BUG: test_dunder_dlpack: accept spec-mandated BufferError for a no-copy cross-device export

[antonwolfy]

Closes #456.

This PR proposes to tolerate BufferError only when copy is False and the requested dl_device differs from the array's own __dlpack_device__().

CPU-native backends are unaffected (cross_device is False). The copy=None/True paths must still copy and succeed, and non-BufferError exceptions still propagate.

The change is verified and passed with dpnp.

[ev-br]

Thanks @antonwolfy.
I tested it locally with the usual gamut of numpy, cupy, jax and torch, and it shows the following glitch for torch defaulting to CUDA as the device (in fact, I was not testing this mode previously, so thank you for pushing me towards doing the right thing) :

$ ARRAY_API_TESTS_SKIP_DTYPES=uint16,uint32,uint64 ARRAY_API_TESTS_MODULE=$VAR   pytest array_api_tests/test_dlpack.py -v --max-examples=10_000 -x
=========================================================================================== test session starts ===========================================================================================
platform linux -- Python 3.12.0, pytest-9.0.2, pluggy-1.6.0 -- /home/ev-br/.conda/envs/array-api-cupy13/bin/python3.12
cachedir: .pytest_cache
metadata: {'Python': '3.12.0', 'Platform': 'Linux-6.8.0-136-generic-x86_64-with-glibc2.39', 'Packages': {'pytest': '9.0.2', 'pluggy': '1.6.0'}, 'Plugins': {'metadata': '3.1.1', 'json-report': '1.5.0', 'hypothesis': '6.151.9'}, 'array_api_tests_module': 'array_api_tests._array_module', 'array_api_tests_version': '2026.02.26+35.ge95e8a0.dirty'}
hypothesis profile 'array-api-tests' -> max_examples=10000, deadline=timedelta(milliseconds=800)
Environment variables:
----------------------
ARRAY_API_TESTS_SKIP_DTYPES = uint16,uint32,uint64
ARRAY_API_TESTS_MODULE = exec('import torch; from array_api_compat import torch as xp; torch.set_default_device("cuda")')

Array API Tests Module: array_api_compat.torch (version unknown). API Version: 2025.12. Enabled Extensions: fft, linalg
rootdir: /home/ev-br/repos/array-api-tests
configfile: pytest.ini
plugins: metadata-3.1.1, json-report-1.5.0, hypothesis-6.151.9
collected 3 items                                                                                                                                                                                         

array_api_tests/test_dlpack.py::test_dlpack_device PASSED                                                                                                                                           [ 33%]
array_api_tests/test_dlpack.py::test_dunder_dlpack FAILED                                                                                                                                           [ 66%]

