Solution for CUDA Installation Issues

I noticed there were a few issues regarding CUDA, and I had ran into the same issue while installing on Google Colab.

It's worth noting that there's often multiple versions of CUDA installed when using pre-installed VMs or Containers, and symbolic links may exist that aren't correctly pointing to what nvidia-smi is stating as the driver CUDA version.

In my case, nvidia-smi was showing CUDA version 10.1 while installing apex was trying to use CUDA 10.0, which was symbolically linked to /usr/local/cuda

Here's how I was able to resolve it on Colab-GPU (Linux) with Latest Pytorch with CUDA enabled

# Check Nvidia-smi
nvidia-smi

# Checks which CUDA is Symlinked
ls -al /usr/local/cuda

# Unlinks Symbolic Link
sudo unlink /usr/local/cuda

# Creates New System Link to CUDA 10.1
sudo ln -s /usr/local/cuda-10.1 /usr/local/cuda

# Install APEX based on another issue
sudo pip3 install -v --no-cache-dir --global-option="--pyprof" --global-option="--cpp_ext" --global-option="--cuda_ext" . 

EDIT: installing with --user led to apex package not being found when running code in a notebook

Hope this helps!

No repository file or test is identified. Start by reproducing the reported Colab/Linux setup with nvidia-smi and ls -al /usr/local/cuda; done would require a maintainer-defined change or documentation target for handling mismatched CUDA links.