There's a CUDA backend for MLX now. Not sure about the maturity.