OSError: libcudart.so.10.2: cannot open shared object file: No such file or directory

The reason is a mismatch of CUDA versions. I ran into this issue because the preinstalled version of pytorch did match the default version which I installed using %pip install torchaudio (CUDA 10.2). print(torch.__version__) gives 1.10.0+cu111 (CUDA 11.1).

So I reinstalled pytorch, torchaudio and torch vision with the command stated on the pytorch website

%pip install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio==0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

After restarting the environment, it should work.

This method uninstalls pytorch and reinstalls another version, it would be faster to just install the matching version of pytorch, in my case:

%pip install -q torchaudio==0.10.0+cu111 -f https://download.pytorch.org/whl/cu111/torch_stable.html

I don't know if it would be better to install the cu113 variant.

Also, I would suggest to check the error logs to find out the python package that causes the error. In my case, it was generated in torch-cluster and it simply resolved by downgrading torch-cluster to 1.5.9 (recent version is 1.6.0 which is release just couple of weeks back and was installed by default)

I've solved it by replacing the version of torchaudio installed by pip with the one from conda.

pip uninstall torchaudio
conda install torchaudio -c pytorch

Notice the message of conda, it installs the version with bundled CUDA lib:

The following NEW packages will be INSTALLED:

torchaudio         pytorch/linux-64::torchaudio-0.11.0-py38_cu113