Skip to content

[Klaud Cold] Update Qwen3.5 FP8 B200 Dynamo-SGLang AgentX MTP image to nightly-dev-cu13-20261025-65633129 / [Klaud Cold] 将 Qwen3.5 FP8 B200 Dynamo-SGLang AgentX MTP 镜像更新至 nightly-dev-cu13-20261025-65633129 - #4477

Open
Klaud-Cold wants to merge 56 commits into
mainfrom
klaud/auto-e920fad623895413-9cc27cc9b9acae03

Conversation

@Klaud-Cold

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

Copy link
Copy Markdown
Collaborator

Goal: Update SGLang image from lmsysorg/sglang:v0.5.22-cu130 to lmsysorg/sglang:nightly-dev-cu13-20261025-65633129.
Baseline: 2026-10-21 · lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae
AgentX · PTP4x1/DTP4x1 · Mean latency · Sources: API 1, API 2, API 3, API 4

Concurrency Total tok/s/GPU ↑ Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
8 5,653.52 56.11 443.89 2.48
16 12,443.5 106.2 508.2 3.04
24 17,501.8 145.56 763.97 3.58
32 20,354.07 182.27 943.99 4.05
48 28,884.31 293.26 895.27 4.98
64 34,683.09 375.89 3,877.7 7.23
Eval Score ↑ Samples
gsm8k/em_strict · c16 97.42% 1,319
gsm8k/em_strict · c24 97.88% 1,319
gsm8k/em_strict · c32 97.8% 1,319
gsm8k/em_strict · c48 97.8% 1,319
gsm8k/em_strict · c64 97.65% 1,319

Stop: Close this PR at any time to stop this candidate now; Klaud cancels its runs and deletes its branch.

中文

**目标:**将 SGLang 镜像从 lmsysorg/sglang:v0.5.22-cu130 更新为 lmsysorg/sglang:nightly-dev-cu13-20261025-65633129。
**基线:**2026-10-21 · lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae
AgentX · PTP4x1/DTP4x1 · 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:**随时关闭此 PR 即可立即停止该候选:Klaud 会取消其运行并删除其分支。

…htly-dev-cu13-20261025-65633129

Pin lmsysorg/sglang:nightly-dev-cu13-20261025-65633129 by digest
(sgl-project/sglang@65633129) for
qwen3.5-fp8-b200-dynamo-sglang-agentic-disagg-mtp.

将 qwen3.5-fp8-b200-dynamo-sglang-agentic-disagg-mtp 的 SGLang 镜像更新为
按摘要固定的 lmsysorg/sglang:nightly-dev-cu13-20261025-65633129
(sgl-project/sglang@65633129)。

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 · Passed · Run 40149594245 / attempt 1 · 2026-10-25 03:07 UTC
lmsysorg/sglang:nightly-dev-cu13-20261025-65633129@sha256:069efa1d6a2d58309529f53eeee34c39d69b1cdb455eaf7fb8de83e3fd0be727 · 367d181c38c9 · AgentX · PTP4x1/DTP4x1 · Mean latency
Change: Master image only: v0.5.22-cu130 (SGLang commit baa9befa, flashinfer 0.7.1) to digest-pinned nightly-dev-cu13-20261025-65633129 (SGLang commit 65633129, flashinfer 0.7.2, sgl-kernel 0.4.9, CUDA 13.0.3); every recipe flag and enum value still exists in arg_groups at 65633129 (sgl-project/sglang@v0.5.22...6563312); Dynamo 805a77f0 source install against the new SGLang started and served all points.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
8 55.05 (-1.9%) 387.57 (-12.7%) 2.45 (-1.2%)
16 106.28 (+0.1%) 500.3 (-1.6%) 3.01 (-1%)
24 147.97 (+1.7%) 598.88 (-21.6%) 3.7 (+3.4%)
32 185.21 (+1.6%) 727.12 (-23%) 3.96 (-2.2%)
48 291.94 (-0.5%) 896.88 (+0.2%) 5.09 (+2.2%)
64 412.56 (+9.8%) 1,393.02 (-64.1%) 5.99 (-17.2%)
Eval Score ↑ Samples
gsm8k/em_strict · c16 97.73% (+0.30 pp) 1,319 each
gsm8k/em_strict · c24 97.42% (-0.45 pp) 1,319 each
gsm8k/em_strict · c32 97.88% (+0.08 pp) 1,319 each
gsm8k/em_strict · c48 98.03% (+0.23 pp) 1,319 each
gsm8k/em_strict · c64 97.95% (+0.30 pp) 1,319 each

Next: Append the perf-changelog entry and dispatch the full-sweep-fail-fast final sweep.

