0.0.36

add token shifting

0.0.35

0.0.35

0.0.34

use dataset.repeat()

0.0.33

make mixed precision policy customizable

0.0.32

make data parallelism magically work for any batch size

0.0.32

0.0.29

make sure mixed precision works

0.0.28

0.0.28a

cleanup

0.0.27

update optax to 0.0.9 and use weight decay

0.0.26

make sure data parallelism can be toggled on through command line

0.0.25

add joblib in preparation for parallel processing