[Klaud Cold] Update dsr1-fp8-mi325x-sglang-mtp SGLang ROCm image to v0.5.19-rocm700-mi30x / 将 dsr1-fp8-mi325x-sglang-mtp 的 SGLang ROCm 镜像更新至 v0.5.19-rocm700-mi30x

[Klaud-Cold]

Summary / 摘要

Klaud Cold image refresh for the master family dsr1-fp8-mi325x-sglang-mtp (configs/amd-master.yaml): update image from lmsysorg/sglang:v0.5.12-rocm700-mi30x to lmsysorg/sglang:v0.5.19-rocm700-mi30x, plus the required perf-changelog.yaml entry. Model, precision, topology (TP8/EP1), MTP speculative decoding, 8k/1k workload, benchmark script, and eval selection are unchanged. Only this family is touched; the non-MTP sibling dsr1-fp8-mi325x-sglang and all other families stay on their current images.

Klaud Cold 为主配置家族 dsr1-fp8-mi325x-sglang-mtp（configs/amd-master.yaml）刷新镜像：将 image 从 lmsysorg/sglang:v0.5.12-rocm700-mi30x 更新为 lmsysorg/sglang:v0.5.19-rocm700-mi30x，并追加必需的 perf-changelog.yaml 条目。模型、精度、拓扑（TP8/EP1）、MTP 投机解码、8k/1k 负载、基准脚本与评测选择均保持不变。仅改动该家族；非 MTP 兄弟家族 dsr1-fp8-mi325x-sglang 及其他家族保持原镜像。

Candidate evidence / 候选证据

• Candidate id: aac067bd5766d07d-d7346b01615deec4; base SHA e52162e5c4805ccf6933199034dbf2f7ba9f5b55; branch klaude/auto-aac067bd5766d07d-d7346b01615deec4.
• Public observation (/api/v1/latest-images): dsr1 / mi325x / sglang / fp8 / mtp / single-node / 8192→1024 on lmsysorg/sglang:v0.5.12-rocm700-mi30x, published 2026-05-18. Release feed (/api/v1/framework-releases): sglang v0.5.19.
• Upstream verification: lmsysorg/sglang:v0.5.19-rocm700-mi30x exists on Docker Hub (linux/amd64, pushed 2026-09-04T21:28Z, digest sha256:590a815c128d7d5c83ef771ad768c9f8be82f64f7d0dbd98dc7b9f085b268a58). It keeps the same ROCm 7.0 / MI30x (gfx942) base line as the current pin. Old image digest: sha256:4548e936fd76e707184c449ed4720739220ca53734944e1d330c8c4c1c5215ed. The image runs as shipped; no engine patches, wheel overlays, or script changes.
• Routing: runner: mi325x expands in configs/runners.yaml only to the mi325x-amds_* fleet; the single telemetry cluster mi325x passed check-capacity before edits, before branch/PR creation and before dispatch. Generated matrix (test-config --config-files configs/amd-master.yaml --config-keys dsr1-fp8-mi325x-sglang-mtp): 5 single-node points (conc 4/8/16/32/64, TP8 EP1, MTP), evals at conc 32 and 64; identical to base except image.
• Open-PR recheck before claiming the branch: no open PR modifies this key or benchmarks/single_node/fixed_seq_len/dsr1_fp8_mi325x{,_mtp}.sh.

Published baseline (frozen) / 已发布基线（冻结）

Source: GET /api/v1/benchmarks?model=DeepSeek-R1-0528&date=2026-05-18&exact=true filtered to hardware=mi325x, framework=sglang, model=dsr1, precision=fp8, spec_method=mtp, disagg=false, isl=8192, osl=1024, image=lmsysorg/sglang:v0.5.12-rocm700-mi30x; GET /api/v1/workflow-info?date=2026-05-18; GET /api/v1/evaluations. Producer run for every point: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/26048239142/attempts/2 (PR #1500; curve snapshot id 1899 is not a producer id). Published date: 2026-05-18.

conc	benchmark id	tput/GPU (tok/s)	output tput/GPU (tok/s)	median TTFT (s)	median TPOT (s)	median E2EL (s)
4	412362	224.80	25.01	0.7110	0.01853	17.789
8	412360	381.95	42.94	0.4547	0.02130	21.038
16	412357	621.20	68.83	0.4439	0.02655	25.354
32	412361	842.61	94.23	0.4705	0.03997	37.805
64	412356	1027.76	114.02	0.4779	0.06736	62.961

Baseline evals (gsm8k, 2026-05-18, same producer run): conc 32 em_strict 0.9575 / em_flexible 0.9583; conc 64 em_strict 0.9575 / em_flexible 0.9568.

Baseline comparability caveat / 基线可比性说明: the retained baseline server log (GET /api/v1/server-log-search?id=412356&q=EAGLE, also q=speculative_num_steps) shows speculative_algorithm=None, speculative_num_steps=None, mem_fraction_static=0.68 and max_running_requests=None→4096, i.e. the published "mtp" curve was served without speculative decoding and with the non-MTP launch flags (its metrics also match the non-MTP dsr1-fp8-mi325x-sglang curve of the same day within noise). The updated-image run below really ran EAGLE/MTP (accept length ≈2.0 on random benchmark traffic, ≈2.3–2.7 on gsm8k). The per-point deltas are therefore reported as informational only and are not claimed as image improvements; they mix the engine update with the MTP-on vs MTP-off difference. / 基线保留的服务端日志显示 speculative_algorithm=None（未启用投机解码，且为非 MTP 启动参数），而本次新镜像运行确实启用了 EAGLE/MTP。因此下列逐点变化仅供参考，不作为镜像本身的性能提升声明。

