Multiple CUDA Single Machine

multi_cuda_single_machine.md

How to maintain multiple CUDA versions in a single computer

Suppose we have already installed Nvidia GPU driver which should be equal to or newer than the required.

1. Download cuda runfile (installer type: runfile (local)) from https://developer.nvidia.com/cuda-downloads
   • For example: https://developer.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.105_**418.39**_linux.run
   • Make sure the GPU driver is equal to or newer than 418.39
2. Extract what we want: toolkit following the interactive prompt or sudo sh cuda-10.1.run --silent --toolkit --toolkitpath=/usr/local/cuda-10.1
   • To maintain a cudatoolkit with different versions of cudnn, we can name the dicrectory cuda-10.1_cudnn-v7.5
3. Check the installation status
   • If /usr/local/cuda-10.1_cudnn-v7.5 contains only bin and pkgconfig, the installation failed. InstallUtils.pm is needed.
     1. ./cuda_8.0.61_375.26_linux.run --tar mxvf to extract InstallUtils.pm
     2. sudo cp InstallUtils.pm /usr/lib/x86_64-linux-gnu/perl-base/ (ubuntu)
     3. export $PERL5LIB
     4. sudo sh cuda-10.1.run --silent --toolkit --toolkitpath=/usr/local/cuda-10.1_cudnn_v7.5
     5. sudo rm /usr/lib/x86_64-linux-gnu/perl-base/InstallUtils.pm
4. Remove or recreate /usr/local/cuda, which is the default cuda path
5. Extract cuDNN file: tar -xzvf cudnn-10.1-linux-x64-v7.5.tgz
   • sudo cp ~/Gode/cudnn-10.1-linux-x64-v7.5/include/cudnn.h /usr/local/cuda-10.1_cudnn_v7.5/include
   • sudo cp ~/Gode/cudnn-10.1-linux-x64-v7.5/lib64/libcudnn* /usr/local/cuda-10.1_cudnn_v7.5/lib64
   • sudo chmod a+r /usr/local/cuda-10.1_cudnn_v7.5/lib64/libcudnn*
6. Manually set environment variables for different virtual environments
   • export PATH=/usr/local/cuda-10.1_cudnn_v7.5/bin:$PATH
   • export LD_LIBRARY_PATH=/usr/local/cuda-10.1_cudnn_v7.5/lib64:$LD_LIBRARY_PATH
7. Or follow this blog, update conda's activate and deactivate command.