decode worker 在加载权重前就会因 ModuleNotFoundError 退出。PR #2909 曾在 gb300-nv 上以同一镜像实际观察到该错误（运行 34303559979），并且只能通过更改 Dynamo 固定版本解决；GB200 上相同的配方在 PR #3061 中得出了同样的结论。上游 Dynamo 已在 #14054（提交 f3d0ed0，2026-09-02）修复。更改 Dynamo 固定版本会同时改动七个配方的 dynamo.hash 与主配置的 router.version，超出仅更新镜像的范围，且禁止修补已安装的 Dynamo，因此未调度冒烟测试。其余检查均通过：配方使用的全部 sglang_config 参数与 SGLANG_* 环境变量在 v0.5.19 中仍然存在（两个参数在 v0.5.14 时已是弃用别名，映射未变）；依赖版本变化见上文英文列表，均读取自构建提交而非镜像层，运行时影响尚未测量。

**下一步：**由维护者将该配置族的 Dynamo 固定版本（七个配方的 dynamo.hash 与主配置的 router.version）调整为包含 f3d0ed0 的提交或 wheel，并通过 release-candidate 重新放行此候选。

[Klaud-Cold]

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

中文

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

[Klaud-Cold]

incompatible · Repairs: 0 · Runs: —
All owned runs ended. PR closed; exact-candidate branch retained pending maintainer review.

中文

incompatible · 修复次数：0 · 运行：—
所有自有运行均已结束。PR 已关闭；保留精确候选分支，等待维护者检查。