This
method takes care of encoding the input and feeding the encoded hidden states via cross-attention layers to the decoder
and auto-regressively generates the decoder output.