Skip to content

[Klaud Cold] Update GLM-5.2 FP4 B300 SGLang AgentX MTP image to v0.5.22-cu130 / [Klaud Cold] 将 GLM-5.2 FP4 B300 SGLang AgentX MTP 镜像更新至 v0.5.22-cu130 - #4325

Closed
Klaud-Cold wants to merge 3 commits into
mainfrom
klaud/auto-50375757c31a0f3e-9e6710099ed6010e
Closed

Klaud-Cold wants to merge 3 commits into
mainfrom
klaud/auto-50375757c31a0f3e-9e6710099ed6010e

Conversation

@Klaud-Cold

@Klaud-Cold Klaud-Cold commented Oct 19, 2026 •

Copy link
Copy Markdown
Collaborator

Goal: Update SGLang image from lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729 to lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae.
Baseline: 2026-10-12 · lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729
AgentX · TP8/EP1 · Mean latency · Sources: API 1, API 2, API 3

Concurrency Total tok/s/GPU ↑ Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 2,619.53 18.69 839.81 3.02
4 4,030.82 27.93 634.37 3.35
8 7,209.85 52.08 564.95 3.84
12 8,068.12 62.87 642.8 4.99
16 12,935.8 83.94 817.81 6.02
20 15,900.02 107.49 965.97 7.01
24 18,357.36 117.62 1,140.81 8.34
28 18,351.52 122.37 2,192.07 11.52
32 16,714.45 110.93 6,116.77 13.62
40 13,371.91 101.33 17,850.93 18.34
48 8,537.44 61.62 56,083.03 27.41
Eval Score ↑ Samples
gsm8k/em_strict · c48 97.19% 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-cu13-20260901-07c8f729 更新至 lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae。
**基线:**2026-10-12 · lmsysorg/sglang:nightly-dev-cu13-20260901-07c8f729
AgentX · TP8/EP1 · 平均延迟 · 来源: API 1, API 2, API 3;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。

Pin lmsysorg/sglang:v0.5.22-cu130 by digest and pass the CUDA graph
batch as cuda-graph-max-bs-decode, since v0.5.22 removed the deprecated
cuda-graph-max-bs alias.

将 GLM-5.2 B300 FP4 SGLang AgentX MTP 镜像按 digest 固定为
lmsysorg/sglang:v0.5.22-cu130,并改用 cuda-graph-max-bs-decode 传递
CUDA graph 批大小,因为 v0.5.22 移除了已弃用的 cuda-graph-max-bs 别名。

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 19, 2026 •

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Running · Run 39305481768 / attempt 1 · 2026-10-19 19:49 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319
Change: Measure PR commit 74b52b2: pin SGLang v0.5.22-cu130 (sgl-project/sglang baa9befa, a descendant of old nightly 07c8f729; FlashInfer 0.6.18 to 0.7.1, sgl-kernel 0.4.6.post1 to 0.4.9, CUDA 13.0.3 unchanged) by digest and rename cuda-graph-max-bs to cuda-graph-max-bs-decode in all eleven overrides because v0.5.22 removed the deprecated alias that 07c8f729 mapped to the same dest, while bf16-gemm-backend cutedsl, hicache, EAGLE, kv-cache-dtype, glm47/glm45 parsers and SGLANG_SIMULATE_ACC_* keep matching choices in arg_groups/fields and environ.py.

Next: All 11 smoke points and the c48 GSM8K eval have been running on b300-dsxe for about an hour with no failures; wait for completion.

中文

初次尝试 · 运行中 · Run 39305481768 / attempt 1 · 2026-10-19 19:49 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319
**变更:*测量 PR 提交 74b52b2:按 digest 固定 SGLang v0.5.22-cu130(sgl-project/sglang baa9befa,是旧 nightly 07c8f729 的后续提交;FlashInfer 0.6.18 升至 0.7.1,sgl-kernel 0.4.6.post1 升至 0.4.9,CUDA 13.0.3 不变),并在全部十一个 override 中将 cuda-graph-max-bs 改为 cuda-graph-max-bs-decode,因为 v0.5.22 移除了 07c8f729 中映射到同一目标的弃用别名;bf16-gemm-backend cutedsl、hicache、EAGLE、kv-cache-dtype、glm47/glm45 解析器和 SGLANG_SIMULATE_ACC_ 在 arg_groups/fields 与 environ.py 中保持一致。实测数值及异常说明见上表。
**下一步:**全部 11 个 smoke 点和 c48 GSM8K 评测已在 b300-dsxe 上运行约一小时且无失败;等待完成。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Queued · Run 39305481768 / attempt 2 · 2026-10-19 20:44 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319
Change: Measure PR commit 74b52b2: pin SGLang v0.5.22-cu130 (sgl-project/sglang baa9befa, a descendant of old nightly 07c8f729; FlashInfer 0.6.18 to 0.7.1, sgl-kernel 0.4.6.post1 to 0.4.9, CUDA 13.0.3 unchanged) by digest and rename cuda-graph-max-bs to cuda-graph-max-bs-decode in all eleven overrides because v0.5.22 removed the deprecated alias that 07c8f729 mapped to the same dest, while bf16-gemm-backend cutedsl, hicache, EAGLE, kv-cache-dtype, glm47/glm45 parsers and SGLANG_SIMULATE_ACC_* keep matching choices in arg_groups/fields and environ.py.

