Default installation causes CUDA error in Google Colab

[cap-jmk]

😵 Describe the installation problem

When trying to install the library with

pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric

I receive the following error in Google Colab

/usr/local/lib/python3.7/dist-packages/torch_sparse/_convert_cuda.so: undefined symbol: _ZNK2at6Tensor5zero_Ev

I experienced the error before with PyTorch 1.10 on a local cluster. I was searching the issues but could not find someone with the same problem, yet. The error is known in the site package rusty1s/pytorch_sparse#180

Environment

• PyG version: NameError: name 'torch_geometric' is not defined
• PyTorch version: 11.1
• OS: Ubuntu
• Python version: 3.7
• CUDA/cuDNN version: 1.10
• How you installed PyTorch and PyG (conda, pip, source): pip
• Any other relevant information (e.g., version of torch-scatter): PyG is not recognized properly as a package and CUDA bindings seem to be broken.

[rusty1s]

Can you try following the installation we utilize in our Google Colabs? See here.

[cap-jmk]

It works, thank you. Do you know why the error occurs?

[rusty1s]

I haven't yet tried to reproduce the error but it looks like Colab only installs the CPU version of torch-sparse while it has a GPU available. Ensuring that you install from binaries rather than via manual installation fixes this.