Should fix #6985.

@gchanan The same problem exists for torch.tensor with a CPU tensor:

In [6]: import torch

In [7]: x = torch.randn(3, 3, requires_grad=True)

In [8]: torch.tensor(x, device='cuda')
Out[8]:
tensor([[-0.3721,  0.5129,  0.6696],
        [-0.8509,  0.4349,  1.1458],
        [ 0.7075, -3.0103, -0.6825]])

In [9]: torch.tensor(x, device='cuda').device
Out[9]: device(type='cpu')

I think this line

@zou3519 is correct; I'll fix that before merging.

@gchanan Can you please briefly explain how this fixed the problem?

type_inference ? type.toScalarType(var.type().scalarType()) : type;

@karandwivedi42 type inference means in this case means dtype (scalar type) inference, while the ATen type holds (scalar type, layout, device type) e.g. CUDASparseFloatType. So, if we do type inference, we only want to use the scalar type of the incoming variable, not the device-type / layout as well.

Actually, the issue with Variable is more complicated; you want to do type-inference on cuda-ness if device isn't specified, otherwise you don't.