Next: Nine smoke points passed; the scheduler preempted c8, c48 and the c48 GSM8K eval and requeued them as run attempt 2, so wait for that rerun.

中文

初次尝试 · 排队中 · Run 39305481768 / attempt 2 · 2026-10-19 20:44 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319
**变更:*测量 PR 提交 74b52b2:按 digest 固定 SGLang v0.5.22-cu130(sgl-project/sglang baa9befa,是旧 nightly 07c8f729 的后续提交;FlashInfer 0.6.18 升至 0.7.1,sgl-kernel 0.4.6.post1 升至 0.4.9,CUDA 13.0.3 不变),并在全部十一个 override 中将 cuda-graph-max-bs 改为 cuda-graph-max-bs-decode,因为 v0.5.22 移除了 07c8f729 中映射到同一目标的弃用别名;bf16-gemm-backend cutedsl、hicache、EAGLE、kv-cache-dtype、glm47/glm45 解析器和 SGLANG_SIMULATE_ACC_ 在 arg_groups/fields 与 environ.py 中保持一致。实测数值及异常说明见上表。
**下一步:**九个 smoke 点已通过;调度器抢占了 c8、c48 和 c48 GSM8K 评测,并以运行第 2 次尝试重新排队,等待该重跑完成。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Queued · Run 39305481768 / attempt 3 · 2026-10-19 21:49 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319
Change: Measure PR commit 74b52b2: pin SGLang v0.5.22-cu130 (sgl-project/sglang baa9befa, a descendant of old nightly 07c8f729; FlashInfer 0.6.18 to 0.7.1, sgl-kernel 0.4.6.post1 to 0.4.9, CUDA 13.0.3 unchanged) by digest and rename cuda-graph-max-bs to cuda-graph-max-bs-decode in all eleven overrides because v0.5.22 removed the deprecated alias that 07c8f729 mapped to the same dest, while bf16-gemm-backend cutedsl, hicache, EAGLE, kv-cache-dtype, glm47/glm45 parsers and SGLANG_SIMULATE_ACC_* keep matching choices in arg_groups/fields and environ.py.

Next: Nine smoke points passed; the scheduler preempted c8, c48 and the c48 GSM8K eval twice and requeued them as run attempt 3, so wait for that rerun.

中文

初次尝试 · 排队中 · Run 39305481768 / attempt 3 · 2026-10-19 21:49 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319
**变更:*测量 PR 提交 74b52b2:按 digest 固定 SGLang v0.5.22-cu130(sgl-project/sglang baa9befa,是旧 nightly 07c8f729 的后续提交;FlashInfer 0.6.18 升至 0.7.1,sgl-kernel 0.4.6.post1 升至 0.4.9,CUDA 13.0.3 不变),并在全部十一个 override 中将 cuda-graph-max-bs 改为 cuda-graph-max-bs-decode,因为 v0.5.22 移除了 07c8f729 中映射到同一目标的弃用别名;bf16-gemm-backend cutedsl、hicache、EAGLE、kv-cache-dtype、glm47/glm45 解析器和 SGLANG_SIMULATE_ACC_ 在 arg_groups/fields 与 environ.py 中保持一致。实测数值及异常说明见上表。
**下一步:**九个 smoke 点已通过;调度器两次抢占了 c8、c48 和 c48 GSM8K 评测,并以运行第 3 次尝试重新排队,等待该重跑完成。

