Evgeni Burovski

> but if you want me to actually give it a go, I can try moving that out of `special` and into `_lib`. Would be great indeed. I would worry...

cross-ref https://github.com/scipy/scipy/pull/21091 which takes the full ndimage submodule and only adds cupy delegation. Working through ndimage is probably a bit simpler, at least conceptually: the whole module is compiled code;...

I wonder if we can get away with deprecating object arrays in `correlate` and `lfilter`. Based on tests, the intended use is to work with arrays of `Decimal`s. I seriously...

> Perhaps, but as I commented above, it is enough for this PR to skip the tests with skip_xp_invalid_arg. Given that object arrays aren't mentioned in the docs for correlate,...

FWIW, I'd work through https://github.com/scipy/scipy/pull/21150 first, both technical details and downstream effects.

You need to import `fwind1` in the docstring example: ``` 1331 -------- 1332 scipy.signal.firwin, scipy.signal.get_window 1333 1334 Examples 1335 -------- 1336 Generate a 5x5 low-pass filter with cutoff frequency 0.1....

cf https://docs.scipy.org/doc/scipy/dev/contributor/debugging_linalg_issues.html#creating-reproducers-in-c-or-fortran

> I am still confused if we should add locking in NumPy or CPython should make sure to protect us. Seconded. I'm confused too. Naively, it feels like CPython should...

What we can do at the `array-api-compat` level is rather limited. We could _probably_ add implementations for `matrix_rank` and `pinv` since they both are rather thin layers on top of...

Comparing what Dask implements natively, https://docs.dask.org/en/latest/array-api.html, with the Array API extension, https://data-apis.org/array-api/2024.12/extensions/linear_algebra_functions.html, here are the missing functions | function | assessment | |-------------------------------------|----------------------------| |cross(x1, x2, /, *, axis=-1) | |...