Not sure how much effort it would take to support compiling CUDA code. Writing a new model for `nvcc` and friends?