[Klaud Cold] Update Kimi-K3 FP4 B200 Dynamo-vLLM AgentX DSpark image to v0.32.0 / [Klaud Cold] 将 Kimi-K3 FP4 B200 Dynamo-vLLM AgentX DSpark 镜像更新至 v0.32.0 by Klaud-Cold · Pull Request #4369 · SemiAnalysisAI/InferenceX · GitHub
Skip to content

[Klaud Cold] Update Kimi-K3 FP4 B200 Dynamo-vLLM AgentX DSpark image to v0.32.0 / [Klaud Cold] 将 Kimi-K3 FP4 B200 Dynamo-vLLM AgentX DSpark 镜像更新至 v0.32.0 - #4369

Draft
Klaud-Cold wants to merge 1 commit into
mainfrom
klaud/auto-fa61042cf1fdc1f2-429ed3bfcab1acd8
Draft

[Klaud Cold] Update Kimi-K3 FP4 B200 Dynamo-vLLM AgentX DSpark image to v0.32.0 / [Klaud Cold] 将 Kimi-K3 FP4 B200 Dynamo-vLLM AgentX DSpark 镜像更新至 v0.32.0#4369
Klaud-Cold wants to merge 1 commit into
mainfrom
klaud/auto-fa61042cf1fdc1f2-429ed3bfcab1acd8

Conversation

@Klaud-Cold

@Klaud-Cold Klaud-Cold commented Oct 21, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

Goal: Update vLLM image from vllm/vllm-openai:nightly-ac9126e58aa7bbab1856ba6593ba4d5003fea516 to vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34.
Baseline: 2026-10-02 · vllm/vllm-openai:nightly-ac9126e58aa7bbab1856ba6593ba4d5003fea516
AgentX · PTP8x1/DTP8x0 · Mean latency · Sources: API 1, API 2

Concurrency Total tok/s/GPU ↑ Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 985.99 7.46 1,103.93 3.12
4 1,618.7 12.21 739.05 4.12
8 2,881.21 21.88 706.38 4.68
14 4,191.69 33.95 744.27 7.44
24 7,110.95 49.77 958.46 11.19
32 8,518.09 56.34 1,050.03 13.53
48 9,819.57 68.99 1,294.8 20.55
56 9,976.39 76.49 1,290.31 24.78
72 10,584.85 85.93 1,510.47 35.02
80 10,451 87.2 1,946.15 43.09
96 N/A N/A N/A N/A

Note: c96: unavailable.

Eval Score ↑ Samples
kimi_tool_call_schema_full/em_strict · c1 97.55% 408
kimi_tool_call_schema_full/em_strict · c4 97.55% 408
kimi_tool_call_schema_full/em_strict · c8 95.59% 408
kimi_tool_call_schema_full/em_strict · c14 96.81% 408
kimi_tool_call_schema_full/em_strict · c24 97.55% 408
kimi_tool_call_schema_full/em_strict · c32 97.79% 408
kimi_tool_call_schema_full/em_strict · c48 98.04% 408
kimi_tool_call_schema_full/em_strict · c56 98.77% 408
kimi_tool_call_schema_full/em_strict · c72 98.53% 408
kimi_tool_call_schema_full/em_strict · c80 97.55% 408
kimi_tool_call_schema_full/em_strict · c96 99.02% 408

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 vLLM 镜像从 vllm/vllm-openai:nightly-ac9126e58aa7bbab1856ba6593ba4d5003fea516 更新至 vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34。
**基线：**2026-10-02 · vllm/vllm-openai:nightly-ac9126e58aa7bbab1856ba6593ba4d5003fea516
AgentX · PTP8x1/DTP8x0 · 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。

@github-actions @claude
Update Kimi-K3 FP4 B200 Dynamo-vLLM AgentX DSpark image to v0.32.0
2e09b0c 
Move kimik3-fp4-b200-dynamo-vllm-agentic-dspark from vLLM nightly-ac9126e5 to
the v0.32.0 release, pinned by digest.

