Suppose we want to train text prediction models in email clients or word
processors. The models must preserve the privacy of user data and adhere to a
specific fixed size to meet memory and inference time requirements. We
introduce a generic framework to solve this problem. Specificall