0.4.5

add a learned null query output token, for those query tokens not sel…

0.4.4

fix issue with masking, but gradient error is still happening very in…

0.4.3

oops, logic was not correct, autoregressive CoLT5 now works..

0.4.2

fix an issue with masking of initial b in coor descent, update triton…

0.4.1

simplify initial b for easier backwards in coor descent, still an iss…

0.4.0

wire up ability to use triton optimized version of coordinate descent…

0.3.8

make sure it works with masking

0.3.7

a little refactoring

0.3.6

allow for passing in k as int or float

0.3.4

cleanup