Why can't I use my graphic card with PyTorch?

1
votes

I am trying to use PyTorch with GPU in my Ubuntu 18.04. The GPU is a GeForce GTX 1070.

nvidia-smi:

| NVIDIA-SMI 460.67       Driver Version: 460.67       CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 1070    Off  | 00000000:0B:00.0 Off |                  N/A |
| 21%   49C    P2    60W / 180W |   4598MiB /  8119MiB |     17%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX 1070    Off  | 00000000:42:00.0 Off |                  N/A |
|  0%   48C    P8     7W / 180W |     20MiB /  8117MiB |      0%      Default |
|                               |                      |                  N/A |

nvcc --version:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Wed_Oct_23_19:24:38_PDT_2019
Cuda compilation tools, release 10.2, V10.2.89

.bashrc file:

export PATH="/usr/local/cuda-10.2/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda-10.2/lib64:$LD_LIBRARY_PATH"

I installed pytorch using the command below (from here):

pip install torch torchvision torchaudio

Torch version:

PyTorch Version: 1.8.0

Python version:

Python 3.8.7

gcc/g++ versions:

gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
g++ (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0

If I try to get the GPU, I get the following:

>>> import torch
>>> print(torch.cuda.is_available())
False

Can anyone advice me please

1
pythonpytorch
What is your CUDA_HOME variable set to? See: discuss.pytorch.org/t/cuda-path-not-correctly-configured/63695/… – jhso
Also, might be a naive suggestion but a reboot is always worth a try, see: discuss.pytorch.org/t/… – jhso
And... finally, make sure that you have cuDNN downloaded and installed: docs.nvidia.com/deeplearning/cudnn/install-guide/index.html – jhso
Actually, there is no need to have CUDA/cuDNN in the host if you install via pip/conda. Both binaries have both libs inside. Are you sure this pip install is installing in the correct Python? – Berriel
BTW, one thing that's always helpful is look at the actual syscalls. strace python -c 'import torch; print(torch.cuda.is_available())' will tell you what python is actually doing to ask the kernel to check in on the hardware; if there's something like a permission error trying to access a something in /dev or /sys it'll show up there. – Charles Duffy

1 Answers

0
votes

If you look at pytorch page, they advise to use special command to install torch with cuda, so probably, you would like to use this one:

pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html