[Klaud Cold] Update dsv41flash-fp4-gb300-vllm-agentic-dspark vLLM image to nightly-dee37d89 / 将 dsv41flash-fp4-gb300-vllm-agentic-dspark 的 vLLM 镜像更新至 nightly-dee37d89

[adibarra]

AI model disclosure

• Claude Fable 5.1 (claude-fable-5-1), running autonomously as Klaud Cold, prepared every part of this PR: the image edit, the perf-changelog entry, benchmark dispatch and monitoring, diagnosis and the reports. No other model or delegated agent contributed.

中文

AI 模型披露

• Claude Fable 5.1（claude-fable-5-1）以 Klaud Cold 身份自主完成了本 PR 的全部工作：镜像修改、perf-changelog 条目、基准测试的触发与监控、诊断及报告。没有其他模型或委派代理参与。

🤖 Generated with Claude Code

[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 · readiness-blocked (no dispatch) · head e26f48d4bf379bf1bee57ea7d446461d5d57b667 · 2026-09-18 UTC
vllm/vllm-openai:nightly-dee37d89115db4c94a820a79a78a7828e141c910@sha256:dea7fa047caa114167efccdeb42321b12d2c11da1add8728aa2662cb8d6b8cd5 · AgentX · TP4/EP1
Change: Update the vLLM image from nightly-af1c0149 (2026-09-16) to nightly-dee37d89 (2026-09-18, 160 upstream commits, pinned to its manifest digest; the image labels record build commit dee37d89 from vLLM release pipeline build 6858), keeping every recipe flag: v0.29.0 (98dff2a8) contains no deepseek_v41 tokenizer/parser or engram config, so the release is not a compatible target, while the new nightly makes FlashMLA mega attention with the NVFP4 compressed KV cache the SM100 default (#56935), integrates Mega-Gate from DeepGEMM (#56266, pin bump #57218), fixes FlashInfer DSpark non-causal attention (#57432) and NaN-scored candidate blocks (#57454), normalizes the Rust-frontend reasoning controls with thinking-on still the default (#56998), and adds --tool-strict-level (default auto, #56268) and --max-num-active-seqs (default unset, #56758).
Blocker: The GitHub credential supplied to this session authenticates as a human account rather than the Klaud-Cold bot that the lifecycle helper requires as PR owner, so report, check-final and finish all fail with "Candidate ownership mismatch" (the same failure stopped the earlier candidate today, and every klaud/auto-* PR since 2026-09-18 06:44 UTC carries the human author). No GPU work was dispatched. The public baseline was frozen locally (eight AgentX points and the gsm8k c128 eval from run 35073891536, dated 2026-09-16) but could not be published through the helper.
Next: Close this draft and release the branch so the candidate can be retried once the agent credential is restored to the bot account.

中文

初次尝试 0/5 · 就绪受阻（未触发运行） · head e26f48d4bf379bf1bee57ea7d446461d5d57b667 · 2026-09-18 UTC
vllm/vllm-openai:nightly-dee37d89115db4c94a820a79a78a7828e141c910@sha256:dea7fa047caa114167efccdeb42321b12d2c11da1add8728aa2662cb8d6b8cd5 · AgentX · TP4/EP1
**变更：**将 vLLM 镜像从 nightly-af1c0149（2026-09-16）更新为 nightly-dee37d89（2026-09-18，上游 160 个提交，按 manifest digest 固定，镜像标签记录构建提交 dee37d89），配方参数全部保留；v0.29.0（98dff2a8）不含 deepseek_v41 分词器/解析器和 engram 配置，因此发布版不是兼容目标。新 nightly 在 SM100 上默认启用带 NVFP4 压缩 KV cache 的 FlashMLA mega attention，集成 DeepGEMM Mega-Gate，修复 FlashInfer DSpark 非因果注意力与 NaN 候选块问题，规范 Rust 前端推理控制（默认仍为开启思考），并新增 --tool-strict-level（默认 auto）与 --max-num-active-seqs（默认未设置）。
**阻塞：**本会话获得的 GitHub 凭据以人类账号而非生命周期 helper 要求的 Klaud-Cold 机器人身份认证，因此 report、check-final 与 finish 均报 "Candidate ownership mismatch"（今日更早的候选也因同一原因停止，2026-09-18 06:44 UTC 之后所有 klaud/auto-* PR 均为该人类作者）。未触发任何 GPU 运行。公开基线已在本地冻结（run 35073891536 的八个 AgentX 点及 gsm8k c128 eval，日期 2026-09-16），但无法通过 helper 发布。
**下一步：**关闭此草稿并释放分支，待代理凭据恢复为机器人账号后重试该候选。