[Klaud Cold] Update minimaxm3-fp8-h200-vllm-agentic-mtp vLLM image to nightly-3df4ae15 / 将 minimaxm3-fp8-h200-vllm-agentic-mtp 的 vLLM 镜像更新至 nightly-3df4ae15

[adibarra]

AI model disclosure

Prepared by Claude Fable 5.1 (claude-fable-5-1) running as Klaud Cold in the klaud-candidate.yml workflow: upstream image and source research, config edit, benchmark dispatch and reporting. No other models or delegated agents.

中文

本 PR 由 Claude Fable 5.1（claude-fable-5-1）以 Klaud Cold 身份在 klaud-candidate.yml 工作流中完成：上游镜像与源码调研、配置修改、基准调度与报告。未使用其他模型或委派代理。

[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 only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• 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：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **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。

[adibarra]

Initial attempt 0/5 · running (evals passed, throughput running) · Run 35570778795 / attempt 1 · 2026-09-21T06:58Z
vllm/vllm-openai:nightly-3df4ae153eb385e27b52f26c81f8edb9e20b9984@sha256:1e1f56a164a3dfdf87a57168983602421a295b0b3570d27c5b49c54b97b341b4 · e7b92e1c89825630be2d95075b01474091bcad95 · AgentX · TP8/EP1 · Mean latency
Change: Move minimaxm3-fp8-h200-vllm-agentic-mtp from vllm/vllm-openai:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 (tag commit d9105ea8, 2026-09-07; the tag now returns 404 on Docker Hub, so fresh nodes cannot import it) to the 2026-09-21 nightly (tag commit 3df4ae15, image label org.opencontainers.image.revision matches, manifest-list digest sha256:1e1f56a1…, CUDA 13.0.2, torch 2.13.0 unchanged). benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh, the c1–c12 grid and the golden AL 2.78 are unchanged.

Why not v0.29.0: the release branch forked at f5c3cc24, which is 300 commits behind the current nightly commit, so the release would roll the bundled engine backwards (same reason #3131 stayed on a nightly).

Source review, d9105ea8..3df4ae15 (748 commits): every recipe flag/env still parses (--linear-backend humming, --language-model-only, --max-cudagraph-capture-size, --default-chat-template-kwargs, --enable-prompt-tokens-details, rejection_sample_method/synthetic_acceptance_length, VLLM_MARLIN_USE_ATOMIC_ADD, MooncakeStoreConnector with load_async, minimax_m3 tool/reasoning parsers). The 0.29 deprecation sweep (#55353) removed only use_fp4_indexer_cache, VLLM_PREFIX_CACHE_RETENTION_INTERVAL and VLLM_MM_HASHER_ALGORITHM, none used here. Relevant behaviour changes: Humming MXFP8 linear path refactor and humming-kernels 0.1.12→0.1.15 (#56685, #57421 shared Marlin/Humming workspaces); MooncakeStore now offloads only prefix-cacheable KV groups and fixes the align-mode check (#55027, affects the c12 DRAM point); EAGLE fixes (#56734, #57356, #53458: draft_id_to_target_id only when draft vocab differs); flashinfer 0.6.18→0.6.18.post1, nvidia-cutlass-dsl 4.6.2→4.7.1, quack-kernels 0.6.4→0.6.5 (MSA decode kernels). MiniMax-M3 NVIDIA model code changed only through the indexer SparseCacheRole plumbing; the other MiniMax-M3 commits are ROCm-only. The draft Inferact/MiniMax-M3-EAGLE3-GQA (LlamaForCausalLMEagle3, bf16) still maps to llama_eagle3. Provenance: image ai.vllm.build.commit label and Buildkite release-v2 build 6971.

Smoke: test-config --config-files configs/nvidia-master.yaml --config-keys minimaxm3-fp8-h200-vllm-agentic-mtp --smoke on ref e7b92e1c → c1 (resident) and c12 (Mooncake DRAM) throughput plus their minimax_m3_full evals, fail-fast, Klaud background priority.

Frozen public baseline (2026-09-15, nightly-d9105ea8, producer run 34896669482 at 5b9f51d4, dataset semianalysis_cc_traces_weka_062126):

Concurrency	Total tok/s/GPU	Output tok/s/GPU	TTFT ms	TPOT ms
1	2,170.5	16.28	1,318.6	3.61
2	2,351.1	19.49	926.8	3.98
4	3,433.5	25.36	932.3	4.53
6	4,110.3	33.13	920.6	5.28
8	5,839.3	44.47	892.0	6.71
10	6,757.4	51.57	1,370.1	9.21
12 (Mooncake DRAM)	7,023.4	56.01	1,064.2	8.11

Published evals for that run: minimax_m3_smoke/em_strict 100% at c1–c12, 1 sample each (the family now runs minimax_m3_full, so eval deltas will be N/A).

