Speedup bert model inference

DeepPavlov Library
#1

What are good and simple ways to make it fast enough on a server?