While doing compiler development, we’re always thinking about how useful it would be to have the dimensions of various for loops and data collections in order to optimize and generate better code. The downside of that is that we would then need to recompile the model every time the data sizes changed. I was talking with someone the other day who asked when it was useful to use the same model on differently sized data, and I responded that sometimes people develop models with a subset of the data. This person then pointed out that in that phase of model development a Stan user is still iterating and recompiling a lot, so they’re recompiling typically every iteration anyway.
Does anyone have compelling use-cases for compiling models that are agnostic to data sizes?