Summary
Currently the model compression still need the training dataset to compute the range of the embedding matrix input, and to determine the range of tabulation. Saving the range of the embedding matrix input in the DP model will remove this dependency, so the user can still compress the model in the case that the original training data is not available.