should the multihead_attention of decoder use causality=False?

[gitfourteen]

In my understanding, Harvard Annotated Transformer only apply _subsequentmask, so called causality here, to decoder self-attention part.