When we install PyTorch using conda (e.g.,conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch), it incompletely installs the cudatoolkit, which means that we cannot use nvcc provided by the cudatoolkit.

The following commands will help for completely installing the cudatoolkit:

$ wget https://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run
$ sh ./cuda_10.2.89_440.33.01_linux.run --no-opengl-libs --toolkit --silent --override

Note: Make sure to set export CUDA_HOME=/usr/local/cuda in the shell configuration file (i.e., .bashrc or .zshrc).

Leave a comment