[BUG] DLPackCompatible protocol members are declared as attributes, so np.ndarray fails to match — mx.array(np_arr) no longer type-checks #4371

Describe the bug

The DLPackCompatible protocol in the type stubs declares its members as mutable attributes rather than methods:

https://github.com/ml-explore/mlx/blob/main/python/mlx/_stub_patterns.txt#L5-L7

class DLPackCompatible(Protocol):
  __dlpack__: Callable[..., Any]
  __dlpack_device__: Callable[..., Any]

A protocol member declared as a variable is invariant and settable, so a class that defines __dlpack__ as a normal (read-only) method — like numpy.ndarray — does not match structurally. mypy states it plainly:

note: Protocol member DLPackCompatible.__dlpack__ expected settable variable, got read-only attribute

The result is that passing a NumPy array to any of the three entry points that accept DLPackCompatible fails type checking, even though all three work perfectly at runtime:

• mx.array — python/src/array.cpp
• mx.asarray — python/src/ops.cpp
• mx.from_dlpack — python/src/ops.cpp

This is a regression against 0.31.2, where array.__init__ was annotated with numpy.ndarray directly:

# 0.31.2
def __init__(self: array, val: Union[scalar, list, tuple, numpy.ndarray, array], dtype: Optional[Dtype] = None): ...
# 0.32.1
def __init__(self: array, val: scalar | list | tuple | DLPackCompatible | array, dtype: Dtype | None = None): ...

(This is separate from #3916 — the stubs are back in 0.32.1, thanks. This is about their content.)

To Reproduce

# probe.py
import mlx.core as mx
import numpy as np

a: np.ndarray = np.zeros((2, 3), dtype=np.float32)
mx.array(a)
mx.asarray(a)
mx.from_dlpack(a)

uv venv --python 3.12 && uv pip install "mlx==0.32.1" numpy mypy
uv run mypy probe.py

Result on 0.32.1 (identical failures from ty):

probe.py:5: error: Argument 1 to "array" has incompatible type "ndarray[tuple[Any, ...], dtype[Any]]";
  expected "int | float | bool | complex | list[Any] | tuple[Any, ...] | DLPackCompatible | array"  [arg-type]
probe.py:6: error: Argument 1 to "asarray" has incompatible type ...  [arg-type]
probe.py:7: error: Argument 1 to "from_dlpack" has incompatible type "ndarray[...]";
  expected "DLPackCompatible"  [arg-type]
Found 3 errors in 1 file (checked 1 source file)

The same script on mlx==0.31.2:

Success: no issues found in 1 source file

And all three calls succeed at runtime on 0.32.1:

array    (2, 3)
asarray  (2, 3)
dlpack   (2, 3)

Expected behavior

mx.array(np_array) type-checks, as it did in 0.31.2. Runtime behaviour is already correct — only the stub is wrong.

Suggested fix

Declare the protocol members as methods, which makes them read-only and matched covariantly:

class DLPackCompatible(Protocol):
  def __dlpack__(self, *args: Any, **kwargs: Any) -> Any: ...
  def __dlpack_device__(self, *args: Any, **kwargs: Any) -> Any: ...

Verified: with that patch applied to the installed mlx/core/__init__.pyi, the repro above is clean under both mypy and ty (Success: no issues found / All checks passed!), on the same 0.32.1 wheel.

If you want a CI gate for this class of regression, the probe.py above is a good end-to-end canary — it exercises the stubs the way downstream projects actually do, and it would have caught both #3916 (stubs missing) and this one (stubs wrong).

Desktop (please complete the following information):

• OS Version: macOS 15 (Darwin 25.5.0), Apple Silicon
• Version: mlx 0.32.1 (mlx-0.32.1-cp312-cp312-macosx_15_0_arm64.whl), Python 3.12, mypy 2.3.1, ty 0.0.73