None dtype comparison not supported in MLX

[aaishwarymishra]

In reporting.py we have

def to_json_serializable(o):
    if o in dtype_to_name:
        return dtype_to_name[o]

While running tests all the test failed because this method was being called at each test and the main problem is if the o is of the type None

MlLX throws an exception for equality like

if None == mlx.core.float16:
    print("success")

We either have to modify the method to something like

def to_json_serializable(o):
    if o is not None and o in dtype_to_name:
        return dtype_to_name[o]

I think it's a minor problem and can be fixed in test itself not sure though.

[ev-br]

PR welcome, I'd say. A minimal workaround to unblock running the test suite with MLX would be best.

[prady0t]

MLX is missing __array_namespace_info__, this prevents pytest from collecting tests. Is this being tackled in the compat layer?

I tried running tests by skipping the __array_namespace_info__ attribute.

if not hasattr(xp, "__array_namespace_info__"):
    pytest.skip(
        "__array_namespace_info__ is not defined in the array module",
        allow_module_level=True,
    )

But there were a few more blockers:

• Calls to xp.__array_namespace_info__().devices() were put inside try block, returning None when exception (our MLX case)

• We should put this line inside the try block too to avoid test suite failures when the device is not an accepted argument (our MLX case):

array-api-tests/array_api_tests/dtype_helpers.py

Line 206 in 4b42caf

x = xp.empty(2, device=device)

After these changes, I ran with ARRAY_API_TESTS_SKIP_DTYPES=float64 (since float64 is not supported), and the tests are running now.

409 failed, 312 passed, 7 skipped, 64 warnings, 716 errors in 191.65s (0:03:11)

[ev-br]

Great! Not sure you saw it @prady0t , there's an MLX tracker at data-apis/array-api-compat#450, which links several sub-issues.
I think it'd be best to use this tracker, and follow a similar style: run tests and note failures per group of functions (ideally, with repro examples from the test suite), then sort triage the failures into roughly three buckets: "submit fixes to MLX itself", "test suite needs improving", and "hard, leave to the end".

[ev-br]

It'd be great to do a breadth-first kind of summary of what are these 409 failures, similar to data-apis/array-api-compat#452

For deciding what to submit upstream and in which order, the general guidance is data-apis/array-api-compat#452 (comment)