Jetson Orin Nano Setup disaster

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin Nano
1

Hello,

I’m unable to find proper versions and guidance for cuda enabled pytorch. I have deployed/flashed JetPack 6.2, but no way to get the GPU up and running out-of-the box. I have tried multiple links, forums, guides, etc, but all of them leads to some version mismatch errors in the end. I want my money back, this product is a joke! It is sold as a GPU solution, but no way to make this thing run.. I am using: torch-2.8.0-cp310-cp310-linux_aarch64.whl. It is still possible that I am missing something, even after 3 days of troubleshooting..

Python 3.10.12 (main, Nov 6 2024, 20:22:13) [GCC 11.4.0] on linux
Type “help”, “copyright”, “credits” or “license” for more information.

import torch
Traceback (most recent call last):
File “”, line 1, in
File “/home/mdegrauw/.local/lib/python3.10/site-packages/torch/init.py”, line 416, in
from torch._C import * # noqa: F403
ImportError: libcudss.so.0: cannot open shared object file: No such file or directory

2

*** Please note that this reply is generated by LLM automatically ***

I understand your frustration, and I’m here to help you troubleshoot the issue.

Firstly, I’d like to acknowledge that you’ve tried multiple links, forum guides, and versions, but still, you’re facing version mismatch errors. I’ll provide you with a step-by-step guide to help you install CUDA-enabled PyTorch on your Jetson Orin Nano.

Prerequisites:

  1. Ensure you have a clean installation of JetPack 6.2 on your Jetson Orin Nano.
  2. Verify that your Python version is 3.10.12 (as mentioned in your error message).

Step 1: Install CUDA

  1. Download the CUDA toolkit for Jetson Orin Nano from the official NVIDIA website: https://developer.nvidia.com/cuda-downloads
  2. Follow the installation instructions for the CUDA toolkit. Make sure to select the correct architecture (aarch64) and CUDA version (11.4 or later) for your Jetson Orin Nano.

Step 2: Install cuDNN

  1. Download the cuDNN library for Jetson Orin Nano from the official NVIDIA website: https://developer.nvidia.com/rdp/cudnn-archive
  2. Follow the installation instructions for cuDNN. Make sure to select the correct architecture (aarch64) and cuDNN version (8.2 or later) for your Jetson Orin Nano.

Step 3: Install PyTorch

  1. Install PyTorch using the following command:
pip3 install torch torchvision torchaudio --index-url https://pypi.jetson-ai-lab.dev/sbsa/cu128

This command installs PyTorch with CUDA support using the --index-url flag, which points to the NVIDIA PyTorch repository for Jetson devices.

Step 4: Verify PyTorch Installation

  1. Run the following command to verify that PyTorch is installed correctly:
python3 -c "import torch; print(torch.__version__)"

This should print the PyTorch version (e.g., 1.12.0a0+git20221124).

Step 5: Verify CUDA Support

  1. Run the following command to verify that CUDA is enabled for PyTorch:
python3 -c "import torch; print(torch.cuda.is_available())"

This should print True if CUDA is enabled.

If you’ve followed these steps and still encounter issues, please provide more details about the errors you’re facing, and I’ll be happy to help you troubleshoot further.

Regarding your comment about wanting your money back, I understand your frustration, but I’d like to assure you that the Jetson Orin Nano is a powerful device capable of delivering excellent performance for AI and computer vision applications. With the correct guidance and support, you can unlock its full potential.

Please let me know if there’s anything else I can help you with.

*** Please give the thumbs up if you get answers you like. Or provide feedback to help us improve the answer. ***

4

Consider adding these instructions to your physical box, or to create a setup website. It is very confusion, to say at least, how to setup the device.

1 Like
5

I give up. I followed your instructions step-by-step. same error. Your Jetson product.. wow! You should go back to the drawing boards. You take developer’s kit too literally! I don’t understand why the card isn’t flashed already with the correct wheel out-of-the box.. It is a huge headache to set this up, and counting.. The product is useless without GPU.

$ python
Python 3.10.12 (main, Nov 6 2024, 20:22:13) [GCC 11.4.0] on linux
Type “help”, “copyright”, “credits” or “license” for more information.

import torch
Traceback (most recent call last):
File “”, line 1, in
File “/home/mdegrauw/.local/lib/python3.10/site-packages/torch/init.py”, line 416, in
from torch._C import * # noqa: F403
ImportError: libcudss.so.0: cannot open shared object file: No such file or directory

6

Is this what we are looking for?

Python 3.10.12 (main, Nov 6 2024, 20:22:13) [GCC 11.4.0] on linux

Type “help”, “copyright”, “credits” or “license” for more information.

>>> import torch

>>> print(torch.version.cuda)

12.6

>>> print(torch._version_)

2.8.0

7

You have it installed or it would not echo the versions.

8

Sounds frustrating. One thing to try is to install the CUDA sparse solver library which appears to be missing on your system : cuDSS | NVIDIA Developer

1 Like
9

Just wanted to drop this here — if you haven’t checked out jetson-containers, it’s worth a look. Basically, with Docker you get a prebuilt Python + CUDA + PyTorch environment that just works on Jetson.

There’s definitely a learning curve (matching JetPack / CUDA / PyTorch versions, building images, etc.), but when you think about what this little box can do, it’s pretty amazing.

I’m keeping my 64 GB unit — the extra RAM makes it so much easier to build and run these containers without hitting walls.

10

Sure! The point is that people buy these units to experiment with the GPU inside, at least I did.. I finally got it work, without the Docker containers. Some PATH references where missing, that’s all.

1 Like
11

Yeah a headache containers seems the way to go why didn’t you wanna use them?

12

I don’t use any containers, life is much simpler and more robust. Containers might be fine if your board is in a bubble, our stuff has to talk to other devices and the best way is directly.

13

I’m facing the exact same situation. 3 days lost playing with package versions. Ending with pytorch 2.8.0 but missing libcudss.so.0 file, I will give a try with the above solution. Could you share the missing PATH references please ?
Thanks

14

I had the same error. Following the instructions and installing the cuDSS works for me!