search

data-apis / array-api-strict

Strict implementation of the Python array API (previously numpy.array_api)
http://data-apis.org/array-api-strict/
Other
7 stars 4 forks source link

Preliminary 2023.12 support #35

Closed asmeurer closed 2 months ago

asmeurer commented 4 months ago

This will not be enabled by default by can be enabled with set_array_api_strict_flags(api_version='2023.12') or using __array_namespace__(api_version='2023.12'). It will issue a warning that not all functions are implemented and it hasn't been tested.

My plan here is to just wrap the functions that are already in NumPy, and maybe add some of the straightforward inspection APIs. There's no test suite support yet, so any features missing from NumPy will probably need to wait.

This issue has a list of the new APIs that need to be added for 2023.12. https://github.com/data-apis/array-api-tests/issues/249

asmeurer commented 4 months ago

@honno the tests are failing here because I have this behavior:

>>> import array_api_strict as xp
>>> xp.signbit(xp.asarray([0., 1.]))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/aaronmeurer/Documents/array-api-strict/array_api_strict/_flags.py", line 293, in wrapper
    raise RuntimeError(
RuntimeError: The function signbit requires API version 2023.12 or later, but the current API version for array-api-strict is 2022.12

It can be enabled by setting the version flag to 2023.12, but I'd like to be able to test all supported versions here.

But the test suite uses hasattr(xp, 'signbit') and assumes that if it has signbit then it can use it. Can we append a simple sanity check to this test? Trying to make disabled functions actually raise AttributeError is a lot more complicated and I'd prefer not to do it if it isn't necessary.

asmeurer commented 4 months ago

This is ready for review. However, we are blocked on the test suite issue I noted above.

The only new API I'm missing is the new device and copy keywords to from_dlpack. Is it possible for me to implement those here without upstream NumPy support?

asmeurer commented 4 months ago

Test suite should be fixed by https://github.com/data-apis/array-api-tests/pull/257. But I also need to add test runs for 2023.12. Apparently the signature and special cases are already running against 2023.12, so we can go ahead and add that.

asmeurer commented 2 months ago

This should be good to go. The dlpack and astype changes can be updated once NumPy 2.1 is released.

The only question is whether I should leave 2022.12 as the default for now or go ahead and switch it.

asmeurer commented 2 months ago

A lot of this failure on CI, which I've never seen before:

FAILED array_api_tests/test_searching_functions.py::test_argmin - hypothesis.errors.FailedHealthCheck: Examples routinely exceeded the max allowable size. (20 examples overran while generating 9 valid ones). Generating examples this large will usually lead to bad results. You could try setting max_size parameters on your collections and turning max_leaves down on recursive() calls.
See https://hypothesis.readthedocs.io/en/latest/healthchecks.html for more information about this. If you want to disable just this health check, add HealthCheck.data_too_large to the suppress_health_check settings for this test.

I can't reproduce it locally. I guess we should just disable that health check.

asmeurer commented 2 months ago

I opened https://github.com/data-apis/array-api-tests/issues/277 about the data_too_large health check issue. For now, I've just disabled that health check in the test suite so this should pass now.