Skip to content

[Klaud Cold] Update DeepSeek-V4-Pro FP8 H200 Dynamo-SGLang AgentX image to nightly-dev-20261025-65633129 / [Klaud Cold] 将 DeepSeek-V4-Pro FP8 H200 Dynamo-SGLang AgentX 镜像更新至 nightly-dev-20261025-65633129 - #4478

Closed
Klaud-Cold wants to merge 1 commit into
mainfrom
klaud/auto-0cb636efc3513aaa-353812587e13d7bc

Conversation

@Klaud-Cold

@Klaud-Cold Klaud-Cold commented Oct 25, 2026 •

Copy link
Copy Markdown
Collaborator

Goal: Update SGLang image from lmsysorg/sglang:nightly-dev-20260914-4358a161 to lmsysorg/sglang:nightly-dev-20261025-65633129.
Baseline: 2026-09-16 · lmsysorg/sglang:nightly-dev-20260914-4358a161
AgentX · PTP8x1/DTP8x0 · Mean latency · Sources: API 1, API 2, API 3, API 4

Concurrency Total tok/s/GPU ↑ Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 2,006.78 15.07 1,352.42 4.82
2 2,115.35 18.4 1,091.41 5.12
4 3,137.32 23.69 1,406.32 5.86
8 5,256.86 41.28 2,265.84 7.44
16 6,033.93 47.91 20,053.53 13.44
Eval Score ↑ Samples
gsm8k/em_strict · c16 97.42% 1,319

Stop: Add the klaud-abort label at any time to stop this candidate now; Klaud cancels its runs, closes this PR and does not retry it. Remove the label to let the planner pick it up again.

中文

**目标:**将 SGLang 镜像从 lmsysorg/sglang:nightly-dev-20260914-4358a161 更新至 lmsysorg/sglang:nightly-dev-20261025-65633129。
**基线:**2026-09-16 · lmsysorg/sglang:nightly-dev-20260914-4358a161
AgentX · PTP8x1/DTP8x0 · 平均延迟 · 来源: API 1, API 2, API 3, API 4;数值及异常说明见上表。
**停止:**随时添加 klaud-abort 标签即可立即停止该候选:Klaud 会取消其运行、关闭此 PR 且不再重试。移除该标签后,规划器会在下次扫描时重新考虑它。

…v-20261025-65633129

Move dsv4-fp8-h200-dynamo-sglang-agentic-agg to the digest-pinned SGLang
nightly 65633129.

将 dsv4-fp8-h200-dynamo-sglang-agentic-agg 更新到按摘要固定的
SGLang nightly 65633129 镜像。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor

Thanks for the contribution!

  • Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
  • PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
  • After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.
中文

感谢你的贡献!

  • **审阅:**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER,请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md,再通过 Slack 联系核心维护者。必须严格遵循模板,并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have,才能触发签核验证。
  • **PR 验证:**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目,并且恰好带有一个主标签时,才会运行扫描:full-sweep-fail-fast(强烈推荐;canary 加逐矩阵 fail-fast)、full-sweep-enabled(有 canary;矩阵任务在失败后继续运行)或 non-canary-full-sweep-enabled(无 canary,也无 fail-fast)。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
  • **合并后:**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复;参见重新运行失败任务的说明。

@Klaud-Cold

Klaud-Cold commented Oct 25, 2026 •

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Failed · Run 40149896305 / attempt 1 · 2026-10-25 02:13 UTC
lmsysorg/sglang:nightly-dev-20261025-65633129@sha256:069efa1d6a2d58309529f53eeee34c39d69b1cdb455eaf7fb8de83e3fd0be727 · 367d181c38c9 · AgentX · PTP8x1/DTP8x0 · Mean latency
Change: Pin the image to SGLang nightly 65633129 (digest sha256:069efa1d) with no recipe change; sgl-project/sglang compare 4358a161...65633129 keeps every used server arg and the DSPARK/hicache/marlin choices, and _handle_dspark's draft-path check in arg_groups/speculative_hook.py is unchanged between both commits.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 N/A N/A N/A

Note: All rows: failed.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.

