[Klaud Cold] Update dsr1-fp8-h200-sglang-mtp SGLang image to v0.5.19-cu130 / 将 dsr1-fp8-h200-sglang-mtp 的 SGLang 镜像升级至 v0.5.19-cu130

[Klaud-Cold]

Update the dsr1-fp8-h200-sglang-mtp master image from lmsysorg/sglang:v0.5.12-cu130 to the current SGLang release lmsysorg/sglang:v0.5.19-cu130 (Docker Hub digest sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9, tag commit sgl-project/sglang@0bcd822). The recipe script, model, TP8/EP1 topology, EAGLE/MTP settings and the 8k1k concurrency range are unchanged.

Baseline

• Published date: 2026-05-20 (benchmarks?model=DeepSeek-R1-0528&date=2026-05-20&exact=true, workflow-info?date=2026-05-20, evaluations?model=DeepSeek-R1-0528&date=2026-05-20&exact=true)
• Old image: lmsysorg/sglang:v0.5.12-cu130 (digest sha256:42194170546745092e74cd5f81ad32a7c6e944c7111fe7bf13588152277ff356, tag commit sgl-project/sglang@127b9e3)
• Workload / topology: single-node H200, deepseek-ai/DeepSeek-R1-0528, SGLang FP8, TP8 EP1, EAGLE MTP (2 steps, top-k 1, 3 draft tokens), fixed-seq-len 8k1k (ISL 8192 / OSL 1024), concurrency 4 / 8 / 16 / 32 / 64, random dataset with chat template
• Producer: run 26135241056 (head 7ec590983c089c6262b1958c8325e9d06ef7f62c, changelog PR #1523); benchmark result IDs 413852, 413854, 413845, 413847, 413850

Conc	Total tok/s/GPU	Output tok/s/GPU	Median TTFT (s)	Median TPOT (ms)	Median E2E (s)
4	431.2	48.0	0.325	10.09	9.34
8	650.4	73.1	0.343	13.39	12.39
16	932.9	103.4	0.372	18.74	17.70
32	1237.8	138.4	0.372	28.21	26.25
64	1622.4	180.0	0.605	44.23	40.98

• Published eval: gsm8k at concurrency 32 and 64, em_strict 0.9545 / em_flexible 0.9560 (evaluation IDs 6795 and 6794, same producer run). The public evaluations feed labels these rows disagg: true, which does not match the single-node TP8 recipe; the benchmark rows above carry the correct disagg: false identity.

---

将 dsr1-fp8-h200-sglang-mtp 的主配置镜像从 lmsysorg/sglang:v0.5.12-cu130 升级到当前 SGLang 发布版 lmsysorg/sglang:v0.5.19-cu130（Docker Hub 摘要 sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9，标签提交 sgl-project/sglang@0bcd822）。配方脚本、模型、TP8/EP1 拓扑、EAGLE/MTP 设置以及 8k1k 并发范围均保持不变。

基线

• 发布日期： 2026-05-20（benchmarks?model=DeepSeek-R1-0528&date=2026-05-20&exact=true，workflow-info?date=2026-05-20，evaluations?model=DeepSeek-R1-0528&date=2026-05-20&exact=true）
• 旧镜像： lmsysorg/sglang:v0.5.12-cu130（摘要 sha256:42194170546745092e74cd5f81ad32a7c6e944c7111fe7bf13588152277ff356，标签提交 sgl-project/sglang@127b9e3）
• 工作负载 / 拓扑： 单节点 H200，deepseek-ai/DeepSeek-R1-0528，SGLang FP8，TP8 EP1，EAGLE MTP（2 步、top-k 1、3 个草稿 token），固定序列长度 8k1k（ISL 8192 / OSL 1024），并发 4 / 8 / 16 / 32 / 64，随机数据集并使用聊天模板
• 数据来源： 运行 26135241056（head 7ec590983c089c6262b1958c8325e9d06ef7f62c，changelog PR #1523）；基准结果 ID 413852、413854、413845、413847、413850

