Jetson Orin Nano support for PyTorch 2.14 with Jetpack 7.1.2

Robotics & Edge Computing Jetson Systems Jetson Orin Nano
1

Jetpack 7.1.2 provides compatibility with CUDNN 9.20.

image
image1417×324 17.5 KB

However, PyTorch 2.14 requires CUDNN 9.24.

image
image1530×471 59.6 KB
,

According to cuDNN (v9.24) documentation there is support for the Jetson platform.

image
image1470×736 81.3 KB

Who should I believe here? 😄

I want to use the torch.switch functionality in PyTorch 2.14 as opposed to PyTorch 2.13 which is in line with the Jetpack 7.2.1 spec.

4

Hi,

CUDA-X libraries are upgradable.
So you can upgrade to meet the PyTorch requirement.

Or we also have a container with all dependencies inside so you don’t need to manually handle one by one:

Thanks.

5

Thanks for your response @AastaLLL !

CUDA-X libraries are upgradable.
So you can upgrade to meet the PyTorch requirement.

Where in the documentation can I find this? On the SDK download page it mentions “Supported Libraries and SDK versions”. How would we know that a CUDA-X library upgrade would not break the Jetpack software stack? Is there a CI/CD report somewhere?

Or we also have a container with all dependencies inside so you don’t need to manually handle one by one:

Some questions regarding the containers:

  • As part of the prerequisites, it is mentioned that the NVIDIA Container Toolkit is required to be installed on the host system. However, the linked repo is depreciated.
  • According to the Framework Containers Support Matrix, Release 26.08 is based on CUDA 13.4.1. Jetpack 7.2.1 says that the latest support is for CUDA 13.2.2. Is this supported on the Jetson Orin Nano? I assume that since at the bottom of the Framework Containers Support matrix it mentions PyTorch for Jetson the answer is yes?
  • The Framework Containers Support Matrix mentions the releases as 26.08, is this the same as 26.08-py3?
  • How can we confirm that the correct version of the Docker Engine, NVIDIA GPU Drivers and NVIDIA Container Toolkit is installed? I assume that we are not required to have, e.g. jtop nor TensorRT installed in the host system?

Thanks in advance. 🙂

6

Hi,

You will need to ensure that the installed or upgraded package is compatible.
For example, if you want to use the latest TensorRT 11.3.0, it depends on:

  • CUDA 13.4 dependency upgrade: TensorRT 11.3.0 packages are built against CUDA 13.4 (Debian, RPM, tar, and zip package filenames use cuda-13.4). TensorRT 11.3.1 DriveOS packages use CUDA Toolkit 13.4. Refer to the TensorRT Support Matrix for supported CUDA releases per platform and to Prerequisites for installer prerequisites.

So you will need to upgrade CUDA to 13.4 manually as well.
Please note that these packages are verified independently; only JetPack is tested with all components together.

However, the container comes with CUDA 13.4 pre-installed and supports Orin since Orin now use SBSA-based drivers.
You do not need to manually handle the compatibility inside the container, as all requirements are already pre-installed and verified.

Thanks.

7

I used the container, and so far the functionality in PyTorch 2.14 works on the Jetson Orin Nano.

Thanks for sharing the rationale behind using the containers, and the tight integration between the versions @AastaLLL . :)