On first launch the app sets up its environment and downloads the right build of PyTorch for your GPU - this takes a few minutes and needs an internet connection.
Training uses your GPU: NVIDIA (CUDA) on Windows and Linux, AMD (ROCm) on Linux. No compatible GPU falls back to CPU (small models / testing only).
The installers are not yet code-signed, so Windows SmartScreen may warn "unknown publisher" - this is expected for a new app. Choose "More info" → "Run anyway".
macOS build is not offered here: Apple GPUs are not a supported training target yet.