并发	总吞吐 tok/s/GPU	输出吞吐 tok/s/GPU	TTFT 中位数 (s)	TPOT 中位数 (ms)	端到端中位数 (s)
4	431.2	48.0	0.325	10.09	9.34
8	650.4	73.1	0.343	13.39	12.39
16	932.9	103.4	0.372	18.74	17.70
32	1237.8	138.4	0.372	28.21	26.25
64	1622.4	180.0	0.605	44.23	40.98

• 已发布评测： 并发 32 与 64 的 gsm8k，em_strict 0.9545 / em_flexible 0.9560（评测 ID 6795 与 6794，同一数据来源运行）。公共评测接口将这些行标记为 disagg: true，与单节点 TP8 配方不符；上表基准行的 disagg: false 身份是正确的。

🤖 Generated with Claude Code

---

Note

Low Risk
Benchmark config and changelog only; no application runtime or security-sensitive code paths.

Overview
Bumps the SGLang container image for the dsr1-fp8-h200-sglang-mtp benchmark recipe from lmsysorg/sglang:v0.5.12-cu130 to v0.5.19-cu130 in configs/nvidia-master.yaml.

Adds a matching perf-changelog.yaml entry (PR #2955) noting the new digest and that the launch script, TP8/EP1 topology, EAGLE/MTP settings, and 8k/1k workload are unchanged.

^{Reviewed by Cursor Bugbot for commit 735b758. 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 (digest sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9) at PR head 63c66bccd36b7443c2a27dba9e66d44778688257
• Change: master image only (configs/nvidia-master.yaml, key dsr1-fp8-h200-sglang-mtp); benchmarks/single_node/fixed_seq_len/dsr1_fp8_h200_mtp.sh unchanged, no runtime patching
• Targeted run: e2e run 34446190942 (e2e-tests.yml from main, ref=63c66bcc, test-config --trim-conc: TP8 EP1 MTP at concurrency 4 plus the smoke gsm8k eval; dispatched 2026-09-10T06:39Z) — completed, success: benchmark job on h200-cw_00, eval-only job on h200-cw_01, collectors and success-rate check all green. Server log shows The server is fired up and ready to roll! on both nodes with the resolved settings attention_backend=flashinfer, speculative_algorithm=EAGLE, draft path auto-set to the target model, steps 2 / top-k 1 / draft tokens 3, mem_fraction_static=0.82, disable_radix_cache=True; the only new log line is the expected SGLANG_ENABLE_SPEC_V2 has been removed warning.
• Benchmark result (concurrency 4, 8k1k, vs published 2026-05-20 baseline, result ID 413852):

Metric	Baseline v0.5.12	v0.5.19	Delta
Total tok/s/GPU	431.2	586.7	+36.1%
Output tok/s/GPU	48.0	65.3	+36.1%
Median TTFT (s)	0.325	0.311	-4.3%
Median TPOT (ms)	10.09	7.23	-28.3%
Median E2E (s)	9.34	6.65	-28.8%

Mean speculative accept length over the smoke's 4838 decode batches was 2.49 of a maximum 3 (baseline accept length N/A: not published). power_valid=1 in the aggregate.

