[Klaud Cold] Update Qwen3.5 FP4 GB200 Dynamo-SGLang AgentX MTP image to v0.5.22-cu130 / [Klaud Cold] 将 Qwen3.5 FP4 GB200 Dynamo-SGLang AgentX MTP 镜像更新至 v0.5.22-cu130 by Klaud-Cold · Pull Request #4278 · SemiAnalysisAI/InferenceX

Klaud-Cold · 2026-10-18T12:47:43Z

Goal: Update SGLang image from lmsysorg/sglang:nightly-dev-20260818-c0b6474b to lmsysorg/sglang:v0.5.22-cu130.
Baseline: 2026-09-11 · lmsysorg/sglang:nightly-dev-20260818-c0b6474b
Mean latency · Sources: API 1, API 2, API 3

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 PTP4x1/DTP4x0 dd714f	5,484.03	55.53	209.24	1.81
AgentX c2 PTP4x1/DTP4x0 f99308	5,791.67	61.28	204.99	1.84
AgentX c4 PTP4x1/DTP4x0 e25615	7,231.56	76.95	210.13	1.97
AgentX c8 PTP4x1/DTP4x0 adb59a	11,969.19	121.42	229.1	2.19
AgentX c12 PTP4x1/DTP4x0 2a9acc	20,649.69	179.87	263.53	2.54
AgentX c20 PTP4x1/DTP4x0 36ba59	30,438.01	252.95	290.4	3.15
AgentX c22 PTP2x1/DTP2x0 ddd472	N/A	N/A	N/A	N/A
AgentX c24 PTP4x1/DTP4x0 95bb75	35,547.87	299.67	296.48	3.5
AgentX c32 PTP2x1/DTP2x0 474621	N/A	N/A	N/A	N/A
AgentX c40 PTP2x1/DTP2x0 2aa262	N/A	N/A	N/A	N/A
AgentX c48 PTP2x1/DTP2x0 d5567a	N/A	N/A	N/A	N/A

Note: AgentX c22 PTP2x1/DTP2x0 ddd472, AgentX c32 PTP2x1/DTP2x0 474621, AgentX c40 PTP2x1/DTP2x0 2aa262, AgentX c48 PTP2x1/DTP2x0 d5567a: unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c20	97.42%	1,319
gsm8k/em_strict · c22	97.5%	1,319
gsm8k/em_strict · c24	97.73%	1,319
gsm8k/em_strict · c32	97.57%	1,319
gsm8k/em_strict · c48	97.73%	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-20260818-c0b6474b 更新至 lmsysorg/sglang:v0.5.22-cu130。
**基线：**2026-09-11 · lmsysorg/sglang:nightly-dev-20260818-c0b6474b
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. v0.5.22 removed the --cuda-graph-max-bs and --mamba-scheduler-strategy aliases, so use --cuda-graph-max-bs-decode and --mamba-radix-cache-strategy. Dynamo 1.4.2 imports sglang.srt.server_args_config_parser, which moved to sglang.srt.utils; Dynamo 1.5.1 resolves either location. 将 GB200 FP4 Dynamo-SGLang AgentX 更新到 SGLang v0.5.22，并按摘要固定 lmsysorg/sglang:v0.5.22-cu130。v0.5.22 移除了 --cuda-graph-max-bs 和 --mamba-scheduler-strategy 别名，因此改用 --cuda-graph-max-bs-decode 和 --mamba-radix-cache-strategy。Dynamo 1.4.2 导入的 sglang.srt.server_args_config_parser 已移至 sglang.srt.utils；Dynamo 1.5.1 可兼容两个位置。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

github-actions · 2026-10-18T12:47:50Z

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 · 2026-10-18T12:49:59Z

Initial attempt · Failed · Run 39128853814 / attempt 1 · 2026-10-18 12:58 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 9a132463cf69
Change: Measure PR head 57ef027: pin SGLang v0.5.22-cu130 (commit baa9befa, image labels verified), rename the removed --cuda-graph-max-bs and --mamba-scheduler-strategy aliases to --cuda-graph-max-bs-decode and --mamba-radix-cache-strategy, and bump the Dynamo wheel from 1.4.2 to 1.5.1 because 1.4.2 imports sglang.srt.server_args_config_parser, which v0.5.22 moved to sglang.srt.utils.

