Comment by bonoboTP

It's very possible but somewhat slower. PyTorch and CUDA have flags for determinism. It won't work across all different GPU models though, but it will get you bitwise equal results on the same GPU.