[Klaud Cold] Update dsr1-fp4-b200-sglang SGLang image to v0.5.19-cu130 / 将 dsr1-fp4-b200-sglang 的 SGLang 镜像更新至 v0.5.19-cu130

[Klaud-Cold]

Update the dsr1-fp4-b200-sglang master image from lmsysorg/sglang:v0.5.16-cu130 to the current SGLang release lmsysorg/sglang:v0.5.19-cu130 (Docker Hub digest sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9, build commit sgl-project/sglang@0bcd822 = tag v0.5.19). Model, TP4/EP1 and TP4/EP4 DP-attention topologies, the 8k1k workload, concurrency ranges and benchmarks/single_node/fixed_seq_len/dsr1_fp4_b200.sh are unchanged.

Baseline

• Published date: 2026-08-06 (workflow-info?date=2026-08-06; benchmarks?model=DeepSeek-R1-0528&date=2026-08-06&exact=true&sequence=8k/1k, filtered to hardware b200, framework sglang, precision fp4, spec none, non-disagg, ISL/OSL 8192/1024; evaluations?model=DeepSeek-R1-0528&date=2026-08-06&exact=true)
• Old image: lmsysorg/sglang:v0.5.16-cu130 (Docker Hub digest sha256:7b6a35df9839fd593a94a1eaee82d7777f472225d9f3ad1f8a2e0cb2bd1785d0, build commit sgl-project/sglang@fdebc93 = tag v0.5.16, CUDA 13.0.1, FlashInfer 0.6.14, sgl-kernel 0.4.5)
• Workload / topology: single-node B200 (cluster:b200-nscale), nvidia/DeepSeek-R1-0528-FP4-V2, SGLang NVFP4, no speculative decoding, fixed-seq-len 8k1k (ISL 8192 / OSL 1024), random dataset; TP4/EP1 at concurrency 1–32 and TP4/EP4 DP-attention at concurrency 64–256 (9 points)
• Producer: all 9 points come from run 30952773184 attempt 1 (head d4363bd7fd5bda1391d2d0a46d834cecc363c0d6, changelog PR #2492); benchmark result IDs 438719 … (curve snapshot curve_workflow_run_id 2261 is a logical snapshot, not the producer)
• Published evals: N/A — the evaluations feed has no gsm8k row for this identity dated 2026-08-06; the newest published gsm8k for this identity is 2026-03-28 on an older image and is not comparable

Point	tput/GPU (tok/s)	output tput/GPU (tok/s)	median TPOT (ms)	median TTFT (ms)
TP4/EP1 c1	389.7	43.5	5.52	207
TP4/EP1 c2	672.2	75.4	6.37	245
TP4/EP1 c4	1134.2	126.2	7.48	245
TP4/EP1 c8	1778.0	199.9	9.49	269
TP4/EP1 c16	2558.7	283.5	13.27	494
TP4/EP1 c32	3428.7	383.4	19.75	769
TP4/EP4 DPA c64	4603.2	510.7	29.73	857
TP4/EP4 DPA c128	6249.9	692.3	44.39	831
TP4/EP4 DPA c256	7955.2	884.2	66.38	4092

---

将 dsr1-fp4-b200-sglang 的主配置镜像从 lmsysorg/sglang:v0.5.16-cu130 更新至当前 SGLang 发布版 lmsysorg/sglang:v0.5.19-cu130（Docker Hub 摘要 sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9，构建提交 sgl-project/sglang@0bcd822 = 标签 v0.5.19）。模型、TP4/EP1 与 TP4/EP4 DP-attention 拓扑、8k1k 工作负载、并发范围以及 benchmarks/single_node/fixed_seq_len/dsr1_fp4_b200.sh 均保持不变。