Next: Repair 1 pins the Dynamo wheel to 1.6.0.dev20261015, the version bundled in the v0.5.22 image.

中文

初次尝试 · 失败 · Run 39128853814 / attempt 1 · 2026-10-18 12:58 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 9a132463cf69
**变更：**测量 PR 提交 57ef027：固定 SGLang v0.5.22-cu130（提交 baa9befa，已核对镜像标签），将已移除的 --cuda-graph-max-bs 和 --mamba-scheduler-strategy 别名改为 --cuda-graph-max-bs-decode 和 --mamba-radix-cache-strategy，并将 Dynamo wheel 从 1.4.2 升级到 1.5.1，因为 1.4.2 导入的 sglang.srt.server_args_config_parser 在 v0.5.22 中已移至 sglang.srt.utils。实测数值及异常说明见上表。
**下一步：**修复 1 将 Dynamo wheel 固定为 v0.5.22 镜像自带的 1.6.0.dev20261015。

Dynamo 1.5.1 calls ServerArgs.get_model_config(), which SGLang v0.5.22 removed. Pin 1.6.0.dev20261015, the version shipped in the v0.5.22 image. Dynamo 1.5.1 调用了 SGLang v0.5.22 已移除的 ServerArgs.get_model_config()。改为固定 v0.5.22 镜像自带的 1.6.0.dev20261015。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Klaud-Cold · 2026-10-18T12:58:45Z

Repair 1/5 · Failed · Run 39129509704 / attempt 1 · 2026-10-18 13:18 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 9a132463cf69
Change: Measure PR head 65a89ac: pin the Dynamo wheel to 1.6.0.dev20261015, the version bundled in the v0.5.22 image, whose _compat.py resolves the moved config parser and the removed ServerArgs.get_model_config().

Next: Repair 2 sets linear-replayssm-cache-len to 16, the v0.5.22 default and the minimum the commit kernel compiles with, in the four ReplaySSM variants.

中文

修复 1/5 · 失败 · Run 39129509704 / attempt 1 · 2026-10-18 13:18 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 9a132463cf69
**变更：**测量 PR 提交 65a89ac：将 Dynamo wheel 固定为 v0.5.22 镜像自带的 1.6.0.dev20261015，其 _compat.py 可处理已迁移的配置解析器和已移除的 ServerArgs.get_model_config()。实测数值及异常说明见上表。
**下一步：**修复 2 在四个 ReplaySSM 变体中将 linear-replayssm-cache-len 设为 16，即 v0.5.22 默认值，也是提交内核可编译的最小值。

SGLang v0.5.22 sizes the GDN ReplaySSM ring by linear-replayssm-cache-len, and its commit kernel's tl.dot needs an inner dimension of at least 16. Use 16, the v0.5.22 default. SGLang v0.5.22 按 linear-replayssm-cache-len 分配 GDN ReplaySSM 环形缓冲区，其提交内核中的 tl.dot 要求内维至少为 16。改用 v0.5.22 默认值 16。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Klaud-Cold · 2026-10-18T13:19:14Z

