Installation Issues with cuDNN on Nvidia Jetson Orin Nano

Issue Overview

Users are experiencing difficulties with the installation of cuDNN on the Nvidia Jetson Orin Nano Dev board, particularly in achieving compatibility with PyTorch. The primary symptoms include:

• Errors Encountered: Users report an ImportError stating libcudnn.so.8: cannot open shared object file: No such file or directory when attempting to run PyTorch after manually installing cuDNN 9.4.

• Installation Context: The issue arises after users have flashed their Jetson Orin Nano and manually installed CUDA and other packages, but not cuDNN, which leads to version compatibility issues.

• Hardware and Software Specifications:

  • CUDA version: 12.4
  • L4T version: 36.2.0
  • Kernel Release: 5.15.122-tegra

• Frequency of Issue: This problem appears to be consistent among users who manually install packages instead of using the default JetPack installation.

• Impact on User Experience: The inability to run PyTorch due to incompatible cuDNN versions significantly hampers users’ ability to utilize the Jetson Orin Nano for machine learning tasks.

Possible Causes

The following potential causes for the issue have been identified:

• Version Incompatibility: Installing cuDNN 9.4 when PyTorch requires a different version (specifically, cuDNN 8) can lead to runtime errors.

• Manual Installation Errors: Users who manually install packages may inadvertently choose incompatible versions or miss dependencies.

• Incomplete Package Installation: If the initial flashing process does not include all necessary packages, it can result in missing libraries required for proper functionality.

• Configuration Errors: Incorrect environment variables or paths that do not point to the correct library locations can prevent the system from finding the required shared objects.

Troubleshooting Steps, Solutions & Fixes

To resolve the issues related to cuDNN installation on the Nvidia Jetson Orin Nano, follow these steps:

1. Uninstall cuDNN 9.4:

   • To uninstall cuDNN properly, use:

     sudo apt-get remove --purge libcudnn*
     

2. Install Default JetPack Packages:

   • Reflash your device if necessary and ensure all default packages are installed:

     sudo apt-get update
     sudo apt-get install nvidia-jetpack
     

3. Install Compatible cuDNN Version:

   • For compatibility with PyTorch, install cuDNN version 8.x (specific version may depend on your PyTorch installation). You can find compatible versions in the cuDNN Archive.
   • Example command to install a specific version (replace <version> with the desired version):

     sudo dpkg -i libcudnn8_<version>_arm64.deb
     

4. Install PyTorch:

   • After ensuring that cuDNN is correctly installed, install PyTorch using:

     wget https://developer.download.nvidia.com/compute/redist/jp/v60/pytorch/torch-1.12.0+cu124-cp310-cp310-linux_aarch64.whl -O torch-1.12.0+cu124-cp310-cp310-linux_aarch64.whl
     pip install torch-1.12.0+cu124-cp310-cp310-linux_aarch64.whl
     

5. Verify Installation:

   • To check if PyTorch and cuDNN are correctly installed and compatible, run:

     import torch
     print(torch.cuda.is_available())
     print(torch.backends.cudnn.version())
     

6. Best Practices for Future Installations:

   • Always refer to official documentation for compatibility matrices between CUDA, cuDNN, TensorRT, and PyTorch.
   • Consider using JetPack for initial installations to avoid manual conflicts.
   • Regularly check for updates and patches from NVIDIA that may address compatibility issues.

By following these troubleshooting steps and solutions, users should be able to resolve the installation issues with cuDNN on their Nvidia Jetson Orin Nano Dev board effectively.