Installing PyTorch v2.1.0 and Torchvision v0.16.1 on Jetson Orin Nano with JetPack 5.1.2

Issue Overview

Users are experiencing difficulties installing PyTorch v2.1.0 and Torchvision v0.16.1 on the Nvidia Jetson Orin Nano Developer Kit running JetPack 5.1.2. The main issues include:

1. Incorrect installation of PyTorch version (1.8.0 instead of 2.1.0)
2. Error when installing Torchvision due to missing shared object files
3. Compatibility issues between PyTorch, Torchvision, and the Jetson Orin Nano’s software environment

These problems prevent users from setting up the desired deep learning environment on their Jetson Orin Nano devices.

Possible Causes

1. Incorrect wheel file: The user initially downloaded and attempted to install the wrong PyTorch wheel file, which led to the installation of an older version (1.8.0) instead of the desired 2.1.0 version.

2. Missing dependencies: The system lacks some required shared object files (libmpi_cxx.so.20, libcufft.so.11, libcublas.so.*[0-9]), which are necessary for PyTorch and Torchvision to function properly.

3. Incompatible versions: There might be compatibility issues between the installed PyTorch version, the Torchvision version being installed, and the JetPack version running on the Jetson Orin Nano.

4. Incorrect system path: The required libraries may not be in the system path, causing Python to fail in locating them during the installation process.

5. JetPack version mismatch: The user might have attempted to use a PyTorch wheel file intended for a different JetPack version, leading to compatibility issues.

Troubleshooting Steps, Solutions & Fixes

1. Install the correct PyTorch wheel:

   • Download the appropriate PyTorch wheel file for JetPack 5.1.2:

     wget https://developer.download.nvidia.cn/compute/redist/jp/v512/pytorch/torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl
     

   • Install the wheel file:

     pip3 install numpy torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl
     

2. Verify PyTorch installation:
   After installation, check the installed PyTorch version:

   python3 -c "import torch; print(torch.__version__)"
   

3. Install Torchvision dependencies:

   sudo apt-get install libjpeg-dev zlib1g-dev libpython3-dev libopenblas-dev libavcodec-dev libavformat-dev libswscale-dev
   

4. Clone and install Torchvision:

   git clone --branch v0.16.1 https://github.com/pytorch/vision torchvision
   cd torchvision
   python3 setup.py install --user
   

5. Locate missing shared object files:
   If you encounter errors about missing .so files, locate them on your system:

   find /usr/lib -name 'libcufft.so*'
   find /usr/lib -name 'libcublas.so*'
   

   If found, add their directory to your LD_LIBRARY_PATH:

   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path/to/directory
   

6. Check CUDA installation:
   Ensure that CUDA is properly installed and configured on your Jetson Orin Nano. You can verify this by running:

   nvcc --version
   

7. Update JetPack:
   If issues persist, consider updating to the latest compatible JetPack version for your Jetson Orin Nano. Check the NVIDIA website for the most recent version and installation instructions.

8. Rebuild PyTorch from source:
   If all else fails, you may need to build PyTorch from source to ensure compatibility with your specific Jetson Orin Nano setup. Follow the official PyTorch documentation for building on Jetson platforms.

9. Seek community support:
   If problems continue, consider posting detailed error messages and your system configuration on the NVIDIA Developer Forums or PyTorch community channels for more specialized assistance.

Remember to reboot your Jetson Orin Nano after making significant changes to ensure all updates take effect. Always backup important data before making major system changes.