don't need attn weight in decoder

[buaahsh]

Attention weight was originally designed for alignment models, so it is not necessary to include it in torchscale.