In newer versions of pytorch, you can set the device by using torch.device. I need to use torch version 0.3.1.post2 for some legacy code. How do I set the device for this version?
As far as I know, you can use the set_device
function. But this is not encouraged. Please see the reference.
The suggested method is, you just set the CUDA_VISIBLE_DEVICES
environmental variable. You can run your script as follows.
CUDA_VISIBLE_DEVICES=GPU_ID python script_name.py
In your program, you can just simply use .cuda()
to use GPUs. (e.g., model=model.cuda()
)