将 kimik3-fp4-b200-dynamo-vllm-agentic-dspark 从 vLLM nightly-ac9126e5 更新至
v0.32.0 正式版，并按 digest 固定。

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

github-actions Bot commented Oct 21, 2026

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 21, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Running · Run 39601503435 / attempt 1 · 2026-10-21 14:34 UTC
vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34 · 6046414d2d1a · AgentX · PTP8x1/DTP8x0 · Mean latency
Change: Measure 2e09b0c with only the master image moved from vLLM nightly-ac9126e5 to v0.32.0 (digest-pinned) and the recipe unchanged, because every flag, env var, attention backend and parser it uses exists unchanged in v0.32.0 source, which contains ac9126e5 as an ancestor (633 commits ahead), keeps CUDA 13.0.3 and flashinfer-python 0.7.0.post1 and bumps tokenspeed-mla from 0.1.8 to 0.2.15 (source: github.com/vllm-project/vllm/compare/ac9126e58aa7bbab1856ba6593ba4d5003fea516...v0.32.0).

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 N/A N/A N/A
4 N/A N/A N/A
8 N/A N/A N/A
14 N/A N/A N/A
24 49.28 (N/A) 897.36 (N/A) 11.05 (N/A)
32 55.9 (N/A) 1,005.11 (N/A) 13.43 (N/A)
48 N/A N/A N/A
56 N/A N/A N/A
72 N/A N/A N/A
80 N/A N/A N/A
96 68.07 (N/A) 1,900.54 (N/A) 83.01 (N/A)

Note: c1, c4, c8, c14, c48, c56, c72, c80: pending.
Note: c1, c4, c8, c14, c48, c56, c72, c80: request errors unavailable.
Note: c1, c4, c8, c14, c48, c56, c72, c80, c96: Δ N/A: point unavailable.
Note: c24: 220 request errors.
Note: c24, c32: Δ N/A: dataset/scenario/statistic mismatch.
Note: c32: 284 request errors.
Note: c96: 863 request errors.

Eval Score ↑ Samples
kimi_tool_call_schema_full/em_strict · c72 99.02% (N/A) N/A/408 (old/new)

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

Next: Wait for the 25 queued smoke jobs; c24, c32, c96 and the c72 eval already passed, and eval deltas show N/A because the new eval rows carry the PP2/DCP8 topology fields that the 2026-10-02 producer rows lacked, so their identity keys differ.

中文

初次尝试 · 运行中 · Run 39601503435 / attempt 1 · 2026-10-21 14:34 UTC
vllm/vllm-openai:v0.32.0@sha256:befccf0e7e4089aa23926bb6f887a85c0d46c2a99e6de5c5ffdf5cb7957c5b34 · 6046414d2d1a · AgentX · PTP8x1/DTP8x0 · 平均延迟
**变更：**测量 2e09b0c：仅将主配置镜像从 vLLM nightly-ac9126e5 更新至 v0.32.0（按 digest 固定），配方未改动，因为其使用的所有参数、环境变量、注意力后端和解析器在 v0.32.0 源码中均保持不变；该版本以 ac9126e5 为祖先（领先 633 个提交），CUDA 仍为 13.0.3，flashinfer-python 仍为 0.7.0.post1，tokenspeed-mla 从 0.1.8 升至 0.2.15（来源：github.com/vllm-project/vllm/compare/ac9126e58aa7bbab1856ba6593ba4d5003fea516...v0.32.0）。实测数值及异常说明见上表。
**下一步：**等待其余 25 个排队中的冒烟任务；c24、c32、c96 与 c72 评测已通过；由于新评测行带有 2026-10-02 生产者数据缺少的 PP2/DCP8 拓扑字段，身份键不同，评测差值显示为 N/A。

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@Ankur-singh Ankur-singh Awaiting requested review from Ankur-singh Ankur-singh will be requested when the pull request is marked ready for review Ankur-singh is a code owner
@kedarpotdar-nv kedarpotdar-nv Awaiting requested review from kedarpotdar-nv kedarpotdar-nv will be requested when the pull request is marked ready for review kedarpotdar-nv is a code owner
@wzhao18 wzhao18 Awaiting requested review from wzhao18 wzhao18 will be requested when the pull request is marked ready for review wzhao18 is a code owner
@nvpohanh nvpohanh Awaiting requested review from nvpohanh nvpohanh will be requested when the pull request is marked ready for review nvpohanh is a code owner
@RohitNagraj RohitNagraj Awaiting requested review from RohitNagraj RohitNagraj will be requested when the pull request is marked ready for review RohitNagraj is a code owner

At least 1 approving review is required to merge this pull request.

Assignees

No one assigned

Labels

None yet

Projects

InferenceX Board
Status: No status

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@Klaud-Cold