• Eval result: gsm8k at concurrency 4 (smoke selection), em_strict 0.9606 / em_flexible 0.9621, n = 1319, infrastructure_success=true. Published baseline eval was 0.9545 / 0.9560 at concurrency 32 and 64, so the numbers are indicative, not a like-for-like point.
• Next step: append the perf-changelog.yaml entry, recheck capacity and start the final full sweep (all five concurrency points plus the default evals) with full-sweep-enabled on the draft PR.
• Upstream source comparison: v0.5.12 @ 127b9e3 (2026-05-16) → v0.5.19 @ 0bcd822 (2026-09-04). Provenance: both images carry ai.sglang.build.commit / org.opencontainers.image.revision labels equal to those tag commits, and v0.5.19 and v0.5.19-cu130 share one Docker Hub digest pushed 2026-09-04 (docker/Dockerfile clones --branch v${SGL_VERSION}).
• Coupled dependencies: torch 2.11.0 → 2.13.0, flashinfer 0.6.11.post1 → 0.6.18, CUDA base 13.0.1 → 13.0.3 (same CUDA 13.0 / driver ≥ 535 requirement as the old cu130 image).
• Flag audit: every launch flag used by the script (--speculative-algorithm EAGLE, --speculative-num-steps, --speculative-num-draft-tokens, --speculative-eagle-topk, --attention-backend flashinfer, --mem-fraction-static, --disable-radix-cache, --chunked-prefill-size, --max-prefill-tokens, --cuda-graph-max-bs, --max-running-requests, --stream-interval, --decode-log-interval, --ep-size, --data-parallel-size, --trust-remote-code, --context-length) is still defined in v0.5.19 (server_args.py plus the new arg_groups/ hooks). For DeepseekV3ForCausalLM, speculative_hook.py still auto-sets the MTP draft path and honors the explicit steps/top-k/draft-token values. SGLANG_ENABLE_SPEC_V2=1 (set by the script) is now a removal warning only: spec V2 is always on since v0.5.13 (#25464), whereas v0.5.12 gated V2 behind that env var, so the executed speculative path is unchanged.
• Behavior deltas noted: CUTLASS FP8 blockwise GEMM was deleted for SM90 in v0.5.16 (#30438) and v0.5.19 adds an SM90 FP8 decode routing fix (#37018), so H200 FP8 GEMM kernels differ from the baseline image; unified radix tree is the default in v0.5.19 (#35081; moot under --disable-radix-cache); compiled-kernel caches moved under SGLANG_CACHE_DIR in v0.5.18 (#32434). The server log also reports no tuned Triton MoE config for E=257,N=256 fp8 on H200 under Triton 3.7.1 and falls back to the previous Triton version's config; this is informational, not an error. The sibling qwen3.5-fp8-h100-sglang-mtp ([Klaud Cold] Update qwen3.5-fp8-h100-sglang-mtp SGLang image to v0.5.19-cu130 / 将 qwen3.5-fp8-h100-sglang-mtp 的 SGLang 镜像升级至 v0.5.19-cu130 #2952) already ran its concurrency-4 EAGLE MTP smoke on v0.5.19-cu130 on Hopper (SM90).
• Decision: image-only bump; no in-scope flag changes required.

---

初始尝试

• 镜像 / SHA： lmsysorg/sglang:v0.5.19-cu130（摘要 sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9），PR head 63c66bccd36b7443c2a27dba9e66d44778688257
• 改动： 仅主配置镜像（configs/nvidia-master.yaml 中的 dsr1-fp8-h200-sglang-mtp）；benchmarks/single_node/fixed_seq_len/dsr1_fp8_h200_mtp.sh 未改动，无任何运行时补丁
• 定向运行： e2e 运行 34446190942（从 main 触发 e2e-tests.yml，ref=63c66bcc，test-config --trim-conc：TP8 EP1 MTP 并发 4 加冒烟 gsm8k 评测；于 2026-09-10T06:39Z 触发）—— 已完成，成功：基准任务在 h200-cw_00，仅评测任务在 h200-cw_01，收集器与成功率检查全部通过。两台节点的服务日志均出现 The server is fired up and ready to roll!，解析后的设置为 attention_backend=flashinfer、speculative_algorithm=EAGLE、草稿模型路径自动设为目标模型、步数 2 / top-k 1 / 草稿 token 3、mem_fraction_static=0.82、disable_radix_cache=True；唯一新增日志为预期中的 SGLANG_ENABLE_SPEC_V2 has been removed 提示。
• 基准结果（并发 4，8k1k，对比 2026-05-20 已发布基线，结果 ID 413852）：

指标	基线 v0.5.12	v0.5.19	变化
总吞吐 tok/s/GPU	431.2	586.7	+36.1%
输出吞吐 tok/s/GPU	48.0	65.3	+36.1%
TTFT 中位数 (s)	0.325	0.311	-4.3%
TPOT 中位数 (ms)	10.09	7.23	-28.3%
端到端中位数 (s)	9.34	6.65	-28.8%

