iframe-proxy

Klaud-Cold · 2026-10-21T06:50:52Z

Goal: Update SGLang image from lmsysorg/sglang:nightly-dev-cu13-20261014-684fea8a to lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4.
Baseline: 2026-10-16 · lmsysorg/sglang:nightly-dev-cu13-20261014-684fea8a@sha256:ddae0e399d3d94f7827f16183301cf8dac4cf0eae0ef25590b7dac5881d713d8
Mean latency · Sources: API 1, API 2

Eval: N/A

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-cu13-20261014-684fea8a 更新至 lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4。
**基线：**2026-10-16 · lmsysorg/sglang:nightly-dev-cu13-20261014-684fea8a@sha256:ddae0e399d3d94f7827f16183301cf8dac4cf0eae0ef25590b7dac5881d713d8
平均延迟 · 来源： API 1, API 2；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。

…21-c1c287a4 将 dsv41flash B200 SGLang AgentX 镜像更新至 nightly-dev-cu13-20261021-c1c287a4。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

github-actions · 2026-10-21T06:51:01Z

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 · 2026-10-21T06:53:08Z

Initial attempt · Running · Run 39558093929 / attempt 1 · 2026-10-21 06:53 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · Mean latency
Change: Image-only bump, measured at branch commit a147eb9, from SGLang 684fea8a to c1c287a4 (502 upstream commits, FlashInfer 0.7.2, sgl-deep-gemm 0.2.3, sglang-kernel 0.4.10, CUDA 13.0.3); every recipe flag and env var still parses, and upstream now defaults decoder SWA bounded replay on and resolves dsv4-attn-backend auto to flashmla_mega on SM100, which changes the c16/c24 points that leave the backend unset (sources: github.com/sgl-project/sglang/compare/684fea8a...c1c287a4, python/sglang/srt/arg_groups/deepseek_v4_hook.py, image label ai.sglang.build.commit=c1c287a4).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 26ea6c	N/A	N/A	N/A
AgentX c4 TP4 EP1 814118	N/A	N/A	N/A
AgentX c8 TP4 EP1 625e7c	N/A	N/A	N/A
AgentX c16 TP4 EP4 3bdc51	N/A	N/A	N/A
AgentX c24 TP4 EP4 cc37c9	N/A	N/A	N/A
AgentX c48 TP4 EP4 c7b307	N/A	N/A	N/A
AgentX c64 TP4 EP4 c3000e	N/A	N/A	N/A
AgentX c96 TP4 EP4 0ac2f8	N/A	N/A	N/A
AgentX c128 TP4 EP4 60dd36	N/A	N/A	N/A
AgentX c192 TP4 EP4 f5ec49	N/A	N/A	N/A
AgentX c256 TP4 EP4 088e19	N/A	N/A	N/A

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

Next: Wait for the smoke run's throughput points and default evals to finish, then inspect the first server error if any job fails.

中文

初次尝试 · 运行中 · Run 39558093929 / attempt 1 · 2026-10-21 06:53 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · 平均延迟
**变更：**仅更新镜像（在分支提交 a147eb9 上测量）：SGLang 从 684fea8a 升级到 c1c287a4（上游 502 个提交，FlashInfer 0.7.2、sgl-deep-gemm 0.2.3、sglang-kernel 0.4.10、CUDA 13.0.3）；配方中的所有参数和环境变量仍可解析，上游现在默认开启 decoder SWA bounded replay，并在 SM100 上将 dsv4-attn-backend auto 解析为 flashmla_mega，这会影响未设置该后端的 c16/c24 点（来源：github.com/sgl-project/sglang/compare/684fea8a...c1c287a4、python/sglang/srt/arg_groups/deepseek_v4_hook.py、镜像标签 ai.sglang.build.commit=c1c287a4）。实测数值及异常说明见上表。
**下一步：**等待冒烟运行的吞吐点和默认评测完成；若有任务失败，则检查第一个服务端错误。

