Evgeni Burovski

IIUC the classic approach would be to test on torch CPU + array-api-strict in some CI. This will ensure that the array api compliant code runs on torch GPU and...

Sorry for dropping the ball here! All in all all workflows you're discussing do sound exactly like a use case for `array-api-strict`. It indeed currently does two things: 1) check...

> I also want (3) a way to temporarily bypass strictness on a single call, while testing and porting. So, from the example above: ... At a danger of sounding...

> We should probably use a different place for a discussion on typing. Seconded. Let's move the typing discussion to the mailing list. As of today, typing is optional for...

What is the status here, this is both approved and a draft, and has a conflict, too. What's your plan here @j-bowhay ?

Matti fixed this in https://github.com/numpy/numpy/pull/27294, so on main check-docs runs in CircleCI, https://github.com/numpy/numpy/blob/main/.circleci/config.yml#L90

cupy errors are likely gone once 14.0.0 is out with https://github.com/cupy/cupy/pull/8430

OK, so it's not a memory leak in some C function from scipy, it's that python in general does not guarantee when things are deallocated, and python 3.12 does something...

Just to make sure, when you say you tried 1.14, the version you tried included https://github.com/scipy/scipy/pull/20404, right? If unsure, retry with the current main or a nightly wheel from https://anaconda.org/scientific-python-nightly-wheels/scipy/files....

OK, this is starting to look like a refcounting error somewhere indeed. The next step to find where this _somewhere_ is to trim down the reproducer. Does it fail if...