Pytorch is a good option here - ultimately it's just a library for efficiently manipulating tensors and supports cuda GPUs.