Local Attention does not introduce any new parameters
to the model.