Klaud-Cold · 2026-10-21T07:48:17Z

Initial attempt · Running · Run 39558093929 / attempt 2 · 2026-10-21 08:08 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · Mean latency
Change: Image-only bump, measured at branch commit a147eb9, from SGLang 684fea8a to c1c287a4 (502 upstream commits, FlashInfer 0.7.2, sgl-deep-gemm 0.2.3, sglang-kernel 0.4.10, CUDA 13.0.3); every recipe flag and env var still parses, and upstream now defaults decoder SWA bounded replay on and resolves dsv4-attn-backend auto to flashmla_mega on SM100, which changes the c16/c24 points that leave the backend unset (sources: github.com/sgl-project/sglang/compare/684fea8a...c1c287a4, python/sglang/srt/arg_groups/deepseek_v4_hook.py, image label ai.sglang.build.commit=c1c287a4).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 26ea6c	N/A	N/A	N/A
AgentX c4 TP4 EP1 814118	N/A	N/A	N/A
AgentX c8 TP4 EP1 625e7c	N/A	N/A	N/A
AgentX c16 TP4 EP4 3bdc51	N/A	N/A	N/A
AgentX c24 TP4 EP4 cc37c9	N/A	N/A	N/A
AgentX c48 TP4 EP4 c7b307	N/A	N/A	N/A
AgentX c64 TP4 EP4 c3000e	N/A	N/A	N/A
AgentX c96 TP4 EP4 0ac2f8	N/A	N/A	N/A
AgentX c128 TP4 EP4 60dd36	N/A	N/A	N/A
AgentX c192 TP4 EP4 f5ec49	N/A	N/A	N/A
AgentX c256 TP4 EP4 088e19	N/A	N/A	N/A

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

Eval	Score ↑	Samples
gsm8k/em_strict · c24	97.5% (N/A)	N/A/1,319 (old/new)
gsm8k/em_strict · c256	97.5% (N/A)	N/A/1,319 (old/new)

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

Next: The scheduler cancelled and requeued the 11 throughput jobs as run attempt 2 after both default evals passed; wait for that attempt without using a repair.

中文

初次尝试 · 运行中 · Run 39558093929 / attempt 2 · 2026-10-21 08:08 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · 平均延迟
**变更：**仅更新镜像（在分支提交 a147eb9 上测量）：SGLang 从 684fea8a 升级到 c1c287a4（上游 502 个提交，FlashInfer 0.7.2、sgl-deep-gemm 0.2.3、sglang-kernel 0.4.10、CUDA 13.0.3）；配方中的所有参数和环境变量仍可解析，上游现在默认开启 decoder SWA bounded replay，并在 SM100 上将 dsv4-attn-backend auto 解析为 flashmla_mega，这会影响未设置该后端的 c16/c24 点（来源：github.com/sgl-project/sglang/compare/684fea8a...c1c287a4、python/sglang/srt/arg_groups/deepseek_v4_hook.py、镜像标签 ai.sglang.build.commit=c1c287a4）。实测数值及异常说明见上表。
**下一步：**两个默认评测通过后，调度器取消了 11 个吞吐任务并以运行第 2 次尝试重新排队；等待该尝试完成，不消耗修复次数。

Klaud-Cold · 2026-10-21T08:26:36Z

Initial attempt · Running · Run 39558093929 / attempt 3 · 2026-10-21 09:24 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · Mean latency
Change: Image-only bump, measured at branch commit a147eb9, from SGLang 684fea8a to c1c287a4 (502 upstream commits, FlashInfer 0.7.2, sgl-deep-gemm 0.2.3, sglang-kernel 0.4.10, CUDA 13.0.3); every recipe flag and env var still parses, and upstream now defaults decoder SWA bounded replay on and resolves dsv4-attn-backend auto to flashmla_mega on SM100, which changes the c16/c24 points that leave the backend unset (sources: github.com/sgl-project/sglang/compare/684fea8a...c1c287a4, python/sglang/srt/arg_groups/deepseek_v4_hook.py, image label ai.sglang.build.commit=c1c287a4).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 26ea6c	N/A	N/A	N/A
AgentX c4 TP4 EP1 814118	N/A	N/A	N/A
AgentX c8 TP4 EP1 625e7c	N/A	N/A	N/A
AgentX c16 TP4 EP4 3bdc51	N/A	N/A	N/A
AgentX c24 TP4 EP4 cc37c9	N/A	N/A	N/A
AgentX c48 TP4 EP4 c7b307	N/A	N/A	N/A
AgentX c64 TP4 EP4 c3000e	N/A	N/A	N/A
AgentX c96 TP4 EP4 0ac2f8	N/A	N/A	N/A
AgentX c128 TP4 EP4 60dd36	N/A	N/A	N/A
AgentX c192 TP4 EP4 f5ec49	N/A	N/A	N/A
AgentX c256 TP4 EP4 088e19	N/A	N/A	N/A

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

Eval	Score ↑	Samples
gsm8k/em_strict · c24	97.5% (N/A)	N/A/1,319 (old/new)
gsm8k/em_strict · c256	97.5% (N/A)	N/A/1,319 (old/new)

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

Next: All 11 throughput jobs of run attempt 3 have been running since about 08:55 UTC with no failures; wait for them to finish.

