No attribute "device" when using Numpy backend

[pfisterj]

Describe the bug

When using Numpy backend, it seems that Tensorly tensors have no attribute "device".

Steps or Code to Reproduce

import tensorly as tl

# works OK
tl.set_backend('pytorch')
A = tl.tensor((3,3),dtype=tl.float64,device="cpu")
print("torch|A:",A)
print("torch|A.dtype:",A.dtype)
print("torch|A.device:",A.device)

# raises an AttributeError: 'numpy.ndarray' object has no attribute 'device' 
# when printing A.device
tl.set_backend('numpy')
A = tl.tensor((3,3),dtype=tl.float64,device="cpu")
print("numpy|A:",A)
print("numpy|A.dtype:",A.dtype)
print("numpy|A.device:",A.device)

Expected behavior

It would be nice to have a "device" attribute set to "cpu" when Numpy is used as backend. The reason behind this is that sometimes we may want to create a new tensor, having e.g. similar 'device' and 'dtype' than another tensor. Even if a tensor with Numpy backend will obviously live on the cpu, this would prevent from having to add an if: else: block in the code.

Note also that A = tl.tensor((3,3),dtype=tl.float64,device="cuda") after tl.set_backend('numpy') raises no warning/error, perhaps it should?

Versions

Windows-10-10.0.19045-SP0
Python 3.11.4 | packaged by Anaconda, Inc. | (main, Jul 5 2023, 13:38:37) [MSC v.1916 64 bit (AMD64)]
NumPy 1.26.4
SciPy 1.10.1
TensorLy 0.8.1
PyTorch 2.2.0+cu121

[JeanKossaifi]

Yes, currently the context of a tensor depends on the backend. I did think about making it backend agnostic, I guess that is one option. We then need to normalize the device across all backends then.

We could introduce e.g. tl.cpu and tl.gpu if necessary. What do you think @aarmey @cohenjer @MarieRoald @yngvem ?

[yngvem]

I think we should wait a bit with that, NumPy will soon add support for the array API in NumPy v2 later this year.

[JeanKossaifi]

I agree @yngvem - actually really excited about the cleaned up API, it's amazing to have such a big project do such a large cleanup!