[Klaud Cold] Update dsv4-fp4-b200-dynamo-sglang SGLang image to v0.5.19-cu130 / 将 dsv4-fp4-b200-dynamo-sglang 的 SGLang 镜像更新至 v0.5.19-cu130

[Klaud-Cold]

Status / 状态

Current status: Deferred for capacity. The image change was committed and this draft PR opened, but the mandatory check-capacity --cluster b200-nscale recheck run immediately before the first targeted dispatch failed twice (2026-09-09T02:30:01Z and 02:30:18Z), after passing at branch-creation time. No benchmark or eval run was dispatched, so no GPU time was used and there are no owned runs to cancel.
Next step: None in this session. Per the Klaud Cold capacity rule the PR is returned to draft (it never left draft), closed, and its remote branch deleted so a later auto-sweep can reselect this candidate when b200-nscale is eligible again. No recovery is awaited or promised.

Green targeted benchmarks and default evals do not prove that global repository checks pass.

Change / 变更

• Family: configs/nvidia-master.yaml:dsv4-fp4-b200-dynamo-sglang (DeepSeek-V4-Pro FP4, B200, Dynamo + SGLang, disaggregated, 8k/1k, no speculative decoding, runner cluster:b200-nscale).
• Image: lmsysorg/sglang:nightly-dev-cu13-20260821-f825d729 → lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9 in the master entry and its ten referenced benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-*.yaml recipes (model.container equals the master image).
• Unchanged: model, precision, topology, Dynamo revision 86f84b94, NIXL KV transfer, workloads, resources, environment flags and recipe references. The generated matrix differs from main only in the image (10 recipes, 12 concurrency points, node counts 2/2/2/5/3/5/5/6/7/8).

Supporting evidence / 支持证据

• Public observation: latest-images reports dsv4 / b200 / dynamo-sglang / fp4 / spec_method=none / disagg / 8192-1024 on lmsysorg/sglang:nightly-dev-cu13-20260821-f825d729, dated 2026-09-01; framework-releases reports SGLang v0.5.19.
• Compatibility: the sgl-project/sglang tag v0.5.19 (commit 59f20bff, released 2026-09-05) is 631 commits ahead of and contains the nightly commit f825d729, so the release is a superset of the code currently pinned. Docker Hub tag v0.5.19-cu130 (CUDA 13.0, linux/amd64 + arm64) was pushed 2026-09-04 and is pinned by its manifest-list digest. Digest-pinned lmsysorg/sglang:<tag>@sha256:... images already have published Dynamo-SGLang results on GB200/GB300 (v0.5.14-cu130@sha256:5027e95b…), so the reference form is proven on srt-slurm.
• No open PR modifies the ten B200 DeepSeek-V4 8k/1k recipes or the family key in configs/nvidia-master.yaml (rechecked immediately before the branch was claimed; PR [NV] Refresh GB300 DeepSeek-V4-Pro AgentX with SGLang DSpark6 / 使用 SGLang DSpark6 更新 GB300 DeepSeek-V4-Pro AgentX #2623's removal of the same tag is in the GB300 agentic families only).
• Local checks: YAML parses; generate_sweep_configs.py test-config --config-keys dsv4-fp4-b200-dynamo-sglang output is identical to main except the image; utils/matrix_logic test_validation.py + test_generate_sweep_configs.py pass (291 tests).

