[DSv4] Refresh stale H200 FP8 submissions / 刷新过期的 H200 FP8 数据 - #2191
Conversation
|
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 |
1 similar comment
|
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 |
- dsv4-fp8-h200-vllm(+mtp): bump image v0.21.0 -> v0.25.0 (match #2169) - dsv4-fp8-h200-sglang(+mtp): bump deepseek-v4-hopper digest to 2026-05-13 push 中文：刷新过期的 H200 FP8 DeepSeek-V4-Pro 数据：vLLM 镜像升级至 v0.25.0， SGLang hopper 镜像更新至 2026-05-13 摘要。
32fcd76 to
2d08ffb
Compare
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
LGTM — clean image refresh for H200 DSv4 FP8, no search-space changes.
Extended reasoning...
Overview
This PR touches only configs/nvidia-master.yaml and perf-changelog.yaml. It bumps the vLLM image for dsv4-fp8-h200-vllm/-mtp from v0.21.0 to v0.25.0 (matching the prior B200/B300 dsv4 vLLM bump in #2169) and bumps the pinned SGLang digest for dsv4-fp8-h200-sglang/-mtp from a 2026-05-02 push to a 2026-05-13 push. The two search-space blocks and all other fields (runner, precision, framework, scenarios) are unchanged, and the perf-changelog entries correctly reference the four affected config keys and this PR number.
Security risks
None. This is a data-plane config change (image tags/digests, changelog metadata) with no code paths touching auth, secrets, or user input.
Level of scrutiny
Low. This is a mechanical, single-purpose image refresh following an already-established pattern from a merged sibling PR (#2169), split per-SKU per the stated process. The bug hunting system found no issues, and I independently confirmed the one candidate finding it ruled out (a claimed gpu-memory-utilization: 0.95 OOM risk) doesn't apply — that field isn't set anywhere in the affected dsv4-fp8-h200-* blocks.
Other factors
The PR author states validation was run (generate_sweep_configs.py, validate_perf_changelog.py), and the actual CI sweep (gated behind the full-sweep-fail-fast label per the repo's automated reminder) is the real correctness check for whether v0.25.0/the new SGLang digest work on H200 — that's an operational verification step outside the scope of a static code review.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29314418376 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29316772188 |
|
/reuse-sweep-run 29316772188 |
origin/main advanced (H200 refresh #2191 merged, adding dsv4-fp8-h200-* entries) so the branch's changelog was stale and check-changelog flagged the main-only entries as deletions. Reset perf-changelog.yaml to origin/main and re-append only the dsv4-fp8-h100-dynamo-vllm entry. 中文：将 H100 分支的 perf-changelog 重置到 origin/main 并仅重新追加 H100 条目， 修复 origin/main 推进（#2191 合并）导致的 append-only 校验删除报错。
origin/main advanced (H200 refresh #2191 merged); reset perf-changelog.yaml to origin/main and re-append the dsv4-fp8-mi300x-vllm and dsv4-fp8-mi300x-vllm-mtp entries so check-changelog's append-only rule passes. 中文：将 MI300X 分支 perf-changelog 重置到 origin/main 并重新追加其条目。
origin/main advanced (#2191 merged); reset perf-changelog.yaml to origin/main and re-append the dsv4-fp8-mi325x-vllm and dsv4-fp8-mi325x-vllm-mtp entries. 中文：将 MI325X 分支 perf-changelog 重置到 origin/main 并重新追加其条目。
Picks up PR #2191's perf-changelog.yaml entries (dsv4-fp8-h200-vllm, dsv4-fp8-h200-sglang H200 refresh) that were merged to main after this branch diverged. Without this merge, check-changelog's base_ref=main vs head_ref=<this branch> diff sees those entries as "deleted" (present on main, absent here) and fails with "Deletions are not allowed in perf-changelog.yaml", even though this branch never touched them. Co-authored-by: Cursor <cursoragent@cursor.com>
1 participant
Summary
Split out of the closed #2189 — H200 only, one PR per SKU.
DSv4 (DeepSeek-V4-Pro) FP8 on H200 was stale (last run 2026-05-21). Pure image refresh, no search-space change:
dsv4-fp8-h200-vllm(+mtp): bump vLLM imagev0.21.0→v0.25.0, matching the B200/B300 dsv4 vLLM bump ([chore] Bump dsv4-fp4 vLLM image to v0.25.0 on B200 and B300 #2169). No script change — the H200 recipe is the FP8/Hopper path and doesn't use the fp4/mega_moe flags that needed--prefill-schedule-intervalon B200/B300.dsv4-fp8-h200-sglang(+mtp): bump the pinnedlmsysorg/sglang:deepseek-v4-hopperdigest from the 2026-05-02 push (7f19c6dc) to the current 2026-05-13 push (1bf5d508).Validation:
generate_sweep_configs.py --model-prefix dsv4 --runner-type h200resolves to the v0.25.0 image;validate_perf_changelog.py --base-ref mainpasses.中文说明
从已关闭的 #2189 拆分而来——仅 H200，按 SKU 拆分为独立 PR。
H200 上的 DSv4（DeepSeek-V4-Pro）FP8 数据已过期（最后一次运行 2026-05-21）。仅刷新镜像，不改动扫描空间：
dsv4-fp8-h200-vllm(+mtp)：vLLM 镜像从v0.21.0升级到v0.25.0，与 B200/B300 的 dsv4 vLLM 升级（[chore] Bump dsv4-fp4 vLLM image to v0.25.0 on B200 and B300 #2169）一致。无需修改脚本——H200 走 FP8/Hopper 路径，不使用 B200/B300 上需要--prefill-schedule-interval的 fp4/mega_moe 参数。dsv4-fp8-h200-sglang(+mtp)：将固定的lmsysorg/sglang:deepseek-v4-hopper镜像摘要从 2026-05-02 推送（7f19c6dc）更新为当前 2026-05-13 推送（1bf5d508）。验证：
generate_sweep_configs.py --model-prefix dsv4 --runner-type h200解析到 v0.25.0 镜像；validate_perf_changelog.py --base-ref main通过。