[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

[Klaud-Cold]

Update the qwen3.5-fp8-h100-sglang-mtp master image from lmsysorg/sglang:v0.5.14-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/EP8 topology, EAGLE/MTP settings and the 8k1k concurrency range are unchanged.

Baseline

• Published date: 2026-07-05 (benchmarks?model=Qwen-3.5-397B-A17B&date=2026-07-05&exact=true, workflow-info?date=2026-07-05)
• Old image: lmsysorg/sglang:v0.5.14-cu130 (digest sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3, tag commit sgl-project/sglang@49e384c)
• Workload / topology: single-node H100, Qwen/Qwen3.5-397B-A17B-FP8, SGLang FP8, TP8 EP8, EAGLE MTP (3 steps, top-k 1, 4 draft tokens), fixed-seq-len 8k1k (ISL 8192 / OSL 1024), concurrency 4 / 8 / 16 / 32, random dataset with chat template
• Producer: run 28719988201 (head 2eaf828b84ebe5cac3c68fabd8a18be99fc6d924, changelog PR #2060); benchmark result IDs 432614, 432616, 432593, 432601

Conc	Total tok/s/GPU	Output tok/s/GPU	Median TTFT (s)	Median TPOT (ms)	Median E2E (s)
4	686.1	76.3	0.410	5.93	5.79
8	662.5	74.5	0.623	12.45	12.51
16	1297.7	143.8	0.445	13.25	12.42
32	1618.7	181.0	0.668	21.23	20.20

• Published eval: gsm8k at concurrency 32, em_strict 0.9697 / em_flexible 0.9598 (evaluation ID 8335, same producer run). The public evaluations feed labels this row disagg: true with 64 prefill/decode GPUs, which does not match the single-node TP8 recipe; the benchmark rows above carry the correct disagg: false identity.

---

将 qwen3.5-fp8-h100-sglang-mtp 的主配置镜像从 lmsysorg/sglang:v0.5.14-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/EP8 拓扑、EAGLE/MTP 设置以及 8k1k 并发范围均保持不变。

基线

• 发布日期： 2026-07-05（benchmarks?model=Qwen-3.5-397B-A17B&date=2026-07-05&exact=true，workflow-info?date=2026-07-05）
• 旧镜像： lmsysorg/sglang:v0.5.14-cu130（摘要 sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3，标签提交 sgl-project/sglang@49e384c）
• 工作负载 / 拓扑： 单节点 H100，Qwen/Qwen3.5-397B-A17B-FP8，SGLang FP8，TP8 EP8，EAGLE MTP（3 步、top-k 1、4 个草稿 token），固定序列长度 8k1k（ISL 8192 / OSL 1024），并发 4 / 8 / 16 / 32，随机数据集并使用聊天模板
• 数据来源： 运行 28719988201（head 2eaf828b84ebe5cac3c68fabd8a18be99fc6d924，changelog PR #2060）；基准结果 ID 432614、432616、432593、432601

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	686.1	76.3	0.410	5.93	5.79
8	662.5	74.5	0.623	12.45	12.51
16	1297.7	143.8	0.445	13.25	12.42
32	1618.7	181.0	0.668	21.23	20.20

• 已发布评测： 并发 32 的 gsm8k，em_strict 0.9697 / em_flexible 0.9598（评测 ID 8335，同一数据来源运行）。公共评测接口将该行标记为 disagg: true 且 64 个 prefill/decode GPU，与单节点 TP8 配方不符；上表基准行的 disagg: false 身份是正确的。

---

Note

Low Risk
Config-only SGLang container image version bump for a single benchmark key; no application logic or security-sensitive code changes.

Overview
Bumps the qwen3.5-fp8-h100-sglang-mtp benchmark recipe in nvidia-master.yaml from lmsysorg/sglang:v0.5.14-cu130 to lmsysorg/sglang:v0.5.19-cu130.

Adds a matching perf-changelog.yaml entry noting the digest and that the recipe script, TP8/EP8 topology, and EAGLE/MTP settings are unchanged.

