Jan 3, 2022
Thanks! I've not tried PTQ before. I've just used dynamic and it has usually worked out pretty well.
What I find interesting is that the 22M extreme distilled model only has a slightly accuracy impact (92.8 to 91.2) while the 13M drops from 92.66 to 81.95). https://huggingface.co/bergum/xtremedistil-l6-h384-emotion