中文

初次尝试 · 运行中 · Run 39558093929 / attempt 3 · 2026-10-21 09:24 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · 平均延迟
**变更：**仅更新镜像（在分支提交 a147eb9 上测量）：SGLang 从 684fea8a 升级到 c1c287a4（上游 502 个提交，FlashInfer 0.7.2、sgl-deep-gemm 0.2.3、sglang-kernel 0.4.10、CUDA 13.0.3）；配方中的所有参数和环境变量仍可解析，上游现在默认开启 decoder SWA bounded replay，并在 SM100 上将 dsv4-attn-backend auto 解析为 flashmla_mega，这会影响未设置该后端的 c16/c24 点（来源：github.com/sgl-project/sglang/compare/684fea8a...c1c287a4、python/sglang/srt/arg_groups/deepseek_v4_hook.py、镜像标签 ai.sglang.build.commit=c1c287a4）。实测数值及异常说明见上表。
**下一步：**运行第 3 次尝试的 11 个吞吐任务自约 08:55 UTC 起全部在运行，尚无失败；等待其完成。

Klaud-Cold · 2026-10-21T09:34:30Z

Initial attempt · Running · Run 39558093929 / attempt 4 · 2026-10-21 11:11 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · Mean latency
Change: Image-only bump, measured at branch commit a147eb9, from SGLang 684fea8a to c1c287a4 (502 upstream commits, FlashInfer 0.7.2, sgl-deep-gemm 0.2.3, sglang-kernel 0.4.10, CUDA 13.0.3); every recipe flag and env var still parses, and upstream now defaults decoder SWA bounded replay on and resolves dsv4-attn-backend auto to flashmla_mega on SM100, which changes the c16/c24 points that leave the backend unset (sources: github.com/sgl-project/sglang/compare/684fea8a...c1c287a4, python/sglang/srt/arg_groups/deepseek_v4_hook.py, image label ai.sglang.build.commit=c1c287a4).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 26ea6c	N/A	N/A	N/A
AgentX c4 TP4 EP1 814118	N/A	N/A	N/A
AgentX c8 TP4 EP1 625e7c	N/A	N/A	N/A
AgentX c16 TP4 EP4 3bdc51	N/A	N/A	N/A
AgentX c24 TP4 EP4 cc37c9	N/A	N/A	N/A
AgentX c48 TP4 EP4 c7b307	N/A	N/A	N/A
AgentX c64 TP4 EP4 c3000e	N/A	N/A	N/A
AgentX c96 TP4 EP4 0ac2f8	N/A	N/A	N/A
AgentX c128 TP4 EP4 60dd36	N/A	N/A	N/A
AgentX c192 TP4 EP4 f5ec49	N/A	N/A	N/A
AgentX c256 TP4 EP4 088e19	N/A	N/A	N/A

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

Eval	Score ↑	Samples
gsm8k/em_strict · c24	97.5% (N/A)	N/A/1,319 (old/new)
gsm8k/em_strict · c256	97.5% (N/A)	N/A/1,319 (old/new)

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

Next: Ten throughput jobs of run attempt 4 have been running since about 09:50 UTC with one still queued and no failures; wait for them to finish.

中文

初次尝试 · 运行中 · Run 39558093929 / attempt 4 · 2026-10-21 11:11 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · 平均延迟
**变更：**仅更新镜像（在分支提交 a147eb9 上测量）：SGLang 从 684fea8a 升级到 c1c287a4（上游 502 个提交，FlashInfer 0.7.2、sgl-deep-gemm 0.2.3、sglang-kernel 0.4.10、CUDA 13.0.3）；配方中的所有参数和环境变量仍可解析，上游现在默认开启 decoder SWA bounded replay，并在 SM100 上将 dsv4-attn-backend auto 解析为 flashmla_mega，这会影响未设置该后端的 c16/c24 点（来源：github.com/sgl-project/sglang/compare/684fea8a...c1c287a4、python/sglang/srt/arg_groups/deepseek_v4_hook.py、镜像标签 ai.sglang.build.commit=c1c287a4）。实测数值及异常说明见上表。
**下一步：**运行第 4 次尝试中有 10 个吞吐任务自约 09:50 UTC 起在运行，1 个仍在排队，尚无失败；等待其完成。