Baseline (published 2026-09-01) / 基线（发布于 2026-09-01）

• Source: GET /api/v1/benchmarks?model=DeepSeek-V4-Pro&date=2026-09-01&exact=true (no view) and GET /api/v1/workflow-info?date=2026-09-01; GET /api/v1/evaluations?model=DeepSeek-V4-Pro filtered to the same identity.
• Identity verified on all 12 rows: dsv4 / b200 / dynamo-sglang / fp4 / spec_method=none / disagg / multinode / 8192-1024 / single_turn / image lmsysorg/sglang:nightly-dev-cu13-20260821-f825d729 / router dynamo-router@86f84b94 / KV nixl.
• Producer (all 12 points and 9 evals): run https://github.com/SemiAnalysisAI/InferenceX/actions/runs/33447526958 (attempt 1, "Run Sweep - Add DSV4 B200 disaggregated Dynamo SGLang STP configuration", PR Add DSV4 B200 disaggregated Dynamo SGLang STP configuration / 新增 DSV4 B200 分离式 Dynamo SGLang STP 配置 #2560, head SHA a430c17bab40fe1bf9c6e94d8b69c3d032dd0143, created 2026-08-31T22:43Z). curve_workflow_run_id=2391 is a logical curve snapshot, not the producer.
• Baseline published points (per-GPU throughput = tput_per_gpu total tok/s/GPU; output = output_tput_per_gpu; latencies in seconds):

conc	topology (P×TP/EP → D×TP/EP)	GPUs	tput/GPU	out tput/GPU	mean TPOT	mean TTFT	mean E2E
1	1×TP8 → 1×TP8	16	52.67	11.77	0.0097	0.868	9.88
32	1×DEP8 → 1×TP8	16	1049.89	234.81	0.0146	1.634	15.12
64	1×DEP8 → 1×TP8	16	1719.59	381.55	0.0179	2.138	18.57
64	1×DEP8 → 4×TP8	40	927.70	128.65	0.0120	2.613	13.68
256	1×DEP8 → 1×DEP8	16	4539.28	1009.02	0.0252	4.807	28.00
256	1×DEP8 → 2×DEP8	24	3268.09	544.84	0.0223	5.406	25.96
256	1×DEP8 → 4×DEP8	40	2029.05	281.89	0.0202	6.460	25.08
512	1×DEP8 → 1×DEP8	16	6402.47	1422.76	0.0268	14.093	38.81
2048	3×DEP8 → 2×DEP8	40	7717.78	2145.78	0.0316	21.706	50.80
4096	4×DEP8 → 2×DEP8	48	8667.01	2888.37	0.0357	42.404	75.25
6144	5×DEP8 → 2×DEP8	56	9332.22	3628.30	0.0382	54.472	89.64
8192	6×DEP8 → 2×DEP8	64	9733.98	4325.50	0.0432	60.463	100.24

• Baseline published evals (gsm8k, n_eff 1319, same producer run): em_strict 0.964–0.969 across conc 64 (1×DEP8→1×TP8 and 1×DEP8→4×TP8), 256 (1×DEP8→2×DEP8, 1×DEP8→4×DEP8), 512, 2048, 4096, 6144, 8192. No published eval for conc 1, 32 or the 1×DEP8→1×DEP8 conc 256 point.

Initial attempt / 初始尝试

• Image/commit: lmsysorg/sglang:v0.5.19-cu130@sha256:d6e72886… at commit d0a00ef89457fea20fbf9f94da0ff6f06a9d3d38.
• Run: none. Outcome: cancelled before dispatch (capacity deferral). The check-capacity --cluster b200-nscale recheck required immediately before dispatch exited non-zero twice at 2026-09-09T02:30Z; the same check had passed before edits and branch creation a few minutes earlier.
• Changes: image refresh only (see Change).
• Benchmark/eval results: N/A (never dispatched).
• Deltas vs baseline: N/A (never dispatched).
• Diagnosis / next step: no image-related finding; the image was never exercised. Candidate released for a later auto-sweep retry.

Repairs / 修复

• None used (0/5). The repair budget was not consumed because no attempt ran.

Final full sweep / 最终完整 sweep

• Not started. No perf-changelog.yaml entry was appended, no sweep label was applied, and the PR never left draft.