^{Reviewed by Cursor Bugbot for commit 9cda017. 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 66b3f0f4a8a0604a79479903d9e652c1700ed0b1
• Change: master image only (configs/nvidia-master.yaml, key qwen3.5-fp8-h100-sglang-mtp); benchmarks/single_node/fixed_seq_len/qwen3.5_fp8_h100_mtp.sh unchanged, no runtime patching
• Targeted run: e2e run 34422693420 (e2e-tests.yml from main, ref=66b3f0f4, test-config --trim-conc: TP8 EP8 MTP at concurrency 4 plus the smoke gsm8k eval; dispatched 2026-09-10T00:46Z) — completed, mixed: benchmark job passed on h100-cw_01 (attempt 1), eval-only job failed on h100-cw_00
• Benchmark result (concurrency 4, 8k1k, vs published 2026-07-05 baseline):

Metric	Baseline v0.5.14	v0.5.19	Delta
Total tok/s/GPU	686.1	816.1	+18.9%
Output tok/s/GPU	76.3	90.8	+18.9%
Median TTFT (s)	0.410	0.257	-37.3%
Median TPOT (ms)	5.93	5.21	-12.1%
Median E2E (s)	5.79	4.97	-14.2%

• Eval result: N/A — the eval-only server never became ready. First error: on rank TP2, during prefill CUDA-graph capture, Triton's CompiledKernel.__init__ raised FileNotFoundError for /root/.cache/sglang/triton/.../fused_moe_kernel.cubin (the compile-cache group metadata existed but the cubin was not yet present). This is a transient compile-cache race between the 8 TP ranks sharing the per-node SGLANG_CACHE_DIR (kernel caches moved there in v0.5.18, #32434); the identical launch command on h100-cw_01 captured all prefill graphs and served the benchmark, and no flag or model incompatibility was involved.
• Next step: Repair 1/5 — rerun only the failed eval job (no code change).
• Upstream source comparison: v0.5.14 @ 49e384c (2026-06-25) → v0.5.19 @ 0bcd822 (2026-09-03). Provenance: release-docker.yml tags v{version}-cu130 from a git clone --branch v{version} build (docker/Dockerfile); Docker Hub v0.5.19 and v0.5.19-cu130 share one digest, pushed 2026-09-04.
• Coupled dependencies: torch 2.11.0 → 2.13.0, flashinfer 0.6.12 → 0.6.18, sgl-kernel 0.4.4 → 0.4.6.post1, transformers 5.8.1 → 5.12.1, CUDA base 13.0.1 → 13.0.3 (same CUDA 13.0 driver requirement as the old cu130 image).
• Flag audit: all 23 launch flags used by the script are still defined (server_args.py plus the new arg_groups/ hooks) with unchanged choices for --quantization fp8, --attention-backend flashinfer, --kv-cache-dtype fp8_e4m3, --mamba-ssm-dtype bfloat16 and --speculative-algorithm EAGLE. --enable-flashinfer-allreduce-fusion is now deprecated and maps to --flashinfer-allreduce-fusion-backend auto with a warning (serving_hook.py); SGLANG_ENABLE_SPEC_V2 is a removal warning in both tags (spec V2 always on). The bf16 SSM-state requirements in attention_hook.py are SM100-only and do not apply to H100 (SM90), where linear-attention decode stays on Triton.
• Behavior deltas noted: unified radix tree is the default for every configuration in v0.5.19 (#35081; moot under --disable-radix-cache); CUTLASS FP8 blockwise was deleted for SM90 in v0.5.16 (#30438), so H100 FP8 GEMMs use the remaining SM90 paths — the agentic qwen3.5-fp8-h100-sglang-agentic-mtp family already published on v0.5.16-cu130 (2026-08-07) with the same model and MTP settings. qwen3.5-fp8-b200-sglang runs v0.5.19-cu130 on main (PR WIP - SGL B200 FP8 8k1k  #2866, full sweep 34175132645 green) as SM100 evidence.
• 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 66b3f0f4a8a0604a79479903d9e652c1700ed0b1
• 改动： 仅主配置镜像（configs/nvidia-master.yaml 中的 qwen3.5-fp8-h100-sglang-mtp）；benchmarks/single_node/fixed_seq_len/qwen3.5_fp8_h100_mtp.sh 未改动，无任何运行时补丁
• 定向运行： e2e 运行 34422693420（从 main 触发 e2e-tests.yml，ref=66b3f0f4，test-config --trim-conc：TP8 EP8 MTP、并发 4，附带 gsm8k 冒烟评测；2026-09-10T00:46Z 触发）—— 已完成，结果不一：基准任务在 h100-cw_01 通过（第 1 次尝试），仅评测任务在 h100-cw_00 失败
• 基准结果（并发 4，8k1k，对比 2026-07-05 已发布基线）：

