[Klaud Cold] Update dsv4-fp4-b300-dynamo-sglang SGLang image to v0.5.19-cu130 (digest-pinned) and refresh the Dynamo pin / 将 dsv4-fp4-b300-dynamo-sglang 的 SGLang 镜像更新为 v0.5.19-cu130（按 digest 固定）并刷新 Dynamo 提交

[Klaud-Cold]

Summary / 摘要

Refresh the dsv4-fp4-b300-dynamo-sglang family (DeepSeek-V4-Pro FP4, B300, Dynamo + SGLang disaggregated, 8k/1k, no speculative decoding) from the removed nightly lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4 to the digest-pinned release lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9, matching the public framework release sglang v0.5.19.

将 dsv4-fp4-b300-dynamo-sglang 系列（DeepSeek-V4-Pro FP4、B300、Dynamo + SGLang 分离式部署、8k/1k、无投机解码）的镜像从已被删除的 nightly lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4 更新为按 digest 固定的正式版 lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9，与公开的框架版本 sglang v0.5.19 一致。

Files / 文件

• configs/nvidia-master.yaml: dsv4-fp4-b300-dynamo-sglang.image (one line).
• benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-{1p1d-tp4-tp4,1p1d-dep4-dep8,2p1d-dep4-dep8,4p1d-dep4-dep8,6p1d-dep4-dep8}.yaml: model.container equals the master image; dynamo.hash bumped from 41882ae9b07232eed4850fb1daf8c958abb2556a (2026-07-14) to b1c5147f6c9eb036d98e0e25e2a980375f491e78 (2026-09-03); the four DEP recipes replace the deprecated SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS / SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND env vars with enable-w4a4-mxfp4-megamoe: true on prefill and decode.
• No other family, script, launcher or workflow is touched. Model, precision, topology, concurrency list, workloads, resources and recipe references are unchanged. Only these five recipes reference the family's CONFIG_FILE paths; the -mtp sibling uses separate *-mtp.yaml recipes and is untouched.
• 未改动其他系列、脚本、启动器或工作流。模型、精度、拓扑、并发列表、负载、资源和配方引用保持不变。只有这五个配方被该系列的 CONFIG_FILE 引用；-mtp 兄弟系列使用独立的 *-mtp.yaml 配方，未做改动。

Why these exact changes / 变更依据

• Old tag is gone. Docker Hub returns no manifest for lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4 (dated nightlies are cleaned up), so the family can no longer be reproduced from its pinned image. / 旧标签已不存在：Docker Hub 已无 nightly-dev-cu13-20260721-8905cbd4 的 manifest（带日期的 nightly 会被清理），该系列已无法从其固定镜像复现。
• Same image lineage. v0.5.19-cu130 (pushed 2026-09-04, tag commit 59f20bff) is built by release-docker.yml with the same Dockerfile target (framework_final), CUDA_VERSION=13.0.3, BUILD_TYPE=all and INSTALL_FLASHINFER_JIT_CACHE=1 as the nightly-dev-cu13-* images from release-docker-dev.yml; only the source commit differs. The @sha256 pin follows the existing dsv4-fp4-gb300-dynamo-sglang recipes, and the B300 launcher already normalises @ in squash-file names. / 镜像谱系一致：v0.5.19-cu130 与 nightly-dev-cu13-* 使用同一 Dockerfile target、CUDA 13.0.3 与构建参数，仅源码提交不同；@sha256 固定方式沿用现有 dsv4-fp4-gb300-dynamo-sglang 配方。
• Dynamo pin must move with SGLang. SGLang v0.5.19 contains config: constructing a config no longer resolves it sgl-project/sglang#35907 (constructing ServerArgs no longer resolves it), #36255, #36972 (server_args.get_model_config() / use_mla_backend() removed) and the ConfigArgumentMerger move to sglang.srt.utils.server_args_config_parser. Dynamo 41882ae9 (2026-07-14) predates all of them; fix(sglang): use resolved args after engine init ai-dynamo/dynamo#13905, #14054 and #14234 add the compatibility shims, and b1c5147f6c9e (2026-09-03) is the first commit containing all three. srt-slurm still builds Dynamo from this pinned commit at run time, as before; no engine or container patching is involved. / Dynamo 提交必须随 SGLang 更新：v0.5.19 包含 #35907、#36255、#36972 以及 ConfigArgumentMerger 模块迁移，旧 Dynamo 提交早于这些改动；dynamo#13905、#14054、#14234 提供兼容层，b1c5147f6c9e 是同时包含三者的首个提交。srt-slurm 仍按原有方式从该提交构建 Dynamo，不涉及引擎或容器补丁。
• MegaMoE FP4 activations. [DeepSeek V4] Add W4A4 MegaMoE server flag sgl-project/sglang#35918 (in v0.5.19) turned SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS / _USE_MXF4_KIND into warn-only _DeprecatedEnv entries; mega_moe.py now selects mxf4xmxf4 only from --enable-w4a4-mxfp4-megamoe. The flag is added to both prefill and decode of the four megamoe recipes so the recipe keeps its FP4-activation path instead of silently falling back to FP8 activations (see dsv4-fp4-b300-sglang-agentic-hicache-mtp: restore MegaMoE FP4 acts, retarget to cluster:b300-dsxe / 恢复 MegaMoE FP4 激活路径，并迁移至 cluster:b300-dsxe #2878 for the single-node sibling). SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK and every other env var used by the recipes are still live in v0.5.19 environ.py. / [DeepSeek V4] Add W4A4 MegaMoE server flag sgl-project/sglang#35918 已将上述两个环境变量改为仅告警，FP4 激活路径只能通过 --enable-w4a4-mxfp4-megamoe 选择；四个 megamoe 配方的 prefill 与 decode 均加入该标志以保持原行为。其余环境变量在 v0.5.19 中仍然有效。

