In order to decrease a GC pressure during training of machine learning models, we override the default julia allocator for arrays, such that these arrays are created inside the pre- allocated buffer, which is scrapped after each iteration. It seems to work for us rather good, but I am afraid this solution to be a bit fragile. For example I do not know, how this might interfere with Julia compiler, if I by accident hit a function which needs to be compiled ( our data are not tensors and they can come in many types due to missing values). I would like to therefore ask, if there is a some way to verify, that Julia has already seen this function with such combination of arguments.
Thanks for the answer,