Skip to content

[Klaud Cold] Update DeepSeek-V4.1-Flash FP4 B300 SGLang AgentX DSpark image to nightly-dev-cu13-20261107-e45e1d0c / [Klaud Cold] 将 DeepSeek-V4.1-Flash FP4 B300 SGLang AgentX DSpark 镜像更新至 nightly-dev-cu13-20261107-e45e1d0c - #4900

Merged
adibarra merged 8 commits into
mainfrom
klaud/auto-1048dffeb5568a29-831c83d20293c280
Nov 11, 2026

Conversation

@Klaud-Cold

@Klaud-Cold Klaud-Cold commented Nov 7, 2026 •

Copy link
Copy Markdown
Collaborator

Goal: Update SGLang image to lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c for dsv41flash-fp4-b300-sglang-agentic-dspark (from lmsysorg/sglang:nightly-dev-cu13-20261020-ab580b2e) and dsv41flash-fp4-b300-sglang-agentic-dspark-lowconc (from lmsysorg/sglang:nightly-dev-cu13-20261014-684fea8a).
Baseline: dsv41flash-fp4-b300-sglang-agentic-dspark 2026-10-29 · lmsysorg/sglang:nightly-dev-cu13-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576; dsv41flash-fp4-b300-sglang-agentic-dspark-lowconc no published baseline · lmsysorg/sglang:nightly-dev-cu13-20261014-684fea8a@sha256:ddae0e399d3d94f7827f16183301cf8dac4cf0eae0ef25590b7dac5881d713d8
Mean latency · Sources: API 1, API 2, API 3, API 4

Point Total tok/s/GPU ↑ Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
AgentX c8 TP4 EP1 c78c78 16,646.59 115.7 416.41 1.73
AgentX c16 TP2 EP1 efd2e7 56,676.71 358.04 683.2 3.05
AgentX c16 TP4 EP1 995237 29,058.54 186.85 837.53 2.23
AgentX c20 TP4 EP1 e9dc7e 38,273.69 237.4 650.07 2.06
AgentX c32 TP2 EP1 2ced24 111,975.37 660.24 2,321.25 3.8
AgentX c48 TP2 EP1 3c9430 155,738.45 1,034.23 1,926.06 4.84
AgentX c64 TP2 EP2 15f44f 221,548.35 1,547.56 1,207.24 6.15
AgentX c256 TP2 EP2 d6027e 445,156.65 3,170.5 7,162.15 15.7
Eval Score ↑ Samples
gsm8k/em_strict · c4 97.57% 1,319
gsm8k/em_strict · c48 96.82% 1,319
gsm8k/em_strict · c256 97.57% 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 dsv41flash-fp4-b300-sglang-agentic-dspark(原为 lmsysorg/sglang:nightly-dev-cu13-20261020-ab580b2e)和 dsv41flash-fp4-b300-sglang-agentic-dspark-lowconc(原为 lmsysorg/sglang:nightly-dev-cu13-20261014-684fea8a)的 SGLang 镜像更新为 lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c。
基线:dsv41flash-fp4-b300-sglang-agentic-dspark 2026-10-29 · lmsysorg/sglang:nightly-dev-cu13-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576; dsv41flash-fp4-b300-sglang-agentic-dspark-lowconc 无已发布基线 · lmsysorg/sglang:nightly-dev-cu13-20261014-684fea8a@sha256:ddae0e399d3d94f7827f16183301cf8dac4cf0eae0ef25590b7dac5881d713d8
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。

…dev-cu13-20261107-e45e1d0c

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

github-actions Bot commented Nov 7, 2026

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 Nov 7, 2026 •

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Passed · Run 42391462778 / attempt 1 · 2026-11-07 18:29 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · 7a2db9d088b1 · Mean latency
Change: Moved both family keys to the SGLang nightly-dev-cu13-20261107-e45e1d0c image (sglang e45e1d0c, FlashInfer 0.7.3, sgl-deep-gemm 0.2.6) and updated the recipe sglang identity to 0.0.0.dev1+ge45e1d0c5; source diff ab580b2e..e45e1d0c and 684fea8a..e45e1d0c keeps every used flag, env var and default (SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT default changed to per_rank, which the recipes already set explicitly).

