Repository navigation
BUG: Should np.from_dlpack raise BufferError instead of ValueError on unsupported dl_device values? #30937
Description
Describe the issue:
In the from_dlpack docs it says:
BufferError – The dlpack and dlpack_device methods on the input array may raise BufferError when the data cannot be exported as DLPack (e.g., incompatible dtype, strides, or device).
I read the meaning as: incompatible dtype, incompatible strides, or incompatible device - which would suggest a BufferError should be raised when device is incompatible?
I understand BufferError may seem like an odd choice (#30341 (comment)) but just raising the question from my interpretation of the docs.
Reproduce the code example:
x_torch = torch.arange(5, device='cuda')
np.from_dlpack(x_torch)Error message:
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
RuntimeError: Unsupported device in DLTensor.Python and NumPy Versions:
2.4.2
3.14.2 | packaged by conda-forge | (main, Feb 4 2026, 15:49:49) [GCC 14.3.0]
Runtime Environment:
No response
Activity
seberg commentedon Mar 5, 2026
Yeah, I agree, would be better/correct to raise a
BufferErrorhere, seems there is always one more path...seberg commentedon Mar 11, 2026
Ah, now that I see the code change, I realize why it wasn't the case.
BufferErroris vital for the exporter, but this is on the consumer side it is a bit less clear cut (and not actually specified at all, I think).But it sounds like it helps you, so happy to change it.
lucyleeow commentedon Mar 11, 2026
Now that I read the docs more carefully it does talk about exporting: "data cannot be exported as DLPack" - but doing some digging, the spec also says:
seberg commentedon Mar 11, 2026
Ah, you are right, so that hints to it, good to know. And it makes sense anyway, I think. (Sorry for the noise, I think it mostly explains why we missed it, fixing it for the export part at some point, but not here.)
lucyleeow commentedon Mar 11, 2026
Honestly, I am always confused with DLPack 🤔 (I also do not have a good sense if it makes sense... 😅 )