How to properly set up an conda environment with PyTorch

1

I want to use PyTorch with cuda support, and here is what I am usually doing when starting a new project:

conda create -n ‘myproject’
conda activate ‘myproject’

Then, in order to install PyTorch, I go to the site

and choose what I require:

Stable, Linux, Conda, Python, Cuda 11.8 (I got a 4090, and with 11.7 it was somehow slow). I copy paste the command and install everything. Is this the correct way? Am I missing something?

Because, with that approach I just got the error :
Could not load library libcudnn_cnn_infer.so.8. Error: libnvrtc.so: cannot open shared object file: No such file or directory

Do I still need to install the nvidia cuda toolkit ?

2

Your workflow is correct and you are running into this issue. You could apply a workaround by creating the missing symbolic link as described here.

3

Thanks! But it is actually recommended to use cuda 11.8 instead of 11.7 for the 4090?

4

Yes, as the CUDA 11.8 builds ship with a newer cuDNN version (8.7) and added the heuristics for your 4090. Older GPUs should not see a huge difference.

1 Like