冒烟运行 4838 个解码批次的平均投机接受长度为 2.49（上限 3）（基线接受长度 N/A：未发布）。聚合结果中 power_valid=1。

• 评测结果： 并发 4 的 gsm8k（冒烟选择），em_strict 0.9606 / em_flexible 0.9621，n = 1319，infrastructure_success=true。已发布基线评测为并发 32 与 64 的 0.9545 / 0.9560，因此该数值仅供参考，不是同一并发点的对比。
• 下一步： 追加 perf-changelog.yaml 条目，重新检查容量，并在草稿 PR 上以 full-sweep-enabled 启动最终完整 sweep（全部五个并发点及默认评测）。
• 上游源码对比： v0.5.12 @ 127b9e3（2026-05-16）→ v0.5.19 @ 0bcd822（2026-09-04）。来源确认：两个镜像的 ai.sglang.build.commit / org.opencontainers.image.revision 标签均等于对应标签提交，且 v0.5.19 与 v0.5.19-cu130 在 Docker Hub 上共享同一摘要（2026-09-04 推送）；docker/Dockerfile 以 --branch v${SGL_VERSION} 克隆。
• 耦合依赖： torch 2.11.0 → 2.13.0，flashinfer 0.6.11.post1 → 0.6.18，CUDA 基础镜像 13.0.1 → 13.0.3（与旧 cu130 镜像相同的 CUDA 13.0 / 驱动 ≥ 535 要求）。
• 参数审计： 脚本使用的全部启动参数（--speculative-algorithm EAGLE、--speculative-num-steps、--speculative-num-draft-tokens、--speculative-eagle-topk、--attention-backend flashinfer、--mem-fraction-static、--disable-radix-cache、--chunked-prefill-size、--max-prefill-tokens、--cuda-graph-max-bs、--max-running-requests、--stream-interval、--decode-log-interval、--ep-size、--data-parallel-size、--trust-remote-code、--context-length）在 v0.5.19 中仍有定义（server_args.py 及新的 arg_groups/ 钩子）。对于 DeepseekV3ForCausalLM，speculative_hook.py 仍会自动设置 MTP 草稿模型路径，并沿用显式的步数 / top-k / 草稿 token 数。脚本设置的 SGLANG_ENABLE_SPEC_V2=1 现仅触发移除提示：自 v0.5.13 起 spec V2 始终开启（#25464），而 v0.5.12 由该环境变量控制 V2，因此实际执行的投机解码路径不变。
• 已注意到的行为差异： v0.5.16 删除了 SM90 的 CUTLASS FP8 blockwise GEMM（#30438），v0.5.19 加入 SM90 FP8 解码路由修复（#37018），因此 H200 FP8 GEMM 内核与基线镜像不同；v0.5.19 默认使用统一 radix 树（#35081，在 --disable-radix-cache 下无影响）；v0.5.18 起编译内核缓存迁移至 SGLANG_CACHE_DIR（#32434）。服务日志还提示 H200 上 Triton 3.7.1 没有 E=257,N=256 fp8 的调优 MoE 配置并回退到上一 Triton 版本的配置；这只是信息提示，不是错误。同类配方 qwen3.5-fp8-h100-sglang-mtp（[Klaud Cold] Update qwen3.5-fp8-h100-sglang-mtp SGLang image to v0.5.19-cu130 / 将 qwen3.5-fp8-h100-sglang-mtp 的 SGLang 镜像升级至 v0.5.19-cu130 #2952）已在 Hopper（SM90）上用 v0.5.19-cu130 完成并发 4 的 EAGLE MTP 冒烟。
• 决定： 仅升级镜像；无需在范围内调整参数。

[Klaud-Cold]