@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

Klaud-Cold commented Oct 20, 2026 •

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Running · Run 39305481768 / attempt 4 · 2026-10-20 08:22 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319 · AgentX · TP8/EP1 · Mean latency
Change: Measure PR commit 74b52b2: pin SGLang v0.5.22-cu130 (sgl-project/sglang baa9befa, a descendant of old nightly 07c8f729; FlashInfer 0.6.18 to 0.7.1, sgl-kernel 0.4.6.post1 to 0.4.9, CUDA 13.0.3 unchanged) by digest and rename cuda-graph-max-bs to cuda-graph-max-bs-decode in all eleven overrides because v0.5.22 removed the deprecated alias that 07c8f729 mapped to the same dest, while bf16-gemm-backend cutedsl, hicache, EAGLE, kv-cache-dtype, glm47/glm45 parsers and SGLANG_SIMULATE_ACC_* keep matching choices in arg_groups/fields and environ.py.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 18.95 (N/A) 468.55 (N/A) 3.01 (N/A)
4 29.19 (N/A) 353.8 (N/A) 3.32 (N/A)
8 52.57 (N/A) 353.38 (N/A) 3.76 (N/A)
12 63.03 (N/A) 411.07 (N/A) 5.38 (N/A)
16 84.69 (N/A) 492.89 (N/A) 6.18 (N/A)
20 109.38 (N/A) 580.19 (N/A) 7.76 (N/A)
24 121.59 (N/A) 680.1 (N/A) 8.87 (N/A)
28 133.06 (N/A) 945.81 (N/A) 10.65 (N/A)
32 123.2 (N/A) 3,552.33 (N/A) 13.62 (N/A)
40 125.74 (N/A) 10,552.42 (N/A) 18.02 (N/A)

Note: All rows: Δ N/A: dataset/scenario/statistic mismatch.
Note: c12: 3 request errors.
Note: c20: 1 request error.
Note: c32, c40: 2 request errors.

Eval Score ↑ Samples
gsm8k/em_strict · c48 96.82% (-0.38 pp) 1,319 each

Next: Ten smoke points and the c48 GSM8K eval passed; the c48 throughput rerun in run attempt 4 started on b300-dsxe after about seven and a half hours queued, so wait for it.

中文

初次尝试 · 运行中 · Run 39305481768 / attempt 4 · 2026-10-20 08:22 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319 · AgentX · TP8/EP1 · 平均延迟
**变更:*测量 PR 提交 74b52b2:按 digest 固定 SGLang v0.5.22-cu130(sgl-project/sglang baa9befa,是旧 nightly 07c8f729 的后续提交;FlashInfer 0.6.18 升至 0.7.1,sgl-kernel 0.4.6.post1 升至 0.4.9,CUDA 13.0.3 不变),并在全部十一个 override 中将 cuda-graph-max-bs 改为 cuda-graph-max-bs-decode,因为 v0.5.22 移除了 07c8f729 中映射到同一目标的弃用别名;bf16-gemm-backend cutedsl、hicache、EAGLE、kv-cache-dtype、glm47/glm45 解析器和 SGLANG_SIMULATE_ACC_ 在 arg_groups/fields 与 environ.py 中保持一致。实测数值及异常说明见上表。
**下一步:**十个 smoke 点和 c48 GSM8K 评测已通过;运行第 4 次尝试中的 c48 吞吐量重跑在排队约七个半小时后已在 b300-dsxe 上开始运行,等待其完成。

@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。

@Klaud-Cold

Klaud-Cold commented Oct 20, 2026 •

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Passed · Run 39305481768 / attempt 5 · 2026-10-20 12:30 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319 · AgentX · TP8/EP1 · Mean latency
Change: Measure PR commit 74b52b2: pin SGLang v0.5.22-cu130 (sgl-project/sglang baa9befa, a descendant of old nightly 07c8f729; FlashInfer 0.6.18 to 0.7.1, sgl-kernel 0.4.6.post1 to 0.4.9, CUDA 13.0.3 unchanged) by digest and rename cuda-graph-max-bs to cuda-graph-max-bs-decode in all eleven overrides because v0.5.22 removed the deprecated alias that 07c8f729 mapped to the same dest, while bf16-gemm-backend cutedsl, hicache, EAGLE, kv-cache-dtype, glm47/glm45 parsers and SGLANG_SIMULATE_ACC_* keep matching choices in arg_groups/fields and environ.py.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 18.95 (N/A) 468.55 (N/A) 3.01 (N/A)
4 29.19 (N/A) 353.8 (N/A) 3.32 (N/A)
8 52.57 (N/A) 353.38 (N/A) 3.76 (N/A)
12 63.03 (N/A) 411.07 (N/A) 5.38 (N/A)
16 84.69 (N/A) 492.89 (N/A) 6.18 (N/A)
20 109.38 (N/A) 580.19 (N/A) 7.76 (N/A)
24 121.59 (N/A) 680.1 (N/A) 8.87 (N/A)
28 133.06 (N/A) 945.81 (N/A) 10.65 (N/A)
32 123.2 (N/A) 3,552.33 (N/A) 13.62 (N/A)
40 125.74 (N/A) 10,552.42 (N/A) 18.02 (N/A)
48 84.11 (N/A) 31,622.52 (N/A) 27.69 (N/A)

Note: All rows: Δ N/A: dataset/scenario/statistic mismatch.
Note: c12: 3 request errors.
Note: c20: 1 request error.
Note: c32, c40, c48: 2 request errors.

