Question 37610 in Ubuntu Question posted in
The official Ubuntu documentation can be found here.

What does it mean if -1 is returned for .get_device() for torch tensor? – Ubuntu

I am using pytorch geometric to train a graph neural network. The problem that led to this question is the following error:

RuntimeError: Expected all tensors to be on the same device, but found
at least two devices, cpu and cuda:0! (when checking argument for
argument mat1 in method wrapper_addmm)

So, I am trying to check which device the tensors are loaded on, and when I run data.x.get_device() and data.edge_index.get_device(), I get -1 for each. What does -1 mean?

In general, I am a bit confused as to when I need to transfer data to the device (whether cpu or gpu), but I assume for each epoch, I simply use .to(device) on my tensors to add to the proper device (but as of now I am not using .to(device) since I am just testing with cpu).

Additional context:

I am running ubuntu 20, and I didn’t see this issue until installing cuda (i.e., I was able to train/test the model on cpu but only having this issue after installing cuda and updating nvidia drivers).
I have cuda 11.7 installed on my system with an nvidia driver compatible up to cuda 12 (e.g., cuda 12 is listed with nvidia-smi), and the output of torch.version.cuda is 11.7. Regardless, I am simply trying to use the cpu at the moment, but will use the gpu once this device issue is resolved.

Tags:

2

Answers


  1. 0

    It means the data is still on the CPU

    -1 indicates not on any GPU device.

    To move it to the GPU, use .to()

    data.x = data.x.to(device)
data.edge_index = data.edge_index.to(device)
    Login or Signup to reply.
  2. 0

    -1 means the tensors are on CPU. when you do .to(device), what is your device variable initialized as? If you want to use only CPU, I suggest initializing device as device = torch.device("cpu") and running your python code with CUDA_VISIBLE_DEVICES='' python your_code.py ...

    Typically, if you are passing your tensors to a model, PyTorch expects your tensors to be on the same device as your model. If you are passing multiple tensors to a method, such as your loss function “nn.CrossEntropy()“`, PyTorch expects both your tensors (predictions and labels) to be on the same device.

    Login or Signup to reply.
Please signup or login to give your own answer.

Back To Top
Search