Installing/Building CUDA enabled pytorch >= 2.9 for python 3.14 on the Jetson Orin Nano

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

Hello! Jetson noob here.

I have a Jetson Orin Nano (JetPack version 6.2 & CUDA version 12.6).

My goal is to use torch >= 2.9 with CUDA and python 3.14

My understanding is:

  1. Currently, the jetson-ai index (https://pypi.jetson-ai-lab.io/jp6/cu126) only offers: torch-2.9.1-cp310-cp310-linux_aarch64.whl

  2. If I want something that looks like (torch-2.9.1-cp314-cp314-linux_aarch64.whl or torch-2.10-cp314-cp314-linux_aarch64.whl, …), I am going to have to build it myself.

  3. Jetson containers are the go-to for building custom wheels. Meaning, I will have to run something that looks like this: LSB_RELEASE=24.04 CUDA_VERSION=12.6 PYTHON_VERSION=3.14 PYTORCH_VERSION=2.10 jetson-containers build pytorch to build the image. And then extract the wheels.

  4. The simplest option is to build the container on my Jetson Orin Nano, but doing so will push it to the edge of its capacity. To help, I must provide lots of swap (~20 GB) and limit the parallelism (MAX_JOBS=4).

I have two questions:

  1. Is there a critical mistake or oversight in the above steps?

  2. Is there a roadmap for bringing python 3.14 wheels to the jetson-ai index ? I expect there will be a growing appetite for 3.14 wheels due to the multithreading upgrades.

Any feedback would be helpful,

Thank you

3

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

Based on the provided context, I’ll address the questions and provide a comprehensive answer.

Question 1: Is there a critical mistake or oversight in the above steps?

The steps outlined seem to be correct for building a custom wheel for PyTorch 1.12.1 with CUDA 11.6 and Python 3.14 on the Jetson Orin Nano. However, there are a few potential issues to consider:

  1. Swap space: Providing 20 GB of swap space is a good idea, as building the container can be resource-intensive. However, it’s essential to ensure that the swap space is properly configured and that the system is not thrashing (i.e., constantly swapping between disk and RAM).
  2. Parallelism: Limiting the parallelism to MAXJOBS=4 is a good idea, as this can help prevent the system from becoming overwhelmed. However, you may need to adjust this value depending on the specific requirements of your build process.
  3. Container build: Building the container on the Jetson Orin Nano itself may not be the most efficient approach, as it can push the system to its limits. Consider building the container on a more powerful machine and then transferring it to the Jetson Orin Nano.
  4. Wheel extraction: After building the container, you’ll need to extract the wheels. Make sure to follow the correct procedure for extracting the wheels, as incorrect extraction can lead to issues with the wheel installation.

Question 2: Is there a roadmap for bringing Python 3.14 wheels to the jetsonai index?

Unfortunately, there is no explicit roadmap mentioned in the provided context for bringing Python 3.14 wheels to the jetsonai index. However, it’s likely that the NVIDIA team will provide support for Python 3.14 wheels in the future, given the growing demand for multithreading upgrades.

In the meantime, you can consider the following options:

  1. Build your own wheels: As you’ve already outlined, building your own wheels using the Jetson containers is a viable option. This approach allows you to create custom wheels that meet your specific requirements.
  2. Request support: You can reach out to the NVIDIA team or the PyTorch community to request support for Python 3.14 wheels on the jetsonai index. This may help raise awareness and prioritize the development of Python 3.14 wheels.
  3. Monitor the jetsonai index: Keep an eye on the jetsonai index for updates on Python 3.14 wheels. You can also monitor the NVIDIA forums and PyTorch community for announcements or discussions related to Python 3.14 support.

Comparison with other options:

If you’re looking for alternative options, you can consider the following:

  1. Use a different Python version: If you’re not strictly tied to Python 3.14, you can consider using an earlier version of Python that is already supported on the jetsonai index.
  2. Use a different wheel source: You can explore other wheel sources, such as the official PyTorch wheels or third-party wheel repositories. However, be cautious when using unofficial wheel sources, as they may not be optimized for the Jetson platform.
  3. Build from source: If you’re comfortable with building from source, you can try building PyTorch from the official repository. This approach allows you to customize the build process and create a version that meets your specific requirements.

In summary, while there may not be a critical mistake or oversight in the outlined steps, it’s essential to be aware of the potential issues and limitations. Additionally, there is no explicit roadmap for bringing Python 3.14 wheels to the jetsonai index, but you can consider alternative options or request support from the NVIDIA team or PyTorch community.

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

4

Hi,

We usually build the PyTorch package with the default Python version.
For JetPack 6, it’s Python 3.10.

Thanks.