Meekail Zain

> One more point on index dtypes I'd like to bring up, can the `indptr` a `indices` have different dtypes? E.g. `indices_dtype` and `indptr_dtype`? As it stands right now, I...

Using `py-spy` helped me figure out that `ArgKminLabels` was not using any euclidean specialization 🤦 , whereas `ArgKmin` was, therefore it was an unfair comparison. I've added a new `EuclideanArgKminLabels`...

> Still, remapping this case to the `metric="euclidean"`-case in `PairwiseDistancesReductions` to use `EuclideanPairwiseDistancesReductions` is doable: is this what you are suggesting, @Micky774? Indeed, I think that the expectation would be...

Thanks @jjerphan for your PR. @ogrisel do you think it is ready for merger? Edit: I believe I'm still missing a changelog entry, but will add one soon.

@jjerphan Updated with your suggestions! Good catch on the extra sort -- it is a vestige from the original copy-and-paste :) Edit: Still need to revisit validation and re-run benchmarks...

> Never downcast indices types based on values in the indices array I wonder if we could preserve that as existing behavior with e.g. `indices_dtype=None` It is helpful in many...

I see, I think I had misunderstood the scope of the change. So to cement my understanding, this is meant as a first-step which essentially enforces that sparse arrays "trust...

We've added an [FAQ section](https://jax.readthedocs.io/en/latest/faq.html#why-do-some-cuda-libraries-fail-to-load-initialize) addressing various CUDA library loading issues and solutions/workarounds, and have (hopefully) made it easier to find by including it in some error messages that often...

This seems to be an instance of a more general problem w/ trying to load CUDA libraries with insufficient memory due to JAX, TF, or other libraries pre-allocating too much...

@hawkinsp LMK if you still have changes you'd like to see