TST/BUG: run all tests on all backends; fix backend-specific bugs #88
Conversation
| @@ -54,7 +54,9 @@ def in1d( | |||
| order = xp.argsort(ar, stable=True) | |||
| reverse_order = xp.argsort(order, stable=True) | |||
| sar = xp.take(ar, order, axis=0) | |||
| if sar.size >= 1: | |||
| ar_size = _compat.size(sar) | |||
| assert ar_size is not None, "xp.unique*() on lazy backends raises" | |||
There was a problem hiding this comment.
Caveat: this is missing data-apis/array-api-compat#231
|
|
||
| # mypy: disable-error-code=no-untyped-usage |
There was a problem hiding this comment.
The return type of xp.anything is untyped, because xp is a ModuleType
src/array_api_extra/testing.py
Outdated
There was a problem hiding this comment.
could you detail how these functions deviate from the ones in scipy._lib._array_api? If we are going to expose these here, it would be good to be able to remove them from SciPy.
Also, consider #17 (comment). But given we have since changed the scope of array-api-extra to include paths specific to known backends, I think they probably are in-scope now.
EDIT: given
Note that I have no plans to use these in scipy, as there are enough extra nuances
there to warrant bespoke variants.
perhaps there is a way to make these extensible for library-specific nuances, to remove code-duplication somewhat?
There was a problem hiding this comment.
- no default_ctx context variable
- no ability to cherry-pick disabling shape, dtype, and/or namespace validation
- no special casing for 0d
All of these features could be catered for here, but I feel they're all severely overdesigned for ensuring the health of array-api-extra.
My goal was to write something that serves well array-api-extra, and as a nice bonus allows other people to use it.
I'm open to making the whole new module private if you'd prefer it that way.
There was a problem hiding this comment.
let's keep it private for now and explain the situation in gh-17. If there are requests for them to be made public in the future, we should reconsider.
|
Also, see #72 for my vision of how I would like to include delegation to library-specific implementations more broadly. The testing introduced here should be enough to push that PR over the line too 🎉 |
crusaderky
force-pushed
the
test_all_xps
branch
from
4c2ec53 to
7445917
Compare
|
Made the functions private. |
create_diagonalon JAXkronon JAX, CuPy, and PyTorchpadon JAXin1don PyTorchpublicprivate functionsxp_assert_equalandxp_assert_close.Note that I have no plans to use these in scipy, as there are enough extra nuances
there to warrant bespoke variants.