Progress 07:54Z: both eval-only jobs passed on the new image. Server log: engine v0.29.1rc1.dev452+g3df4ae153, HummingMxfp8LinearKernel for MXFP8 GEMM, MARLIN MxFp8 MoE backend, EAGLE3 auxiliary layers (2, 30, 57), MooncakeStoreConnector embedded mode with a 93.4 GB global segment at c12; cold weight load took ~890 s. Throughput jobs for c1 and c12 are still running.

Eval	Score ↑	Samples
minimax_m3_full/em_strict · c1	99% (Δ N/A: baseline suite was minimax_m3_smoke)	102
minimax_m3_full/em_strict · c12	99% (Δ N/A: baseline suite was minimax_m3_smoke)	102

Blocker (needs a maintainer): python -m infx.klaud report --kind baseline fails with Candidate ownership mismatch. infx/klaud/lifecycle.py requires the candidate PR author and dispatch actor to be the Klaud-Cold account, but every GitHub token supplied to this candidate job (AGENT_PAT) resolves to a different user, so this PR and its run are not recognised as owned. The previous wave hit the same error (run 35547715376, candidate for #3325). Consequences: the typed baseline/attempt reports cannot be published, finish cannot verify or clean up, and recovery cannot release klaud/claim-b3fff31b6f9a6956. Restoring AGENT_PAT to the Klaud-Cold account (or relaxing the owner check) is required before any Klaud candidate can complete.

Next: Watch the smoke run and post its results here, then apply the documented unsuccessful-session cleanup (close, delete branch) by hand since finish cannot.

中文

初始尝试 0/5 · 运行中（评测已通过，吞吐运行中） · Run 35570778795 / attempt 1 · 2026-09-21T06:58Z
vllm/vllm-openai:nightly-3df4ae153eb385e27b52f26c81f8edb9e20b9984@sha256:1e1f56a1… · e7b92e1c · AgentX · TP8/EP1 · 平均延迟
**变更：**将 minimaxm3-fp8-h200-vllm-agentic-mtp 的镜像从 2026-09-07 的 nightly（tag 提交 d9105ea8，该 tag 在 Docker Hub 已返回 404）更新为 2026-09-21 的 nightly（tag 提交 3df4ae15，按 manifest-list digest 固定，CUDA 13.0.2，torch 2.13.0 不变）。配方脚本、c1–c12 网格与 golden AL 2.78 均不变。未选择 v0.29.0：其发布分支从落后当前 nightly 300 个提交的 f5c3cc24 分出，会使引擎回退。源码对比（748 个提交）中所有配方参数与环境变量仍存在；相关变化见上文英文段落与链接。
阻塞（需维护者处理）：report --kind baseline 报 Candidate ownership mismatch。生命周期助手要求 PR 作者与调度账号为 Klaud-Cold，但本次任务提供的 AGENT_PAT 解析为其他用户，因此本 PR 与运行不被识别为自有；上一波候选（#3325，run 35547715376）出现相同错误。类型化报告、finish 与恢复流程均无法执行，也无法释放 klaud/claim-b3fff31b6f9a6956。需将 AGENT_PAT 恢复为 Klaud-Cold 账号后才能完成任何 Klaud 候选。
**进展 07:54Z：**两个仅评测任务已在新镜像上通过（见上表）；c1 与 c12 吞吐任务仍在运行。
**下一步：**跟踪 smoke 运行并在此发布结果，然后手动执行文档规定的失败会话清理（关闭 PR、删除分支）。数值及基线见上表。

[adibarra]

Initial attempt 0/5 · passed (smoke) · Run 35570778795 / attempt 1 · completed 2026-09-21T09:03Z
vllm/vllm-openai:nightly-3df4ae153eb385e27b52f26c81f8edb9e20b9984@sha256:1e1f56a164a3dfdf87a57168983602421a295b0b3570d27c5b49c54b97b341b4 · measured e7b92e1c89825630be2d95075b01474091bcad95 · AgentX · TP8/EP1 · Mean latency · dataset semianalysis_cc_traces_weka_062126
Change: as in the first comment (image bump only; recipe unchanged). All four smoke jobs succeeded on h200-dgxc-slurm_00/01/02.