基线

• 发布日期： 2026-08-06（workflow-info?date=2026-08-06；benchmarks?model=DeepSeek-R1-0528&date=2026-08-06&exact=true&sequence=8k/1k，按硬件 b200、框架 sglang、精度 fp4、无投机解码、非分离式、ISL/OSL 8192/1024 过滤；evaluations?model=DeepSeek-R1-0528&date=2026-08-06&exact=true）
• 旧镜像： lmsysorg/sglang:v0.5.16-cu130（Docker Hub 摘要 sha256:7b6a35df9839fd593a94a1eaee82d7777f472225d9f3ad1f8a2e0cb2bd1785d0，构建提交 sgl-project/sglang@fdebc93 = 标签 v0.5.16，CUDA 13.0.1，FlashInfer 0.6.14，sgl-kernel 0.4.5）
• 工作负载 / 拓扑： 单节点 B200（cluster:b200-nscale），nvidia/DeepSeek-R1-0528-FP4-V2，SGLang NVFP4，无投机解码，固定序列长度 8k1k（ISL 8192 / OSL 1024），随机数据集；TP4/EP1 并发 1–32，TP4/EP4 DP-attention 并发 64–256（共 9 个点）
• 生产运行： 全部 9 个点均来自 run 30952773184 attempt 1（head d4363bd7fd5bda1391d2d0a46d834cecc363c0d6，changelog PR #2492）；曲线快照 curve_workflow_run_id 2261 仅为逻辑快照，不是生产运行
• 已发布评测： N/A —— 评测接口中没有该配置在 2026-08-06 的 gsm8k 记录；该配置最新的已发布 gsm8k 为 2026-03-28、基于更旧镜像，不可比较

点位	每 GPU 吞吐 (tok/s)	每 GPU 输出吞吐 (tok/s)	TPOT 中位数 (ms)	TTFT 中位数 (ms)
TP4/EP1 c1	389.7	43.5	5.52	207
TP4/EP1 c2	672.2	75.4	6.37	245
TP4/EP1 c4	1134.2	126.2	7.48	245
TP4/EP1 c8	1778.0	199.9	9.49	269
TP4/EP1 c16	2558.7	283.5	13.27	494
TP4/EP1 c32	3428.7	383.4	19.75	769
TP4/EP4 DPA c64	4603.2	510.7	29.73	857
TP4/EP4 DPA c128	6249.9	692.3	44.39	831
TP4/EP4 DPA c256	7955.2	884.2	66.38	4092

🤖 Generated with Claude Code

---

Note

Low Risk
Config and changelog-only change; no benchmark script or topology edits, only a pinned container image version update.

Overview
Bumps the dsr1-fp4-b200-sglang perf config container image from lmsysorg/sglang:v0.5.16-cu130 to lmsysorg/sglang:v0.5.19-cu130 in configs/nvidia-master.yaml. The new image brings CUDA 13.0.3, FlashInfer 0.6.18, and sgl-kernel 0.4.6.post1 (vs 13.0.1 / 0.6.14 / 0.4.5 on the old tag).

A matching perf-changelog.yaml entry records the digest, build commit, and states that model, B200 topology (TP4/EP1 and TP4/EP4 DP-attention), 8k1k fixed-seq-len grid, and benchmarks/single_node/fixed_seq_len/dsr1_fp4_b200.sh are unchanged—so this is a runtime refresh for comparable re-benchmarking only.

^{Reviewed by Cursor Bugbot for commit 25f8f98. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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]

