Add array.data_ptr() Python binding for low-level interop

[Aristide021]

Proposed changes

Exposes array.data<void>() in Python as data_ptr() -> int, returning the address of the first element after evaluation.

Implementation:

• Release GIL before eval() to avoid blocking during synchronization
• Return uintptr_t cast to Python int
• Pointer validity is tied to underlying storage lifetime

Tests:

• Add test_data_ptr in python/tests/test_array.py
• Cover base contiguous array, positive-offset slice, transpose view, and negative-stride view
• Validate pointers against NumPy __array_interface__ and verify pointed values via ctypes

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works

[zcbenz]

Thanks for the PR but conversion of arrays is already supported via buffer protocol and dlpack protocol: https://ml-explore.github.io/mlx/build/html/usage/numpy.html . We should not add another way to do the same thing.

[Aristide021]

Thanks for the PR but conversion of arrays is already supported via buffer protocol and dlpack protocol: https://ml-explore.github.io/mlx/build/html/usage/numpy.html . We should not add another way to do the same thing.

@zcbenz The linked documentation covers the read direction. The use case here was the write direction for zero-copy writes from external kernels into unified memory. Neither DLPack nor the buffer protocol exposes that cleanly.

I also want to note that array.data<void>() already exists in the C++ API. So this isn't another way to do the same thing, it's a Python binding of something that's already there. Closing it as "use DLPack" conflates two different things.

[zcbenz]

The dlpack protocol does support zero-copy writing, from the linked doc:

a = mx.arange(3)
a_view = np.array(a, copy=False)
print(a_view.flags.owndata) # False
a_view[0] = 1
print(a[0].item()) # 1

We don't just expose C++ interface to Python with a 1:1 mapping, for example in C++ we use begin/end for iterators but __iter__ in Python.