[Klaud Cold] Update DSv4 FP4 B200 vLLM AgentX DSpark image to upstream nightly 2eb491e / [Klaud Cold] 将 DSv4 FP4 B200 vLLM AgentX DSpark 镜像更新至上游 nightly 2eb491e

[Klaud-Cold]

Goal: Update vLLM image from vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-591bb95 to vllm/vllm-openai:nightly-2eb491e18b1ce5b8a25e4c88f32dbb5d028d525c@sha256:514cd42702d29a7b8efe20b0a916523ac473ecebdec8d53b1a0d0f5439a48634.
Baseline: 2026-09-19 · vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-591bb95
Mean latency · Sources: API 1, API 2, API 3, API 4, API 5

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP8 EP1 9c1d23	2,201.3	16.41	1,860.04	3.48
AgentX c4 TP8 EP1 f81c5a	3,874.04	27.14	764.45	5.03
AgentX c6 TP8 EP1 0d8f91	4,483.15	35.42	855.8	4.76
AgentX c10 TP8 EP1 da6c15	8,029.38	59.46	736.62	6.12
AgentX c14 TP8 EP1 68c217	9,418.59	74.89	803.57	6.91
AgentX c16 TP8 EP1 eb2826	11,719.42	78.68	932.12	8.98
AgentX c32 TP8 EP8 658e6f	17,480.3	113.95	6,331.27	13.31
AgentX c64 TP8 EP8 ac9a9a	34,561.72	261.21	7,683.01	11.9
AgentX c96 TP8 EP8 a7529d	47,863.05	385.49	9,231	13.57
AgentX c128 TP8 EP8 5f0930	54,814.79	455.61	10,673.35	14.83
AgentX c160 TP8 EP8 51422e	64,489.73	499.22	12,379.88	17.17
AgentX c192 TP8 EP8 68a170	69,323.46	522.64	14,684.94	22.13

Eval	Score ↑	Samples
gsm8k/em_strict · c192	97.04%	1,319

中文

**目标：**将 vLLM 镜像从 vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-591bb95 更新为 vllm/vllm-openai:nightly-2eb491e18b1ce5b8a25e4c88f32dbb5d028d525c@sha256:514cd42702d29a7b8efe20b0a916523ac473ecebdec8d53b1a0d0f5439a48634。
**基线：**2026-09-19 · vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-591bb95
平均延迟 · 来源： API 1, API 2, API 3, API 4, API 5；数值及异常说明见上表。

[github-actions]

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]

Initial attempt · Passed · Run 38164192033 / attempt 1 · 2026-10-11 20:28 UTC
vllm/vllm-openai:nightly-2eb491e18b1ce5b8a25e4c88f32dbb5d028d525c@sha256:514cd42702d29a7b8efe20b0a916523ac473ecebdec8d53b1a0d0f5439a48634 · c801c9f2dbbd · Mean latency
Change: Measured ref 953d8d6: Pin upstream vLLM nightly 2eb491e by digest; the unpublished 591bb95 fork build's deep_gemm_amxf4_mega_moe and VLLM_DSV4_MEGA_FP8_COMBINE are absent upstream (kernel.py), so DEP8 uses deep_gemm_mega_moe (FP8 activations, BF16 combine); other flags parse upstream, and fork build 20ebb28 was skipped after repeated c192 crashes.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP8 EP1 9c1d23	16.74 (+2%)	1,511.11 (-18.8%)	3.55 (+2%)
AgentX c32 TP8 EP8 658e6f	127.42 (+11.8%)	4,325.22 (-31.7%)	11.07 (-16.8%)

Eval	Score ↑	Samples
gsm8k/em_strict · c16	97.12% (N/A)	N/A/1,319 (old/new)
gsm8k/em_strict · c192	97.27% (+0.23 pp)	1,319 each

Note: gsm8k/em_strict · c16: Δ N/A: no matched eval baseline.

Next: Append the changelog entry and run the final full sweep.

