4.2.3
- 新增单文件执行超时与托管恢复。WebUI、Router 和 API Server 支持
--task-timeout,默认3600秒,从取得解析名额开始计时,覆盖解析和服务端产物打包。超时任务返回task_timeout,独立 API 停止接收新任务,由部署方重建服务。Router 通过父进程监督回收超时 worker 及模型子进程,并重新启动服务,即使执行线程、事件循环或取消清理挂起也能触发恢复。 - 改善 WebUI 恢复体验与进程清理。本地 WebUI 由 Router 托管唯一 worker,恢复期间显示状态,尚未提交的排队请求在服务就绪后继续。旧 worker 上其他未完成请求返回
worker_restarted,问题文件可手动重试。完善 Windows 进程树、临时目录与控制通道回收,明确报告模型预加载失败,并在连续三次启动恢复失败后停止自动重试。 - 修复 Doclib 全文检索被图片载荷挤占。构造检索文本时提取正文、公式和图表说明,再应用字符预算,避免内嵌图片的 base64 内容占满索引而遗漏正文。升级后从解析缓存重建已有索引,并保留跨页内容续接。缓存不完整的文档会记录迁移原因,补齐缓存或重新解析后重启 Doclib 可继续迁移。
- 提前检查 vLLM 与 Transformers 兼容性。启动引擎前识别
vllm<0.29.0与transformers>=5.23.0的不兼容组合,并给出升级 vLLM 或调整 Transformers 版本范围的明确提示,减少模型加载阶段难以定位的失败。 - 更新 DocVortex 依赖。最低版本提升至
0.6.24,同步原生文档解析与共享渲染能力的改进。
配置提示。Router 的 --worker-tier 默认改为 auto,由各 worker 根据自己的设备环境确定档位上限。需要固定上限时可显式指定 flash、basic、standard 或 advanced。--task-timeout 必须为正整数,上传、排队、客户端下载及模型启动预加载各自使用独立的等待与生命周期设置。
4.2.3
- Add per-file execution timeouts and managed recovery. WebUI, Router, and API Server support
--task-timeout, defaulting to3600seconds from acquiring an execution slot through parsing and server-side artifact packaging. Timed-out tasks returntask_timeout. A standalone API stops accepting new work and relies on its deployment host to recreate the service. Router supervises local workers from the parent process, reclaims timed-out workers and model subprocesses, and restarts the service even when execution threads, the event loop, or cancellation cleanup stall. - Improve WebUI recovery and process cleanup. Local WebUI uses Router to manage a single worker, displays recovery status, and resumes queued requests awaiting submission once the service is ready. Other unfinished requests on the replaced worker return
worker_restarted, and affected files can be retried manually. Improve Windows process-tree, temporary-directory, and control-channel cleanup, surface model preload failures, and stop automatic recovery after three consecutive startup failures. - Keep image payloads out of Doclib full-text indexes. Build search text from body content, formulas, and figure or table descriptions before applying the character budget, preventing embedded image base64 data from crowding out searchable text. Rebuild existing indexes from cached parse results during upgrade while preserving cross-page continuations. Documents with incomplete caches record a migration reason and can resume migration after restoring or reparsing the cache and restarting Doclib.
- Check vLLM and Transformers compatibility before startup. Detect the incompatible combination of
vllm<0.29.0andtransformers>=5.23.0before engine initialization, with actionable guidance to upgrade vLLM or select a compatible Transformers range. - Update DocVortex. Raise the minimum version to
0.6.24for native document parsing and shared rendering improvements.
Configuration note. Router now defaults --worker-tier to auto, allowing each worker to select its tier ceiling from its device environment. Specify flash, basic, standard, or advanced for a fixed ceiling. --task-timeout requires a positive integer. Uploads, queueing, client downloads, and model startup preloading use their own waiting and lifecycle settings.
升级 / Upgrade
pip install -U "mineru==4.2.3"