TF32 torch.backends.cudnn.enabled: it enables cudnn for some operations such as conv layers and RNNs, which can yield a significant speedup.