Tensorflow's AdditiveAttention implementation does not have weights

Question

I am trying to understand how new (custom) Layers are implemented in Tensorflow/Keras. Bahdanau\'s Additive Attention seems pretty straight forward. Pa