Repository navigation
Folio 0.1.4
Folio 0.1.4
- OCR automatically reduces GPU memory use when full GPU loading or inference fails: memory fitting places language layers and their attention cache in RAM while retaining GPU image encoding, followed by GPU image encoding alone and CPU-only recognition if needed. The 16,384-token context is preserved.
- Faster OCR preparation: verified assets and Vulkan device discovery are reused across recognition requests, cancellation and server restarts.
- Recognition status identifies memory fitting, GPU image encoding and CPU execution. Mixed modes can run more slowly than full GPU inference; performance depends on the hardware and available memory.
Windows x64 setup EXE and portable ZIP, Linux x86_64 Flatpak, checksums and complete corresponding source are published together. Updates remain signed and user-initiated.
Database schema remains 6 and document format remains 4. See UPDATES.md for installation and recovery.
Validation: Linux and native Windows CI passed (353 and 342 Rust tests respectively, plus updater, packaging and runtime checks). Real local OCR passed on NVIDIA and AMD hardware, including fitted and GPU-image-encoder-only modes. A forced fitting budget kept 11 of 18 language layers on the GPU with the full 16,384-token context. The signed Flatpak upgrade from 0.1.3 preserved a disposable library; packaged offline math, PDF previews, AppStream metadata and a private native Wayland smoke test also passed.
Validation limits: the constrained-memory test used the runtime fitting budget rather than physically restricting VRAM. Intel GPU execution and Windows OCR on physical GPU hardware were not verified. The Windows installer GUI and full in-app notification/download/relaunch flow were not exercised for this release.