Skip to content

[Klaud Cold] Update DeepSeek-V4-Pro FP4 B200 vLLM AgentX MTP image to nightly-83c078d9 / [Klaud Cold] 将 DeepSeek-V4-Pro FP4 B200 vLLM AgentX MTP 镜像更新至 nightly-83c078d9 - #4457

Closed
Klaud-Cold wants to merge 5 commits into
mainfrom
klaud/auto-b7e2ccf9e45c74b9-1995ad3bffd832b3
Closed

Klaud-Cold wants to merge 5 commits into
mainfrom
klaud/auto-b7e2ccf9e45c74b9-1995ad3bffd832b3

Conversation

@Klaud-Cold

@Klaud-Cold Klaud-Cold commented Oct 24, 2026 •

Copy link
Copy Markdown
Collaborator

Goal: Update vLLM image from vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28 to vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370.
Baseline: 2026-10-21 · vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f
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 c1 TP8 EP1 9c1d23 2,202.99 16.42 1,833.4 3.49
AgentX c4 TP8 EP1 f81c5a 3,764.85 26.3 794.31 4.61
AgentX c6 TP8 EP1 0d8f91 4,584.36 36.09 876.64 7.77
AgentX c10 TP8 EP1 da6c15 7,992 59.23 755.98 6.1
AgentX c14 TP8 EP1 68c217 9,405.86 74.74 770.08 6.94
AgentX c16 TP8 EP1 eb2826 11,596.55 78.59 892.89 8.32
AgentX c32 TP8 EP8 385cf0 17,794.79 115.3 6,114.91 13.09
AgentX c64 TP8 EP8 658947 34,457.51 259.03 7,686.35 12.01
AgentX c96 TP8 EP8 a5e86a 48,138.17 388.68 9,083.98 13.46
AgentX c128 TP8 EP8 f5201b 53,831.76 446 11,098.2 15.49
AgentX c160 TP8 EP8 a1f501 65,658.3 509.52 12,065.44 16.7
AgentX c192 TP8 EP8 0ff5ab 67,069.84 505.75 15,571.19 23.25
Eval Score ↑ Samples
gsm8k/em_strict · c16 97.88% 1,319
gsm8k/em_strict · c192 97.19% 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 vLLM 镜像从 vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28 更新至 vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370。
**基线:**2026-10-21 · vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-20ebb28@sha256:dfdee3dbba3d8d30554d90f723a3a18dac562f8442f2c2e8f81804acd1ec673f
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。

Move dsv4-fp4-b200-vllm-agentic-mtp from the DSV4 dev build
nightly-dev-x86_64-cu13.0.1-20ebb28 to upstream vLLM nightly 83c078d9,
pinned by digest. Upstream names the DeepSeek-V4 MegaMoE backend
deep_gemm_mega_moe and has no VLLM_DSV4_MEGA_FP8_COMBINE, so the DEP8
recipe uses that backend and drops the env var.

将 dsv4-fp4-b200-vllm-agentic-mtp 从 DSV4 开发构建
nightly-dev-x86_64-cu13.0.1-20ebb28 更新为上游 vLLM nightly 83c078d9,
并按 digest 固定。上游的 DeepSeek-V4 MegaMoE 后端名为
deep_gemm_mega_moe,且不存在 VLLM_DSV4_MEGA_FP8_COMBINE,因此 DEP8
配方改用该后端并移除该环境变量。

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 24, 2026 •

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Failed · Run 40048229064 / attempt 1 · 2026-10-24 01:51 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 80dff120d27f
Change: At measured ref 67bc85f, pin upstream vLLM nightly 83c078d9 (0.32.1rc1.dev361+g83c078d95, torch 2.14.0, flashinfer 0.7.2) by digest, replacing the unpublished DSV4 dev build 0.17.2rc1.dev6312+g20ebb2850 (torch 2.13.0, flashinfer 0.6.18.post1; commit absent from vllm-project/vllm, so old-side provenance comes only from the image wheels); upstream has no deep_gemm_amxf4_mega_moe (MXFP4-activation MegaMoE dispatch) or VLLM_DSV4_MEGA_FP8_COMBINE, so DEP8 uses upstream deep_gemm_mega_moe (fp8xfp4, BF16 combine) and drops that env; all other flags, DSpark defaults and SimpleCPU offload keys verified in vllm/config and arg_utils at 83c078d9.

Next: Repair 1: lower the DEP8 kv-cache-memory-bytes by 4 GiB to restore the activation headroom lost to the larger upstream graph pools and weights, then rerun smoke.

