[cs702]

Yes, it makes sense to apply this only to the Softmax we use to compute attention. It makes no sense to apply it to the output Softmax, which must compute a probability distribution over the vocabulary.