Tensorflow Install

Procedure to install the most recent tensorflow package compatible with your nvidia machine:

1. Make sure your nvidia drivers is up to date.

   You can check that the Nvidia drivers is properly install with the command:

   nvidia-smi
   

   It should print something like this:

   +-----------------------------------------------------------------------------+
   | NVIDIA-SMI 515.65.01    Driver Version: 515.65.01    CUDA Version: 11.7     |
   |-------------------------------+----------------------+----------------------+
   | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
   | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
   |                               |                      |               MIG M. |
   |===============================+======================+======================|
   |   0  NVIDIA GeForce ...  Off  | 00000000:01:00.0 Off |                  N/A |
   | N/A   44C    P3    13W /  N/A |      4MiB /  4096MiB |      0%      Default |
   |                               |                      |                  N/A |
   +-------------------------------+----------------------+----------------------+
   
   +-----------------------------------------------------------------------------+
   | Processes:                                                                  |
   |  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
   |        ID   ID                                                   Usage      |
   |=============================================================================|
   |    0   N/A  N/A      1090      G   /usr/lib/xorg/Xorg                  4MiB |
   +-----------------------------------------------------------------------------+
   

2. Check the most recent CUDA Drivers version you can install using this cuda link.

   

   (source: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html)

3. Now that you know the latest version of CUDA you can install, check which latest version of Tensorflow (GPU) you can install using this tf link.

   

   (source: https://www.tensorflow.org/install/source#gpu)

4. Install the CUDA version you choose in section 3.

   CUDA Download link: https://developer.nvidia.com/cuda-toolkit-archive CUDA Installation guide: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html

   Warning

   if you know your nvidia drivers matches with this CUDA version, when running the runfile installation of CUDA add --toolkit (sudo sh cuda_11.2.2_460.32.03_linux.run --toolkit) in the install command to make sure nvidia drivers won’t be updated.

   You will need to update environment variables PATH and LD_LIBRARY_PATH in your .bashrc: (It is not mandatory to define these variable globally, you can also set them only when activating your conda environment, see below)

   # These are symlinks ! be sure it point to the desired CUDA version you installed
   export PATH="/usr/local/cuda/bin:$PATH"
   export LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH"
   export LD_LIBRARY_PATH="/usr/local/cuda/include:$LD_LIBRARY_PATH"
   

   You can check if CUDA is properly installed by running the command (won’t work if you prefer to define environment variables for conda only):

   nvcc -V
   

5. Download the cuDNN version you choose in section 3.

   cuDNN: https://docs.nvidia.com/deeplearning/cudnn/install-guide/index.html

   tar -xf cudnn-linux-x86_64-*.tar.xz
   sudo cp cudnn-linux-x86_64-*/include/* /usr/local/cuda/include/
   sudo cp cudnn-linux-x86_64-*/lib/* /usr/local/cuda/lib64/
   

   Warning

   You may encounter problem if you don’t have the official kernel version as cuDNN works with specific version of kernel: https://docs.nvidia.com/deeplearning/cudnn/support-matrix/index.html. (In my case cuDNN 8.5 is properly working with kernel 5.19 although 5.15 version is officialy supported)

6. Install tensorflow on conda:

   As of now, I’ve not being able to install tensorflow with GPU capability using conda: conda install -c conda-forge tensorflow=2.9 (from v2.0, the tensorflow package already include gpu capbility). Using pip did work:

   pip install tensorflow-gpu==2.9
   # If you want to define the environment variables only for the current conda environment
   conda env config vars set PATH="/usr/local/cuda/bin:$PATH"
   conda env config vars set LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/local/cuda/include:$LD_LIBRARY_PATH"
   

   You can check tensorflow properly detect your GPU using the command:

   import tensorflow as tf
   tf.test.gpu_device_name()