中文

初次尝试 · 失败 · Run 40048229064 / attempt 1 · 2026-10-24 01:51 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 80dff120d27f
**变更:**在实测提交 67bc85f 上,按 digest 固定上游 vLLM nightly 83c078d9(0.32.1rc1.dev361+g83c078d95,torch 2.14.0,flashinfer 0.7.2),替换未发布的 DSV4 开发构建 0.17.2rc1.dev6312+g20ebb2850(torch 2.13.0,flashinfer 0.6.18.post1;该提交不在 vllm-project/vllm 中,旧版来源仅能由镜像 wheel 证明);上游没有 deep_gemm_amxf4_mega_moe(MXFP4 激活的 MegaMoE 分发)和 VLLM_DSV4_MEGA_FP8_COMBINE,因此 DEP8 改用上游 deep_gemm_mega_moe(fp8xfp4,BF16 combine)并移除该环境变量;其余参数、DSpark 默认值和 SimpleCPU 卸载配置已在 83c078d9 的 vllm/config 与 arg_utils 中核实。实测数值及异常说明见上表。
**下一步:**修复 1:将 DEP8 的 kv-cache-memory-bytes 降低 4 GiB,以恢复被上游更大的 CUDA graph 内存池和权重占用的激活余量,然后重新运行 smoke。

Upstream nightly 83c078d9 uses about 3 GiB more CUDA graph memory and
0.4 GiB more weight memory per DEP8 rank, so the fixed KV pool left too
little headroom and DEP8 c96 hit OOM in the sparse indexer logits.

上游 nightly 83c078d9 每个 DEP8 rank 的 CUDA graph 内存多约 3 GiB、
权重内存多 0.4 GiB,固定 KV 池留下的余量不足,DEP8 c96 在稀疏索引
logits 分配时发生 OOM,因此将 DEP8 KV 缓存降低 4 GiB。

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

Klaud-Cold commented Oct 24, 2026 •

Copy link
Copy Markdown
Collaborator Author

Repair 1/5 · Passed · Run 40051674447 / attempt 1 · 2026-10-24 06:11 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 80dff120d27f · Mean latency
Change: At measured ref 6b1b098, lower DEP8 kv-cache-memory-bytes from 35 GiB to 31 GiB (c32-c128) and from 34 GiB to 30 GiB (c160, c192) so the upstream image's larger CUDA graph pools (4.08 vs 0.96 GiB) and weights (130.5 vs 130.1 GiB) leave room for sparse-indexer prefill buffers; TP8 is unchanged.

Point Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
AgentX c1 TP8 EP1 2f3181 17.28 (N/A) 1,362.31 (N/A) 3.33 (N/A)
AgentX c4 TP8 EP1 e18886 27.45 (N/A) 673.84 (N/A) 4.15 (N/A)
AgentX c6 TP8 EP1 4d9730 37.91 (N/A) 506.44 (N/A) 4.62 (N/A)
AgentX c10 TP8 EP1 aa6a29 60.14 (N/A) 602.12 (N/A) 5.36 (N/A)
AgentX c14 TP8 EP1 f16b18 75.95 (N/A) 567.22 (N/A) 5.83 (N/A)
AgentX c16 TP8 EP1 21025a 82.84 (N/A) 711.31 (N/A) 7 (N/A)
AgentX c32 TP8 EP8 1073d8 129.76 (N/A) 3,977.22 (N/A) 10.07 (N/A)
AgentX c64 TP8 EP8 7ffbf1 266.37 (N/A) 6,911.41 (N/A) 11.77 (N/A)
AgentX c96 TP8 EP8 deb33a 399.33 (N/A) 8,402.51 (N/A) 12.67 (N/A)
AgentX c128 TP8 EP8 6beebb 472.84 (N/A) 9,806.27 (N/A) 14.13 (N/A)
AgentX c160 TP8 EP8 49fc80 518.03 (N/A) 11,325.81 (N/A) 16.9 (N/A)
AgentX c192 TP8 EP8 14f7af 534.25 (N/A) 13,270.23 (N/A) 22.15 (N/A)

Note: All rows: Δ N/A: no matched baseline.
Note: AgentX c16 TP8 EP1 21025a: 2 request errors.
Note: AgentX c192 TP8 EP8 14f7af: 1 request error.

