OSError: libcufft.so.10: cannot open shared object file: No such file or directory

[nano-25131]

Install Ubuntu 20.04 on Jetson Nano

when i use torch and install and then

Python 3.8.10 (default, Mar 18 2025, 20:04:55)
[GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.

import torch
Traceback (most recent call last):
File "", line 1, in
File "/usr/local/lib/python3.8/dist-packages/torch/init.py", line 191, in
_load_global_deps()
File "/usr/local/lib/python3.8/dist-packages/torch/init.py", line 153, in _load_global_deps
ctypes.CDLL(lib_path, mode=ctypes.RTLD_GLOBAL)
File "/usr/lib/python3.8/ctypes/init.py", line 373, in init
self._handle = _dlopen(self._name, mode)
OSError: libcufft.so.10: cannot open shared object file: No such file or directory

why?

[Qengineering]

A few showstoppers.
This is a very old PyTorch version. Currently, you have 2.9.0. This version is 1.13.0. It all has to do with CUDA. The Jetson Nano has CUDA version 10. A higher version can't be installed simply because of the lack of the right low-level drivers.
Hence, every piece of software, PyTorch or Tensorflow, that needs a higher CUDA version can not be used.

In practice, many models require a higher PyTorch version than our 1.13.0. They cannot be applied to the Jetson Nano.

Another important point is the OS. You have an ARM64 system under the hood. Not a x86_64. In other words, most pip installations will fail because they install x86_64 software. Be very careful!

In practice:
If you installed packages with pip, remove them first.
Next, install the supplied wheels to install PyTorch and TorchVision, as described in the Readme.

[nano-25131]

I followed your previous guide to install Ubuntu 20.04 on the Jetson Nano. I didn’t directly flash your prebuilt image — instead, I went through the installation steps one by one.

After the installation, running nvcc -V showed that the CUDA version was 13.0. Then, I tried installing PyTorch, but it failed. Later, I reinstalled the CUDA version that comes with the NVIDIA SDK, and the CUDA version changed to 10.2.

After that, I was able to successfully install PyTorch and torchvision, but when I ran my code, it kept showing the error:

RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED

Then I executed the following commands:

sudo apt remove libcudnn8
sudo apt install libcudnn8=8.2.1.*+cuda10.2

and everything worked fine after that.

[Qengineering]

Glad to hear!