xp.eye(..., device=device) has not effect when xp is cupy

[PascalCarrivain]

Dear all,

I run the following command when xp is cupy

# cupy
x = cupy.asarray([1.0] * 3)
xp = array_api_compat.array_namespace(x)
E = xp.eye(3, 3, k=0, dtype='float64', device=1)
print(E.device)
with xp.cuda.Device(1):
    E = xp.eye(3, 3, k=0, dtype='float64', device=1)
    print(E.device)
# torch
x = torch.as_tensor([1.0] * 3)
xp = array_api_compat.array_namespace(x)
E = xp.eye(3, 3, k=0, dtype=torch.float64, device='cuda:1')
print(E.device)

and I get the result:

<CUDA Device 0>
<CUDA Device 1>
cuda:1

It seems that device argument of xp.eye has no effect here when xp is cupy. It seems to work well when xp is torch. Did I miss something?

Thank you.

array-api-compat 1.12.0 cupy-cuda12x 13.4.1

[ev-br]

Yes, this is unfortunately a known issue with CuPy that it requires a context manager. NOt sure if there are plans to change that, maybe @leofang could weight in?

[betatim]

Naively I'd have thought that array-api-compat is the place where we should be inserting that context manager for users when they pass device= as argument. Basically, to smooth out the differences between the different libraries. A bit like what is done in asarray

https://github.com/data-apis/array-api-compat/blob/6c708d13e826fb850161babf34f306fe80cae875/array_api_compat/cupy/_aliases.py#L65-L87

[ev-br]

Potentially, yes. The key question ATM is what's the cupy team plan. If this is something they plan to fix, we can either wait or potentially help materializing for cupy 14.x.