Eval Score ↑ Samples
gsm8k/em_strict · c16 97.57% (-0.30 pp) 1,319 each
gsm8k/em_strict · c192 97.27% (+0.08 pp) 1,319 each

Next: Append the perf-changelog entry and run the full sweep with full-sweep-fail-fast.

中文

修复 1/5 · 已通过 · Run 40051674447 / attempt 1 · 2026-10-24 06:11 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 80dff120d27f · 平均延迟
**变更:**在实测提交 6b1b098 上,将 DEP8 的 kv-cache-memory-bytes 从 35 GiB 降至 31 GiB(c32-c128),从 34 GiB 降至 30 GiB(c160、c192),使上游镜像更大的 CUDA graph 内存池(4.08 vs 0.96 GiB)和权重(130.5 vs 130.1 GiB)之外仍有稀疏索引 prefill 缓冲区的空间;TP8 不变。实测数值及异常说明见上表。
**下一步:**追加 perf-changelog 条目,并使用 full-sweep-fail-fast 运行完整扫描。

github-actions Bot and others added 2 commits October 24, 2026 06:11
合并 main 分支。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
新增 dsv4 B200 vLLM AgentX 镜像更新的 perf-changelog 条目。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@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 added the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Oct 24, 2026
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Running · Run 40070408546 / attempt 1 · 2026-10-24 06:30 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 262b1bc436c5
Change: Merge main and append the dsv4-fp4-b200-vllm-agentic-mtp perf-changelog entry at the tail; the recipe and image are unchanged from repair 1, and check-final covers all 12 baseline AgentX points plus the c16 and c192 evals.

Next: Wait for the full-sweep-fail-fast sweep to finish, then verify every point and eval artifact.

中文

最终完整 sweep · 运行中 · Run 40070408546 / attempt 1 · 2026-10-24 06:30 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 262b1bc436c5
**变更:**合并 main 并在末尾追加 dsv4-fp4-b200-vllm-agentic-mtp 的 perf-changelog 条目;recipe 与镜像与修复 1 相同,check-final 覆盖全部 12 个基线 AgentX 点位以及 c16 和 c192 评测。实测数值及异常说明见上表。
**下一步:**等待 full-sweep-fail-fast 完整扫描完成,然后核验每个点位和评测的产物。

@github-actions

github-actions Bot commented Oct 24, 2026 •

Copy link
Copy Markdown
Contributor

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Queued · Run 40070408546 / attempt 2 · 2026-10-24 07:20 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 262b1bc436c5
Change: Merge main and append the dsv4-fp4-b200-vllm-agentic-mtp perf-changelog entry at the tail; the recipe and image are unchanged from repair 1, and check-final covers all 12 baseline AgentX points plus the c16 and c192 evals.

Next: Wait for the full-sweep-fail-fast sweep to finish, then verify every point and eval artifact.

中文

最终完整 sweep · 排队中 · Run 40070408546 / attempt 2 · 2026-10-24 07:20 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 262b1bc436c5
**变更:**合并 main 并在末尾追加 dsv4-fp4-b200-vllm-agentic-mtp 的 perf-changelog 条目;recipe 与镜像与修复 1 相同,check-final 覆盖全部 12 个基线 AgentX 点位以及 c16 和 c192 评测。实测数值及异常说明见上表。
**下一步:**等待 full-sweep-fail-fast 完整扫描完成,然后核验每个点位和评测的产物。

@Klaud-Cold

Klaud-Cold commented Oct 24, 2026 •

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Passed · Run 40070408546 / attempt 3 · 2026-10-24 14:07 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 262b1bc436c5 · Mean latency
Change: Merge main and append the dsv4-fp4-b200-vllm-agentic-mtp perf-changelog entry at the tail; the recipe and image are unchanged from repair 1, and check-final covers all 12 baseline AgentX points plus the c16 and c192 evals.

