You need to compile from source but PyTorch and Tensorflow now support ROCm in their build configuration. Numpy never supported GPUs and the most advanced Nump

You need to compile from source but PyTorch and Tensorflow now support ROCm in their build configuration. Numpy never supported GPUs and the most advanced Numpy-like was Cupy supported by the Chainer team which stopped development in favor of PyTorch. Now I'm unsure how well Rocm is supported with Apple drivers as Apple always customized their drivers.