Local checks / 本地检查

• generate_sweep_configs.py test-config --config-files configs/nvidia-master.yaml --config-keys dsv4-fp4-b300-dynamo-sglang: same five matrix rows, node-count 2/2/2/3/4 unchanged, only image differs from base.
• generate_sweep_configs.py full-sweep --config-files configs/nvidia-master.yaml --framework dynamo-sglang: exit 0.
• pytest utils/matrix_logic/test_generate_sweep_configs.py utils/test_gb300_power_official_contract.py: 121 passed.
• srtctl dry-run -f <recipe> for all five recipes with the dcgm-power srt-slurm fork pinned by runners/launch_b300-dsxe.sh: exit 0.
• Capacity check (python -m utils.klaud check-capacity --cluster b300-dsxe) passed before edits and before each dispatch.

Baseline / 基线（published 2026-07-30 / 发布日期 2026-07-30）

Source: GET /api/v1/benchmarks?model=DeepSeek-V4-Pro&date=2026-07-30&exact=true, filtered to hardware=b300, framework=dynamo-sglang, precision=fp4, spec_method=none, isl=8192, osl=1024, benchmark_type=single_turn, image lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4. All eight points come from producer run https://github.com/SemiAnalysisAI/InferenceX/actions/runs/30431262611/attempts/1 (Run Sweep - feat: add DSV4 FP4 B300 Dynamo-SGLang STP configuration, from GET /api/v1/workflow-info?date=2026-07-30); logical curve snapshot id 2240 is not a producer id. Time metrics are seconds.

conc / 并发	topology / 拓扑	tput/GPU (tok/s)	output tput/GPU (tok/s)	median TTFT (s)	median TPOT (s)	median E2EL (s)	median interactivity (tok/s/user)	benchmark id
1	1P x TP4/EP1 + 1D x TP4/EP1 (4+4 GPUs)	105.7	23.6	0.319	0.0102	9.70	97.7	438001
8	1P x TP4/EP1 + 1D x TP4/EP1 (4+4 GPUs)	629.2	141.5	0.368	0.0130	12.63	76.7	438003
16	1P x TP4/EP1 + 1D x TP4/EP1 (4+4 GPUs)	1091.7	241.9	0.578	0.0150	14.51	66.5	438000
32	1P x TP4/EP1 + 1D x TP4/EP1 (4+4 GPUs)	1786.7	399.6	0.886	0.0178	17.56	56.3	438002
256	1P x TP4/EP4/DP-attn + 1D x TP8/EP8/DP-attn (4+8 GPUs)	5051.4	842.1	11.162	0.0238	32.80	42.0	438004
768	2P x TP4/EP4/DP-attn + 1D x TP8/EP8/DP-attn (8+8 GPUs)	7702.0	1710.8	21.938	0.0286	47.89	34.9	437998
2048	4P x TP4/EP4/DP-attn + 1D x TP8/EP8/DP-attn (16+8 GPUs)	10243.4	3417.6	28.569	0.0377	63.19	26.5	437999
3072	6P x TP4/EP4/DP-attn + 1D x TP8/EP8/DP-attn (24+8 GPUs)	11212.4	4983.0	21.867	0.0463	63.95	21.6	438005