Attempts / 尝试记录

Attempt / 尝试	Image / SHA	Run URLs	Benchmark / eval result / 基准与评测结果	Per-point deltas vs baseline / 逐点相对基线变化	Diagnosis / 诊断
Baseline (published 2026-05-18) / 基线（发布于 2026-05-18）	lmsysorg/sglang:v0.5.12-rocm700-mi30x	https://github.com/SemiAnalysisAI/InferenceX/actions/runs/26048239142/attempts/2	5/5 points published; gsm8k conc32 em_strict 0.9575, conc64 em_strict 0.9575	reference / 参考	Published dashboard data, not rerun. Retained log shows speculative decoding was off (see caveat) / 公开面板数据，未重跑；日志显示未启用投机解码
Update 1 / 更新 1 (targeted e2e, klaud-run=true, fail-fast=true)	lmsysorg/sglang:v0.5.19-rocm700-mi30x @ 97045041f5c0dbe62b2d1baef9b56e05b84a5958	https://github.com/SemiAnalysisAI/InferenceX/actions/runs/33991903593	success: 5/5 benchmark points with non-empty results_bmk/agg_bmk.json (all power_valid=1); gsm8k (1319 samples) conc32 em_strict 0.9553 / flexible 0.9553 (baseline 0.9575 / 0.9583, Δ −0.2 pt); conc64 em_strict 0.9507 / flexible 0.9515 (baseline 0.9575 / 0.9568, Δ −0.7 / −0.5 pt, ≈1.2 SE)	See per-point table below (informational) / 见下表（仅供参考）	Image works as shipped. Engine notes: SGLANG_ENABLE_SPEC_V2 is now a removed no-op (spec always runs the V2 worker); the engine now resets max_running_requests to 48 under speculative decoding, so the conc-64 point queues 16 requests server-side and its median TTFT rises to 13.85 s (regression, reported not fixed; recipe flags are out of Klaud scope) / 镜像可直接运行；新引擎在投机解码下将 max_running_requests 重置为 48，conc-64 点排队导致 TTFT 中位数升至 13.85 s（回归，仅报告）
Final full sweep (full-sweep-enabled) / 最终全量 sweep	lmsysorg/sglang:v0.5.19-rocm700-mi30x @ 1fceee6e753e4cf7959fee2e52bd46fcc276feec (image + changelog)	https://github.com/SemiAnalysisAI/InferenceX/actions/runs/33993631221 (labeled event) and https://github.com/SemiAnalysisAI/InferenceX/actions/runs/33993560455 (synchronize event)	N/A — not executed: both run-sweep.yml runs concluded skipped with every job skipped. run-sweep.yml gates check-changelog, reuse-sweep-gate and setup on !github.event.pull_request.draft, and this PR must stay a draft per Klaud Cold policy, so no GPU sweep ran on the PR head / 不适用 — 未执行：两次 run-sweep 运行均为 skipped（所有作业跳过），因为该工作流要求 PR 非草稿，而 Klaud Cold 须保持草稿状态	N/A (see result)	Blocked by workflow draft gating, not by the image. The ready_for_review event is a run-sweep.yml trigger, so a maintainer converting this draft to ready will start the labeled full sweep on the PR head; nothing else is needed from the recipe side. The perf-changelog entry was validated locally with utils/validate_perf_changelog.py --base-ref e52162e5c --head-ref 1fceee6e7 (exit 0) / 受工作流草稿门控阻止，与镜像无关；维护者将 PR 转为 ready 后即会触发已加标签的全量 sweep

Update 1 per-point deltas vs frozen baseline (informational) / 更新 1 逐点变化（仅供参考）

Δ = (new − baseline) / baseline. Lower is better for TTFT/TPOT/E2EL. Matched per point on tp8 / ep1 / dp-attn off / isl 8192 / osl 1024 / concurrency.

conc	new tput/GPU	Δ tput/GPU	new output tput/GPU	Δ output tput/GPU	new median TTFT s	Δ TTFT	new median TPOT s	Δ TPOT	new median E2EL s	Δ E2EL	power_valid
4	392.20	+74.5%	43.62	+74.4%	0.7377	+3.8%	0.01016	-45.2%	10.002	-43.8%	1
8	594.03	+55.5%	66.77	+55.5%	0.4908	+7.9%	0.01404	-34.1%	13.377	-36.4%	1
16	837.15	+34.8%	92.75	+34.7%	0.4600	+3.6%	0.02103	-20.8%	19.971	-21.2%	1
32	1023.81	+21.5%	114.48	+21.5%	0.5449	+15.8%	0.03431	-14.1%	32.076	-15.2%	1
64	1228.57	+19.5%	136.28	+19.5%	13.8537	+2798.8%	0.04418	-34.4%	53.458	-15.1%	1

Regressions to note / 需注意的回归: conc-64 median TTFT (0.48 s → 13.85 s, p99 37.9 s) from the engine's new default max_running_requests=48 under speculative decoding; small TTFT increases at conc 8 and 32; gsm8k conc-64 exact-match −0.7 pt. No rejection threshold is applied.

Limitations / 限制

• A green targeted benchmark proves the updated image serves this family's points and default evals; it does not prove global PR checks, CODEOWNER review, or other families pass. / 目标基准通过仅证明新镜像可运行该家族的点与默认评测，不代表全局检查、CODEOWNER 审核或其他家族通过。
• Baseline numbers are published dashboard data from 2026-05-18 served without speculative decoding (see caveat); deltas are indicative and no improvement is claimed. / 基线为 2026-05-18 未启用投机解码的公开数据；变化仅供参考，不声明性能提升。
• Raw artifacts remain on the e2e run; nothing was staged, reused, reviewed or merged by Klaud Cold. / 原始产物保留在 e2e 运行上；Klaud Cold 未执行暂存、复用、评审或合并。

[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