0.2.1

fix a bug with the final norm in palm, thanks to @conceptofmind and @…

…dmahan93

0.2.0

address https://github.com/lucidrains/PaLM-rlhf-pytorch/issues/41 , b…

…e faithful to the paper

0.1.4

old action log probs should be the true distribution in the kl div lo…

…ss, addressing https://github.com/lucidrains/PaLM-rlhf-pytorch/issues/43

0.1.2

flash attention sdp context config only needs to be done once

0.1.1

fix assert

0.1.0

add ability to use flash attention if using pytorch 2.0, thanks to @c…

…onceptofmind for the initial PR!

0.0.68

0.0.68

0.0.67

fix silly error in masked kl div loss, thanks to @taynoel84

0.0.66

allow for setting critic palm from rlhftrainer

0.0.65

fix an error with the way action log prob is collected during the epi…

…sode rollouts, addressing https://github.com/lucidrains/PaLM-rlhf-pytorch/issues/31 and thanks to @kisseternity