Eval Score ↑ Samples
gsm8k/em_strict · c48 96.82% (-0.38 pp) 1,319 each

Next: All 11 smoke points and the c48 GSM8K eval passed on b300-dsxe; append the perf-changelog entry and dispatch the full-sweep-fail-fast final sweep.

中文

初次尝试 · 已通过 · Run 39305481768 / attempt 5 · 2026-10-20 12:30 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 0e5b07c78319 · AgentX · TP8/EP1 · 平均延迟
**变更:*测量 PR 提交 74b52b2:按 digest 固定 SGLang v0.5.22-cu130(sgl-project/sglang baa9befa,是旧 nightly 07c8f729 的后续提交;FlashInfer 0.6.18 升至 0.7.1,sgl-kernel 0.4.6.post1 升至 0.4.9,CUDA 13.0.3 不变),并在全部十一个 override 中将 cuda-graph-max-bs 改为 cuda-graph-max-bs-decode,因为 v0.5.22 移除了 07c8f729 中映射到同一目标的弃用别名;bf16-gemm-backend cutedsl、hicache、EAGLE、kv-cache-dtype、glm47/glm45 解析器和 SGLANG_SIMULATE_ACC_ 在 arg_groups/fields 与 environ.py 中保持一致。实测数值及异常说明见上表。
**下一步:**全部 11 个 smoke 点和 c48 GSM8K 评测已在 b300-dsxe 上通过;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

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 Bot and others added 2 commits October 20, 2026 12:30
…P v0.5.22-cu130

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…1a0f3e-9e6710099ed6010e

# Conflicts:
#	inferencex-e2e/perf-changelog.yaml
@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 20, 2026
@Klaud-Cold

Klaud-Cold commented Oct 20, 2026 •

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Running · Run 39423009783 / attempt 1 · 2026-10-20 13:03 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 8b715ccf448e
Change: Append the perf-changelog entry, merge current main so the entry is a pure tail append, and run the full-sweep-fail-fast final sweep with the same image and recipe as the passed smoke attempt.

Next: The c48 GSM8K eval passed; all 11 throughput points have been running on b300-dsxe for about 40 minutes, so wait for them.

中文

最终完整 sweep · 运行中 · Run 39423009783 / attempt 1 · 2026-10-20 13:03 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 8b715ccf448e
**变更:**追加 perf-changelog 条目,合并当前 main 使该条目成为纯尾部追加,并以与已通过 smoke 尝试相同的镜像和配方运行 full-sweep-fail-fast 最终扫描。实测数值及异常说明见上表。
**下一步:**c48 GSM8K 评测已通过;全部 11 个吞吐量点已在 b300-dsxe 上运行约 40 分钟,等待其完成。

@github-actions

Copy link
Copy Markdown
Contributor

@Klaud-Cold

Klaud-Cold commented Oct 20, 2026 •

Copy link
Copy Markdown
Collaborator Author

Infrastructure retry 1 · Cancelled · Run 39423009783 / attempt 2 · 2026-10-20 14:34 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 8b715ccf448e
Change: Rerun the failed and fail-fast-cancelled final-sweep jobs on the same head; c20 finished its benchmark with 0/3007 request errors but GitHub rejected its server-log artifact with FinalizeArtifact 404, a transient artifact-service failure, and fail-fast cancelled c16, c24, c28, c32, c40 and c48.

Next: Stopped because a maintainer closed this PR at 14:32 UTC, which cancelled the queued rerun of the seven remaining final-sweep points; Klaud does not reopen it.

中文

基础设施重试 1 · 已取消 · Run 39423009783 / attempt 2 · 2026-10-20 14:34 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 8b715ccf448e
**变更:**在同一提交上重跑失败及被 fail-fast 取消的最终扫描任务;c20 基准测试已完成且请求错误为 0/3007,但 GitHub 以 FinalizeArtifact 404 拒绝了其服务器日志 artifact,属于临时性 artifact 服务故障,随后 fail-fast 取消了 c16、c24、c28、c32、c40 和 c48。实测数值及异常说明见上表。
**下一步:**已停止:维护者于 14:32 UTC 关闭了此 PR,导致排队中的其余七个最终扫描点重跑被取消;Klaud 不会重新打开它。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

aborted · Cleanup pending. Stop and confirm owned runs before closing.

中文

aborted · 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 removed the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Oct 20, 2026
@Klaud-Cold
Klaud-Cold deleted the klaud/auto-50375757c31a0f3e-9e6710099ed6010e branch October 20, 2026 14:35
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

aborted · Repairs: 0 · Runs: 39305481768, 39422947988, 39423009783
Reason: Stopped by the klaud-abort label.
All owned runs ended. PR closed; branch deleted for retry.

中文

aborted · 修复次数:0 · 运行:39305481768, 39422947988, 39423009783
**原因:**已由 klaud-abort 标签停止。
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.

2 participants