YORU v1.1.2

Improvements

• Build the uv environment on macOS and Linux as well as Windows, from a single lockfile. uv sync previously aborted before resolving a single package on anything but Windows (#41)
• Pick the compute device automatically -- CUDA, then Apple MPS, then CPU -- with a YORU_DEVICE override, a --device flag on both training scripts, and a selector in the training GUI. Apple Silicon users were silently training and inferring on the CPU before this (#41)
• Add CI on Windows, Linux and macOS, and turn the smoke tests into a real train/detect round trip through YORU's own code paths (#41)
• Write runtime errors and work logs to a ~/.yoru user directory, relocatable with YORU_HOME (#38, #40)
• Launch the GUIs asynchronously and surface their errors instead of swallowing them (#32, #36)
• Add train_ultralytics.py (#33)

Bug Fixes

• Fix the error-handling gaps found reviewing the exception-handling work (#39)
• Bound numpy to <2: opencv-python 4.10.0.82 is built against the numpy 1.x ABI, and an unbounded numpy broke import cv2 on Python 3.10+ (#41)
• Tell hatchling that the package directory is yoru, which it could not infer from the distribution name yoru_uv (#41)
• Stop the vendored YOLOv5 from gating MPS on the deprecated torch.has_mps (#41)

Documentation

• Document what the installer cannot install: a Chromium browser for the launcher, the NVIDIA driver (the CUDA toolkit is not required -- the PyTorch wheels carry their own runtime), and the Apple Silicon toolchain and macOS permission prompts (#41)
• Give the uv route its own instructions, including how to install uv itself (#41)
• Fix typos and broken links in the docs (#37)

Platform support

macOS is supported for the first time: macOS 14 (Sonoma) or later on Apple Silicon. Windows and Linux resolve exactly as before (torch 2.6.0+cu124, torchvision 0.21.0+cu124).

Full Changelog: v1.1.1...v1.1.2