I’m concerned about this as well, and the generate() method in the transformers library explicitly suggests that decoder-only models should use the left padding method. I would also like to know the reason for this
1 Like