Install CUDA and TensorFlow in WSL2

Install directly in the system without using Anaconda.

Install CUDA

First, find the corresponding CUDA version according to the TensorFlow you want to install, and install the corresponding CUDA. For example, I want to install TensorFlow version 2.12, which uses CUDA 11.8, so I download CUDA from here. I use Linux-X86-64-WSL-Ubuntu-2.0, and I choose the installation package of runfile(local). This is the most convenient and recommended!

In the prompt box that pops up, select Install:

image-20230602232528589

After installation, you need to manually add environment variables in ~/.bashrc, add:

export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda-11.8/lib64
export PATH=/usr/local/cuda-11.8/bin${PATH:+:${PATH}}

relaunch terminal or source ~/.bashrc.

At this point, enter nvcc --versions to check whether the installation is successful.

image-20230602232739577

Install TensorFlow

The installation instructions of the HPCC GNN Bundle mention that

Tensorflow should be installed using su so that all users can see it, and must be installed using the same version of Python3 as is embedded in the HPCC Systems platform.

1. Enter root mode, su root

2. Install cuDNN: pip install nvidia-cudnn-cu11==8.6.0.163

3. Add environment variables in ~/.bashrc:

   CUDNN_PATH=$(dirname $(python -c "import nvidia.cudnn;print(nvidia.cudnn.__file__)"))
   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDNN_PATH/lib

4. Upgrade pip: pip install --upgrade pip

5. Install TensorFlow using pip: pip install tensorflow==2.12.*

6. Check whether the installation is successful: exit root mode and run python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

7. Or you can use this to show the name of GPU(s):

   import tensorflow as tf
    
   gpu_device_name = tf.test.gpu_device_name()
   print(gpu_device_name)