Problem with PyTorch on Jetson Orin Nano

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

Hi,

I have Jetson Orin Nano with Ubuntu 22.04, JetPack 6.2, CUDA 12.6, Python 3.10 and I have a problem with installing PyTorch. I work in my miniconda3 environment.

I tried:

wget https://pypi.jetson-ai-lab.io/jp6/cu126/+f/d44/80f94e78eb0cb/torch-2.9.0.dev20250825+cu126.g4651aaa-cp310-cp310-linux_aarch64.whl#sha256=d4480f94e78eb0cbf2073621fef4cfb495ffffd978c38fa4fd4abcbc18e335ab

pip install torch-2.9.0-cp310-cp310-linux_aarch64.whl

wget https://pypi.jetson-ai-lab.io/jp6/cu126/+f/70c/f279206c90470/torchvision-0.24.0-cp310-cp310-linux_aarch64.whl#sha256=70cf279206c90470a55fad52c2f17e5e7f98c709fa64c1cf0c1ff03a6bd274a8

pip install torchvision-0.24.0-cp310-cp310-linux_aarch64.whl

Installation went well, but after trying to import torch i got an error message:

>>import torch
Traceback (most recent call last):
File “”, line 1, in
File “/home/bat/miniconda3/envs/py310/lib/python3.10/site-packages/torch/init.py”, line 415, in
from torch._C import * # noqa: F403
ImportError: /lib/aarch64-linux-gnu/libnccl.so.2: file too short

I know that Jetson doesn’t use libnccl, but i downloaded this .whl file from official site. I also tried to use older version of torch downloaded from this website: https://developer.download.nvidia.com/compute/redist/jp/v61/pytorch/torch-2.5.0a0+872d972e41.nv24.08.17622132-cp310-cp310-linux_aarch64.whl

It seems to be working but i couldn’t find a torchvision version (20.0.0) compatible with torch 2.5.0. I looked through this forum looking for an answer, but every link i tried to use, lead to not-working websites. For example: “https://pypi.jetson-ai-lab.dev“

5

Hi,

NCCL is supported on JetPack 6.

We have tested the PyTorch package on JetPack 6.2.1 and it can work as expected.
Could you check if the CUDA package is well installed in your environment first?

$ python3
Python 3.10.12 (main, Feb  4 2025, 14:57:36) [GCC 11.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.__version__
'2.9.0.dev20250825+cu126.g4651aaa'
>>> torch.cuda.is_available()
True
>>> torch.version.cuda
'12.6'
>>> torch.distributed.is_available()
True
>>> torch.distributed.is_nccl_available()
True

Thanks.

6

Thanks for your response!

I have already found a little older version of torch (2.8.0) witch compatible torchvision (0.23.0) and it is working now.

Python 3.10.18 (main, Jun 5 2025, 13:08:10) [GCC 11.2.0] on linux
Type “help”, “copyright”, “credits” or “license” for more information.

import torch
torch.version
‘2.8.0’
torch.cuda.is_available()
True
torch.version.cuda
‘12.6’
torch.distributed.is_available()
True
torch.distributed.is_nccl_available()
True

For interested i downloaded these from:

torch: https://pypi.jetson-ai-lab.io/jp6/cu126/+f/62a/1beee9f2f1470/torch-2.8.0-cp310-cp310-linux_aarch64.whl#sha256=62a1beee9f2f147076a974d2942c90060c12771c94740830327cae705b2595fc

torchvision: https://pypi.jetson-ai-lab.io/jp6/cu126/+f/907/c4c1933789645/torchvision-0.23.0-cp310-cp310-linux_aarch64.whl#sha256=907c4c1933789645ebb20dd9181d40f8647978e6bd30086ae7b01febb937d2d1

I still have no idea why newer version of torch didnt’t work for me.

7

Hello,
nccl was bugged from pytorch some days ago. We decide to remove beta whl at this moment from pypi

9

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.