中文

初次尝试 · 已通过 · Run 38164192033 / attempt 1 · 2026-10-11 20:28 UTC
vllm/vllm-openai:nightly-2eb491e18b1ce5b8a25e4c88f32dbb5d028d525c@sha256:514cd42702d29a7b8efe20b0a916523ac473ecebdec8d53b1a0d0f5439a48634 · c801c9f2dbbd · 平均延迟
**变更：**测量 ref 953d8d6：按摘要固定上游 vLLM nightly 2eb491e；未公开的 591bb95 fork 构建所用 deep_gemm_amxf4_mega_moe 和 VLLM_DSV4_MEGA_FP8_COMBINE 在上游不存在，因此 DEP8 改用 deep_gemm_mega_moe（FP8 激活，BF16 combine）；其余参数均可被上游解析；fork 构建 20ebb28 因 c192 反复崩溃而跳过。实测数值及异常说明见上表。
**下一步：**追加变更日志条目并运行最终完整 sweep。

[Klaud-Cold]

Final full sweep · Failed · Run 38172422803 / attempt 1 · 2026-10-11 21:41 UTC
vllm/vllm-openai:nightly-2eb491e18b1ce5b8a25e4c88f32dbb5d028d525c@sha256:514cd42702d29a7b8efe20b0a916523ac473ecebdec8d53b1a0d0f5439a48634 · 4fac5a774660 · Mean latency
Change: Run the full fail-fast sweep on the smoke-validated image; DEP8 c128 hit torch.OutOfMemoryError on all ranks in DSv4 attention _fused_qnorm_rope_kv_insert about 5 minutes into AgentX warmup (weights 130.23 GiB, KV 35 GiB and graphs match the old image, so upstream runtime buffers are larger), and fail-fast cancelled the siblings.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP8 EP1 9c1d23	N/A	N/A	N/A
AgentX c4 TP8 EP1 f81c5a	N/A	N/A	N/A
AgentX c6 TP8 EP1 0d8f91	N/A	N/A	N/A
AgentX c10 TP8 EP1 da6c15	N/A	N/A	N/A
AgentX c14 TP8 EP1 68c217	N/A	N/A	N/A
AgentX c16 TP8 EP1 eb2826	N/A	N/A	N/A
AgentX c32 TP8 EP8 658e6f	N/A	N/A	N/A
AgentX c64 TP8 EP8 ac9a9a	N/A	N/A	N/A
AgentX c96 TP8 EP8 a7529d	N/A	N/A	N/A
AgentX c128 TP8 EP8 5f0930	N/A	N/A	N/A
AgentX c160 TP8 EP8 51422e	N/A	N/A	N/A
AgentX c192 TP8 EP8 68a170	N/A	N/A	N/A

Note: AgentX c1 TP8 EP1 9c1d23, AgentX c4 TP8 EP1 f81c5a, AgentX c6 TP8 EP1 0d8f91, AgentX c10 TP8 EP1 da6c15, AgentX c14 TP8 EP1 68c217, AgentX c16 TP8 EP1 eb2826, AgentX c32 TP8 EP8 658e6f, AgentX c64 TP8 EP8 ac9a9a, AgentX c96 TP8 EP8 a7529d, AgentX c160 TP8 EP8 51422e, AgentX c192 TP8 EP8 68a170: cancelled.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.
Note: AgentX c128 TP8 EP8 5f0930: failed.

Eval	Score ↑	Samples
gsm8k/em_strict · c16	97.12% (N/A)	N/A/1,319 (old/new)
gsm8k/em_strict · c192	97.65% (+0.61 pp)	1,319 each

Note: gsm8k/em_strict · c16: Δ N/A: no matched eval baseline.

Next: Rerun the failed and cancelled jobs on the same head to test whether the c128 OOM recurs.

中文

最终完整 sweep · 失败 · Run 38172422803 / attempt 1 · 2026-10-11 21:41 UTC
vllm/vllm-openai:nightly-2eb491e18b1ce5b8a25e4c88f32dbb5d028d525c@sha256:514cd42702d29a7b8efe20b0a916523ac473ecebdec8d53b1a0d0f5439a48634 · 4fac5a774660 · 平均延迟
**变更：**在冒烟验证通过的镜像上运行完整 fail-fast sweep；DEP8 c128 在 AgentX 预热约 5 分钟时所有 rank 于 DSv4 注意力 _fused_qnorm_rope_kv_insert 中出现 torch.OutOfMemoryError（权重 130.23 GiB、KV 35 GiB 和 graph 占用与旧镜像一致，说明上游运行时缓冲更大），fail-fast 取消了其余点。实测数值及异常说明见上表。
**下一步：**在同一 head 上重跑失败和取消的作业，检验 c128 OOM 是否复现。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=38172422803