指标	基线 v0.5.14	v0.5.19	变化
总吞吐 tok/s/GPU	686.1	816.1	+18.9%
输出吞吐 tok/s/GPU	76.3	90.8	+18.9%
TTFT 中位数 (s)	0.410	0.257	-37.3%
TPOT 中位数 (ms)	5.93	5.21	-12.1%
端到端中位数 (s)	5.79	4.97	-14.2%

• 评测结果： N/A —— 仅评测的服务端未能就绪。首个错误：TP2 rank 在 prefill CUDA graph 捕获期间，Triton 的 CompiledKernel.__init__ 对 /root/.cache/sglang/triton/.../fused_moe_kernel.cubin 抛出 FileNotFoundError（编译缓存的分组元数据已存在，但 cubin 尚未落盘）。这是 8 个 TP rank 共享节点级 SGLANG_CACHE_DIR（v0.5.18 起内核缓存迁移至此，#32434）时的瞬时编译缓存竞争；相同的启动命令在 h100-cw_01 上完整捕获了全部 prefill graph 并完成了基准测试，不涉及任何参数或模型不兼容。
• 下一步： 修复 1/5 —— 仅重跑失败的评测任务（不改代码）。
• 上游源码对比： v0.5.14 @ 49e384c（2026-06-25）→ v0.5.19 @ 0bcd822（2026-09-03）。来源确认：release-docker.yml 以 git clone --branch v{version} 构建并打 v{version}-cu130 标签（docker/Dockerfile）；Docker Hub 上 v0.5.19 与 v0.5.19-cu130 共享同一摘要，推送于 2026-09-04。
• 耦合依赖： torch 2.11.0 → 2.13.0，flashinfer 0.6.12 → 0.6.18，sgl-kernel 0.4.4 → 0.4.6.post1，transformers 5.8.1 → 5.12.1，CUDA 基础镜像 13.0.1 → 13.0.3（与旧 cu130 镜像相同的 CUDA 13.0 驱动要求）。
• 参数审计： 脚本使用的全部 23 个启动参数仍然存在（server_args.py 及新的 arg_groups/ 钩子），--quantization fp8、--attention-backend flashinfer、--kv-cache-dtype fp8_e4m3、--mamba-ssm-dtype bfloat16、--speculative-algorithm EAGLE 的可选值未变。--enable-flashinfer-allreduce-fusion 已弃用，会告警并映射为 --flashinfer-allreduce-fusion-backend auto（serving_hook.py）；SGLANG_ENABLE_SPEC_V2 在两个标签中都只是移除告警（spec V2 始终开启）。attention_hook.py 中的 bf16 SSM 状态要求仅针对 SM100，不适用于 H100（SM90），其线性注意力解码仍使用 Triton。
• 已注意到的行为变化： v0.5.19 对所有配置默认启用统一 radix tree（#35081；在 --disable-radix-cache 下无影响）；v0.5.16 删除了 SM90 的 CUTLASS FP8 blockwise（#30438），H100 的 FP8 GEMM 走剩余的 SM90 路径 —— agentic 家族 qwen3.5-fp8-h100-sglang-agentic-mtp 已在 v0.5.16-cu130 上以相同模型和 MTP 设置发布（2026-08-07）。qwen3.5-fp8-b200-sglang 在 main 上运行 v0.5.19-cu130（PR WIP - SGL B200 FP8 8k1k  #2866，完整扫描 34175132645 全绿），可作为 SM100 佐证。
• 决定： 仅升级镜像；无需在范围内修改参数。

[Klaud-Cold]