Final full sweep

• Image / SHA: lmsysorg/sglang:v0.5.19-cu130 at PR head 7db0a243f3dec67134c66fd7362fcf11152b6e5e (adds the perf-changelog.yaml entry for dsr1-fp8-h200-sglang-mtp on top of the image bump; validated locally with utils/validate_perf_changelog.py against base 01cea4db)
• Trigger: full-sweep-enabled applied to the draft PR at 2026-09-10T07:35Z after a passing capacity check. Run 34450728410 (run-sweep.yml, labeled event, attempt 1) is the final sweep; the push-triggered run 34450701759 was cancelled by the workflow's concurrency group before running any benchmark.
• Scope / outcome: full matrix, TP8 EP1 MTP 8k1k at concurrency 4 / 8 / 16 / 32 / 64 plus the default gsm8k evals at concurrency 32 and 64 — completed, success. All five benchmark jobs ran on h200-dgxc-slurm_00–_04, both eval jobs on h200-cw_00 / _01; check-changelog, setup, collect-results, collect-evals, compare-results, upload-changelog-metadata and calc-success-rate are green. results_bmk/agg_bmk.json holds all five points with power_valid=1 and one recipe fingerprint (f5fa3b86…); the klaud-sweep-manifest records full-sweep: true on this head.
• Benchmark results vs published 2026-05-20 baseline (same TP8 EP1 MTP 8k1k points, result IDs 413852 / 413854 / 413845 / 413847 / 413850):

Conc	Total tok/s/GPU	Output tok/s/GPU	Median TTFT (s)	Median TPOT (ms)	Median E2E (s)
4	431.2 → 571.9 (+32.6%)	48.0 → 63.6 (+32.6%)	0.325 → 0.328 (+0.7%)	10.09 → 7.42 (-26.4%)	9.34 → 6.91 (-26.0%)
8	650.4 → 791.0 (+21.6%)	73.1 → 88.9 (+21.6%)	0.343 → 0.320 (-6.8%)	13.39 → 10.80 (-19.3%)	12.39 → 10.37 (-16.3%)
16	932.9 → 1081.7 (+15.9%)	103.4 → 119.8 (+15.9%)	0.372 → 0.348 (-6.4%)	18.74 → 16.37 (-12.6%)	17.70 → 15.29 (-13.6%)
32	1237.8 → 1401.9 (+13.3%)	138.4 → 156.8 (+13.3%)	0.372 → 0.365 (-1.9%)	28.21 → 25.21 (-10.6%)	26.25 → 23.60 (-10.1%)
64	1622.3 → 1677.0 (+3.4%)	180.0 → 186.0 (+3.4%)	0.605 → 3.318 (+449%)	44.23 → 39.69 (-10.3%)	40.98 → 39.53 (-3.6%)

• Regression to note: median TTFT at concurrency 64 rose from 0.60 s to 3.32 s (mean 1.84 s → 4.33 s; p99 19.6 s → 18.3 s is unchanged). Its server log shows the scheduler holding 57–60 running requests with 4–7 requests queued through steady state (#queue-req peaked at 53 during the ramp), so new prompts wait behind the 32k-token chunked prefill batches before their first token. Throughput (+3.4%), TPOT (-10.3%) and E2E (-3.6%) at that point still improved, and the other four points improved on every metric except a +0.7% TTFT at concurrency 4. Mean speculative accept length at concurrency 64 was 2.38 of a maximum 3 (baseline N/A: not published). No flag change is in scope for this image refresh; the regression is reported for reviewer judgment.
• Eval results: gsm8k em_strict / em_flexible: concurrency 32 → 0.9560 / 0.9560, concurrency 64 → 0.9591 / 0.9606 (n = 1319 each, infrastructure_success=true). Published baseline: 0.9545 / 0.9560 at both points.
• Status: targeted smoke and exact-head final validation both passed; global approval remains with the automatic reviews and CODEOWNER sign-off. Next step: run the Klaud completion check, which verifies matrix/result coverage and marks the PR ready. Smoke evidence is in the Initial attempt.

