I've found that there is usually very little code to be shared within training and inference when most of the pre and post processing is pulled into the model itself with custom operations. Relatively few people are aware that you can write training code in C++ today with torch to share a common language with inference and training, but very few take advantage of this in practice.