recurrent decoding example

[Howuhh]

Hi! Very nice and clean implementation. In the paper they state that the model preserves fast inference with constant memory. Unfortunately, I don't quite understand how this is possible with attention and local convolution.

Would it be difficult to add an example?