Repair 2/5 · Passed · Run 39130985168 / attempt 1 · 2026-10-18 14:42 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 9a132463cf69 · Mean latency
Change: Measure PR head 97df4cb: set linear-replayssm-cache-len from 8 to 16 in the four ReplaySSM variants, because v0.5.22 sizes the GDN ReplaySSM ring by this value and gdn_replayssm_compact_commit_kernel only compiles with a ring of at least 16.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 PTP4x1/DTP4x0 dd714f	56.42 (N/A)	265.77 (N/A)	1.78 (N/A)
AgentX c2 PTP4x1/DTP4x0 f99308	62.24 (N/A)	236.36 (N/A)	1.79 (N/A)
AgentX c22 PTP2x1/DTP2x0 ddd472	527.42 (N/A)	406.79 (N/A)	5.68 (N/A)
AgentX c24 PTP4x1/DTP4x0 95bb75	299.81 (N/A)	320.14 (N/A)	3.45 (N/A)
AgentX c32 PTP2x1/DTP2x0 474621	676.6 (N/A)	506.63 (N/A)	8.09 (N/A)
AgentX c40 PTP2x1/DTP2x0 2aa262	823.99 (N/A)	546.77 (N/A)	9.4 (N/A)

Note: AgentX c1 PTP4x1/DTP4x0 dd714f, AgentX c2 PTP4x1/DTP4x0 f99308, AgentX c24 PTP4x1/DTP4x0 95bb75: Δ N/A: dataset/scenario/statistic mismatch.
Note: AgentX c22 PTP2x1/DTP2x0 ddd472, AgentX c32 PTP2x1/DTP2x0 474621, AgentX c40 PTP2x1/DTP2x0 2aa262: Δ N/A: point unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c20	97.19% (-0.23 pp)	1,319 each
gsm8k/em_strict · c22	97.73% (+0.23 pp)	1,319 each
gsm8k/em_strict · c24	97.42% (-0.30 pp)	1,319 each
gsm8k/em_strict · c32	97.5% (-0.08 pp)	1,319 each
gsm8k/em_strict · c48	97.5% (-0.23 pp)	1,319 each

Next: Smoke benchmarks and evals passed, so append the perf-changelog entry and run the full-sweep-fail-fast final sweep.

中文

修复 2/5 · 已通过 · Run 39130985168 / attempt 1 · 2026-10-18 14:42 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 9a132463cf69 · 平均延迟
**变更：**测量 PR 提交 97df4cb：在四个 ReplaySSM 变体中将 linear-replayssm-cache-len 从 8 改为 16，因为 v0.5.22 按该值分配 GDN ReplaySSM 环形缓冲区，而 gdn_replayssm_compact_commit_kernel 仅在长度至少为 16 时才能编译。实测数值及异常说明见上表。
**下一步：**冒烟基准测试与评测均已通过，因此追加 perf-changelog 条目并运行 full-sweep-fail-fast 最终扫描。

新增 Qwen3.5 GB200 SGLang v0.5.22 的 perf-changelog 条目。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Klaud-Cold · 2026-10-18T14:43:46Z

Final full sweep · Queued · Run 39137200943 / attempt 1 · 2026-10-18 14:43 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 4d8bf0112f4a
Change: Add the perf-changelog entry and run the full-sweep-fail-fast final sweep over all 11 points and 5 default evals at this head.

Next: Wait for complete final-sweep coverage, then verify artifacts against the frozen baseline.

中文

最终完整 sweep · 排队中 · Run 39137200943 / attempt 1 · 2026-10-18 14:43 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 4d8bf0112f4a
**变更：**添加 perf-changelog 条目，并在此提交上对全部 11 个点和 5 个默认评测运行 full-sweep-fail-fast 最终扫描。实测数值及异常说明见上表。
**下一步：**等待最终扫描完整覆盖，然后对照冻结基线核验产物。

github-actions · 2026-10-18T15:06:58Z

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=39145685012

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=39145685012

Klaud-Cold · 2026-10-18T15:08:21Z

Final full sweep · Failed · Run 39137200943 / attempt 2 · 2026-10-18 15:14 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 4d8bf0112f4a
Change: Add the perf-changelog entry and run the full-sweep-fail-fast final sweep over all 11 points and 5 default evals at this head.

Next: Repair 3 sets dynamo install to false so workers use the identical ai-dynamo 1.6.0.dev20261015 already in the v0.5.22 image and skip the flaky host download.

