Today I happened to notice that for bart_large.save_pretrained(x) the model is 1.6GB on disk but the pretrained size (both in the local cache and on the model Hub) is 972M. However, for bart-base the trained size and pretrained sizes are the same. Looking at the difference between the model sizes, I think the 1.6GB is probably the “right” size. Anyone know how they are compressing the bart-large down to 972M for storage on the Hub and is there a way to do this to save space for my bigger trained models?