Point Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
AgentX c8 TP4 EP1 974719 107.03 (N/A) 476.24 (N/A) 1.61 (N/A)
AgentX c16 TP2 EP1 8209e7 298.69 (N/A) 858.47 (N/A) 2.41 (N/A)
AgentX c64 TP2 EP2 ca0723 1,228.5 (N/A) 1,537.04 (N/A) 5.57 (N/A)
AgentX c1 TP4 EP1 cde61e 46.18 (N/A) 535.61 (N/A) 1.42 (N/A)

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

Eval Score ↑ Samples
gsm8k/em_strict · c256 97.5% (-0.08 pp) 1,319 each
gsm8k/em_strict · c48 97.19% (+0.38 pp) 1,319 each

Next: Smoke benchmarks (0 request errors) and both gsm8k evals passed; append the perf-changelog entry and dispatch the full final sweep.

中文

初次尝试 · 已通过 · Run 42391462778 / attempt 1 · 2026-11-07 18:29 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · 7a2db9d088b1 · 平均延迟
**变更:**将两个 family key 迁移到 SGLang nightly-dev-cu13-20261107-e45e1d0c 镜像(sglang e45e1d0c、FlashInfer 0.7.3、sgl-deep-gemm 0.2.6),并将配方中的 sglang identity 更新为 0.0.0.dev1+ge45e1d0c5;ab580b2e..e45e1d0c 与 684fea8a..e45e1d0c 的源码对比显示所用参数、环境变量和默认值均保留(SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT 默认值改为 per_rank,配方已显式设置)。实测数值及异常说明见上表。
**下一步:**smoke 基准(0 个请求错误)和两个 gsm8k 评测均已通过;追加 perf-changelog 条目并启动完整的最终扫描。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 1 is waiting on 42391462778. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 42391462778。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 github-actions Bot added the klaud-parked Klaud parked until its queued or running benchmark runs end label Nov 7, 2026
@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。

@github-actions github-actions Bot added klaud-in-progress Klaud is working on this candidate or will resume it and removed klaud-parked Klaud parked until its queued or running benchmark runs end labels Nov 7, 2026
@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 Nov 7, 2026
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Running · Run 42401592549 / attempt 1 · 2026-11-07 18:31 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · 85d157e3fcf1 · Mean latency
Change: Appended the perf-changelog entry for both family keys and merged main; the image and recipe settings are unchanged from the passed smoke attempt.

Point Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
AgentX c8 TP4 EP1 58c3da N/A N/A N/A
AgentX c16 TP4 EP1 bc85a5 N/A N/A N/A
AgentX c20 TP4 EP1 46a763 N/A N/A N/A
AgentX c16 TP2 EP1 0f8a3b N/A N/A N/A
AgentX c32 TP2 EP1 015911 N/A N/A N/A
AgentX c48 TP2 EP1 2f06fa N/A N/A N/A
AgentX c64 TP2 EP2 6f35c1 N/A N/A N/A
AgentX c256 TP2 EP2 14c7d3 N/A N/A N/A
AgentX c1 TP4 EP1 1383c7 N/A N/A N/A
AgentX c4 TP4 EP1 e38378 N/A N/A N/A

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

Next: Wait for the full final sweep (10 AgentX points and 2 gsm8k evals) on b300-dsxe.

中文

最终完整 sweep · 运行中 · Run 42401592549 / attempt 1 · 2026-11-07 18:31 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · 85d157e3fcf1 · 平均延迟
**变更:**为两个 family key 追加 perf-changelog 条目并合并 main;镜像和配方设置与已通过的 smoke 尝试相同。实测数值及异常说明见上表。
**下一步:**等待 b300-dsxe 上的完整最终扫描(10 个 AgentX 点和 2 个 gsm8k 评测)完成。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 2 is waiting on 42401592549. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 42401592549。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 github-actions Bot added klaud-parked Klaud parked until its queued or running benchmark runs end and removed klaud-in-progress Klaud is working on this candidate or will resume it labels Nov 7, 2026
@github-actions

github-actions Bot commented Nov 7, 2026 •

Copy link
Copy Markdown
Contributor

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

resumed · The session ended before finishing; session 3 continues it in a new run of its planner.

