Drop Python 3.9; test on Python 3.13; drop NumPy 1.21; skip CUDA install #304

crusaderky · 2025-04-10T11:54:32Z

Follow-up from #266

Bump up minimum NumPy version from 1.21 to 1.22, in alignment with scikit-learn minimum
Bump up minimum Python version from 3.9 to 3.10; again this is possible thanks to a recent dependency bump in scikit-learn (closes Drop Python 3.9 #230)
Run tests on Python 3.13
Only test on lowest and highest Python versions
Don't install CUDA stack (closes CI installs GPU libraries #284)
Don't install alternative backends when testing numpy 1.x and dev
Closes post 1.11 cleanup tasks #258
Run array-api-tests locally with --max-examples 100000 on NumPy 1.22
Run array-api-tests locally with --max-examples 1000 on all other numpy versions
Disable fail-fast (I lost count of how many times it caused attrition)

crusaderky · 2025-04-10T15:42:40Z

docs/supported-array-libraries.md

@@ -100,8 +95,6 @@ version.
 - As with NumPy, type annotations and positional-only arguments may not
  exactly match the spec for functions that are not wrapped at all.

-The minimum supported PyTorch version is 1.13.


Do not advertise what you don't test.

crusaderky · 2025-04-10T15:43:27Z

docs/supported-array-libraries.md

@@ -131,8 +124,6 @@ For `linalg`, several methods are missing, for example:
 - `matrix_rank`
 Other methods may only be partially implemented or return incorrect results at times.

-The minimum supported Dask version is 2023.12.0.


Do not advertise what you don't test. Also helplessly obsolete; we know for sure that less than 2024.9.0 won't work.

crusaderky · 2025-04-10T15:44:33Z

pyproject.toml

@@ -24,11 +23,14 @@ classifiers = [

 [project.optional-dependencies]
 cupy = ["cupy"]
-dask = ["dask"]
+dask = ["dask>=2024.9.0"]


It may be a good idea to completely remove all mentions of minimum versions until we have unit tests in place for them (i.e. not anytime soon).

crusaderky · 2025-04-10T15:44:43Z

pyproject.toml

-numpy = ["numpy"]
+# Note: array-api-compat follows scikit-learn minimum dependencies, which support
+# much older versions of NumPy than what SPEC0 recommends.
+numpy = ["numpy>=1.22"]
 pytorch = ["torch"]
 sparse = ["sparse>=0.15.1"]


crusaderky · 2025-04-10T17:14:43Z

numpy-1-22-xfails.txt

 # attributes are np.float32 instead of float
 # (see also https://github.com/data-apis/array-api/issues/405)
 array_api_tests/test_data_type_functions.py::test_finfo[float32]
 array_api_tests/test_data_type_functions.py::test_finfo[complex64]

+# Rounding error near infinity
+array_api_tests/test_operators_and_elementwise_functions.py::test_clip


This seems to crop up only when the test is run on its own

data-apis/array-api-tests#360

crusaderky · 2025-04-10T17:15:14Z

numpy-skips.txt

-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__imod__(x1, x2)]
-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__mod__(x, s)]
-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__mod__(x1, x2)]
-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[remainder(x1, x2)]


Tested these and could reproduce none of the core dumps

Can't say I fancy chasing segfaults down the line. OTOH we won't be able to remove these skips otherwise, so OK, let's live dangerously and see if these reappear.

crusaderky · 2025-04-10T17:15:54Z

torch-skips.txt

-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__imod__(x1, x2)]
-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__mod__(x, s)]
-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__mod__(x1, x2)]
-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[remainder(x1, x2)]


Could reproduce none of these

crusaderky · 2025-04-10T17:37:00Z

@ev-br ready for review

This reverts commit 3bc8ffa.

ev-br

Removing skips which are claimed to have caused segfaults feels a bit dangerous; let's give it a shot and see if segfaults reappear.

The rest is great, thank you @crusaderky

ev-br · 2025-04-15T10:42:35Z

numpy-1-22-xfails.txt

 # type promotion issues
+# NOTE: some of these may not fail until one runs array-api-tests with
+# --max-examples 100000


Huh, this is really trying hard to fish these out :-)

ev-br · 2025-04-15T10:44:21Z

numpy-skips.txt

-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__imod__(x1, x2)]
-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__mod__(x, s)]
-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__mod__(x1, x2)]
-array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[remainder(x1, x2)]


Can't say I fancy chasing segfaults down the line. OTOH we won't be able to remove these skips otherwise, so OK, let's live dangerously and see if these reappear.

crusaderky force-pushed the np122 branch from f1d2fdc to 877ea1a Compare April 10, 2025 15:00

crusaderky commented Apr 10, 2025

View reviewed changes

Drop Python 3.9; Test on Python 3.13; drop NumPy 1.21; skip CUDA install

38d4d56

crusaderky force-pushed the np122 branch from c0ca1c6 to 38d4d56 Compare April 10, 2025 15:46

crusaderky added 6 commits April 10, 2025 17:50

xfails

221ff01

clip

3bc8ffa

revert vecdot on np-dev

b73e179

Remove numpy-skips

fd06e11

torch fix

4bab6fb

Dummy files

aceca8c

crusaderky commented Apr 10, 2025

View reviewed changes

crusaderky added 2 commits April 10, 2025 18:19

revert vecdot

ebeaca2

disable fail-fast

f71a31d

crusaderky changed the title ~~Drop Python 3.9; Test on Python 3.13; drop NumPy 1.21; skip CUDA install~~ Drop Python 3.9; test on Python 3.13; drop NumPy 1.21; skip CUDA install Apr 10, 2025

crusaderky marked this pull request as ready for review April 10, 2025 17:36

Revert "clip"

107c5a5

This reverts commit 3bc8ffa.

crusaderky mentioned this pull request Apr 12, 2025

False positives caused by strict equality comparison on NumPy 1.22? data-apis/array-api-tests#365

Closed

crusaderky added 3 commits April 12, 2025 17:30

reverts

e727b5d

Add note

1297e56

xfail linspace and where

6901fc0

ev-br approved these changes Apr 15, 2025

View reviewed changes

ev-br merged commit bff3bf4 into data-apis:main Apr 15, 2025
23 checks passed

ev-br added this to the 1.12 milestone Apr 15, 2025

crusaderky deleted the np122 branch April 15, 2025 13:15

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Drop Python 3.9; test on Python 3.13; drop NumPy 1.21; skip CUDA install #304

Drop Python 3.9; test on Python 3.13; drop NumPy 1.21; skip CUDA install #304

crusaderky commented Apr 10, 2025 •

edited

Loading

crusaderky Apr 10, 2025

crusaderky Apr 10, 2025

crusaderky Apr 10, 2025 •

edited

Loading

crusaderky Apr 10, 2025

crusaderky Apr 10, 2025

crusaderky Apr 10, 2025

crusaderky Apr 10, 2025

ev-br Apr 15, 2025

crusaderky Apr 10, 2025

crusaderky commented Apr 10, 2025

ev-br left a comment

ev-br Apr 15, 2025

ev-br Apr 15, 2025

Drop Python 3.9; test on Python 3.13; drop NumPy 1.21; skip CUDA install #304

Drop Python 3.9; test on Python 3.13; drop NumPy 1.21; skip CUDA install #304

Conversation

crusaderky commented Apr 10, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

crusaderky Apr 10, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

crusaderky commented Apr 10, 2025

ev-br left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

crusaderky commented Apr 10, 2025 •

edited

Loading

crusaderky Apr 10, 2025 •

edited

Loading