中文

最终完整 sweep · 失败 · Run 39137200943 / attempt 2 · 2026-10-18 15:14 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 4d8bf0112f4a
**变更：**添加 perf-changelog 条目，并在此提交上对全部 11 个点和 5 个默认评测运行 full-sweep-fail-fast 最终扫描。实测数值及异常说明见上表。
**下一步：**修复 3 将 dynamo install 设为 false，使 worker 直接使用 v0.5.22 镜像中相同的 ai-dynamo 1.6.0.dev20261015，跳过不稳定的主机端下载。

Staging the identical 1.6.0.dev20261015 wheel from PyPI on the host failed intermittently with IncompleteRead; skip the install and use the image's copy. 在主机上从 PyPI 预取相同的 1.6.0.dev20261015 wheel 时会间歇性出现 IncompleteRead；跳过安装，直接使用镜像内置的版本。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Klaud-Cold · 2026-10-18T15:14:56Z

Repair 3/5 · Passed · Run 39139577662 / attempt 1 · 2026-10-18 16:34 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 9a132463cf69 · Mean latency
Change: Measure PR head 258f9a6: set dynamo install to false so workers use the ai-dynamo 1.6.0.dev20261015 already in the v0.5.22 image, removing the host-side PyPI wheel download that failed intermittently.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 PTP4x1/DTP4x0 dd714f	56.13 (N/A)	287.1 (N/A)	1.81 (N/A)
AgentX c2 PTP4x1/DTP4x0 f99308	59.97 (N/A)	238.95 (N/A)	1.79 (N/A)
AgentX c22 PTP2x1/DTP2x0 ddd472	527.59 (N/A)	391.52 (N/A)	5.7 (N/A)
AgentX c24 PTP4x1/DTP4x0 95bb75	300.1 (N/A)	314.75 (N/A)	3.44 (N/A)
AgentX c32 PTP2x1/DTP2x0 474621	676.95 (N/A)	489.48 (N/A)	8.15 (N/A)
AgentX c40 PTP2x1/DTP2x0 2aa262	823.29 (N/A)	586.79 (N/A)	9.59 (N/A)

Note: AgentX c1 PTP4x1/DTP4x0 dd714f, AgentX c2 PTP4x1/DTP4x0 f99308, AgentX c24 PTP4x1/DTP4x0 95bb75: Δ N/A: dataset/scenario/statistic mismatch.
Note: AgentX c22 PTP2x1/DTP2x0 ddd472, AgentX c32 PTP2x1/DTP2x0 474621, AgentX c40 PTP2x1/DTP2x0 2aa262: Δ N/A: point unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c20	97.5% (+0.08 pp)	1,319 each
gsm8k/em_strict · c22	97.57% (+0.08 pp)	1,319 each
gsm8k/em_strict · c24	97.8% (+0.08 pp)	1,319 each
gsm8k/em_strict · c32	97.8% (+0.23 pp)	1,319 each
gsm8k/em_strict · c48	97.73% (+0.00 pp)	1,319 each

Next: Rerun the full-sweep-fail-fast final sweep at this head.

中文

修复 3/5 · 已通过 · Run 39139577662 / attempt 1 · 2026-10-18 16:34 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 9a132463cf69 · 平均延迟
**变更：**测量 PR 提交 258f9a6：将 dynamo install 设为 false，使 worker 使用 v0.5.22 镜像中已有的 ai-dynamo 1.6.0.dev20261015，去掉间歇性失败的主机端 PyPI wheel 下载。实测数值及异常说明见上表。
**下一步：**在此提交上重新运行 full-sweep-fail-fast 最终扫描。

Klaud-Cold · 2026-10-18T16:36:35Z

Final full sweep · Queued · Run 39145685012 / attempt 1 · 2026-10-18 16:36 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 258f9a6cef99
Change: Add the perf-changelog entry and run the full-sweep-fail-fast final sweep over all 11 points and 5 default evals at this head.