Initial attempt

• Image / SHA: lmsysorg/sglang:v0.5.19-cu130 (Docker Hub digest sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9, build commit sgl-project/sglang@0bcd822 = tag v0.5.19, image created 2026-09-04) at PR head 2e86dbaa83a43e468d874965c3e16b787f157331
• Change: master image only (configs/nvidia-master.yaml, key dsr1-fp4-b200-sglang); benchmarks/single_node/fixed_seq_len/dsr1_fp4_b200.sh unchanged, no runtime patching; the dsr1-fp4-b200-sglang-mtp sibling is not touched
• Targeted run: e2e run 34548446011 (e2e-tests.yml from main, ref=2e86dbaa, test-config --config-files configs/nvidia-master.yaml --config-keys dsr1-fp4-b200-sglang --trim-conc, fail-fast, Klaud background priority; dispatched 2026-09-11T00:54Z) — completed, success (all jobs terminal: both benchmark jobs and both eval-only jobs succeeded; collect-results, collect-evals and calc-success-rate green; finished 2026-09-11T03:04Z). Points: TP4/EP1 concurrency 1 and TP4/EP4 DP-attention concurrency 64, both with the gsm8k smoke eval. Startup/compatibility check only, not the full curve.
• Upstream source comparison (v0.5.16 fdebc93 → v0.5.19 0bcd822, image labels confirm both build commits):
  • python/sglang/srt/server_args.py: every flag the script passes still exists as a ServerArgs field in both versions (enable_dp_attention, enable_dp_attention_local_control_broadcast, enable_dp_lm_head, schedule_conservativeness, enable_prefill_delayer, scheduler_recv_interval, enable_symm_mem, moe_runner_backend, attention_backend, kv_cache_dtype, mem_fraction_static, max_running_requests, ep_size, quantization, stream_interval, chunked_prefill_size, enable_flashinfer_allreduce_fusion, context_length). --cuda-graph-max-bs and --disable-piecewise-cuda-graph are deprecated aliases for --cuda-graph-max-bs-decode and --cuda-graph-backend-prefill=disabled in both versions, with identical mapping, so no rename affects the recipe. SGLANG_RADIX_FORCE_MISS is still defined in environ.py and consumed in managers/schedule_policy.py in both versions.
  • Bundled stack (python/pyproject.toml and image layers): torch 2.11.0 → 2.13.0, FlashInfer 0.6.14 → 0.6.18, sgl-kernel 0.4.5 → 0.4.6.post1, DeepGEMM 0.1.4.post1 → 0.1.7, CUDA 13.0.1 → 13.0.3, cuDNN 9.13 → 9.14; torchao removed.
  • Behavior changes on this recipe's path, from the v0.5.17, v0.5.18 and v0.5.19 notes: MoE deferred finalize is on by default for NVFP4 + flashinfer_trtllm DeepSeek-V3-family models (#33618); FlashInfer MNNVL pure allreduce is auto-enabled for DeepSeek-V3-family models (#30700); piecewise prefill graph is avoided for trtllm_mla (#32785); breakable prefill CUDA graph became the DP-attention default (#31682) but the recipe pins the prefill graph backend to disabled; the unified radix tree is now the default cache (#35081) while the recipe forces cache misses; DP-attention decode→extend prefix off-by-one fix (#37505). These are default/implementation changes, not option removals, so the launch command is kept as is.
  • Cluster evidence: qwen3.5-fp4-b200-sglang already runs lmsysorg/sglang:v0.5.19-cu130 on cluster:b200-nscale with a green full sweep (#2954).
• Benchmark results (smoke points, results_bmk/agg_bmk.json, image lmsysorg/sglang:v0.5.19-cu130, vs. the frozen 2026-08-06 baseline; per-GPU throughput and median latencies):

Point	tput/GPU new (tok/s)	tput/GPU base	Δ tput	median TPOT new (ms)	base	Δ TPOT	median TTFT new (ms)	base	Δ TTFT
TP4/EP1 c1	406.9	389.7	+4.4%	5.27	5.52	−4.5%	209	207	+0.6%
TP4/EP4 DPA c64	4942.5	4603.2	+7.4%	27.83	29.73	−6.4%	757	857	−11.7%

• Eval results: TP4/EP4 DPA c64 gsm8k exact_match strict 0.9560 (flexible 0.9560, n=1319, stderr 0.0056), meta_env.json matches TP4/EP4/DPA conc 64; published baseline eval N/A (no gsm8k row for this identity dated 2026-08-06). TP4/EP1 c1 gsm8k exact_match strict 0.9515 (flexible 0.9538, n=1319, stderr 0.0059), meta_env.json matches TP4/EP1 conc 1. Both evals ran against the smoke server's own image; eval_results_all/agg_eval_all.json lists both rows with infrastructure_success: true.
• Outcome: targeted smoke passed (startup, both deployment shapes, both smoke evals). This is targeted success only, not final validation.
• Next step: append the perf-changelog entry, recheck capacity and start the final full sweep (full-sweep-enabled) on the new PR head.

