[Klaud Cold] Update dsr1-fp8-mi325x-sglang SGLang ROCm image to v0.5.19-rocm700-mi30x / 将 dsr1-fp8-mi325x-sglang 的 SGLang ROCm 镜像更新至 v0.5.19-rocm700-mi30x - #2949
Conversation
…9-rocm700-mi30x Bump lmsysorg/sglang from v0.5.12-rocm700-mi30x to v0.5.19-rocm700-mi30x (digest sha256:590a815c128d7d5c83ef771ad768c9f8be82f64f7d0dbd98dc7b9f085b268a58) for the DeepSeek-R1-0528 FP8 MI325X SGLang 8k1k recipe. Same ROCm 7.0 base image; model, TP8 topology, concurrencies and launch flags unchanged. 将 dsr1-fp8-mi325x-sglang 的 SGLang ROCm 镜像从 v0.5.12-rocm700-mi30x 更新至 v0.5.19-rocm700-mi30x（同一 ROCm 7.0 基础镜像）。模型、TP8 拓扑、并发点与 启动参数均保持不变。 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
Initial attempt
The smoke eval ran on the trimmed concurrency-4 point while the published evals are at concurrency 32 and 64, so the eval comparison is indicative only. Server log confirms the AITER rule:
初次尝试
冒烟评测在裁剪后的并发 4 点上运行，而已发布评测位于并发 32 和 64，因此评测对比仅供参考。 服务日志证实 AITER 规则生效：
|
…p entry Append the perf-changelog entry for the dsr1-fp8-mi325x-sglang image update (lmsysorg/sglang v0.5.12-rocm700-mi30x -> v0.5.19-rocm700-mi30x), linking PR #2949. Historical entries are unchanged. 为 dsr1-fp8-mi325x-sglang 的镜像更新（lmsysorg/sglang v0.5.12-rocm700-mi30x -> v0.5.19-rocm700-mi30x）追加 perf-changelog 条目，关联 PR #2949。历史条目保持不变。 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Final full sweep
Throughput is at parity to slightly better at every point; the only regression is median TTFT at concurrency 4 (+7.5%, 0.053 s), reported without a rejection threshold.
Both evals: n_eff 1319,
最终全量 sweep
各点吞吐持平或略有提升；唯一回退是并发 4 的 TTFT 中位数（+7.5%，0.053 秒），如实报告，不设拒绝阈值。
两项评测：n_eff 1319，
|
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=34429203486 |
|
Klaud Cold: validated. All owned runs are terminal. Repairs: 0. Runs: 34422505103, 34429182346, 34429203486. 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：34422505103, 34429182346, 34429203486。 完整 sweep 已通过验证；PR 保持就绪，等待审查。 |
|
Claude finished @Klaud-Cold's task in 1m 10s —— View job Review of PR #2949
LGTM - no blocking issues found
|
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Looks good, straightforward config change. Checked the image tag bump in configs/amd-master.yaml for dsr1-fp8-mi325x-sglang and the matching perf-changelog.yaml entry — the description/pr-link format matches prior image-bump entries in the changelog, and the change is appended at the tail with no edits to existing bytes. This recipe is single-node (multinode: false) with no model.container field, so the srt-slurm "model.container must equal image" coupling rule in AGENTS.md doesn't apply here.
Extended reasoning...
Overview
The diff touches exactly two files: configs/amd-master.yaml (a one-line image tag bump for the dsr1-fp8-mi325x-sglang recipe, from sglang v0.5.12 to v0.5.19) and perf-changelog.yaml (a new appended entry documenting that bump). No code, launcher scripts, or benchmark logic is modified.
Security risks
None. No secrets, credentials, or executable logic are introduced. The changed lines are a container image tag string and a YAML changelog record.
Level of scrutiny
This matches the repo's own definition of a mechanical, low-risk change: a version bump with no ambiguity. I verified the perf-changelog.yaml entry is appended strictly at the tail (git diff shows only additions after the last existing entry, no prior bytes touched), satisfying the append-only/byte-sensitive requirement in AGENTS.md. I also checked the "Multi-node srt-slurm changes... For image bumps, model.container must equal image" rule; this recipe has multinode: false and no model.container field in its schema, so that coupling rule is inapplicable. The changelog description/pr-link format is consistent with numerous prior image-bump entries in the file (e.g. PRs #595, #816, #1031, #1041, #1321).