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=38172422803

[Klaud-Cold]

Repair 1/5 · Failed · Run 38172422803 / attempt 2 · 2026-10-11 22:12 UTC
vllm/vllm-openai:nightly-2eb491e18b1ce5b8a25e4c88f32dbb5d028d525c@sha256:514cd42702d29a7b8efe20b0a916523ac473ecebdec8d53b1a0d0f5439a48634 · 4fac5a774660 · Mean latency
Change: No recipe change: reran on the same head and DEP8 c128 reproduced the same all-rank torch.OutOfMemoryError (522 MiB q_fp8 in DSv4 attention _fused_qnorm_rope_kv_insert) during AgentX warmup, so the upstream image's larger runtime memory does not fit the fixed 35 GiB KV and 8416-token batch budgets that this image update must preserve.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
AgentX c1 TP8 EP1 9c1d23	N/A	N/A	N/A
AgentX c4 TP8 EP1 f81c5a	N/A	N/A	N/A
AgentX c6 TP8 EP1 0d8f91	N/A	N/A	N/A
AgentX c10 TP8 EP1 da6c15	N/A	N/A	N/A
AgentX c14 TP8 EP1 68c217	N/A	N/A	N/A
AgentX c16 TP8 EP1 eb2826	N/A	N/A	N/A
AgentX c32 TP8 EP8 658e6f	N/A	N/A	N/A
AgentX c64 TP8 EP8 ac9a9a	N/A	N/A	N/A
AgentX c96 TP8 EP8 a7529d	N/A	N/A	N/A
AgentX c128 TP8 EP8 5f0930	N/A	N/A	N/A
AgentX c160 TP8 EP8 51422e	N/A	N/A	N/A
AgentX c192 TP8 EP8 68a170	N/A	N/A	N/A

Note: AgentX c1 TP8 EP1 9c1d23, AgentX c4 TP8 EP1 f81c5a, AgentX c6 TP8 EP1 0d8f91, AgentX c10 TP8 EP1 da6c15, AgentX c14 TP8 EP1 68c217, AgentX c16 TP8 EP1 eb2826, AgentX c32 TP8 EP8 658e6f, AgentX c64 TP8 EP8 ac9a9a, AgentX c96 TP8 EP8 a7529d, AgentX c160 TP8 EP8 51422e, AgentX c192 TP8 EP8 68a170: cancelled.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.
Note: AgentX c128 TP8 EP8 5f0930: failed.

Eval	Score ↑	Samples
gsm8k/em_strict · c16	97.12% (N/A)	N/A/1,319 (old/new)
gsm8k/em_strict · c192	97.65% (+0.61 pp)	1,319 each

Note: gsm8k/em_strict · c16: Δ N/A: no matched eval baseline.

Next: Stop: the same failure recurred, and fixing it would require resizing frozen memory settings outside an image update.

中文

修复 1/5 · 失败 · Run 38172422803 / attempt 2 · 2026-10-11 22:12 UTC
vllm/vllm-openai:nightly-2eb491e18b1ce5b8a25e4c88f32dbb5d028d525c@sha256:514cd42702d29a7b8efe20b0a916523ac473ecebdec8d53b1a0d0f5439a48634 · 4fac5a774660 · 平均延迟
**变更：**配方不变：在同一 head 上重跑后，DEP8 c128 在 AgentX 预热中再次在所有 rank 出现相同的 torch.OutOfMemoryError（DSv4 注意力 _fused_qnorm_rope_kv_insert 中分配 522 MiB q_fp8），说明上游镜像更大的运行时内存无法容纳本次镜像更新必须保留的固定 35 GiB KV 与 8416 token 批预算。实测数值及异常说明见上表。
**下一步：**停止：同一失败再次出现，修复需要调整冻结的内存设置，超出镜像更新范围。

[Klaud-Cold]

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

中文

failed · 清理待完成。先停止并确认自有运行结束，再关闭 PR。

[Klaud-Cold]

failed · Repairs: 1 · Runs: 38164192033, 38172408598, 38172422803, 38179512684
All owned runs ended. PR closed; branch deleted for retry.

中文

failed · 修复次数：1 · 运行：38164192033, 38172408598, 38172422803, 38179512684
所有自有运行均已结束。PR 已关闭；分支已删除，可重新尝试。