0.7.5

fix renormalizing attention for the causal case

0.7.4

tie embeddings to logits by default

0.7.3

fix local attention head number validation

0.7.2

0.7.2a

1.7.2

0.7.1

cleanup sequential sequence module

0.7.0

separate projections of queries keys values so it is easier for someo…

0.6.0

bump for release

0.5.1

offer way to fix projection matrices at the end of training

0.5.0

add dropouts and GLU variant for feedforward

0.4.0

make it possible for gpt-like architecture