Google Colab: torch cuda is true but No CUDA GPUs are available

Question

I use Google Colab to train the model, but like the picture shows that when I input 'torch.cuda.is_available()' and the ouput is 'true'. And then I run the code but it has the error that RuntimeError: No CUDA GPUs are available.

Answer 1

Try to install cudatoolkit version you want to use "conda install pytorch torchvision cudatoolkit=10.1 -c pytorch"

Answer 2

I guess I have found one solution which fixes mine. I used to have the same error. When you run this:

import tensorflow as tf
tf.test.gpu_device_name()

it will give you the GPU number, which in my case it was

/device:GPU:0

I realized that I was passing the code as:

parser.add_argument('--gpu', type=str, default="1", help='choose GPU')

so I replaced the "1" with "0", the number of GPU that Colab gave me, then it worked.

I hope this fixes your problem too.