0.12.9

fix bug with reversible networks and redrawing projections on forward

0.12.8

update enwik8 example to use automatic mixed precision, and fix a sma…

…ll bug with switching between amp and non-amp

0.12.7

change variable name to amp_enabled

0.12.6

make sure EPFLs cuda code can work with amp autocast, by setting amp_…

…autocast to True

0.12.5

fix context mask in autoregressive wrapper again

0.12.4

fix bug with how autoregressive wrapper handles contexts

0.12.3

cast projection matrix type to be same as query

0.12.2

get around EPFLs CUDA kernel being float only

0.12.1

clean up encoder / decoder code

0.12.0

add encoder / decoder architecture from @gulnazaki