Repair 1/5

• Image / SHA: unchanged — lmsysorg/sglang:v0.5.19-cu130 at PR head 66b3f0f4a8a0604a79479903d9e652c1700ed0b1
• Change: none in the repository. Rerun of only the failed eval-only job of the initial run (run 34422693420, attempt 2, requested 2026-09-10T01:14Z); the passing benchmark job is not repeated.
• Reason: the Initial attempt eval server died in a transient Triton compile-cache race (FileNotFoundError for fused_moe_kernel.cubin on TP2) while the identical launch on another node served the benchmark. Nothing in the recipe or image needs changing; the rerun checks reproducibility.
• Result: passed — eval-only job succeeded on h100-cw_00 (server ready, gsm8k on 1319 samples). Scores vs the published 2026-07-05 eval (baseline measured at concurrency 32, this smoke at concurrency 4):

gsm8k metric	Baseline v0.5.14 (conc 32)	v0.5.19 (conc 4)	Threshold
exact_match strict	0.9697	0.9697	≥ 0.94
exact_match flexible	0.9598	0.9606	≥ 0.94

The Triton compile-cache race did not recur; together with the attempt-1 benchmark point this completes the targeted smoke (startup/compatibility evidence, not a full curve).

• Next step: append the perf-changelog.yaml entry, recheck capacity and start the final full sweep (full-sweep-enabled, PR kept draft).

---

修复 1/5

• 镜像 / SHA： 未变 —— lmsysorg/sglang:v0.5.19-cu130，PR head 66b3f0f4a8a0604a79479903d9e652c1700ed0b1
• 改动： 仓库无改动。仅重跑初始运行中失败的评测任务（运行 34422693420，第 2 次尝试，2026-09-10T01:14Z 发起）；已通过的基准任务不重复执行。
• 原因： 初始尝试的评测服务端因瞬时 Triton 编译缓存竞争（TP2 上 fused_moe_kernel.cubin 的 FileNotFoundError）退出，而另一节点上相同的启动命令完成了基准测试。配方与镜像均无需修改；重跑用于确认可复现性。
• 结果： 通过 —— 仅评测任务在 h100-cw_00 成功（服务端就绪，gsm8k 1319 条样本）。与 2026-07-05 已发布评测对比（基线在并发 32 测得，本次冒烟在并发 4）：

gsm8k 指标	基线 v0.5.14（并发 32）	v0.5.19（并发 4）	阈值
exact_match strict	0.9697	0.9697	≥ 0.94
exact_match flexible	0.9598	0.9606	≥ 0.94

Triton 编译缓存竞争未再出现；结合第 1 次尝试的基准点，定向冒烟测试完成（属于启动/兼容性证据，不是完整曲线）。

• 下一步： 追加 perf-changelog.yaml 条目，重新检查容量，并启动最终完整扫描（full-sweep-enabled，PR 保持草稿）。

[Klaud-Cold]

Final full sweep

• Image / SHA: lmsysorg/sglang:v0.5.19-cu130 at PR head 5cb93a2d98a8194fdb07e63c9dce17f5bf665131 (adds the perf-changelog.yaml entry on top of the image bump; validate_perf_changelog.py against origin/main passed locally)
• Run: run-sweep 34429032159 (full-sweep-enabled applied 2026-09-10T02:20Z with the PR kept draft; the concurrent push-triggered run 34429017350 was superseded and cancelled by the workflow's concurrency group)
• Scope: the complete qwen3.5-fp8-h100-sglang-mtp matrix — 8k1k TP8 EP8 MTP at concurrency 4 / 8 / 16 / 32 plus the default gsm8k eval at concurrency 32 (canary selection may run the lowest point first)
• Result: passed — run-sweep completed success on the exact PR head (attempt 1): 4/4 benchmark points, the gsm8k eval, collect/compare/changelog-metadata jobs and the klaud-sweep-manifest (full-sweep: true) all green; every point reports power_valid: 1. Points ran on h100-cw_00 (c4) and h100-dgxc-slurm (c8/c16/c32); the eval on h100-cw_01.
• Throughput / latency vs the published 2026-07-05 baseline (same TP8 EP8 MTP 8k1k recipe):