中文

resumed · 会话未完成即已结束;第 3 个会话将在其规划运行的新运行中继续。

@github-actions github-actions Bot added klaud-in-progress Klaud is working on this candidate or will resume it and removed klaud-parked Klaud parked until its queued or running benchmark runs end labels Nov 8, 2026
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Cancelled · Run 42401592549 / attempt 2 · 2026-11-08 08:00 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · 85d157e3fcf1 · Mean latency
Change: Rerun of the final sweep after a preemption; the image and recipe settings are unchanged.

Point Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
AgentX c16 TP2 EP1 0f8a3b 366.14 (N/A) 649.19 (N/A) 2.54 (N/A)
AgentX c48 TP2 EP1 2f06fa 1,070 (N/A) 1,662.18 (N/A) 4.48 (N/A)
AgentX c64 TP2 EP2 6f35c1 1,544.59 (N/A) 1,271.99 (N/A) 6.04 (N/A)
AgentX c16 TP4 EP1 bc85a5 192.91 (N/A) 636.91 (N/A) 1.79 (N/A)
AgentX c1 TP4 EP1 1383c7 46.78 (N/A) 574.47 (N/A) 1.41 (N/A)
AgentX c4 TP4 EP1 e38378 67.76 (N/A) 392.3 (N/A) 1.48 (N/A)
AgentX c8 TP4 EP1 58c3da 114.98 (N/A) 528.71 (N/A) 1.65 (N/A)
AgentX c20 TP4 EP1 46a763 N/A N/A N/A
AgentX c32 TP2 EP1 015911 N/A N/A N/A
AgentX c256 TP2 EP2 14c7d3 N/A N/A N/A

Note: All rows: Δ N/A: no matched baseline.
Note: AgentX c64 TP2 EP2 6f35c1: 3 request errors.
Note: AgentX c20 TP4 EP1 46a763, AgentX c32 TP2 EP1 015911, AgentX c256 TP2 EP2 14c7d3: cancelled.
Note: AgentX c20 TP4 EP1 46a763, AgentX c32 TP2 EP1 015911, AgentX c256 TP2 EP2 14c7d3: request errors unavailable.

Eval Score ↑ Samples
gsm8k/em_strict · c256 97.42% (-0.15 pp) 1,319 each

Next: Four jobs (c20, c32, c256 and the c48 gsm8k eval) hit the 8h20m GitHub job limit after about 7 hours pending in the Slurm queue, not a server failure; main also appended perf-changelog entries, so merge main, keep this entry at the tail and dispatch the final sweep again.

中文

最终完整 sweep · 已取消 · Run 42401592549 / attempt 2 · 2026-11-08 08:00 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · 85d157e3fcf1 · 平均延迟
**变更:**抢占后重跑最终扫描;镜像和配方设置保持不变。实测数值及异常说明见上表。
**下一步:**四个任务(c20、c32、c256 和 c48 gsm8k 评测)在 Slurm 队列中等待约 7 小时后触及 GitHub 8 小时 20 分钟的任务时限,并非服务器故障;main 也追加了 perf-changelog 条目,因此合并 main、保持本条目位于末尾并重新启动最终扫描。

…568a29-831c83d20293c280

# Conflicts:
#	inferencex-e2e/perf-changelog.yaml
@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 Nov 8, 2026
@Klaud-Cold

Klaud-Cold commented Nov 8, 2026 •

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Failed · Run 42461137126 / attempt 1 · 2026-11-08 10:54 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · 82df6c224dc4 · Mean latency
Change: Merged main, which had appended perf-changelog entries, and kept this entry at the tail; the image and recipe settings are unchanged.

Point Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
AgentX c8 TP4 EP1 58c3da N/A N/A N/A
AgentX c16 TP4 EP1 bc85a5 N/A N/A N/A
AgentX c20 TP4 EP1 46a763 N/A N/A N/A
AgentX c16 TP2 EP1 0f8a3b N/A N/A N/A
AgentX c32 TP2 EP1 015911 N/A N/A N/A
AgentX c48 TP2 EP1 2f06fa N/A N/A N/A
AgentX c64 TP2 EP2 6f35c1 N/A N/A N/A
AgentX c256 TP2 EP2 14c7d3 N/A N/A N/A
AgentX c1 TP4 EP1 1383c7 46.82 (N/A) 572.86 (N/A) 1.4 (N/A)
AgentX c4 TP4 EP1 e38378 N/A N/A N/A