---

状态

当前状态： 因容量原因推迟。镜像变更已提交并创建了本草稿 PR，但在首次定向派发前必须执行的 check-capacity --cluster b200-nscale 复查连续两次失败（2026-09-09T02:30:01Z 与 02:30:18Z），而该检查在创建分支时曾通过。未派发任何基准或评测运行，因此未使用 GPU 时间，也没有需要取消的运行。
下一步： 本会话内无后续操作。按照 Klaud Cold 容量规则，PR 保持草稿状态（从未转为 ready）、关闭，并删除其远程分支，以便 b200-nscale 恢复可用后由后续 auto-sweep 重新选择该候选。不等待也不承诺恢复。

定向基准测试与默认评测通过并不证明仓库的全局检查通过。

变更

• 系列：configs/nvidia-master.yaml:dsv4-fp4-b200-dynamo-sglang（DeepSeek-V4-Pro FP4、B200、Dynamo + SGLang、分离式、8k/1k、无投机解码，runner cluster:b200-nscale）。
• 镜像：lmsysorg/sglang:nightly-dev-cu13-20260821-f825d729 → lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9，同时更新主配置条目及其引用的十个 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-*.yaml 配方（model.container 与主配置 image 一致）。
• 未变更：模型、精度、拓扑、Dynamo 版本 86f84b94、NIXL KV 传输、负载、资源、环境变量与配方引用。生成的矩阵与 main 仅镜像不同（10 个配方、12 个并发点，节点数 2/2/2/5/3/5/5/6/7/8）。

支持证据

• 公开观测：latest-images 报告 dsv4 / b200 / dynamo-sglang / fp4 / spec_method=none / 分离式 / 8192-1024 使用 lmsysorg/sglang:nightly-dev-cu13-20260821-f825d729，日期 2026-09-01；framework-releases 报告 SGLang v0.5.19。
• 兼容性：sgl-project/sglang 标签 v0.5.19（提交 59f20bff，2026-09-05 发布）领先 nightly 提交 f825d729 631 个提交且包含该提交，因此该发布版是当前所固定代码的超集。Docker Hub 标签 v0.5.19-cu130（CUDA 13.0，linux/amd64 + arm64）于 2026-09-04 推送，并按 manifest-list 摘要固定。按摘要固定的 lmsysorg/sglang:<tag>@sha256:... 镜像已在 GB200/GB300 上有公开发布的 Dynamo-SGLang 结果（v0.5.14-cu130@sha256:5027e95b…），该引用形式在 srt-slurm 上已被验证。
• 没有任何开放 PR 修改这十个 B200 DeepSeek-V4 8k/1k 配方或 configs/nvidia-master.yaml 中的该系列键（在占用分支前已再次核查；PR [NV] Refresh GB300 DeepSeek-V4-Pro AgentX with SGLang DSpark6 / 使用 SGLang DSpark6 更新 GB300 DeepSeek-V4-Pro AgentX #2623 删除同一标签的位置仅在 GB300 agentic 系列中）。
• 本地检查：YAML 可解析；generate_sweep_configs.py test-config --config-keys dsv4-fp4-b200-dynamo-sglang 的输出与 main 仅镜像不同；utils/matrix_logic 的 test_validation.py + test_generate_sweep_configs.py 通过（291 个测试）。