Next: Stopped as readiness-blocked: both smoke jobs failed at startup with ValueError 'DSpark dense speculative decoding requires setting --speculative-draft-model-path' because the h200-dgxc route mounts /models/gharunners/hf-hub-cache/DeepSeek-V4-Pro (no bundled dspark_* config) as /model, not the DeepSeek-V4-Pro-0813 directory the baseline run 34857875250 served; the same check exists in the old image, so fixing the shared route or switching weights is out of scope.

中文

初次尝试 · 失败 · Run 40149896305 / attempt 1 · 2026-10-25 02:13 UTC
lmsysorg/sglang:nightly-dev-20261025-65633129@sha256:069efa1d6a2d58309529f53eeee34c39d69b1cdb455eaf7fb8de83e3fd0be727 · 367d181c38c9 · AgentX · PTP8x1/DTP8x0 · 平均延迟
**变更:*将镜像固定为 SGLang nightly 65633129(digest sha256:069efa1d),不修改 recipe;sgl-project/sglang 对比 4358a161...65633129 保留了全部所用 server 参数及 DSPARK/hicache/marlin 选项,且 arg_groups/speculative_hook.py 中 handle_dspark 的 draft 路径检查在两个提交间没有变化。实测数值及异常说明见上表。
**下一步:**以 readiness-blocked 停止:两个 smoke 任务均在启动时报 ValueError 'DSpark dense speculative decoding requires setting --speculative-draft-model-path',原因是 h200-dgxc 路由将 /models/gharunners/hf-hub-cache/DeepSeek-V4-Pro(无内置 dspark
配置)挂载为 /model,而非基线运行 34857875250 所用的 DeepSeek-V4-Pro-0813 目录;旧镜像中同样存在该检查,因此修改共享路由或更换权重均超出本次范围。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 1 is waiting on 40149896305. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40149896305。恢复流程会重跑被抢占的运行,并在运行结束后启动下一个会话。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

resumed · The session ended before finishing; session 2 continues it in a new run of its planner.

中文

resumed · 会话未完成即已结束;第 2 个会话将在其规划运行的新运行中继续。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

readiness-blocked · Cleanup pending. Stop and confirm owned runs before closing.

中文

readiness-blocked · 清理待完成。先停止并确认自有运行结束,再关闭 PR。

@Klaud-Cold Klaud-Cold closed this Oct 25, 2026
@Klaud-Cold
Klaud-Cold deleted the klaud/auto-0cb636efc3513aaa-353812587e13d7bc branch October 25, 2026 02:13
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

readiness-blocked · Reason: weights-missing · Repairs: 0 · Runs: 40149896305
Reason: Stopped as readiness-blocked: both smoke jobs failed at startup with ValueError 'DSpark dense speculative decoding requires setting --speculative-draft-model-path' because the h200-dgxc route mounts /models/gharunners/hf-hub-cache/DeepSeek-V4-Pro (no bundled dspark_* config) as /model, not the DeepSeek-V4-Pro-0813 directory the baseline run 34857875250 served; the same check exists in the old image, so fixing the shared route or switching weights is out of scope.
All owned runs ended. PR closed; branch deleted for retry.

中文

readiness-blocked · 原因:weights-missing · 修复次数:0 · 运行:40149896305
**原因:*以 readiness-blocked 停止:两个 smoke 任务均在启动时报 ValueError 'DSpark dense speculative decoding requires setting --speculative-draft-model-path',原因是 h200-dgxc 路由将 /models/gharunners/hf-hub-cache/DeepSeek-V4-Pro(无内置 dspark_ 配置)挂载为 /model,而非基线运行 34857875250 所用的 DeepSeek-V4-Pro-0813 目录;旧镜像中同样存在该检查,因此修改共享路由或更换权重均超出本次范围。
所有自有运行均已结束。PR 已关闭;分支已删除,可重新尝试。

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

1 participant