OK, then I could write an app that would go and fill out all the little nooks and crannies and then dump a summary. It sounds like it’d be the easiest to do directly in C to avoid any overheads that might get incurred by using pytorch.
I’m not sure how could that help with detecting fragmentation. Unless you mean looking at mallocs that are lesser than a page size and then accounting the remaining fragments, but that sounds like it’d be very unreliable/difficult to accomplish since it’d require an exact understanding of how CUDA does things internally. Basically this is what I did in my 2nd post, except on the macro-level.
I think this is diverging from the topic now, so I will create a new one for that and keep this one focused.