Klaud-Cold · 2026-10-21T11:25:39Z

Initial attempt · Passed · Run 39558093929 / attempt 5 · 2026-10-21 13:12 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · Mean latency
Change: Image-only bump, measured at branch commit a147eb9, from SGLang 684fea8a to c1c287a4 (502 upstream commits, FlashInfer 0.7.2, sgl-deep-gemm 0.2.3, sglang-kernel 0.4.10, CUDA 13.0.3); every recipe flag and env var still parses, and upstream now defaults decoder SWA bounded replay on and resolves dsv4-attn-backend auto to flashmla_mega on SM100, which changes the c16/c24 points that leave the backend unset (sources: github.com/sgl-project/sglang/compare/684fea8a...c1c287a4, python/sglang/srt/arg_groups/deepseek_v4_hook.py, image label ai.sglang.build.commit=c1c287a4).

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 26ea6c	47.45 (N/A)	244.45 (N/A)	1.46 (N/A)
AgentX c4 TP4 EP1 814118	67.73 (N/A)	454.32 (N/A)	1.59 (N/A)
AgentX c8 TP4 EP1 625e7c	116.38 (N/A)	644.77 (N/A)	1.68 (N/A)
AgentX c16 TP4 EP4 3bdc51	173.79 (N/A)	2,502.11 (N/A)	2.18 (N/A)
AgentX c24 TP4 EP4 cc37c9	263.03 (N/A)	2,834.78 (N/A)	2.44 (N/A)
AgentX c48 TP4 EP4 c7b307	568.74 (N/A)	1,466.64 (N/A)	3.22 (N/A)
AgentX c64 TP4 EP4 c3000e	832.09 (N/A)	1,382.44 (N/A)	3.71 (N/A)
AgentX c96 TP4 EP4 0ac2f8	1,165.55 (N/A)	1,251.63 (N/A)	4.39 (N/A)
AgentX c128 TP4 EP4 60dd36	1,453.8 (N/A)	1,244.22 (N/A)	5.09 (N/A)
AgentX c192 TP4 EP4 f5ec49	1,966.29 (N/A)	1,587.66 (N/A)	6.23 (N/A)
AgentX c256 TP4 EP4 088e19	2,460.21 (N/A)	1,793.14 (N/A)	8.74 (N/A)

Note: All rows: Δ N/A: dataset/scenario/statistic mismatch.

Eval	Score ↑	Samples
gsm8k/em_strict · c24	97.5% (N/A)	N/A/1,319 (old/new)
gsm8k/em_strict · c256	97.5% (N/A)	N/A/1,319 (old/new)

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

Next: All 11 smoke throughput points and both default evals passed on the updated image; append the perf-changelog entry and dispatch the full fail-fast sweep.

中文

初次尝试 · 已通过 · Run 39558093929 / attempt 5 · 2026-10-21 13:12 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · 6046414d2d1a · 平均延迟
**变更：**仅更新镜像（在分支提交 a147eb9 上测量）：SGLang 从 684fea8a 升级到 c1c287a4（上游 502 个提交，FlashInfer 0.7.2、sgl-deep-gemm 0.2.3、sglang-kernel 0.4.10、CUDA 13.0.3）；配方中的所有参数和环境变量仍可解析，上游现在默认开启 decoder SWA bounded replay，并在 SM100 上将 dsv4-attn-backend auto 解析为 flashmla_mega，这会影响未设置该后端的 c16/c24 点（来源：github.com/sgl-project/sglang/compare/684fea8a...c1c287a4、python/sglang/srt/arg_groups/deepseek_v4_hook.py、镜像标签 ai.sglang.build.commit=c1c287a4）。实测数值及异常说明见上表。
**下一步：**更新镜像上的全部 11 个冒烟吞吐点和两个默认评测均已通过；接下来追加 perf-changelog 条目并启动完整的 fail-fast 扫描。