---

初始尝试

• 镜像 / SHA： lmsysorg/sglang:v0.5.19-cu130（Docker Hub 摘要 sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9，构建提交 sgl-project/sglang@0bcd822 = 标签 v0.5.19，镜像创建于 2026-09-04），PR head 2e86dbaa83a43e468d874965c3e16b787f157331
• 改动： 仅主配置镜像（configs/nvidia-master.yaml，键 dsr1-fp4-b200-sglang）；benchmarks/single_node/fixed_seq_len/dsr1_fp4_b200.sh 未改动，无运行时补丁；未触碰 dsr1-fp4-b200-sglang-mtp 同族配方
• 定向运行： e2e run 34548446011（从 main 触发 e2e-tests.yml，ref=2e86dbaa，test-config --config-files configs/nvidia-master.yaml --config-keys dsr1-fp4-b200-sglang --trim-conc，fail-fast，Klaud 后台优先级；2026-09-11T00:54Z 触发）—— 已完成，成功（所有作业均已结束：两个基准作业与两个仅评测作业均成功；collect-results、collect-evals 与 calc-success-rate 均为绿色；2026-09-11T03:04Z 结束）。点位：TP4/EP1 并发 1 与 TP4/EP4 DP-attention 并发 64，均带 gsm8k 冒烟评测。仅为启动/兼容性检查，不是完整曲线。
• 上游源码对比（v0.5.16 fdebc93 → v0.5.19 0bcd822，两个镜像的标签均确认了构建提交）：
  • python/sglang/srt/server_args.py：脚本传入的所有参数在两个版本中都仍是 ServerArgs 字段。--cuda-graph-max-bs 与 --disable-piecewise-cuda-graph 在两个版本中都是 --cuda-graph-max-bs-decode 与 --cuda-graph-backend-prefill=disabled 的弃用别名，映射一致，没有重命名影响本配方。SGLANG_RADIX_FORCE_MISS 在两个版本的 environ.py 中仍有定义，并在 managers/schedule_policy.py 中被使用。
  • 捆绑组件（python/pyproject.toml 与镜像层）：torch 2.11.0 → 2.13.0，FlashInfer 0.6.14 → 0.6.18，sgl-kernel 0.4.5 → 0.4.6.post1，DeepGEMM 0.1.4.post1 → 0.1.7，CUDA 13.0.1 → 13.0.3，cuDNN 9.13 → 9.14；移除了 torchao。
  • 影响本配方路径的行为变化（来自 v0.5.17、v0.5.18、v0.5.19 发布说明）：NVFP4 + flashinfer_trtllm 的 DeepSeek-V3 系列默认开启 MoE deferred finalize（#33618）；DeepSeek-V3 系列自动启用 FlashInfer MNNVL 纯 allreduce（#30700）；trtllm_mla 不再使用分段 prefill 图（#32785）；DP-attention 默认改为可中断 prefill CUDA 图（#31682），但本配方将 prefill 图后端固定为 disabled；统一 radix 树成为默认缓存（#35081），而本配方强制缓存未命中；修复 DP-attention decode→extend 前缀差一错误（#37505）。这些都是默认值/实现层面的变化，并非删除选项，因此启动命令保持不变。
  • 集群证据：qwen3.5-fp4-b200-sglang 已在 cluster:b200-nscale 上使用 lmsysorg/sglang:v0.5.19-cu130 并通过完整扫描（#2954）。
• 基准结果（冒烟点位，results_bmk/agg_bmk.json，镜像 lmsysorg/sglang:v0.5.19-cu130，对比冻结的 2026-08-06 基线；每 GPU 吞吐与中位延迟）：

点位	新每 GPU 吞吐 (tok/s)	基线	Δ 吞吐	新 TPOT 中位数 (ms)	基线	Δ TPOT	新 TTFT 中位数 (ms)	基线	Δ TTFT
TP4/EP1 c1	406.9	389.7	+4.4%	5.27	5.52	−4.5%	209	207	+0.6%
TP4/EP4 DPA c64	4942.5	4603.2	+7.4%	27.83	29.73	−6.4%	757	857	−11.7%