Next: Wait for complete final-sweep coverage, then verify artifacts against the frozen baseline.

中文

最终完整 sweep · 排队中 · Run 39145685012 / attempt 1 · 2026-10-18 16:36 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 258f9a6cef99
**变更：**添加 perf-changelog 条目，并在此提交上对全部 11 个点和 5 个默认评测运行 full-sweep-fail-fast 最终扫描。实测数值及异常说明见上表。
**下一步：**等待最终扫描完整覆盖，然后对照冻结基线核验产物。

Klaud-Cold · 2026-10-18T18:05:49Z

Final full sweep · Passed · Run 39145685012 / attempt 2 · 2026-10-19 00:35 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 258f9a6cef99 · Mean latency
Change: Validate the complete updated-image family at this head.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 PTP4x1/DTP4x0 dd714f	56.16 (N/A)	252.26 (N/A)	1.79 (N/A)
AgentX c2 PTP4x1/DTP4x0 f99308	59.64 (N/A)	229.75 (N/A)	1.76 (N/A)
AgentX c4 PTP4x1/DTP4x0 e25615	78.11 (N/A)	231.54 (N/A)	1.94 (N/A)
AgentX c8 PTP4x1/DTP4x0 adb59a	121.83 (N/A)	246.62 (N/A)	2.11 (N/A)
AgentX c12 PTP4x1/DTP4x0 2a9acc	179.61 (N/A)	264.97 (N/A)	2.5 (N/A)
AgentX c20 PTP4x1/DTP4x0 36ba59	254.03 (N/A)	318.03 (N/A)	2.96 (N/A)
AgentX c22 PTP2x1/DTP2x0 ddd472	527.49 (N/A)	404.66 (N/A)	5.77 (N/A)
AgentX c24 PTP4x1/DTP4x0 95bb75	299.37 (N/A)	321.98 (N/A)	3.4 (N/A)
AgentX c32 PTP2x1/DTP2x0 474621	674.62 (N/A)	500.65 (N/A)	8.03 (N/A)
AgentX c40 PTP2x1/DTP2x0 2aa262	823.15 (N/A)	581.91 (N/A)	9.37 (N/A)
AgentX c48 PTP2x1/DTP2x0 d5567a	910.38 (N/A)	738.09 (N/A)	13.47 (N/A)

Note: AgentX c1 PTP4x1/DTP4x0 dd714f, AgentX c2 PTP4x1/DTP4x0 f99308, AgentX c4 PTP4x1/DTP4x0 e25615, AgentX c8 PTP4x1/DTP4x0 adb59a, AgentX c12 PTP4x1/DTP4x0 2a9acc, AgentX c20 PTP4x1/DTP4x0 36ba59, AgentX c24 PTP4x1/DTP4x0 95bb75: Δ N/A: dataset/scenario/statistic mismatch.
Note: AgentX c22 PTP2x1/DTP2x0 ddd472, AgentX c32 PTP2x1/DTP2x0 474621, AgentX c40 PTP2x1/DTP2x0 2aa262, AgentX c48 PTP2x1/DTP2x0 d5567a: Δ N/A: point unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c20	97.19% (-0.23 pp)	1,319 each
gsm8k/em_strict · c22	97.73% (+0.23 pp)	1,319 each
gsm8k/em_strict · c24	97.88% (+0.15 pp)	1,319 each
gsm8k/em_strict · c32	97.8% (+0.23 pp)	1,319 each
gsm8k/em_strict · c48	97.65% (-0.08 pp)	1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 39145685012 / attempt 2 · 2026-10-19 00:35 UTC
lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae · 258f9a6cef99 · 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 · 2026-10-18T18:17:59Z

resumed · The session ended before finishing; attempt 2 of its planner run continues it.