Point	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓	Total tok/s/GPU ↑	Requests ok/total
c1 resident · new	16.23	1,291.0	3.58	2,168.3	248/259, 0 errors
c1 resident · baseline 2026-09-15	16.28	1,318.6	3.61	2,170.5	errors unavailable
c12 Mooncake DRAM · new	56.14	1,023.9	8.56	7,052.8	1,487/1,618, 2 errors
c12 Mooncake DRAM · baseline 2026-09-15	56.01	1,064.2	8.11	7,023.4	errors unavailable

Eval	Score ↑	Samples
minimax_m3_full/em_strict · c1	99% (Δ N/A: baseline suite was minimax_m3_smoke)	102
minimax_m3_full/em_strict · c12	99% (Δ N/A: baseline suite was minimax_m3_smoke)	102

Deltas are left to the canonical renderer, which cannot run here (see below); values are copied from the bmk_agentic_* and eval_results_all artifacts via infx.klaud.reporting.values. Notes: c1 server log ends with an EngineDeadError emitted during the SIGTERM teardown after the 3600 s window (not a serving failure); c12 logged ten Failed to get … Mooncake keys warnings within 07:58:18–07:58:28, i.e. first-lookup misses right after startup, then none. Power valid on both points.

Changelog: appended the minimaxm3-fp8-h200-vllm-agentic-mtp entry at d98915cf1841eb6e7ea679739b2a0f4f85dccbfc; utils/process_changelog.py --base-ref origin/main --head-ref d98915cf generates the family (7 points c1–c12, 7 minimax_m3_full evals; identical to the base roster apart from the image).

Stop reason · unexpected-error: check-final and finish fail with Candidate ownership mismatch (the lifecycle helper only recognises PRs and dispatches made by the Klaud-Cold account; the token supplied to this job is a different account, same as the previous wave's #3325). Without check-final the final sweep must not be dispatched, and without finish no verified outcome can be recorded, so no full-sweep-fail-fast label was applied and no GPU sweep was started. Following the unsuccessful-session policy, this PR is now closed by hand and its branch deleted (GitHub's Restore branch keeps d98915cf recoverable). The family claim klaud/claim-b3fff31b6f9a6956 is left in place for the helper/maintainers; automated recovery for this parent run will keep failing on the same ownership check.

To resume: restore AGENT_PAT to the Klaud-Cold account, delete refs/heads/klaud/claim-b3fff31b6f9a6956 (and the sibling claims from run 35547715376), then let the next planner wave re-select the family; or restore this branch, reopen, and label full-sweep-fail-fast manually, since the image already passes smoke and the old tag nightly-d9105ea8 is no longer pullable from Docker Hub.

中文

初始尝试 0/5 · smoke 通过 · Run 35570778795 / attempt 1 · 2026-09-21T09:03Z 完成
新镜像 nightly-3df4ae15…@sha256:1e1f56a1… · 测量提交 e7b92e1c · AgentX · TP8/EP1 · 平均延迟 · 数据集 semianalysis_cc_traces_weka_062126
**变更：**同首条评论（仅更新镜像，配方不变）。四个 smoke 任务全部成功；数值及基线对照见上表，评测得分见上表。c1 日志末尾的 EngineDeadError 出现在 3600 s 结束后的 SIGTERM 关停阶段，不是服务故障；c12 在启动后 10 秒内记录了 10 条 Mooncake 首次查找未命中的警告，之后再无。
**变更日志：**已在 d98915cf 追加 minimaxm3-fp8-h200-vllm-agentic-mtp 条目，生成的配置族为 7 个点与 7 项 minimax_m3_full 评测，除镜像外与基线一致。
停止原因 · unexpected-error：check-final 与 finish 均因 Candidate ownership mismatch 失败（生命周期助手仅识别 Klaud-Cold 账号创建的 PR 与调度，本任务的令牌属于其他账号，与上一波 #3325 相同）。缺少 check-final 不得调度完整 sweep，缺少 finish 无法记录已验证结果，因此未添加 full-sweep-fail-fast，未启动 GPU sweep。按失败会话策略手动关闭本 PR 并删除分支（可通过 GitHub 的 Restore branch 恢复 d98915cf）。配置族声明 klaud/claim-b3fff31b6f9a6956 保留给助手/维护者处理。
**恢复方式：**将 AGENT_PAT 恢复为 Klaud-Cold 账号并删除相关 claim 引用后由下一波 planner 重新选择；或恢复分支、重新打开并手动添加 full-sweep-fail-fast，因为新镜像已通过 smoke，且旧 tag nightly-d9105ea8 已无法从 Docker Hub 拉取。