• 评测结果： TP4/EP4 DPA c64 gsm8k exact_match strict 0.9560（flexible 0.9560，n=1319，标准误 0.0056），meta_env.json 与 TP4/EP4/DPA 并发 64 一致；已发布基线评测 N/A（该配置没有 2026-08-06 的 gsm8k 记录）。TP4/EP1 c1 gsm8k exact_match strict 0.9515（flexible 0.9538，n=1319，标准误 0.0059），meta_env.json 与 TP4/EP1 并发 1 一致。两次评测均针对冒烟服务自身的镜像运行；eval_results_all/agg_eval_all.json 包含两行且 infrastructure_success: true。
• 结果： 定向冒烟通过（启动、两种部署形态、两次冒烟评测）。这只是定向成功，不是最终验证。
• 下一步： 追加 perf-changelog 条目，重新检查容量，并在新的 PR head 上启动最终完整扫描（full-sweep-enabled）。

[Klaud-Cold]

Final full sweep

• Image / SHA: lmsysorg/sglang:v0.5.19-cu130 (digest sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9) at PR head 25f8f98fa258680daf485f6d6b60084eb961d24b (rebased onto main a5f23f46 so the perf-changelog diff is purely additive; utils/validate_perf_changelog.py --base-ref origin/main passes and all prior changelog bytes are preserved)
• Changes since the Initial attempt: appended the dsr1-fp4-b200-sglang perf-changelog entry (whole family, no scenario/eval-selection/append-only modifiers); the image line is unchanged
• Run: run-sweep 34557177019 (full-sweep-enabled, PR kept draft; started 2026-09-11T03:06Z, completed 04:32Z) — completed, success: all 9 benchmark jobs, all 3 eval jobs, check-changelog, collect-results, collect-evals, compare-results, calc-success-rate and upload-changelog-metadata succeeded; no failures or retries. klaud-sweep-manifest records head 25f8f98f, run 34557177019 attempt 1, full-sweep: true. An earlier push-triggered run 34557151441 was cancelled at setup by the workflow concurrency group before any GPU work when this labeled run started. Matrix: 9 points on cluster:b200-nscale — TP4/EP1 concurrency 1, 2, 4, 8, 16, 32 and TP4/EP4 DP-attention concurrency 64, 128, 256, with default gsm8k evals at TP4/EP1 c32 and TP4/EP4 c128/c256
• Benchmark results (results_bmk/agg_bmk.json, 9 of 9 points, image lmsysorg/sglang:v0.5.19-cu130, vs. the frozen 2026-08-06 baseline from run 30952773184; same topology, concurrency and 8k1k random dataset per point):

Point	tput/GPU new (tok/s)	base	Δ	median TPOT new (ms)	base	Δ	median TTFT new (ms)	base	Δ	median E2EL new (s)	base	Δ
TP4/EP1 c1	407.4	389.7	+4.6%	5.27	5.52	−4.5%	208	207	+0.1%	4.99	5.22	−4.3%
TP4/EP1 c2	704.4	672.2	+4.8%	6.06	6.37	−4.8%	250	245	+2.3%	5.91	6.26	−5.5%
TP4/EP1 c4	1179.5	1134.2	+4.0%	7.21	7.48	−3.6%	247	245	+0.8%	6.79	7.08	−4.0%
TP4/EP1 c8	1828.3	1778.0	+2.8%	9.26	9.49	−2.5%	267	269	−0.6%	8.88	9.09	−2.2%
TP4/EP1 c16	2584.5	2558.7	+1.0%	13.14	13.27	−0.9%	491	494	−0.5%	12.46	12.57	−0.9%
TP4/EP1 c32	3487.5	3428.7	+1.7%	19.35	19.75	−2.0%	764	769	−0.5%	18.52	19.06	−2.8%
TP4/EP4 DPA c64	4918.8	4603.2	+6.9%	27.91	29.73	−6.1%	780	857	−9.0%	26.67	28.46	−6.3%
TP4/EP4 DPA c128	6637.6	6249.9	+6.2%	41.90	44.39	−5.6%	762	831	−8.3%	39.25	41.73	−5.9%
TP4/EP4 DPA c256	8365.6	7955.2	+5.2%	62.93	66.38	−5.2%	3445	4092	−15.8%	62.28	65.66	−5.1%

Throughput improves on every point (+1.0% to +6.9%); TPOT and E2EL improve everywhere. The only regressions are median TTFT at TP4/EP1 c1 (+0.1%), c2 (+2.3%) and c4 (+0.8%), all within a few milliseconds.