中文

初次尝试 · 已通过 · Run 40149594245 / attempt 1 · 2026-10-25 03:07 UTC
lmsysorg/sglang:nightly-dev-cu13-20261025-65633129@sha256:069efa1d6a2d58309529f53eeee34c39d69b1cdb455eaf7fb8de83e3fd0be727 · 367d181c38c9 · AgentX · PTP4x1/DTP4x1 · 平均延迟
**变更:**仅更新 master 镜像:v0.5.22-cu130(SGLang commit baa9befa,flashinfer 0.7.1)更新为按摘要固定的 nightly-dev-cu13-20261025-65633129(SGLang commit 65633129,flashinfer 0.7.2,sgl-kernel 0.4.9,CUDA 13.0.3);配方使用的全部参数和枚举值在 65633129 的 arg_groups 中仍然存在(sgl-project/sglang@v0.5.22...656331295619a4d0267b40b7da0b5cd5b549fcbc);Dynamo 805a77f0 源码安装在新版 SGLang 上已正常启动并完成全部点位。实测数值及异常说明见上表。
**下一步:**追加 perf-changelog 条目并通过 full-sweep-fail-fast 发起最终扫描。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 1 is waiting on 40149594245. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40149594245。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。

…gentX MTP image update

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Klaud-Cold Klaud-Cold added full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) and removed full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) labels Oct 25, 2026
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Klaud-Cold Klaud-Cold added the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Oct 25, 2026
@Klaud-Cold

Klaud-Cold commented Oct 25, 2026 •

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Passed · Run 40157231416 / attempt 1 · 2026-10-25 04:54 UTC
lmsysorg/sglang:nightly-dev-cu13-20261025-65633129@sha256:069efa1d6a2d58309529f53eeee34c39d69b1cdb455eaf7fb8de83e3fd0be727 · 196dec2c3fba · AgentX · PTP4x1/DTP4x1 · Mean latency
Change: Validate the complete updated-image family at this head.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
8 54.75 (-2.4%) 381.04 (-14.2%) 2.47 (-0.4%)
16 106.4 (+0.2%) 496.73 (-2.3%) 2.97 (-2.3%)
24 148.78 (+2.2%) 595.14 (-22.1%) 3.71 (+3.6%)
32 185.23 (+1.6%) 730.48 (-22.6%) 3.91 (-3.5%)
48 304.35 (+3.8%) 882.89 (-1.4%) 4.37 (-12.2%)
64 386.14 (+2.7%) 2,653.23 (-31.6%) 7.32 (+1.2%)
Eval Score ↑ Samples
gsm8k/em_strict · c16 97.65% (+0.23 pp) 1,319 each
gsm8k/em_strict · c24 97.57% (-0.30 pp) 1,319 each
gsm8k/em_strict · c32 97.8% (+0.00 pp) 1,319 each
gsm8k/em_strict · c48 97.95% (+0.15 pp) 1,319 each
gsm8k/em_strict · c64 97.88% (+0.23 pp) 1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40157231416 / attempt 1 · 2026-10-25 04:54 UTC
lmsysorg/sglang:nightly-dev-cu13-20261025-65633129@sha256:069efa1d6a2d58309529f53eeee34c39d69b1cdb455eaf7fb8de83e3fd0be727 · 196dec2c3fba · AgentX · PTP4x1/DTP4x1 · 平均延迟
**变更:**验证此提交更新镜像后的完整配置族。实测数值及异常说明见上表。
**下一步:**标记为就绪,等待维护者审查。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 2 is waiting on 40157231416. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40157231416。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。

@github-actions

Copy link
Copy Markdown
Contributor

@Klaud-Cold
Klaud-Cold marked this pull request as ready for review October 25, 2026 04:55
@Klaud-Cold
Klaud-Cold requested a review from a team October 25, 2026 04:55
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

validated · Repairs: unknown · Runs: 40149594245, 40157092036, 40157137123, 40157204569, 40157210291, 40157231416
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数:未知 · 运行:40149594245, 40157092036, 40157137123, 40157204569, 40157210291, 40157231416
所有自有运行均已结束。完整 sweep 已验证;已就绪,等待审查。

@claude

claude Bot commented Oct 25, 2026 •

Copy link
Copy Markdown
Contributor

Claude finished @Klaud-Cold's task in 13s —— View job


LGTM - no blocking issues found

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code review found no issues

No high-confidence issues detected in this change.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended)

Projects

Status: No status

Development

Successfully merging this pull request may close these issues.

1 participant