Conc	Total tok/s/GPU (base → new)	Output tok/s/GPU (base → new)	Median TTFT s (base → new)	Median TPOT ms (base → new)	Median E2E s (base → new)
4	686.1 → 815.8 (+18.9%)	76.3 → 90.7 (+18.9%)	0.410 → 0.258 (-37.1%)	5.93 → 5.19 (-12.5%)	5.79 → 4.97 (-14.1%)
8	662.5 → 1177.7 (+77.8%)	74.5 → 132.4 (+77.8%)	0.623 → 0.288 (-53.7%)	12.45 → 7.11 (-42.9%)	12.51 → 6.89 (-44.9%)
16	1297.7 → 1630.6 (+25.6%)	143.8 → 180.6 (+25.6%)	0.445 → 0.302 (-32.1%)	13.25 → 10.38 (-21.7%)	12.42 → 9.77 (-21.4%)
32	1618.7 → 2079.1 (+28.4%)	181.0 → 232.4 (+28.4%)	0.668 → 0.427 (-36.2%)	21.23 → 16.65 (-21.6%)	20.20 → 15.65 (-22.5%)

• Eval (gsm8k, concurrency 32, 1319 samples):

Metric	Baseline v0.5.14	v0.5.19	Delta	Threshold
exact_match strict	0.9697	0.9651	-0.0046 (within ±0.0047 SE)	≥ 0.94
exact_match flexible	0.9598	0.9591	-0.0007	≥ 0.94

• Notes: the conc 8 baseline point (662.5 tok/s/GPU, below the conc 4 value) looks like an outlier in the published curve, so its +77.8% delta overstates the typical gain; the other three points sit at +19% to +28%. No regressions observed; the small gsm8k strict decrease is inside one standard error.
• Next step: finish verifies matrix/result coverage and marks the PR ready for the automatic reviews. Repairs used: 1/5 (an eval rerun with no code change).

---

最终完整扫描

• 镜像 / SHA： lmsysorg/sglang:v0.5.19-cu130，PR head 5cb93a2d98a8194fdb07e63c9dce17f5bf665131（在镜像升级之上追加 perf-changelog.yaml 条目；本地对 origin/main 运行 validate_perf_changelog.py 通过）
• 运行： run-sweep 34429032159（2026-09-10T02:20Z 添加 full-sweep-enabled，PR 保持草稿；由推送触发的并发运行 34429017350 已被工作流并发组取代并取消）
• 范围： 完整的 qwen3.5-fp8-h100-sglang-mtp 矩阵 —— 8k1k TP8 EP8 MTP、并发 4 / 8 / 16 / 32，加上并发 32 的默认 gsm8k 评测（canary 选择可能先运行最低并发点）
• 结果： 通过 —— run-sweep 在 PR 精确 head 上以 success 完成（第 1 次尝试）：4/4 个基准点、gsm8k 评测、collect/compare/changelog-metadata 任务以及 klaud-sweep-manifest（full-sweep: true）全部通过；每个点均报告 power_valid: 1。基准点运行于 h100-cw_00（c4）和 h100-dgxc-slurm（c8/c16/c32）；评测运行于 h100-cw_01。
• 吞吐 / 延迟对比 2026-07-05 已发布基线（同一 TP8 EP8 MTP 8k1k 配方）：

并发	总吞吐 tok/s/GPU（基线 → 新）	输出吞吐 tok/s/GPU（基线 → 新）	TTFT 中位数 s（基线 → 新）	TPOT 中位数 ms（基线 → 新）	端到端中位数 s（基线 → 新）
4	686.1 → 815.8 (+18.9%)	76.3 → 90.7 (+18.9%)	0.410 → 0.258 (-37.1%)	5.93 → 5.19 (-12.5%)	5.79 → 4.97 (-14.1%)
8	662.5 → 1177.7 (+77.8%)	74.5 → 132.4 (+77.8%)	0.623 → 0.288 (-53.7%)	12.45 → 7.11 (-42.9%)	12.51 → 6.89 (-44.9%)
16	1297.7 → 1630.6 (+25.6%)	143.8 → 180.6 (+25.6%)	0.445 → 0.302 (-32.1%)	13.25 → 10.38 (-21.7%)	12.42 → 9.77 (-21.4%)
32	1618.7 → 2079.1 (+28.4%)	181.0 → 232.4 (+28.4%)	0.668 → 0.427 (-36.2%)	21.23 → 16.65 (-21.6%)	20.20 → 15.65 (-22.5%)

