Just FYI, but Torch supports OpenCL - which I have played with since that "should" allow models to be worked with on either card rather than having a CUDA and an AMD version...

Or you can run CPU only - if you have an eternity to let your training loop run.