Add support for from_dlpack with numpy 1.23.0 #284
Conversation
| @@ -97,7 +97,13 @@ def _register_from_dlpack_cpu_to_numpy(): | |||
| @_from_dlpack_cpu.register(np.ndarray) | |||
| def _from_dlpack_cpu_to_numpy(to, array): | |||
| try: | |||
| # private `_from_dlpack` method added in 1.22.0 | |||
| return np._from_dlpack(array) | |||
There was a problem hiding this comment.
I have a vague recollection that the semantics of these methods are actually not quite the same. Do we have a way to test against different versions of NumPy that would allow us to validate that both of these methods work as expected?
There was a problem hiding this comment.
np._from_dlpack(obj) function, where obj supports
__dlpack__, and returns an ndarray
NumPy 1.22.0 Release Notes
numpy.from_dlpack has been added to allow easy exchange of data using the DLPack protocol. It accepts Python objects that implement the
__dlpack__and__dlpack_device__methods and returns a ndarray object which is generally the view of the data of the input object.
NumPy 1.23.0 Release Notes
from the release notes the first private function in 1.22 mentions it supports objects that implement __dlpack__, while the notes of 1.23 mention support for objects with __dlpack__ and __dlpack_device__. There was a bugfix for __dlpack_device__ in numpy/numpy#21138. Which maybe indicates we shouldn't be using this until we have a min version of numpy 1.23 ?
There was a problem hiding this comment.
Do we have a way to test against different versions of NumPy that would allow us to validate that both of these methods work as expected?
For automated tests we could potentially add a new workflow that has a test matrix for supported numpy versions? Combined with some pytest marks to run a subset of the tests? (the core tests run pretty quickly so that might not be strictly needed here)
There was a problem hiding this comment.
Ah yeah, that's the difference. I don't think we do much with __dlpack_device__ since it matters less on CPU. This is probably fine then.
Add support for from_dlpack with numpy 1.23.0
Check
np.from_dlpack(added in1.22.0, renamed from private_from_dlpack), while keeping check ofnp._from_dlpack(added in 1.22.0) for backwards compatibility with numpy 1.22.0