In the nn benchmark, when the profile option is passed, CUDA.compilecache is called which is not defined in the CUDA package. Maybe it was deprecated?