• 评测（gsm8k，并发 32，1319 条样本）：

指标	基线 v0.5.14	v0.5.19	变化	阈值
exact_match strict	0.9697	0.9651	-0.0046（在 ±0.0047 标准误内）	≥ 0.94
exact_match flexible	0.9598	0.9591	-0.0007	≥ 0.94

• 说明： 基线中并发 8 的点（662.5 tok/s/GPU，低于并发 4）疑似已发布曲线中的异常值，因此其 +77.8% 的变化高估了典型收益；其余三个点的提升在 +19% 至 +28% 之间。未观察到回归；gsm8k strict 的微小下降在一个标准误之内。
• 下一步： finish 校验矩阵/结果覆盖并将 PR 标记为可评审，以启动自动评审。已使用修复次数：1/5（一次不改代码的评测重跑）。

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34429032159
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=34429032159

[Klaud-Cold]

Klaud Cold: validated. All owned runs are terminal. Repairs: 1. Runs: 34422693420, 34429017350, 34429032159.

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：1。运行：34422693420, 34429017350, 34429032159。

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 13s —— View job

---

Review: PR #2952

• Gather context (diff vs origin/main, commits, changelog tail)
• Verify master-config change (image public, key/prefix, rest of recipe untouched)
• Verify perf-changelog.yaml entry (appended at end, format, PR link, no append-only misuse)
• Run validate_perf_changelog.py against origin/main — could not execute locally (Python execution is not in this job's allowed tools); relied on CI instead, where the changelog-metadata job in run-sweep 34429032159 passed on the exact PR head 5cb93a2d9
• Post review summary

LGTM - no blocking issues found

The diff is exactly two hunks: the image: line for qwen3.5-fp8-h100-sglang-mtp in configs/nvidia-master.yaml (public Docker Hub tag lmsysorg/sglang:v0.5.19-cu130; model, model-prefix, runner, TP8/EP8 search space and MTP scenario untouched) and a single perf-changelog.yaml entry appended at the very end of the file with the matching config key and PR link. No append-only flag or eval modifiers are involved, and the recipe script is not modified, so the full sweep on this head (4/4 points plus gsm8k, all green) covers the change.
· Branch

[claude]

Looks good, straightforward config change.

What was reviewed: the configs/nvidia-master.yaml hunk (confirmed only the image tag for qwen3.5-fp8-h100-sglang-mtp changed, all other recipe fields untouched) and the perf-changelog.yaml addition (confirmed it is a single append at the file tail with no reordering or byte changes to prior entries). Also checked for a model.container field that would need to mirror the image bump — this recipe schema has no container field anywhere in the file, so that concern doesn't apply.

Extended reasoning...

Overview

The PR changes exactly two files: configs/nvidia-master.yaml, where the qwen3.5-fp8-h100-sglang-mtp recipe's image field is bumped from lmsysorg/sglang:v0.5.14-cu130 to lmsysorg/sglang:v0.5.19-cu130, and perf-changelog.yaml, which gets one new entry appended at the end of the file documenting the bump, referencing the target image digest, and noting the recipe script/topology/EAGLE-MTP settings are unchanged. I ran the diff directly and confirmed both hunks match this description precisely — no other lines in the recipe block changed, and the changelog addition is a clean append with no modification to existing bytes.

Security risks

None. This is a data-only configuration change (a container image tag bump and a changelog entry) with no code paths, credentials, or access-control logic involved.

Level of scrutiny

Low. This is exactly the kind of small, mechanical, self-contained change the approval guidelines call out as approvable — a version bump with no ambiguity and a matching, correctly-formed changelog entry. The repo's own review conventions (append-only changelog, image-bump-must-be-a-real-image-change) are both satisfied per my direct read of the diff.

Other factors

The repository convention note raised the question of whether a model.container field also needs updating alongside image. I checked and this recipe/config schema has no container field anywhere in the file, so that check does not apply here. No CODEOWNERS restriction is evident for these two files, and the bug-hunting pass reported zero findings.

[adibarra]

/reuse-sweep-run 34429032159