to clarify i have been working on https://raw.githubusercontent.com/karl3wm/httptransformer/refs/heads/main/ne... which is part of the git repository https://github.com/karl3wm/httptransformer , the functionality of which is presently centered around test_nettensors.py . This repository is focused on evaluating part or all of large language models that are too large to fit on a local system, using huggingface transformers which usually assumes models are held locally. My current changes are not present in the repository as they do not run yet, although I did recently push a newer version of netsafetensors.py from my in-progress work and i have not tested the current state of the repository (it may or may not need a rollback to some earlier commit for test_nettensors.py to run). this is hobby code, i've produced code like this a lot in recent years. if i engage it persistently enough it's stabilizable. so, `missing_idcs = (next_hole < tails[idx:]).nonzero()[:,0]` was ported from https://github.com/karl3wm/httptransformer/blob/main/netsafetensors.py#L164 ie `if next_hole < tail`. The original check `if next_hole < tail` was focused on a single length of data read by the user, and checking whether or not there was a hole partway through it. I vectorized this check with `(next_hole < tails[idx:])` to perform it for every requested read in a sorted batch. But I don't think this logic is likely to be correct as if the reads are non-overlapping then the start of the next hole will only overlap one of them, and the others will instead be performing an ordering check rather than a check for being uncached.