Assume I have a multi-GPU system. Let tensor “a” be on one of the GPUs, and tensor “b” be on CPU. How can I move “b” to the same GPU that “a” resides?
Unfortunately, b.type_as(a) always moves b to GPU 0.
Thanks.
Assume I have a multi-GPU system. Let tensor “a” be on one of the GPUs, and tensor “b” be on CPU. How can I move “b” to the same GPU that “a” resides?
Unfortunately, b.type_as(a) always moves b to GPU 0.
Thanks.
Thanks @ptrblck. The problem I have with the “Tensor.new” function is that
If “a” is on GPU and “b” on CPU, then “a.new(b)” does not work (error: …constructor received an invalid combination of arguments…). “a.new(b.numpy())” works though, but I am afraid that it is inefficient.
If “a” and “b” are already on the same device, then “a.new(b)” will unnecessarily create a new copy of “b”
I am looking for a function like “b.type_as(a)”, but could automatically move the data to the same device as “a”.
As far as I understand, you would like to move b
to the same device as a
.
This should work:
a = torch.randn(10, 10).cuda()
print(a)
b = a.new(a)
print(b)
c = a.new(10, 10)
print(c)
@ptrblck The problem is that “b” does not necessarily have the same shape and type of “a”. For example “a” could be 10⨉10 float tensor, while “b” 13⨉19⨉23 int Tensor.
You can pass the standard arguments to new
as to a new Tensor
:
a = torch.randn(10, 10).cuda()
print(a)
b = a.new(13, 19, 23).long()
print(b)
Would this work?
EDIT: You could of course pass a numpy array or something else to the constructor.
Could you post your use case? I have the feeling, I’m not really understanding your problem and thus posting useless approaches.
Thanks @ptrblck. I have a huge project, and in most places, in order to move the data to the proper device I used “type_as”. Then I wanted to run several instances of that program on different GPUs of the machine at the same time. The problem was that type_as always uses GPU 0. Right now I am using the approach explained here Select GPU device through env vars, and it solves my problem.
This may be new functionality from the Tensor API, but to move tensor a
to the device of tensor b
I use:
a = a.to(b.device)