中文

resumed · 会话未完成即已结束；其规划运行的第 2 次尝试将继续该会话。

Klaud-Cold · 2026-10-19T00:35:12Z

validated · Repairs: unknown · Runs: 39128853814, 39129509704, 39130985168, 39137102013, 39137200943, 39139559001, 39139574348, 39139577662, 39145685012
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数：unknown · 运行：39128853814, 39129509704, 39130985168, 39137102013, 39137200943, 39139559001, 39139574348, 39139577662, 39145685012
所有自有运行均已结束。完整 sweep 已验证；已就绪，等待审查。

claude

Code review found no issues

No high-confidence issues detected in this change.

Resolve conflicts: take main's refactored recipes and re-apply this PR's non-image edits; take main's perf-changelog.yaml and re-append this PR's entry. 中文：合并 origin/main；配方采用 main 重构后的版本并重新应用本 PR 的非镜像改动，perf-changelog.yaml 采用 main 版本并在末尾重新追加本 PR 的条目。

adibarra · 2026-10-22T01:19:55Z

/use 39145685012

adibarra · 2026-10-22T01:19:59Z

/reuse-sweep-run 39145685012

github-actions · 2026-10-22T01:20:12Z

@adibarra staged run 39145685012.

Qwen-3.5-397B-A17B · agentic-traces · FP4 · gb200_dynamo-sglang · 2026-10-18

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

github-project-automation Bot added this to InferenceX Board Oct 18, 2026

chore(perf-changelog): add Qwen3.5 GB200 SGLang v0.5.22 entry

4d8bf01

新增 Qwen3.5 GB200 SGLang v0.5.22 的 perf-changelog 条目。 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 18, 2026

Klaud-Cold removed the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Oct 18, 2026

Klaud-Cold added the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Oct 18, 2026

Klaud-Cold marked this pull request as ready for review October 19, 2026 00:35

Klaud-Cold requested a review from a team October 19, 2026 00:35

Klaud-Cold requested review from Ankur-singh, RohitNagraj, kedarpotdar-nv, nvpohanh and wzhao18 as code owners October 19, 2026 00:35

claude Bot reviewed Oct 19, 2026

View reviewed changes

chore: refresh PR #4278 for sweep reuse [skip-sweep]

5daad93

adibarra merged commit 1e29fdc into main Oct 22, 2026
24 checks passed

adibarra deleted the klaud/auto-60738ecbf3a926e0-512f37ee2bd1e03d branch October 22, 2026 01:21

github-project-automation Bot moved this to Done in InferenceX Board Oct 22, 2026

adibarra mentioned this pull request Oct 22, 2026

fix(klaud): freeze AgentX datasets and evals so baseline deltas render / fix(klaud)：冻结 AgentX 数据集与评测以显示基线差值 #4390

Merged

nvpohanh mentioned this pull request Oct 22, 2026

[Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs sgl-project/sglang#45749

Merged

Conversation

Klaud-Cold commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Oct 18, 2026

Uh oh!

Klaud-Cold commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 18, 2026

Uh oh!

github-actions Bot commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 18, 2026

Uh oh!

Klaud-Cold commented Oct 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Klaud-Cold commented Oct 18, 2026

Uh oh!

Klaud-Cold commented Oct 19, 2026

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

adibarra commented Oct 22, 2026

Uh oh!

adibarra commented Oct 22, 2026

Uh oh!

github-actions Bot commented Oct 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Klaud-Cold commented Oct 18, 2026 •

edited

Loading

Klaud-Cold commented Oct 18, 2026 •

edited

Loading

Klaud-Cold commented Oct 18, 2026 •

edited

Loading

Klaud-Cold commented Oct 18, 2026 •

edited

Loading

github-actions Bot commented Oct 18, 2026 •

edited

Loading

Klaud-Cold commented Oct 18, 2026 •

edited

Loading

Klaud-Cold commented Oct 18, 2026 •

edited

Loading

Klaud-Cold commented Oct 18, 2026 •

edited

Loading

github-actions Bot commented Oct 22, 2026 •

edited

Loading