Skip to content

[Klaud Cold] Update DeepSeek-V4-Pro FP4 GB200 Dynamo-SGLang AgentX agg image to v0.5.22-cu130 / [Klaud Cold] 将 DeepSeek-V4-Pro FP4 GB200 Dynamo-SGLang AgentX agg 镜像更新至 v0.5.22-cu130 - #4529

Merged
adibarra merged 3 commits into
mainfrom
klaud/auto-e53b88fb3e1e4cb8-38c8fc1ee9032f8d
Oct 26, 2026
Merged

adibarra merged 3 commits into
mainfrom
klaud/auto-e53b88fb3e1e4cb8-38c8fc1ee9032f8d

Conversation

@Klaud-Cold

@Klaud-Cold Klaud-Cold commented Oct 26, 2026 •

Copy link
Copy Markdown
Collaborator

Goal: Update SGLang image from lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 to lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae.
Baseline: 2026-10-21 · lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469
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 2,502.65 18.31 858.07 3.21
4 3,898.22 27.4 660.72 3.66
Eval Score ↑ Samples
gsm8k/em_strict · c4 97.27% 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:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469 更新至 lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae。
**基线:**2026-10-21 · lmsysorg/sglang:v0.5.21-cu130@sha256:b1259f3ea3275f66237c498ea388919729018bc9f01c3d638391e06e2cf3f469
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。

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 26, 2026 •

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Passed · Run 40268541074 / attempt 1 · 2026-10-26 06:19 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 65a6dd197872 · AgentX · PTP8x1/DTP8x0 · Mean latency
Change: Bump the master image to SGLang v0.5.22-cu130 at measured ref d23dd1d (digest-pinned, arm64 manifest present); source diff of v0.5.21 (e00930c5) vs v0.5.22 (baa9befa) shows every recipe flag and env var still parsed with unchanged defaults, and dp-size 1 without DP attention keeps its meaning after the new attn-dp-size split, so no recipe change is needed.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 18.24 (N/A) 918.12 (N/A) 3.19 (N/A)
4 27.79 (N/A) 755.89 (N/A) 3.65 (N/A)

Note: All rows: Δ N/A: no matched baseline.

Eval Score ↑ Samples
gsm8k/em_strict · c4 97.57% (+0.30 pp) 1,319 each

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

中文

初次尝试 · 已通过 · Run 40268541074 / attempt 1 · 2026-10-26 06:19 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 65a6dd197872 · AgentX · PTP8x1/DTP8x0 · 平均延迟
**变更:**在测量引用 d23dd1d 上将主镜像更新为 SGLang v0.5.22-cu130(已固定摘要,含 arm64 清单);对比 v0.5.21(e00930c5)与 v0.5.22(baa9befa)源码,配方中的所有参数和环境变量仍可解析且默认值不变,新增 attn-dp-size 拆分后不启用 DP attention 的 dp-size 1 语义不变,因此无需修改配方。实测数值及异常说明见上表。
**下一步:**追加 perf-changelog 条目,检查最终矩阵并通过 full-sweep-fail-fast 发起最终扫描。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 1 is waiting on 40268541074. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40268541074。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。

…agg v0.5.22-cu130

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 26, 2026
@Klaud-Cold

Klaud-Cold commented Oct 26, 2026 •

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Cancelled · Run 40279186312 / attempt 1 · 2026-10-26 06:21 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 4f555851babf · AgentX · PTP8x1/DTP8x0 · Mean latency
Change: Append the exact-family perf-changelog entry for the v0.5.22-cu130 image after the smoke benchmarks and gsm8k eval passed, with no recipe change.

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

Note: All rows: cancelled.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: no matched baseline.

Next: This push-triggered run was superseded by the full-sweep-fail-fast label run 40279236382, which carries the final sweep.

中文

最终完整 sweep · 已取消 · Run 40279186312 / attempt 1 · 2026-10-26 06:21 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 4f555851babf · AgentX · PTP8x1/DTP8x0 · 平均延迟
**变更:**在冒烟基准测试与 gsm8k 评测通过后,为 v0.5.22-cu130 镜像追加该配置族的 perf-changelog 条目,配方无改动。实测数值及异常说明见上表。
**下一步:**该推送触发的运行已被 full-sweep-fail-fast 标签触发的运行 40279236382 取代,最终扫描由后者执行。

@Klaud-Cold

Klaud-Cold commented Oct 26, 2026 •

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Passed · Run 40279236382 / attempt 1 · 2026-10-26 09:42 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 4f555851babf · AgentX · PTP8x1/DTP8x0 · Mean latency
Change: Validate the complete updated-image family at this head.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 18.24 (-0.4%) 968.69 (+12.9%) 3.2 (-0.3%)
4 27.06 (-1.2%) 780.97 (+18.2%) 3.73 (+1.9%)
Eval Score ↑ Samples
gsm8k/em_strict · c4 97.42% (+0.15 pp) 1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40279236382 / attempt 1 · 2026-10-26 09:42 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 4f555851babf · AgentX · PTP8x1/DTP8x0 · 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 40279236382. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40279236382。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 26, 2026 09:42
@Klaud-Cold
Klaud-Cold requested a review from a team October 26, 2026 09:42
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

validated · Repairs: unknown · Runs: 40268541074, 40279186312, 40279236382
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数:未知 · 运行:40268541074, 40279186312, 40279236382
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 26, 2026 •

Copy link
Copy Markdown
Contributor

Claude finished @Klaud-Cold's task in 19s —— View job


LGTM, no blocking issues found.

