Upgrading cuDNN on Databricks notebook

[Amoozegar]

I'm trying to upgrade Tensorflow version from 2.8 to 2.13 on Databricks notebook that is attached to a cluster with Databricks Runtime 10.4. How can I upgrade cuDNN from 8.0 to at least 8.6 to be compatible with the Tensorflow new version?

[Kaniz]

Hi @Amoozegar, Upgrading cuDNN to a compatible version with TensorFlow 2.13 is essential for optimal performance.

Let’s walk through the steps:

1. Check Current cuDNN Version:

   • First, verify the current cuDNN version installed on your Databricks cluster. You can do this by running the following command in a Databricks Notebook:

     !conda list cudnn
     

   • Note down the version number.

2. Upgrade cuDNN Using Conda:

   • To upgrade cuDNN, you can use conda. Here are a couple of options:
     • Option 1 (Recommended): Install the latest cuDNN version from the conda-forge channel. This channel often provides more up-to-date packages than the default anaconda channel:

       !conda install -c conda-forge cudnn
       

     • Option 2: If you specifically need version 7.4.2 (as mentioned for TensorFlow 2.13), you can force an update to that version:

       !conda install -c anaconda cudnn=7.4.2
       

   • Choose the appropriate option based on your requirements.

3. Alternative Approach (Advanced):

   • Sometimes, conda might not have the exact version you need. In such cases, you can manually download the cuDNN tar file from NVIDIA’s website.
   • Extract the downloaded tar file to find the “cuda” folder containing the cuDNN files.
   • Copy and paste the cuDNN files to your conda environment’s lib and include folders. For example:

     sudo cp cuda/include/cudnn*.h /anaconda3/envs/<your_environment_here>/include
     sudo cp cuda/lib64/libcudnn* /anaconda3/envs/<your_environment_here>/lib
     

   • Replace <your_environment_here> with the actual name of your conda environment.

4. Verify Installation:

   • After upgrading cuDNN, verify that TensorFlow recognizes it. Run the following Python code in your Databricks notebook:

     import tensorflow as tf
     print(tf.config.list_physical_devices('GPU'))
     

   • If a list of GPU devices is returned, you’ve successfully installed TensorFlow with the upgraded cuDNN version.

Remember to adjust the steps based on your specific requirements and environment. Happy TensorFlow-ing! 🚀🔍

For more details, you can refer to the official TensorFlow GPU installation guide.

[Amoozegar]

Hi @Kaniz  , Thanks for your response. When I run '!conda list cudnn' on databricks notebook, I get the following error:    '/bin/bash: conda: command not found'

[Amoozegar]

Also I should mention I'm using Databricks Runtime 10.4 for ML, and my assumption is that I might not have direct control over upgrading CUDA or cuDNN within the runtime environment.