• Eval results (eval_results_all/agg_eval_all.json, 3 of 3 default gsm8k evals, infrastructure_success: true, n=1319 each): TP4/EP1 c32 em_strict 0.9545 (flexible 0.9545); TP4/EP4 DPA c128 em_strict 0.9545 (flexible 0.9553); TP4/EP4 DPA c256 em_strict 0.9538 (flexible 0.9553). Published baseline eval N/A (no gsm8k row for this identity dated 2026-08-06); the smoke evals on this image scored 0.9515 (c1) and 0.9560 (c64).
• Outcome: final validation passed on the exact PR head. This is final validation of the family sweep, not global PR approval; CODEOWNER review and merge remain human steps.
• Next step: run finish to verify matrix/result coverage and mark the PR ready for review.

---

最终完整扫描

• 镜像 / SHA： lmsysorg/sglang:v0.5.19-cu130（摘要 sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9），PR head 25f8f98fa258680daf485f6d6b60084eb961d24b（已重新基于 main a5f23f46，使 perf-changelog 差异纯为追加；utils/validate_perf_changelog.py --base-ref origin/main 通过，所有既有 changelog 字节保持不变）
• 相对初始尝试的改动： 追加了 dsr1-fp4-b200-sglang 的 perf-changelog 条目（整个配方族，不带 scenario/评测选择/append-only 修饰符）；镜像行未变
• 运行： run-sweep 34557177019（full-sweep-enabled，PR 保持草稿；2026-09-11T03:06Z 开始，04:32Z 完成）—— 已完成，成功：9 个基准作业、3 个评测作业、check-changelog、collect-results、collect-evals、compare-results、calc-success-rate 与 upload-changelog-metadata 全部成功；无失败或重试。klaud-sweep-manifest 记录 head 25f8f98f、run 34557177019 attempt 1、full-sweep: true。此前由推送触发的 run 34557151441 在本次带标签运行开始时被工作流并发组于 setup 阶段取消，未产生任何 GPU 工作。矩阵：cluster:b200-nscale 上 9 个点 —— TP4/EP1 并发 1、2、4、8、16、32 与 TP4/EP4 DP-attention 并发 64、128、256，默认 gsm8k 评测位于 TP4/EP1 c32 及 TP4/EP4 c128/c256
• 基准结果（results_bmk/agg_bmk.json，9/9 个点，镜像 lmsysorg/sglang:v0.5.19-cu130，对比来自 run 30952773184 的冻结 2026-08-06 基线；各点拓扑、并发与 8k1k 随机数据集一致）：

点位	新每 GPU 吞吐 (tok/s)	基线	Δ	新 TPOT 中位数 (ms)	基线	Δ	新 TTFT 中位数 (ms)	基线	Δ	新 E2EL 中位数 (s)	基线	Δ
TP4/EP1 c1	407.4	389.7	+4.6%	5.27	5.52	−4.5%	208	207	+0.1%	4.99	5.22	−4.3%
TP4/EP1 c2	704.4	672.2	+4.8%	6.06	6.37	−4.8%	250	245	+2.3%	5.91	6.26	−5.5%
TP4/EP1 c4	1179.5	1134.2	+4.0%	7.21	7.48	−3.6%	247	245	+0.8%	6.79	7.08	−4.0%
TP4/EP1 c8	1828.3	1778.0	+2.8%	9.26	9.49	−2.5%	267	269	−0.6%	8.88	9.09	−2.2%
TP4/EP1 c16	2584.5	2558.7	+1.0%	13.14	13.27	−0.9%	491	494	−0.5%	12.46	12.57	−0.9%
TP4/EP1 c32	3487.5	3428.7	+1.7%	19.35	19.75	−2.0%	764	769	−0.5%	18.52	19.06	−2.8%
TP4/EP4 DPA c64	4918.8	4603.2	+6.9%	27.91	29.73	−6.1%	780	857	−9.0%	26.67	28.46	−6.3%
TP4/EP4 DPA c128	6637.6	6249.9	+6.2%	41.90	44.39	−5.6%	762	831	−8.3%	39.25	41.73	−5.9%
TP4/EP4 DPA c256	8365.6	7955.2	+5.2%	62.93	66.38	−5.2%	3445	4092	−15.8%	62.28	65.66	−5.1%

