I noticed two issues on my Jetson Orin AGX running JetPack 6.2.1:
The PyPI index for jp6/cu128 (https://pypi.jetson-ai-lab.io/jp6/cu128) does not include PyTorch 2.8.0, whereas the jp6/cu126 index does. It would be very helpful if PyTorch 2.8.0 could be added to the cu128 index.
I have installed CUDA 12.8 and cuDNN 9.10.2 manually. As a result, jtop reports the JetPack version as missing. Could you clarify if this has any consequences, and suggest any recommended steps to address or mitigate this?
Thanks for your replies so far. I want to clarify that my use case does not prefer using a container (Docker / Jetson container) — I would like a native / system-level installation of PyTorch 2.8.0 + CUDA 12.8 (cu128) on JetPack 6.2.1 (Orin AGX).
If there is any existing non-container build (wheel, apt package, or binary) for that combination, I’d greatly appreciate a link. Otherwise, could someone share step-by-step instructions or patches to build it from source (with CUDA support)?
Also, there was a second question earlier in this thread that hasn’t gotten a response — could someone help with that as well?
Thank you for sharing the jtop fix — following the guide, jtop now correctly shows JetPack 6.2.1. Really appreciate it! 🙏
Additionally, I’m still hoping someone can help with the first issue: PyTorch 2.8.0 is missing from the JP6/cu128 PyPI index for Jetson Orin AGX with JetPack 6.2.1 and CUDA 12.8. If anyone has a non-container installation method (wheel, apt package, or source build instructions), I would be very grateful if you could share it.
Due to limited resources, we now only support the default JetPack 6 and JetPack 7 environments.
But you should be able to build it from the source.
There is a sharing from @whitesscott that tries to enable the MPI for PyTorch.
You can follow the command to build PyTorch from the source (can skip the MPI part):