Cygwin isn't remotely comparable in terms of compatibility. With WSL2 I can train stuff with PyTorch with GPU acceleration without configuring anything. Try doing that with cygwin.