0.4.2

fix bug with residual attention

0.4.1

for post-normalization, let wrapper take care of last normalization

0.4.0

add residual attention, from Realformer paper

0.3.5

do GLU gating for attention layer output, without queries

0.3.4

create a cross attention only attention layer (CrossAttender)

0.3.3

allow for only cross attention in attention layers

0.3.2

allow encoder to cross attend

0.3.1

default to topk sampling

0.3.0

bump version

0.2.5

fix talking heads