所有点的吞吐均有提升（+1.0% 至 +6.9%）；TPOT 与 E2EL 全部改善。仅有的回退是 TP4/EP1 c1（+0.1%）、c2（+2.3%）、c4（+0.8%）的 TTFT 中位数，均在几毫秒以内。

• 评测结果（eval_results_all/agg_eval_all.json，3/3 个默认 gsm8k 评测，infrastructure_success: true，各 n=1319）：TP4/EP1 c32 em_strict 0.9545（flexible 0.9545）；TP4/EP4 DPA c128 em_strict 0.9545（flexible 0.9553）；TP4/EP4 DPA c256 em_strict 0.9538（flexible 0.9553）。已发布基线评测 N/A（该配置没有 2026-08-06 的 gsm8k 记录）；本镜像的冒烟评测得分为 0.9515（c1）与 0.9560（c64）。
• 结果： 最终验证在精确的 PR head 上通过。这是配方族扫描的最终验证，不是 PR 的全局批准；CODEOWNER 审阅与合并仍由人工完成。
• 下一步： 运行 finish 验证矩阵/结果覆盖并将 PR 标记为可审阅。

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34557177019
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34557177019

[Klaud-Cold]

Klaud Cold: validated. All owned runs are terminal. Repairs: 0. Runs: 34548446011, 34557151441, 34557177019.

The full sweep is verified; this PR remains ready for review.

---

Klaud Cold：validated。所有自有运行均已结束。修复次数：0。运行：34548446011, 34557151441, 34557177019。

完整 sweep 已通过验证；PR 保持就绪，等待审查。

[claude]

Looks good, straightforward image-version bump. Reviewed the config diff (single image field change in the dsr1-fp4-b200-sglang recipe, no other fields touched) and the changelog diff (new entry appended strictly at the tail with no edits to prior bytes, matching the append-only convention).

Extended reasoning...

Overview

The PR touches exactly two files: configs/nvidia-master.yaml, where the dsr1-fp4-b200-sglang recipe's image field is bumped from lmsysorg/sglang:v0.5.16-cu130 to lmsysorg/sglang:v0.5.19-cu130 with no other field changes, and perf-changelog.yaml, which gains one new entry appended at the physical tail of the file describing the bump (old/new build commits, CUDA/FlashInfer/sgl-kernel versions, and a digest for the new image) plus an assertion that the benchmark script's model/backend/dtype/concurrency-grid configuration is unchanged.

Security risks

None identified. This is a data-only YAML change (an image tag string and a changelog text entry) with no code execution, no credentials, no auth/permissions logic, and no injection surface.

Level of scrutiny

Low. This matches the repository's documented "Klaud Cold" automated image-bump pattern: an exact-version image tag bump paired with a changelog entry citing build commits and a digest. I verified via git diff --stat that the config change is a single line and the changelog change is purely additive at the tail (no historical bytes edited), which satisfies the append-only, byte-sensitive requirement documented in AGENTS.md/CONTRIBUTING.md for this changelog file. The recipe is single-node (no multinode/model.container field), so the multi-node model.container == image coupling rule noted in the task context does not apply here.

Other factors

No CODEOWNERS-restricted paths appear to be implicated beyond standard config/changelog ownership, and the PR conversation timeline shows no outstanding CHANGES_REQUESTED or unaddressed third-party objections. No bugs were reported by the bug-hunting system, and my own review of the diff found nothing beyond the mechanical, well-formed change described.