Note: AgentX c8 TP4 EP1 58c3da, AgentX c16 TP4 EP1 bc85a5, AgentX c20 TP4 EP1 46a763, AgentX c16 TP2 EP1 0f8a3b, AgentX c32 TP2 EP1 015911, AgentX c64 TP2 EP2 6f35c1, AgentX c256 TP2 EP2 14c7d3, AgentX c4 TP4 EP1 e38378: cancelled.
Note: AgentX c8 TP4 EP1 58c3da, AgentX c16 TP4 EP1 bc85a5, AgentX c20 TP4 EP1 46a763, AgentX c16 TP2 EP1 0f8a3b, AgentX c32 TP2 EP1 015911, AgentX c48 TP2 EP1 2f06fa, AgentX c64 TP2 EP2 6f35c1, AgentX c256 TP2 EP2 14c7d3, AgentX c4 TP4 EP1 e38378: request errors unavailable.
Note: All rows: Δ N/A: no matched baseline.
Note: AgentX c48 TP2 EP1 2f06fa: failed.

Next: The failure is a transient download error, not a server failure; main appended another perf-changelog entry and now conflicts, so merge main, keep this entry at the tail and dispatch the final sweep again.

中文

最终完整 sweep · 失败 · Run 42461137126 / attempt 1 · 2026-11-08 10:54 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · 82df6c224dc4 · 平均延迟
**变更:**合并了已追加 perf-changelog 条目的 main,并保持本条目位于末尾;镜像和配方设置保持不变。实测数值及异常说明见上表。
**下一步:**该失败是临时下载错误,而非服务器故障;main 又追加了 perf-changelog 条目并产生冲突,因此合并 main,保持本条目位于末尾,并重新启动最终扫描。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 3 is waiting on 42461137126. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 3 个会话正在等待 42461137126。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 github-actions Bot added klaud-parked Klaud parked until its queued or running benchmark runs end and removed klaud-in-progress Klaud is working on this candidate or will resume it labels Nov 8, 2026
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

resumed · The session ended before finishing; session 4 continues it in a new run of its planner.

中文

resumed · 会话未完成即已结束;第 4 个会话将在其规划运行的新运行中继续。

@github-actions github-actions Bot added the klaud-in-progress Klaud is working on this candidate or will resume it label Nov 8, 2026
@github-actions github-actions Bot added klaud-parked Klaud parked until its queued or running benchmark runs end klaud-in-progress Klaud is working on this candidate or will resume it and removed klaud-in-progress Klaud is working on this candidate or will resume it klaud-parked Klaud parked until its queued or running benchmark runs end labels Nov 9, 2026
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Passed · Run 42564578697 / attempt 5 · 2026-11-10 17:14 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · be3b5179346a · Mean latency
Change: Validate the complete updated-image family at this head.

Point Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
AgentX c1 TP4 EP1 360a69 46.82 (N/A) 576.8 (N/A) 1.4 (N/A)
AgentX c4 TP4 EP1 ceab4d 69.54 (N/A) 410.46 (N/A) 1.5 (N/A)
AgentX c8 TP4 EP1 c78c78 116.04 (+0.3%) 527.74 (+26.7%) 1.66 (-4%)
AgentX c16 TP2 EP1 efd2e7 367.59 (+2.7%) 661.99 (-3.1%) 2.54 (-16.7%)
AgentX c16 TP4 EP1 995237 191.64 (+2.6%) 650.49 (-22.3%) 1.81 (-18.8%)
AgentX c20 TP4 EP1 e9dc7e 237.45 (+0%) 764.05 (+17.5%) 2.04 (-1%)
AgentX c32 TP2 EP1 2ced24 682.04 (+3.3%) 1,787.37 (-23%) 3.39 (-10.8%)
AgentX c48 TP2 EP1 3c9430 1,061.36 (+2.6%) 1,779.51 (-7.6%) 4.56 (-5.8%)
AgentX c64 TP2 EP2 15f44f 1,547.78 (+0%) 1,286.19 (+6.5%) 6.02 (-2.1%)
AgentX c256 TP2 EP2 d6027e 3,690.44 (+16.4%) 5,924.13 (-17.3%) 13.79 (-12.2%)

