# # # 3) Similarly, self-attention layers in the decoder allow each # position in the decoder to attend to all positions in the decoder up # to and including that position. We need to prevent leftward ...