---

最终完整 sweep

• 镜像 / SHA： lmsysorg/sglang:v0.5.19-cu130，PR head 7db0a243f3dec67134c66fd7362fcf11152b6e5e（在镜像升级之上追加 dsr1-fp8-h200-sglang-mtp 的 perf-changelog.yaml 条目；已用 utils/validate_perf_changelog.py 对基线 01cea4db 本地校验）
• 触发： 容量检查通过后，于 2026-09-10T07:35Z 在草稿 PR 上添加 full-sweep-enabled。运行 34450728410（run-sweep.yml，labeled 事件，第 1 次尝试）为最终 sweep；由推送触发的 运行 34450701759 在执行任何基准之前被工作流并发组取消。
• 范围 / 结果： 完整矩阵，TP8 EP1 MTP 8k1k 并发 4 / 8 / 16 / 32 / 64，以及并发 32 与 64 的默认 gsm8k 评测 —— 已完成，成功。五个基准任务运行于 h200-dgxc-slurm_00–_04，两个评测任务运行于 h200-cw_00 / _01；check-changelog、setup、collect-results、collect-evals、compare-results、upload-changelog-metadata 与 calc-success-rate 均通过。results_bmk/agg_bmk.json 包含全部五个点，power_valid=1，配方指纹一致（f5fa3b86…）；klaud-sweep-manifest 在该 head 上记录 full-sweep: true。
• 基准结果对比 2026-05-20 已发布基线（相同 TP8 EP1 MTP 8k1k 点，结果 ID 413852 / 413854 / 413845 / 413847 / 413850）：

并发	总吞吐 tok/s/GPU	输出吞吐 tok/s/GPU	TTFT 中位数 (s)	TPOT 中位数 (ms)	端到端中位数 (s)
4	431.2 → 571.9 (+32.6%)	48.0 → 63.6 (+32.6%)	0.325 → 0.328 (+0.7%)	10.09 → 7.42 (-26.4%)	9.34 → 6.91 (-26.0%)
8	650.4 → 791.0 (+21.6%)	73.1 → 88.9 (+21.6%)	0.343 → 0.320 (-6.8%)	13.39 → 10.80 (-19.3%)	12.39 → 10.37 (-16.3%)
16	932.9 → 1081.7 (+15.9%)	103.4 → 119.8 (+15.9%)	0.372 → 0.348 (-6.4%)	18.74 → 16.37 (-12.6%)	17.70 → 15.29 (-13.6%)
32	1237.8 → 1401.9 (+13.3%)	138.4 → 156.8 (+13.3%)	0.372 → 0.365 (-1.9%)	28.21 → 25.21 (-10.6%)	26.25 → 23.60 (-10.1%)
64	1622.3 → 1677.0 (+3.4%)	180.0 → 186.0 (+3.4%)	0.605 → 3.318 (+449%)	44.23 → 39.69 (-10.3%)	40.98 → 39.53 (-3.6%)

• 需注意的回退： 并发 64 的 TTFT 中位数从 0.60 s 升至 3.32 s（均值 1.84 s → 4.33 s；p99 19.6 s → 18.3 s 基本不变）。其服务日志显示调度器在稳态下保持 57–60 个运行请求、4–7 个请求排队（#queue-req 在爬升阶段峰值为 53），因此新提示需等待 32k token 的分块 prefill 批次之后才能产出首个 token。该点的吞吐（+3.4%）、TPOT（-10.3%）与端到端（-3.6%）仍有改善，其余四个点除并发 4 的 TTFT +0.7% 外各项指标均改善。并发 64 的平均投机接受长度为 2.38（上限 3）（基线 N/A：未发布）。本次镜像刷新范围内不包含参数调整；该回退如实报告，供审阅者判断。
• 评测结果： gsm8k em_strict / em_flexible：并发 32 → 0.9560 / 0.9560，并发 64 → 0.9591 / 0.9606（各 n = 1319，infrastructure_success=true）。已发布基线：两点均为 0.9545 / 0.9560。
• 状态： 定向冒烟与精确 head 的最终验证均已通过；全局批准仍取决于自动审阅与 CODEOWNER 签署。下一步：运行 Klaud 完成检查，由其校验矩阵/结果覆盖并将 PR 标记为可审阅。冒烟证据见 初始尝试。

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34450728410
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34450728410

[Klaud-Cold]

Klaud Cold: validated. All owned runs are terminal. Repairs: 0. Runs: 34446190942, 34450701759, 34450728410.

