Hi everyone, I’m trying to run a YOLO model using Ultralytics YOLO on a Jetson Orin Nano, but I keep running into compatibility issues between PyTorch, TorchVision, JetPack, and CUDA, and I can’t seem to get a working setup.
My current setup is Jetson Orin Nano with JetPack 6.2.x and CUDA 12.6 on Ubuntu. The main problem is that when I install PyTorch and TorchVision (usually via pip), I get errors like “RuntimeError: operator torchvision::nms does not exist,” which makes it seem like TorchVision is not built against the correct PyTorch or CUDA version. I understand that JetPack determines the CUDA version, but it’s unclear which exact PyTorch and TorchVision versions are actually compatible with this setup.
I’ve tried multiple approaches but keep ending up in the same loop. First, I tried running Ultralytics directly with PyTorch, installing different combinations of torch and torchvision, but they always end up being incompatible. Then I tried exporting the model to ONNX and using TensorRT, but that led to other issues like missing or incompatible dependencies such as PyCUDA. I also reflashed the system multiple times to start clean, but the same problems keep coming back depending on what versions I install.
At this point, I feel like I’m stuck in a loop where I install one thing, it breaks something else, and fixing that creates another issue. I think the root problem is that I don’t have a correct baseline of versions that actually work together on Jetson with CUDA 12.6. I would really appreciate guidance on what exact combination of JetPack, PyTorch, and TorchVision I should be using, or whether I should be using a completely different approach (for example containers or a TensorRT pipeline) to run YOLO reliably on Jetson.