Fetching the paper…

Train Large, Then Compress: Rethinking Model Size for Efficient Training and Inference of Transformers · Around