0.0.53

add first pass of palm encoder decoder

0.0.52

some reorganization, to ready for encoder / decoder

0.0.51

some reorganization, to ready for encoder / decoder

0.0.50

some reorganization, to ready for encoder / decoder

0.0.48

fix masking logic when using palm as encoder

0.0.47

make sure key padding mask is in effect if training reward model as e…

0.0.46

turn off xpos if using palm as encoder

0.0.45

fix pooled critic values during generation, thanks to @Nightbringers

0.0.44

able to override lora R value when adding a new finetuning scope

0.0.43

make sure xpos scale base value is customizable from palm init