QLoRA - merging weights

#14
by ajmoreno - opened
Cognitive Computations org

Excellent work as always. I was curious how you are merging the QLoRA adapter weights. I know there was some discussion last year around the optimal way to merge the QLoRA adapter weights with the base model but doesn't seem as though a consensus was reached (a lot of results had shown the naive method often gave poor results). Thanks!

Cognitive Computations org
ehartford changed discussion status to closed

Sign up or log in to comment