Point Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
AgentX c1 TP8 EP1 9c1d23 17.24 (+5%) 1,420.39 (-22.5%) 3.34 (-4.3%)
AgentX c4 TP8 EP1 f81c5a 27.4 (+4.2%) 636.94 (-19.8%) 4.09 (-11.3%)
AgentX c6 TP8 EP1 0d8f91 37.84 (+4.9%) 496.67 (-43.3%) 4.31 (-44.5%)
AgentX c10 TP8 EP1 da6c15 60.32 (+1.8%) 558.23 (-26.2%) 5.25 (-13.9%)
AgentX c14 TP8 EP1 68c217 76.1 (+1.8%) 593.12 (-23%) 5.88 (-15.3%)
AgentX c16 TP8 EP1 eb2826 82.81 (+5.4%) 709.47 (-20.5%) 7.24 (-13%)
AgentX c32 TP8 EP8 385cf0 128.49 (+11.4%) 4,095 (-33%) 10.55 (-19.4%)
AgentX c64 TP8 EP8 658947 266.45 (+2.9%) 6,716.75 (-12.6%) 11.58 (-3.6%)
AgentX c96 TP8 EP8 a5e86a 405.17 (+4.2%) 8,179.44 (-10%) 12.29 (-8.7%)
AgentX c128 TP8 EP8 f5201b 474.3 (+6.3%) 9,749.27 (-12.2%) 13.77 (-11.1%)
AgentX c160 TP8 EP8 a1f501 529.3 (+3.9%) 10,875.27 (-9.9%) 15.92 (-4.7%)
AgentX c192 TP8 EP8 0ff5ab 502.25 (-0.7%) 14,188.79 (-8.9%) 25.32 (+8.9%)

Note: AgentX c16 TP8 EP1 eb2826, AgentX c64 TP8 EP8 658947, AgentX c160 TP8 EP8 a1f501: 1 request error.

Eval Score ↑ Samples
gsm8k/em_strict · c16 97.27% (-0.61 pp) 1,319 each
gsm8k/em_strict · c192 97.35% (+0.15 pp) 1,319 each

Next: Merge main, keep the dsv4-fp4-b200-vllm-agentic-mtp perf-changelog entry at the physical tail, and rerun the final full sweep on the merged head.

中文

最终完整 sweep · 已通过 · Run 40070408546 / attempt 3 · 2026-10-24 14:07 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 262b1bc436c5 · 平均延迟
**变更:**合并 main 并在末尾追加 dsv4-fp4-b200-vllm-agentic-mtp 的 perf-changelog 条目;recipe 与镜像与修复 1 相同,check-final 覆盖全部 12 个基线 AgentX 点位以及 c16 和 c192 评测。实测数值及异常说明见上表。
**下一步:**合并 main,将 dsv4-fp4-b200-vllm-agentic-mtp 的 perf-changelog 条目保持在文件末尾,并在合并后的提交上重新运行最终完整 sweep。

@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。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@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 Oct 24, 2026
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Queued · Run 40101218191 / attempt 1 · 2026-10-24 14:10 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 5d935a6fbede · Mean latency
Change: Merge main again so the dsv4-fp4-b200-vllm-agentic-mtp perf-changelog entry stays at the physical tail after main's new entries; the recipe and image are unchanged from repair 1, and check-final covers all 12 baseline AgentX points plus the c16 and c192 evals.

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 385cf0 N/A N/A N/A
AgentX c64 TP8 EP8 658947 N/A N/A N/A
AgentX c96 TP8 EP8 a5e86a N/A N/A N/A
AgentX c128 TP8 EP8 f5201b N/A N/A N/A
AgentX c160 TP8 EP8 a1f501 N/A N/A N/A
AgentX c192 TP8 EP8 0ff5ab 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 full-sweep-fail-fast sweep on the merged head to finish, then verify every point and eval artifact.

中文

最终完整 sweep · 排队中 · Run 40101218191 / attempt 1 · 2026-10-24 14:10 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 5d935a6fbede · 平均延迟
**变更:**再次合并 main,使 dsv4-fp4-b200-vllm-agentic-mtp 的 perf-changelog 条目在 main 新增条目之后仍位于文件末尾;recipe 与镜像与修复 1 相同,check-final 覆盖全部 12 个基线 AgentX 点位以及 c16 和 c192 评测。实测数值及异常说明见上表。
**下一步:**等待合并后提交上的 full-sweep-fail-fast 完整扫描完成,然后核验每个点位和评测的产物。

@Klaud-Cold

Klaud-Cold commented Oct 24, 2026 •

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Passed · Run 40101218191 / attempt 2 · 2026-10-24 19:48 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 5d935a6fbede · Mean latency
Change: Validate the complete updated-image family at this head.

