Installing pytorch on our jetson agx orin fails with: wheel version does not match filename

Robotics & Edge Computing Jetson Systems Jetson AGX Orin
1

Hi everyone,

I am trying to install the pytorch wheel nvidia published under this link:

https://developer.download.nvidia.com/compute/redist/jp/v61/pytorch/torch-2.5.0a0+872d972e41.nv24.08.17622132-cp310-cp310-linux_aarch64.whl

With this command:

uv pip install “torch @ https://developer.download.nvidia.com/compute/redist/jp/v61/pytorch/torch-2.5.0a0+872d972e41.nv24.08.17622132-cp310-cp310-linux_aarc64.whl”

But my installation fails with this error:

Wheel version does not match filename: 2.5.0a0-872d972e41.nv24.8 1= 2.5.0a0-872s972-41.nv24.8.17622132

I was trying to install this wheel because I believe the normal pytorch cuda wheels do not support the jetson GPU architecture?

Does someone maybe know what I am doing wrong here? (I assume this is my fault :) )

Thanks in advance!

3

Hi,

Which JetPack version do you use?
Please try the packages below for the JetPack 6 environment:

Thanks.

4

pytorch now builds wheels with cuda that will do what you want.

https://pytorch.org/get-started/locally/ pick linux, your-cuda-version then copy the block next to Run this Command:

If you don’t want torch-2.13.0 just paste the url from Run this Command box in your browser and scroll down and click torch. You want a wheel with your cuda version, without the . dot, in the filename. Ctrl-f and search for, for example, cu132 for cuda-13.2

5

Hi,
I checked in the terminal with “cat /etc/nv_tegra_release” and it gave me this:
R36 (release) revision 4.4
I believe this corresponds to Jetpack 6.2, but I am still trying to figure out how I match these versions to the output.
The versions you linked are a little newer than the code was written with, so I don’t know if that is worth trying, but I will try anyway. (I believe the code was written with pytorch 2.6.0)

6

Hi whitesscott,
I forgot to add this to my original post, but the reason I am trying to install the nvidia versions is that the PyTorch-provided version failed to run the code with an error something like: “No kernel image is available”. Which I thought meant that that version did not support this architecture.
If the PyTorch versions are supposed to support the Jetson, could this maybe be an issue because I am trying to use a version that is quite old (2.6.0 or 2.5.0)?

7

Hi AastaLLL, Unfortunately this version is too new for the code to work…

I did however figure out how to install the pytorch version on doe developer.download.nvidia.com site (by changing the wheel name), but now I have the issue that there is no torchaudio wheel for that version…

Do you know if this is available somewhere?

8

I managed to fix the issue myself, I found the torchaudio wheel somewhere locally, and used the link in my first post to install the correct pytorch version.