I just installed JetPack 5.1 on AGX Orin, but I don’t see a corresponding binary package with PyTorch. Folder Index of /compute/redist/jp has versions up to v502. Should I use v502 or wait for v51?
Perhaps torch-1.13.0a0+410ce96a.nv22.12-cp38-cp38-linux_aarch64.whl is the correct version for JetPack 5.1? It’s not included in compatibility matrix at PyTorch for Jetson Platform Release Notes :: NVIDIA Deep Learning Frameworks Documentation, so it seems possible.
I installed torch-1.13.0a0+410ce96a.nv22.12-cp38-cp38-linux_aarch64.whl on JetPack 5.1 and it seems to be working. I’d prefer a “proper” version for JetPack 5.1, though.
Thanks @pauljurczak, yes we are working on the PyTorch wheels for JetPack 5.1 and plan to have them available soon hopefully. For now, I think you should be able to use the wheels from JetPack 5.0 (I was able to use those as well).
Hi,
I found the following pytorch wheel file (for Jetpack 5.1) in the Deep Learning Frameworks Documentation:
torch-1.14.0a0+44dac51c.nv23.01-cp38-cp38-linux_aarch64.whl
Does this mean Jetson Orin officially support PyTorch 2.0 now?
Does this mean Jetson Orin officially support PyTorch 2.0 now?
It seems so. I upgraded torch-1.13 to torch-1.14 and it works fine with JetPack 5.1.
I wonder why a link to v51 doesn’t show on Index of /compute/redist/jp page?
Yeah, it’s not updated yet.
I also did not find a supported torchvision version for torch-1.14. I compiled torchvision 0.14.1 from scratch and it seems to load using import torchvision. But need to verify if it actually works or not.
That was the version of torchvision I tried too (it seemed to work fine). You can also build PyTorch from scratch on an AGX Orin if you want a different version.
I’ve built torchvision-0.15.0 from source and it works fine here, as far as I can tell.
@pauljurczak Can you please share the branch name / link of torchvision-0.15.0?
git clone https://github.com/pytorch/vision
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.