[dpnp.ndarray] `data.ptr` usage error

[abagusetty]

Reproducer:

The motivation for this reproducer is to mimic a user application code on aurora with cupy.ndarray constructor with memptr argument which I am not sure how to access it for dpnp. I've tried to mimic memptr from cupy to buffer arg in dpnp.ndarray. @antonwolfy Sorry about tagging in a wrong issue. Here is the reproducer for this [discussion](https://github.com/IntelPython/dpnp/issues/2486#issuecomment-3075387410)

The end goal is to use dpnp.ndarray.data.ptr to cast it to a void_p using cast. The error is this:

Traceback (most recent call last):
  File "/lus/flare/projects/NWChemEx_aesp_CNDA/abagusetty/gpu4pyscf/test_dpnp_dataptr.py", line 58, in <module>
    print(out.data)  # This line will raise the TypeError
  File "/lus/flare/projects/NWChemEx_aesp_CNDA/abagusetty/gpu4pyscf/dpnp/dpnp/dpnp_array.py", line 1030, in data
    return dpm.create_data(self._array_obj)
  File "/lus/flare/projects/NWChemEx_aesp_CNDA/abagusetty/gpu4pyscf/dpnp/dpnp/memory/_memory.py", line 105, in create_data
    raise TypeError(f"Expected USM memory, but got {type(usm_data)}")
TypeError: Expected USM memory, but got <class 'dpnp.memory._memory.MemoryUSMDevice'>

import dpnp
import numpy as np
import types

# --- Monkey patch cupy.ndarray to mimic memory handling via CuPy-like interface ---
class CuPyNdarrayWrapper:
    def __call__(self, shape, dtype=np.float64, memptr=None):
        if memptr is not None:
            if hasattr(memptr, 'get_array'):
                memptr = memptr.get_array()
            # Try to reuse buffer for dpnp
            return dpnp.ndarray(shape, dtype=dtype, buffer=memptr)
        else:
            return dpnp.ndarray(shape, dtype=dtype)

    def __instancecheck__(self, instance):
        return isinstance(instance, dpnp.dpnp_array.dpnp_array)

    def __subclasscheck__(self, subclass):
        return issubclass(subclass, dpnp.dpnp_array.dpnp_array)

# --- Fake "cupy" module ---
import sys
cupy_fake = types.ModuleType("cupy")
cupy_fake.ndarray = CuPyNdarrayWrapper()
sys.modules["cupy"] = cupy_fake

# --- Simulate user logic ---
comp = 1
nao_max = 59
MIN_BLK_SIZE = 4096
nao_sub = 58
ip0 = 0
ip1 = 4096

# Allocate initial buffer in dpnp (this uses USMDevice memory internally)
buf = dpnp.empty((comp, nao_max, MIN_BLK_SIZE), order='C')

# This is where the problem is triggered — we pass a USMDevice buffer directly
out = cupy_fake.ndarray((comp, nao_sub, ip1 - ip0), memptr=buf.data)

# Accessing .data will internally call dpnp → dpm.create_data()
print(out)
print(out.data)  # This line will raise the TypeError

[antonwolfy]

@abagusetty, thank you for the reproducer provided. I'm going to fix that issue.

Also, in case it would be helpful, I'd like to point that there is no need to call in the CuPyNdarrayWrapper class:

            if hasattr(memptr, 'get_array'):
                memptr = memptr.get_array()

because it will be done inside dpnp.ndarray.__init__.