Built-in CUDA not working?

1

On a fresh install of Ubuntu 16.04 LTS, using the recommended install:

conda install pytorch torchvision cuda80 -c soumith

PyTorch works immediately. But a call to my_tensor.cuda() results in:

AssertionError:
Found no NVIDIA driver on your system. Please check that you
have an NVIDIA GPU and installed a driver from
http://www.nvidia.com/Download/index.aspx

Installing CUDA separately works. But according to this thread it seems that PyTorch’s built in CUDA should work without a separate installation. I’ll be allocating and deallocating many machines running PyTorch, so I just wanted to check if the manual CUDA installation is expected at the moment? Or is there something I’m missing? Thank you!

1 Like
2

Pytorch comes with its own runtime cuda libraries, but driver has to be installed on each machine separately.

1 Like
3

See:

4

what do you mean by “driver has to be installed on each machine separately”. I installed cuda 8 and cudnn 6 and having the same error.

5

By driver we mean the kernel driver for your graphics card. Are you sure that it is installed and working correctly? You can check the output of nvidia-smi command.

6

I’m getting this same error with and without installing CUDA separately. I’m completely unable to run pyTorch. Running nvidia-smi gives no errors and says my driver version is 384.98.

7

Hey, did you find any solution to your problem?