PyTorch on Jetson installation instructions gives a bad example

If you followed the instructions from Installing PyTorch for Jetson Platform - NVIDIA Docs, try to install PyTorch on JetPack, you would see some errors.

Python community is pretty amateur at version management. For whatever reason, the document gives this instruction:

python3 -m pip install --upgrade pip; python3 -m pip install numpy==’1.26.1’ python3 -m pip install --no-cache $TORCH_INSTALL

where $TORCH_INSTALL is either:

v511/pytorch/torch-2.0.0+nv23.05-cp38-cp38-linux_aarch64.whl

or

v512/torch-2.2.0a0+81ea7a4+nv23.12-cp38-cp38-linux_aarch64.whl

v511 and v512 are JetPack versions, p38 is Python version. At this point, Jetson Pack 6 is still in preview stage. The problem is, numpy==’1.26.1’ (should be numpy==1.26.1 without quote) requires python 3.9 or later, and these JetPack-PyTorch are only good for python 3.8. If you copy-pasted the command, numpy will fail. If you bumped up python version to 3.9 to please numpy 1.26.1, JetPack-PyTorch will fail.

Fortunately, this JetPack-PyTorch isn’t picky on that particular numpy version as appears. Stay with python 3.8 and whatever the latest numpy you have, JetPack-PyTorch will install.

Hi,

Thanks for sharing.

Usually, the command is for the latest software release, which is JetPack 6 DP.
So the version might be not suitable for other environments.

Thanks.

JetPack 6 DP isn’t available for my setup (Xavier Agx). The document author should really pay attention to the details, it impacts the user’s productivity a lot. In those examples, they are really for v511 and v512 respectively.

Hi,

We will pass your suggestion to the corresponding team.
Thanks.