Point Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
AgentX c1 TP8 EP1 9c1d23 17.3 (+5.3%) 1,397.07 (-23.8%) 3.3 (-5.4%)
AgentX c4 TP8 EP1 f81c5a 26.53 (+0.9%) 607.47 (-23.5%) 4.22 (-8.5%)
AgentX c6 TP8 EP1 0d8f91 37.74 (+4.6%) 482.16 (-45%) 4.33 (-44.3%)
AgentX c10 TP8 EP1 da6c15 59.93 (+1.2%) 549.2 (-27.4%) 5.24 (-14.1%)
AgentX c14 TP8 EP1 68c217 76.25 (+2%) 560.32 (-27.2%) 5.83 (-16%)
AgentX c16 TP8 EP1 eb2826 82.3 (+4.7%) 710.02 (-20.5%) 6.93 (-16.7%)
AgentX c32 TP8 EP8 385cf0 129.01 (+11.9%) 4,128.68 (-32.5%) 10.03 (-23.4%)
AgentX c64 TP8 EP8 658947 273.76 (+5.7%) 6,529.98 (-15%) 10.92 (-9.1%)
AgentX c96 TP8 EP8 a5e86a 404.13 (+4%) 8,170.19 (-10.1%) 12.37 (-8.1%)
AgentX c128 TP8 EP8 f5201b 477.77 (+7.1%) 9,615.33 (-13.4%) 13.79 (-11%)
AgentX c160 TP8 EP8 a1f501 507.26 (-0.4%) 11,279.99 (-6.5%) 16.75 (+0.3%)
AgentX c192 TP8 EP8 0ff5ab 498.78 (-1.4%) 14,002.94 (-10.1%) 24.46 (+5.2%)

Note: AgentX c16 TP8 EP1 eb2826, AgentX c64 TP8 EP8 658947, AgentX c96 TP8 EP8 a5e86a, AgentX c160 TP8 EP8 a1f501: 1 request error.

Eval Score ↑ Samples
gsm8k/em_strict · c16 97.5% (-0.38 pp) 1,319 each
gsm8k/em_strict · c192 97.35% (+0.15 pp) 1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40101218191 / attempt 2 · 2026-10-24 19:48 UTC
vllm/vllm-openai:nightly-83c078d955cd6e0c63d3b12abe669c8de936a370@sha256:cb12c3016938a4ee98ed176ac833984089460689a0982cff4b5a60bb45b3cded · 5d935a6fbede · 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

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。

@Klaud-Cold
Klaud-Cold marked this pull request as ready for review October 24, 2026 19:48
@Klaud-Cold
Klaud-Cold requested a review from a team October 24, 2026 19:48
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

validated · Repairs: 1 · Runs: 40048229064, 40051674447, 40069357943, 40070408546, 40101197357, 40101218191
All owned runs ended. Full sweep verified; ready for review.

中文

validated · 修复次数:1 · 运行:40048229064, 40051674447, 40069357943, 40070408546, 40101197357, 40101218191
所有自有运行均已结束。完整 sweep 已验证;已就绪,等待审查。

@claude

claude Bot commented Oct 24, 2026 •

Copy link
Copy Markdown
Contributor

Claude finished @Klaud-Cold's task in 20s —— View job


LGTM - no blocking issues found

What I checked
  • The image vllm/vllm-openai:nightly-83c078d9…@sha256:cb12c301… is on Docker Hub and pinned by digest.
  • The recipe no longer refers to deep_gemm_amxf4_mega_moe or VLLM_DSV4_MEGA_FP8_COMBINE. All six DEP8 overrides now use deep_gemm_mega_moe.
  • The KV cache sizes convert correctly: 33285996544 is exactly 31 GiB (c32 to c128) and 32212254720 is exactly 30 GiB (c160 and c192). The TP8 sizes are unchanged.
  • The perf-changelog.yaml entry is the last entry in the file and covers dsv4-fp4-b200-vllm-agentic-mtp.
  • I could not run git fetch (it needed approval), so I didn't check whether main has added changelog entries since the last merge.

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

/use 40101218191

@github-actions

github-actions Bot commented Oct 24, 2026 •

Copy link
Copy Markdown
Contributor

@Klaud-Cold staged run 40101218191.

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.

Code review found no issues

No high-confidence issues detected in this change.

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

requeued · Open for over 5 days while the published baseline of this family moved to a newer image. Closed with klaud-requeue so a new candidate can refresh the family from its current image.

中文

requeued · 本 PR 已开放超过 5 天,而该配置族已发布的基线已更新为更新的镜像。已关闭并标记 klaud-requeue,新的候选可以基于当前镜像更新该配置族。

@Klaud-Cold Klaud-Cold closed this Oct 29, 2026
@Klaud-Cold
Klaud-Cold deleted the klaud/auto-b7e2ccf9e45c74b9-1995ad3bffd832b3 branch October 29, 2026 08:52
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-requeue Closed so Klaud re-plans the family; no maintainer hold

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

1 participant