BUG: reject non-scalar fill_value in full{_like}

[lucascolley]

closes gh-55 @ev-br

[lucascolley]

looks like https://github.com/data-apis/array-api-tests/blob/3249a739f8ce9be95e5210ed90ea5d8d62045eec/array_api_tests/test_manipulation_functions.py#L376 needs to be modified somehow, not sure exactly how

[ev-br]

Ah yes indeed, the test is not strictly spec-compliant apparently: indexing an array produces an size-1 array not a python scalar

(Pdb) xp.full((int(count),), x_slice[i], dtype=x.dtype)
*** TypeError: Expected Python scalar fill_value, got type <class 'array_api_strict._array_object.Array'>
(Pdb) x_slice[i]
Array(False, dtype=array_api_strict.bool)
(Pdb) x_slice
Array([False], dtype=array_api_strict.bool)

I think it needs get_scalar_type(x_slice[i]) from https://github.com/data-apis/array-api-tests/blob/master/array_api_tests/dtype_helpers.py#L188 (untested)

The need to send two matching PRs is sadly rather common. Here the required array-api-tests PR is straightforward in a sense that it'll clearly make tests standard-compatible. Would you be able to send a -tests PR?

Since CI cannot easily test a branch vs branch (and runs a very small number of examples), I typically resort to first testing things locally with editable installs. Roughly (I'm sure you've a nicer way with pixi),

# ... in an `array-api-strict` branch
$ pip install -e .
$ pushd ../array-api-tests
# ... fix the test
$ ARRAY_API_TESTS_MODULE=array_api_strict pytest array_api_tests/test_manipulation_functions.py::test_repeat --max-examples=10_000

When it passes locally, push the -tests PR, make sure it passes with the -strict main on CI, merge the -tests PR, and kick the CI for the -strict PR once again.

[ev-br]

Huh, apparently test_repeat is skipped everywhere

$ git grep "array_api_tests/test_manipulation_functions.py::test_repeat"
cupy-xfails.txt:array_api_tests/test_manipulation_functions.py::test_repeat
dask-xfails.txt:array_api_tests/test_manipulation_functions.py::test_repeat
numpy-1-22-xfails.txt:array_api_tests/test_manipulation_functions.py::test_repeat
numpy-1-26-xfails.txt:array_api_tests/test_manipulation_functions.py::test_repeat
numpy-dev-xfails.txt:array_api_tests/test_manipulation_functions.py::test_repeat
numpy-xfails.txt:array_api_tests/test_manipulation_functions.py::test_repeat
torch-xfails.txt:array_api_tests/test_manipulation_functions.py::test_repeat

Comments are "# pytorch/pytorch#151311: torch.repeat_interleave rejects short integers" and
"# uint64 repeats not supported" in numpy.

[ev-br]

Meanwhile, here's a fix for this particular issue: data-apis/array-api-tests#412

[ev-br]

A rebase on main should rid of test_special_cases failures
EDIT: #182

[lucascolley]

yep all green