Other factors
No CODEOWNERS restriction is apparent for configs/ or perf-changelog.yaml beyond normal review. No test suite changes are needed for a pure data/config bump. The PR conversation timeline shows no outstanding CHANGES_REQUESTED review or unresolved objection from another reviewer. The bug hunter reported no findings and the exit reason was dry_streak (a valid completion state). Overall this is simple enough that a human does not need to re-verify it.
|
/reuse-sweep-run 34429203486 |
将 main 合并到 PR #2949，保留已验证的配方并复用完整扫描结果。
2 participants
Update the
dsr1-fp8-mi325x-sglangrecipe image fromlmsysorg/sglang:v0.5.12-rocm700-mi30xtolmsysorg/sglang:v0.5.19-rocm700-mi30x(digestsha256:590a815c128d7d5c83ef771ad768c9f8be82f64f7d0dbd98dc7b9f085b268a58). Model, TP8 topology, concurrency points, launch flags and evals are unchanged.Baseline
lmsysorg/sglang:v0.5.12-rocm700-mi30x(SGLang tag commit127b9e3, AITERa6bb499, baserocm/sgl-dev:rocm7-vllm-20250904)GET /api/v1/workflow-info?date=2026-05-18;GET /api/v1/benchmarks?model=DeepSeek-R1-0528&date=2026-05-18&exact=true&sequence=8k/1kfiltered to mi325x / sglang / fp8 / spec none / disagg false;GET /api/v1/evaluationsfiltered to the producer run0bcd822, AITERc16d44b, same ROCm 7.0 base image), docker/rocm.Dockerfile @ v0.5.19Published evals (gsm8k, n_eff 1319): conc 32 em_strict 0.9530 / em_flexible 0.9553; conc 64 em_strict 0.9545 / em_flexible 0.9553. The published eval rows carry
disagg: truealthough the recipe is aggregated; treated as a metadata quirk of the published data.将
dsr1-fp8-mi325x-sglang配方的镜像从lmsysorg/sglang:v0.5.12-rocm700-mi30x更新至lmsysorg/sglang:v0.5.19-rocm700-mi30x（摘要sha256:590a815c128d7d5c83ef771ad768c9f8be82f64f7d0dbd98dc7b9f085b268a58）。模型、TP8 拓扑、并发点、启动参数与评测均保持不变。基线
lmsysorg/sglang:v0.5.12-rocm700-mi30x（SGLang 标签提交127b9e3，AITERa6bb499，基础镜像rocm/sgl-dev:rocm7-vllm-20250904）GET /api/v1/workflow-info?date=2026-05-18；GET /api/v1/benchmarks?model=DeepSeek-R1-0528&date=2026-05-18&exact=true&sequence=8k/1k，筛选 mi325x / sglang / fp8 / 无投机解码 / 非分离式；GET /api/v1/evaluations筛选至该产出运行0bcd822，AITERc16d44b，ROCm 7.0 基础镜像相同），docker/rocm.Dockerfile @ v0.5.19已发布评测（gsm8k，n_eff 1319）：并发 32 em_strict 0.9530 / em_flexible 0.9553；并发 64 em_strict 0.9545 / em_flexible 0.9553。已发布评测行标记为
disagg: true，而该配方为聚合部署，视为发布数据的元数据差异。🤖 Generated with Claude Code
Note
Low Risk
Config-only container image pin for a single benchmark recipe; no runtime code or topology changes.
Overview
Bumps the
dsr1-fp8-mi325x-sglangAMD master recipe to SGLang ROCmv0.5.19-rocm700-mi30x(fromv0.5.12-rocm700-mi30x). Model, TP8 fixed 8k/1k sweep, concurrency range, and other recipe fields are unchanged.Adds a
perf-changelog.yamlentry for config keydsr1-fp8-mi325x-sglangthat records the image update and links PR #2949.Reviewed by Cursor Bugbot for commit 31c2396. Bugbot is set up for automated code reviews on this repo. Configure here.