  • The new image is a public Docker Hub image pinned by digest.
  • The inferencex-e2e/perf-changelog.yaml entry is added at the end of the file and covers dsv4-fp4-gb200-dynamo-sglang-agentic-agg.
  • The full-sweep-fail-fast sweep (run 40279236382) passed. Throughput and TPOT are about the same as the baseline, and the gsm8k score is 97.42% (+0.15 pp).
  • TTFT went up 13–18% at c1 and c4. Maintainers may want to note that.

@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.

I reviewed this PR and found no bugs. The changed config file has designated code owners, so one of them should still take a look before merge.

What was reviewed:

  • The image bump for dsv4-fp4-gb200-dynamo-sglang-agentic-agg in nvidia-master.yaml. The new tag is digest-pinned like the old one, and no other fields changed.
  • The perf-changelog.yaml entry. It is added only at the end with a blank separator line, earlier bytes are unchanged, and it lists the same config key with a matching description, PR link and [agentic-coding] scenario type.
Extended reasoning...

The change touches two YAML files. It replaces one digest-pinned SGLang image (v0.5.21-cu130 to v0.5.22-cu130) for one GB200 agentic config key and appends one matching perf-changelog block. It touches no security-sensitive surface such as auth, injection or data exposure. The change is small and mechanical and the bug hunt ran until it found nothing more (dry_streak). I deferred rather than approved because .github/CODEOWNERS assigns specific owners to inferencex-e2e/configs/nvidia-master.yaml.

@adibarra

Copy link
Copy Markdown
Collaborator

/use 40279236382

@github-actions

github-actions Bot commented Oct 26, 2026 •

Copy link
Copy Markdown
Contributor

@adibarra staged run 40279236382.

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

Resolve perf-changelog.yaml by taking main's file and re-appending this PR's entry.

中文:合并 origin/main;perf-changelog.yaml 采用 main 版本,并在末尾重新追加本 PR 的条目。
@adibarra
adibarra merged commit ca2be1c into main Oct 26, 2026
26 checks passed
@adibarra
adibarra deleted the klaud/auto-e53b88fb3e1e4cb8-38c8fc1ee9032f8d branch October 26, 2026 23:57
@github-actions

Copy link
Copy Markdown
Contributor

Ingest report — ✅ Results ingested into the production database.

Ingest workflow · Merge Ingest run · Benchmark source run

Category new duplicate skipped unmapped failed
benchmark 2 0 0 0 0
eval 1 1 0 0 0
trace 2 0 0 0 0
power 0 0 2 0 0

benchmarkFiles 2, rawBenchmarkRows 2, coveredByPerJobFiles 0, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 1319, gpuMetricSamplesNew 0, powerPublicationPoints 2, changelogEntriesWritten 1

Counts cover only this run. Rows: benchmark_results, eval_results, agentic rows linked to a trace, gpu_metrics series. This comment updates if the ingest is rerun.

中文

Ingest 报告 — ✅ 结果已写入 production 数据库。

上表仅统计本次运行写入的数据,各行单位依次为 benchmark_results 行、eval_results 行、已关联 trace 的 agentic 行和 gpu_metrics 序列。重新运行 ingest 时本评论会原地更新。

@nvpohanh

Copy link
Copy Markdown
Collaborator

[by Claude Code] The SGLang cookbook integration of this PR's AgentX config (dsv4-fp4-gb200-dynamo-sglang-agentic-agg) will happen in sgl-project/sglang#46807, in the Agentic Long-Context Deployment section of the DeepSeek-V4 cookbook page.

The commands mirror the launch measured in run 40279236382, pinned by this PR's /use comment.

@nvpohanh

Copy link
Copy Markdown
Collaborator

[by Claude Code] sgl-project/sglang#46807 has been merged. This PR's AgentX config (dsv4-fp4-gb200-dynamo-sglang-agentic-agg) is now in the Agentic Long-Context Deployment section of the SGLang DeepSeek-V4 cookbook: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4#agentic-long-context-deployment

The commands mirror the launch measured in run 40279236382, pinned by this PR's /use comment.

@nvpohanh

Copy link
Copy Markdown
Collaborator

[by Claude Code] The SGLang cookbook update of this PR's AgentX config (dsv4-fp4-gb200-dynamo-sglang-agentic-agg) will happen in sgl-project/sglang#46948, in the Agentic Long-Context Deployment section of the DeepSeek-V4 cookbook page. The config has been live in that section since sgl-project/sglang#46807; sgl-project/sglang#46948 refreshes it with the pin below.

The commands mirror the launch measured in run 40279236382, pinned by this PR's /use comment.

@nvpohanh

Copy link
Copy Markdown
Collaborator

[by Claude Code] sgl-project/sglang#46948 has been merged. This PR's AgentX config (dsv4-fp4-gb200-dynamo-sglang-agentic-agg) is now in the Agentic Long-Context Deployment section of the SGLang DeepSeek-V4 cookbook: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4#agentic-long-context-deployment

The commands mirror the launch measured in run 40279236382, pinned by this PR's /use comment.

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: Done

Development

Successfully merging this pull request may close these issues.

4 participants