Published evaluations for the baseline run / 基线运行的已发布评测（GET /api/v1/evaluations, filtered to the same identity and producer run）:

task / 任务	conc / 并发	topology / 拓扑	metrics / 指标	eval id
gsm8k	32	1P TP4/EP1 + 1D TP4/EP1	em_flexible=0.9674, em_flexible_se=0.0049, em_strict=0.9682, em_strict_se=0.0048, n_eff=1319	9030
gsm8k	256	1P TP4/EP4 + 1D TP8/EP8	em_flexible=0.9659, em_flexible_se=0.0050, em_strict=0.9659, em_strict_se=0.0050, n_eff=1319	9031
gsm8k	768	2P TP4/EP4 + 1D TP8/EP8	em_flexible=0.9666, em_flexible_se=0.0049, em_strict=0.9674, em_strict_se=0.0049, n_eff=1319	9027
gsm8k	2048	4P TP4/EP4 + 1D TP8/EP8	em_flexible=0.9644, em_flexible_se=0.0051, em_strict=0.9644, em_strict_se=0.0051, n_eff=1319	9028
gsm8k	3072	6P TP4/EP4 + 1D TP8/EP8	em_flexible=0.9674, em_flexible_se=0.0049, em_strict=0.9674, em_strict_se=0.0049, n_eff=1319	9029

Attempts / 尝试记录

Attempt / 尝试	Image / SHA	Run	Benchmarks / 基准	Evals / 评测	Δ vs baseline (per point) / 相对基线差异	Diagnosis / 诊断
Baseline (published 2026-07-30) / 基线	lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4 (Dynamo 41882ae9)	https://github.com/SemiAnalysisAI/InferenceX/actions/runs/30431262611	8/8 points published (table above)	see evaluation table above / 见上方评测表	—	Reference only; the old image is never rerun. / 仅作参考，不重跑旧镜像。
Attempt 1 (initial update) / 尝试 1（首次更新）	lmsysorg/sglang:v0.5.19-cu130@sha256:d6e72886… (Dynamo b1c5147f) @ 001e9af8413ed8d323b3cd0dd9e5e3925069e308	https://github.com/SemiAnalysisAI/InferenceX/actions/runs/34281233059 (e2e-tests.yml from main, inputs.ref = measured SHA, klaud-run=true, fail-fast=true)	failed: 0/5 matrix rows measured. c768 (2P DEP4 + 1D DEP8) failed; TP4x1D c1x8x16x32, c256, c2048, c3072 cancelled by fail-fast before finishing. results_bmk artifact is empty (142 bytes).	failed: 0/5 eval rows. TP4x1D eval-only job failed; the other four eval-only jobs cancelled by fail-fast. eval_results_all is empty.	N/A: no measured points (infrastructure failure before the server became healthy). / 无可比较数据（服务器健康前即因基础设施故障中止）。	Both failing jobs (Slurm jobs 958 and 961) landed on nodes dsxe-sa-b300-prd0-gpu-00 + gpu-01. On gpu-01 the decode worker and frontend containers aborted at start: pyxis: --container-mounts: source path does not exist: /scratch/models/DeepSeek-V4-Pro (node-local staged weights missing on that node). The prefill workers on gpu-00 started normally with the new image and began the pinned Dynamo source build before being killed by the teardown. Not attributable to the image, Dynamo pin or recipe flags. / 两次失败的作业都落在 gpu-00+gpu-01，gpu-01 缺少节点本地模型目录 /scratch/models/DeepSeek-V4-Pro，decode 与 frontend 容器无法挂载而启动失败；gpu-00 上的 prefill 使用新镜像正常启动。与镜像、Dynamo 提交或配方标志无关。
Attempt 2 (repair 1: retry, no code change) / 尝试 2（修复 1：重试，无代码改动）	same image and Dynamo pin @ 001e9af8413ed8d323b3cd0dd9e5e3925069e308	https://github.com/SemiAnalysisAI/InferenceX/actions/runs/34283344876 (e2e-tests.yml from main, inputs.ref = measured SHA, klaud-run=true, fail-fast=true)	failed: 0/5 matrix rows measured. c768 (2P DEP4 + 1D DEP8) failed; TP4x1D c1x8x16x32, c256, c2048, c3072 cancelled by fail-fast. results_bmk empty (142 bytes).	failed: 0/5 eval rows. c2048 eval-only job failed; the other four eval-only jobs cancelled by fail-fast. eval_results_all empty.	N/A: no measured points (same infrastructure failure). / 无可比较数据（同一基础设施故障）。	Same root cause as attempt 1, now on two nodes. Slurm job 963 (c768) again allocated gpu-00+gpu-01 and the gpu-01 decode/frontend containers failed with pyxis: --container-mounts: source path does not exist: /scratch/models/DeepSeek-V4-Pro. Slurm job 969 (c2048 eval-only, head node dsxe-sa-b300-prd0-gpu-11) failed even earlier in srtctl with FileNotFoundError: Model path does not exist: /scratch/models/DeepSeek-V4-Pro. The launcher documents /scratch/models as node-local NVMe with the same layout on every compute node; at least gpu-01 and gpu-11 currently lack the DeepSeek-V4-Pro directory. Not attributable to the image, Dynamo pin or recipe flags. / 与尝试 1 同一根因，且出现在两个节点：作业 963 再次分到 gpu-00+gpu-01，gpu-01 容器挂载失败；作业 969（c2048 仅评测，头节点 gpu-11）在 srtctl 阶段即报 Model path does not exist: /scratch/models/DeepSeek-V4-Pro。至少 gpu-01 与 gpu-11 缺少节点本地的 DeepSeek-V4-Pro 权重目录。与镜像、Dynamo 提交或配方标志无关。