================================================================================================ FAILURES =================================================================================================
___________________________________________________________________________________________ test_dunder_dlpack ____________________________________________________________________________________________
  + Exception Group Traceback (most recent call last):
  |   File "/home/ev-br/repos/array-api-tests/array_api_tests/test_dlpack.py", line 36, in test_dunder_dlpack
  |     x=hh.arrays(dtype=hh.all_dtypes, shape=hh.shapes(min_dims=1, max_side=2)),
  |                ^^^
  |   File "/home/ev-br/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/hypothesis/core.py", line 2246, in wrapped_test
  |     raise the_error_hypothesis_found
  | ExceptionGroup: Hypothesis found 2 distinct failures. (2 sub-exceptions)
  +-+---------------- 1 ----------------
    | Traceback (most recent call last):
    |   File "/home/ev-br/repos/array-api-tests/array_api_tests/test_dlpack.py", line 60, in test_dunder_dlpack
    |     x.__dlpack__(**copy_kw, **max_version_kw, **dl_device_kw)
    |   File "/home/ev-br/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/torch/_tensor.py", line 1585, in __dlpack__
    |     return handle_torch_function(Tensor.__dlpack__, (self,), *args, **kwargs)
    |            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    |   File "/home/ev-br/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/torch/overrides.py", line 1771, in handle_torch_function
    |     result = mode.__torch_function__(public_api, types, args, kwargs)
    |              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    |   File "/home/ev-br/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/torch/utils/_device.py", line 122, in __torch_function__
    |     return func(*args, **kwargs)
    |            ^^^^^^^^^^^^^^^^^^^^^
    |   File "/home/ev-br/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/torch/_tensor.py", line 1667, in __dlpack__
    |     return _C._to_dlpack_versioned(self, dl_device=dl_device, copy=copy)
    |            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    | ValueError: cannot move (i.e. copy=False) tensor from cuda:0 to cpu without copying.
    | 
    | ========== FAILING CODE SNIPPET:
    | tensor([], device='cuda:0', dtype=torch.bool).__dlpack__() with copy_kw = {'copy': False}, max_version_kw = {'max_version': (1, 0)} and dl_device_kw = {'dl_device': (1, 0)}
    | ====================
    | 
    | Falsifying example: test_dunder_dlpack(
    |     x=tensor([], device='cuda:0', dtype=torch.bool),  # or any other generated value
    |     copy_kw={'copy': False},
    |     max_version_kw={'max_version': (1, 0)},  # or any other generated value
    |     dl_device_kw={'dl_device': (1, 0)},
    |     data=data(...),
    | )
    +---------------- 2 ----------------
    | Traceback (most recent call last):
    |   File "/home/ev-br/repos/array-api-tests/array_api_tests/test_dlpack.py", line 60, in test_dunder_dlpack
    |     x.__dlpack__(**copy_kw, **max_version_kw, **dl_device_kw)
    |   File "/home/ev-br/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/torch/_tensor.py", line 1585, in __dlpack__
    |     return handle_torch_function(Tensor.__dlpack__, (self,), *args, **kwargs)
    |            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    |   File "/home/ev-br/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/torch/overrides.py", line 1771, in handle_torch_function
    |     result = mode.__torch_function__(public_api, types, args, kwargs)
    |              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    |   File "/home/ev-br/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/torch/utils/_device.py", line 122, in __torch_function__
    |     return func(*args, **kwargs)
    |            ^^^^^^^^^^^^^^^^^^^^^
    |   File "/home/ev-br/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/torch/_tensor.py", line 1665, in __dlpack__
    |     return _C._to_dlpack(self, dl_device=dl_device, copy=copy)
    |            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    | ValueError: cannot move (i.e. copy=False) tensor from cuda:0 to cpu without copying.
    | 
    | ========== FAILING CODE SNIPPET:
    | tensor([], device='cuda:0', dtype=torch.bool).__dlpack__() with copy_kw = {'copy': False}, max_version_kw = {} and dl_device_kw = {'dl_device': (1, 0)}
    | ====================
    | 
    | Falsifying example: test_dunder_dlpack(
    |     x=tensor([], device='cuda:0', dtype=torch.bool),  # or any other generated value
    |     copy_kw={'copy': False},
    |     max_version_kw={},  # or any other generated value
    |     dl_device_kw={'dl_device': (1, 0)},
    |     data=data(...),
    | )
    +------------------------------------

From the test output, the failure is

In [1]: import torch

In [2]: x = torch.as_tensor([], device='cuda:0', dtype=torch.bool)

In [3]: x.__dlpack__(copy=False, dl_device=(1,0))
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[3], line 1
----> 1 x.__dlpack__(copy=False, dl_device=(1,0))

File ~/.conda/envs/array-api-cupy13/lib/python3.12/site-packages/torch/_tensor.py:1665, in Tensor.__dlpack__(self, stream, max_version, dl_device, copy)
   1661     return xla_dlpack.to_dlpack(self)
   1663 if max_version is None or max_version[0] < 1:
   1664     # Fallback to the old, unversioned variant.
-> 1665     return _C._to_dlpack(self, dl_device=dl_device, copy=copy)
   1667 return _C._to_dlpack_versioned(self, dl_device=dl_device, copy=copy)

ValueError: cannot move (i.e. copy=False) tensor from cuda:0 to cpu without copying.

To make sure I understand it right. So what it says is that torch is not compliant: it emits a ValueError instead of the spec-mandated BufferError?

[antonwolfy]

Yes, per my understanding. Every spec version mandates BufferError for a copy=False cross-device export:

If False, the function must never copy, and raise a BufferError in case a copy is deemed necessary (e.g. if a cross-device data movement is requested, and it is not possible without a copy).

[ev-br]

Thanks for confirming. I opened pytorch/pytorch#194267

[ev-br]

Let's give it a go. I'll follow up on adding an xfail in the array-api-compat CI.
Thanks @antonwolfy