Klaud-Cold · 2026-10-21T12:25:33Z

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。

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

Klaud-Cold · 2026-10-21T13:15:26Z

Final full sweep · Queued · Run 39606211010 / attempt 1 · 2026-10-21 13:54 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · c1457ce06a76 · Mean latency
Change: No recipe change since the passing smoke run; appended the perf-changelog entry at c1457ce and applied full-sweep-fail-fast for the full sweep on the same updated image.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 26ea6c	N/A	N/A	N/A
AgentX c4 TP4 EP1 814118	N/A	N/A	N/A
AgentX c8 TP4 EP1 625e7c	N/A	N/A	N/A
AgentX c16 TP4 EP4 3bdc51	N/A	N/A	N/A
AgentX c24 TP4 EP4 cc37c9	N/A	N/A	N/A
AgentX c48 TP4 EP4 c7b307	N/A	N/A	N/A
AgentX c64 TP4 EP4 c3000e	N/A	N/A	N/A
AgentX c96 TP4 EP4 0ac2f8	N/A	N/A	N/A
AgentX c128 TP4 EP4 60dd36	N/A	N/A	N/A
AgentX c192 TP4 EP4 f5ec49	N/A	N/A	N/A
AgentX c256 TP4 EP4 088e19	N/A	N/A	N/A

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

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

Next: All 11 throughput jobs and both eval jobs of the full sweep have been queued since 13:14 UTC behind higher-priority work; keep waiting without cancelling.

中文

最终完整 sweep · 排队中 · Run 39606211010 / attempt 1 · 2026-10-21 13:54 UTC
lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb · c1457ce06a76 · 平均延迟
**变更：**自冒烟运行通过以来配方未变；在 c1457ce 追加了 perf-changelog 条目，并为同一更新镜像的完整扫描添加 full-sweep-fail-fast 标签。实测数值及异常说明见上表。
**下一步：**完整扫描的全部 11 个吞吐任务和两个评测任务自 13:14 UTC 起一直在更高优先级任务之后排队；继续等待，不取消。

Update dsv41flash B200 SGLang AgentX image to nightly-dev-cu13-202610…

a147eb9

…21-c1c287a4 将 dsv41flash B200 SGLang AgentX 镜像更新至 nightly-dev-cu13-20261021-c1c287a4。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

github-project-automation Bot added this to InferenceX Board Oct 21, 2026

Add perf-changelog entry for dsv41flash B200 SGLang AgentX image update

c1457ce

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

Klaud-Cold added the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Oct 21, 2026

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP4 EP1 26ea6c	7,858.15	47.19	348.33	1.49
AgentX c4 TP4 EP1 814118	9,319.36	64.61	659.28	1.61
AgentX c8 TP4 EP1 625e7c	16,454.21	114.81	760.14	1.74
AgentX c16 TP4 EP4 3bdc51	27,738.49	174.51	2,413.55	1.98
AgentX c24 TP4 EP4 cc37c9	43,306.03	261.94	2,965.78	2.36
AgentX c48 TP4 EP4 c7b307	84,650.5	555.28	1,806.09	3.55
AgentX c64 TP4 EP4 c3000e	115,368.89	803.64	1,771.86	4.26
AgentX c96 TP4 EP4 0ac2f8	147,277.54	1,107	1,725.87	5.35
AgentX c128 TP4 EP4 60dd36	173,310.86	1,359.03	1,729.95	6.43
AgentX c192 TP4 EP4 f5ec49	254,010.81	1,823.3	2,316.09	7.98
AgentX c256 TP4 EP4 088e19	305,125.74	2,165.79	2,797.61	11.44

Sunbelt Computer Software

PL/B Language Development and Support

Navigation Menu

Repository navigation

[Klaud Cold] Update DeepSeek-V4.1-Flash FP4 B200 SGLang AgentX image to nightly-dev-cu13-20261021-c1c287a4 / [Klaud Cold] 将 DeepSeek-V4.1-Flash FP4 B200 SGLang AgentX 镜像更新至 nightly-dev-cu13-20261021-c1c287a4 - #4365

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 21, 2026

Uh oh!

Klaud-Cold commented Oct 21, 2026

Uh oh!

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Uh oh!

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Uh oh!

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Uh oh!

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Uh oh!

Klaud-Cold commented Oct 21, 2026

Uh oh!

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Sunbelt Computer Software

PL/B Language Development and Support

Conversation

Klaud-Cold commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Oct 21, 2026

Uh oh!

Klaud-Cold commented Oct 21, 2026

Uh oh!

Klaud-Cold commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 21, 2026

Uh oh!

Klaud-Cold commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Klaud-Cold commented Oct 21, 2026 •

edited

Loading

Klaud-Cold commented Oct 21, 2026 •

edited

Loading