基线（发布于 2026-09-01）

• 来源：GET /api/v1/benchmarks?model=DeepSeek-V4-Pro&date=2026-09-01&exact=true（无 view）与 GET /api/v1/workflow-info?date=2026-09-01；GET /api/v1/evaluations?model=DeepSeek-V4-Pro 按同一身份过滤。
• 12 行数据均已核对身份：dsv4 / b200 / dynamo-sglang / fp4 / spec_method=none / 分离式 / 多节点 / 8192-1024 / single_turn / 镜像 lmsysorg/sglang:nightly-dev-cu13-20260821-f825d729 / 路由 dynamo-router@86f84b94 / KV nixl。
• 生产者（全部 12 个点与 9 项评测）：运行 https://github.com/SemiAnalysisAI/InferenceX/actions/runs/33447526958（attempt 1，PR Add DSV4 B200 disaggregated Dynamo SGLang STP configuration / 新增 DSV4 B200 分离式 Dynamo SGLang STP 配置 #2560，head SHA a430c17bab40fe1bf9c6e94d8b69c3d032dd0143，创建于 2026-08-31T22:43Z）。curve_workflow_run_id=2391 是逻辑曲线快照，不是生产者。
• 基线公开数据点见上方英文表格（每 GPU 吞吐 = tput_per_gpu，输出吞吐 = output_tput_per_gpu，延迟单位为秒）。
• 基线公开评测（gsm8k，n_eff 1319，同一生产者运行）：em_strict 0.964–0.969，覆盖并发 64（1×DEP8→1×TP8 与 1×DEP8→4×TP8）、256（1×DEP8→2×DEP8、1×DEP8→4×DEP8）、512、2048、4096、6144、8192。并发 1、32 以及 1×DEP8→1×DEP8 的并发 256 点没有公开评测。

初始尝试

• 镜像/提交：lmsysorg/sglang:v0.5.19-cu130@sha256:d6e72886…，提交 d0a00ef89457fea20fbf9f94da0ff6f06a9d3d38。
• 运行：无。结果：派发前取消（容量推迟）。派发前必须执行的 check-capacity --cluster b200-nscale 复查于 2026-09-09T02:30Z 连续两次返回非零；几分钟前在编辑与创建分支之前该检查曾通过。
• 变更：仅镜像刷新（见“变更”）。
• 基准/评测结果：N/A（从未派发）。
• 相对基线的差异：N/A（从未派发）。
• 诊断/下一步：没有与镜像相关的发现；镜像从未被实际运行。候选已释放，供后续 auto-sweep 重试。

修复

• 未使用（0/5）。由于没有任何尝试运行，修复预算未被消耗。

最终完整 sweep

• 未开始。未追加 perf-changelog.yaml 条目，未应用任何 sweep 标签，PR 从未离开草稿状态。

🤖 Generated with Claude Code

[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]

Status: Capacity deferral, session terminating. No benchmark or eval run was dispatched for this PR, so there is nothing to cancel and no GPU time was used.

Confirmed finding: The required check-capacity --cluster b200-nscale recheck immediately before the first targeted e2e-tests.yml dispatch exited non-zero twice (2026-09-09T02:30:01Z, 02:30:18Z). The same check passed minutes earlier, before edits and branch creation. The updated image itself was never exercised, so this says nothing about its compatibility.

Action / next step: Per the Klaud Cold capacity rule this PR stays in draft with no sweep labels, is closed, and its remote branch klaud/auto-4b07c82a9d375dc9-c90dff266c5a5a8a is deleted so a later auto-sweep can reselect the candidate when the cluster is eligible. No recovery is awaited or promised.

Links: commit d0a00ef · baseline producer run https://github.com/SemiAnalysisAI/InferenceX/actions/runs/33447526958

---

状态： 因容量原因推迟，会话终止。本 PR 未派发任何基准或评测运行，因此无需取消，也未使用 GPU 时间。

已确认的发现： 在首次定向派发 e2e-tests.yml 前必须执行的 check-capacity --cluster b200-nscale 复查连续两次返回非零（2026-09-09T02:30:01Z、02:30:18Z）。几分钟前在编辑与创建分支之前该检查曾通过。更新后的镜像从未被实际运行，因此这不代表其兼容性存在问题。

操作/下一步： 按照 Klaud Cold 容量规则，本 PR 保持草稿状态且不带任何 sweep 标签，随后关闭并删除其远程分支 klaud/auto-4b07c82a9d375dc9-c90dff266c5a5a8a，以便集群恢复可用后由后续 auto-sweep 重新选择该候选。不等待也不承诺恢复。

链接： 提交 d0a00ef · 基线生产者运行 https://github.com/SemiAnalysisAI/InferenceX/actions/runs/33447526958