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：34446190942, 34450701759, 34450728410。

完整 sweep 已通过验证；PR 保持就绪，等待审查。

[claude]

Claude finished @Klaud-Cold's task in 1m 14s —— View job

---

Review of PR #2955

• Gather context: inspect the diff against origin/main
• Validate configs/nvidia-master.yaml change (image accessibility, key prefix, recipe unchanged)
• Validate perf-changelog.yaml entry (appended at end, validator passes)
• Post review

LGTM - no blocking issues found.

Image-only bump of dsr1-fp8-h200-sglang-mtp to the public Docker Hub tag lmsysorg/sglang:v0.5.19-cu130; the perf-changelog.yaml entry is appended at the very end of the file with the correct config key and PR link, and the unchanged recipe script still passes --use-chat-template and --ep-size $EP_SIZE. The full sweep on this exact head (run 34450728410) is green, including check-changelog, so the validator result is covered by CI even though utils/validate_perf_changelog.py could not be executed in this review session (Bash execution of that script was not permitted). The concurrency-64 median TTFT increase the author reported is a runtime observation for CODEOWNER judgment, not a defect in the diff.

[claude]

I reviewed this PR and didn't find any bugs. Because configs/nvidia-master.yaml is CODEOWNER-owned (per .github/CODEOWNERS: @ ankur-singh, @ kedarpotdar-nv, @ InferenceX/core) and image bumps require verifying the tag is an unmodified upstream SGLang build, a human/CODEOWNER look would still be worthwhile.

What was reviewed:

• The image bump itself (v0.5.12-cu130 -> v0.5.19-cu130) for the dsr1-fp8-h200-sglang-mtp recipe; all other fields (model, model-prefix, runner, TP8/EP1 topology, MTP settings) are unchanged.
• The perf-changelog.yaml entry: correctly appended at the physical tail, no historical entries edited, schema (config-keys/description/pr-link) matches surrounding entries.
• No model.container field exists for this single-node recipe, so the container/image parity rule for multi-node srt-slurm recipes does not apply here.

Extended reasoning...

Overview

This PR is a minimal, mechanical two-file change: a single Docker image tag bump for the dsr1-fp8-h200-sglang-mtp recipe in configs/nvidia-master.yaml (v0.5.12-cu130 to v0.5.19-cu130), and a matching append-only entry in perf-changelog.yaml. I confirmed via git diff that no other fields in the recipe changed (model, model-prefix, runner, TP8/EP1 topology, and EAGLE/MTP settings are untouched), and that the changelog entry was appended strictly at the physical tail without modifying any prior entries, using the same config-keys/description/pr-link schema as neighboring entries.

Security risks

None identified. This is a config value change (image tag) plus a documentation-style changelog append; there is no code execution path, credential handling, or auth/permissions logic involved in the diff itself.

Level of scrutiny

This would normally warrant a light-touch review as a version-bump-only change with a satisfied changelog requirement. However, configs/nvidia-master.yaml is explicitly CODEOWNER-owned (@ ankur-singh, @ kedarpotdar-nv, @ InferenceX/core per .github/CODEOWNERS), and the repository's own review checklist calls for verifying that a bumped SGLang image is unmodified upstream and that any corresponding cookbook PR is merged — something I cannot verify from the diff alone. That combination (CODEOWNER path + external verification requirement) is why I'm deferring rather than approving, per the guideline to not auto-approve CODEOWNER-owned paths.

Other factors

The bug-hunting pass reported no findings, and there's no unresolved third-party objection visible in the timeline. The jump spans several SGLang minor versions (0.5.12 to 0.5.19) on an EAGLE/MTP recipe, which is a reasonable magnitude for a human with domain context to sanity-check even though nothing in the repo's known regression notes (KLAUD_DEBUG.md) points at H200-specific issues in that range.

[adibarra]

/reuse-sweep-run 34450728410