Note: AgentX c1 TP4 EP1 360a69, AgentX c8 TP4 EP1 c78c78, AgentX c16 TP2 EP1 efd2e7, AgentX c32 TP2 EP1 2ced24: 1 request error.
Note: AgentX c1 TP4 EP1 360a69, AgentX c4 TP4 EP1 ceab4d: Δ N/A: no published baseline for dsv41flash-fp4-b300-sglang-agentic-dspark-lowconc on lmsysorg/sglang:nightly-dev-cu13-20261014-684fea8a@sha256:ddae0e399d3d94f7827f16183301cf8dac4cf0eae0ef25590b7dac5881d713d8.
Note: AgentX c48 TP2 EP1 3c9430, AgentX c64 TP2 EP2 15f44f: 2 request errors.
Note: AgentX c256 TP2 EP2 d6027e: 8 request errors.

Eval Score ↑ Samples
gsm8k/em_strict · c48 97.42% (+0.61 pp) 1,319 each
gsm8k/em_strict · c256 97.35% (-0.23 pp) 1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 42564578697 / attempt 5 · 2026-11-10 17:14 UTC
lmsysorg/sglang:nightly-dev-cu13-20261107-e45e1d0c@sha256:81eb456abfaa3cf50a2ac463abe02705486f3f1e10929ca306e5929b937d5388 · be3b5179346a · 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
Klaud-Cold marked this pull request as ready for review November 10, 2026 17:14
@Klaud-Cold
Klaud-Cold requested a review from a team November 10, 2026 17:14
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

/use 42564578697

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

validated · Repairs: unknown · Runs: 42391462778, 42401571311, 42401592549, 42461109915, 42461137126, 42473471139, 42473490832, 42564578697
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数:未知 · 运行:42391462778, 42401571311, 42401592549, 42461109915, 42461137126, 42473471139, 42473490832, 42564578697
所有自有运行均已结束。完整 sweep 已验证;已就绪,等待审查。

@github-actions github-actions Bot added klaud-finished Klaud verified the full sweep; ready for review and removed klaud-parked Klaud parked until its queued or running benchmark runs end labels Nov 10, 2026
@github-actions

github-actions Bot commented Nov 10, 2026 •

Copy link
Copy Markdown
Contributor

@Klaud-Cold staged run 42564578697.

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

@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 image bump and found no bugs. The NVIDIA master config and the B300 srt-slurm recipes have designated code owners, so one of them should still take a look.

What was reviewed:

  • Both master entries use the same tag and digest, and that pair matches the merged B200 DSpark entry already on main. The tag's short SHA e45e1d0c agrees with the recipe pin 0.0.0.dev1+ge45e1d0c5 in both agentic.yaml and agentic-lowconc.yaml.
  • The perf-changelog block is added only at the end of the file: earlier entries are unchanged, the blank-line separator is correct, the file ends with a trailing newline, and both config keys are listed in one block.
  • Not checked here: benchmark results for the new image, and whether an accepted /use <run_id> exists.
Extended reasoning...

The change swaps the SGLang image (tag and digest) in two dsv41flash B300 AgentX entries of nvidia-master.yaml, updates the matching sglang framework pin in two single-node srt-slurm recipes, and appends one perf-changelog entry. It touches no security-sensitive surface such as auth, secrets or code execution. The change is small and mechanical, and the image matches the one already merged for B200. I deferred rather than approved because CODEOWNERS assigns both nvidia-master.yaml and the b300-* recipe directories to specific NVIDIA maintainers.

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 1f169e5 into main Nov 11, 2026
33 checks passed
@adibarra
adibarra deleted the klaud/auto-1048dffeb5568a29-831c83d20293c280 branch November 11, 2026 04:36
@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 10 0 10 0 0
eval 2 2 4 0 0
trace 10 0 0 0 0
power 10 0 0 0 0

benchmarkFiles 11, rawBenchmarkRows 20, coveredByPerJobFiles 10, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 2638, gpuMetricSamplesNew 181492, powerPublicationPoints 10, 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。

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) klaud-finished Klaud verified the full sweep; ready for review

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

2 participants