Status / 状态

Stopped: the same failure occurred twice without progress (repair budget used: 1 of 3). Both targeted attempts died before any SGLang server became healthy because /scratch/models/DeepSeek-V4-Pro is missing on at least two b300-dsxe compute nodes (dsxe-sa-b300-prd0-gpu-01, dsxe-sa-b300-prd0-gpu-11). That is cluster weight staging, outside the scope of an image refresh, so no further GPU attempts were made. Both e2e runs (34281233059, 34283344876) are confirmed terminal; every matrix job is failed or cancelled, and the collection jobs completed with empty artifacts.

已停止：同一故障连续出现两次且无进展（已使用修复次数：1/3）。 两次目标尝试都在任何 SGLang 服务器就绪之前中止，原因是 b300-dsxe 至少两个计算节点（gpu-01、gpu-11）缺少 /scratch/models/DeepSeek-V4-Pro。这属于集群权重预置问题，超出镜像刷新范围，因此不再消耗 GPU。两个 e2e 运行（34281233059、34283344876）均已确认结束；所有矩阵作业为失败或已取消，收集作业完成但产物为空。

• No perf-changelog.yaml entry was appended and the PR stays a draft without a sweep label, because no updated-image attempt has passed yet. / 由于尚无通过的目标尝试，未追加 perf-changelog.yaml 条目，PR 保持草稿且未加 sweep 标签。
• The change itself is ready to re-test once the node-local weights are restored (or the affected nodes are drained): re-dispatch e2e-tests.yml with test-config --config-files configs/nvidia-master.yaml --config-keys dsv4-fp4-b300-dynamo-sglang at this SHA. The prefill containers on gpu-00 started with the new image and began the pinned Dynamo source build before teardown, so the image pull and container start are known to work. / 节点本地权重恢复（或问题节点被下线）后即可重新测试：在此 SHA 上重新分发上述 e2e-tests.yml 命令。gpu-00 上的 prefill 容器已使用新镜像启动并开始构建固定的 Dynamo 源码，说明镜像拉取与容器启动正常。
• Image-level compatibility (Dynamo b1c5147f vs SGLang v0.5.19 API, --enable-w4a4-mxfp4-megamoe) has not been exercised on GPU in this PR; it is supported by the source-level evidence above only. / 镜像层面的兼容性（Dynamo 与 SGLang v0.5.19 的 API、--enable-w4a4-mxfp4-megamoe）尚未在 GPU 上实际验证，目前仅有上述源码层面的证据。

Limitations / 局限性

• A green targeted benchmark proves this family runs on the new image; it does not prove that every global repository check passes. / 目标基准通过只能证明该系列可在新镜像上运行，不能证明所有全局检查都通过。
• Deltas are computed only for points whose topology, concurrency and dataset match the published baseline; unmatched or invalid points are excluded from any improvement claim. / 差异只对拓扑、并发和数据集与基线一致的点计算；不匹配或无效的点不计入改进结论。
• Performance improvements are best effort; regressions are reported, not used as a rejection threshold. / 性能提升为尽力而为；回退会如实报告，不作为拒绝阈值。

🤖 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