Hi, I’m currently trying to install PyTorch with CUDA support natively (not using Docker) on my Jetson AGX Orin (64GB), and I am running into issues with missing libcudnn.so.8 when running Ultralytics YOLO pose models. It seems the PyTorch installation I have is CPU-only and not built for Jetson. Here are my system details: JetPack / L4T: 6.1 (R36.4.0)
Board: Jetson AGX Orin Developer Kit 64GB
OS: Ubuntu 22.04 (aarch64)
Python: 3.10.12
CUDA: 12.6.68
cuDNN: 9.3.0
TensorRT: 10.3.0.30
OpenCV: 4.10.0 with CUDA: YES
Output of cat /etc/nv_tegra_release: R36 (release), REVISION: 4.0, GCID: 37537400, BOARD: generic, EABI: aarch64
also when i remove the cpu based pytorch and tried to install pytoch with cuda enabled support, it couldn’t work.
Update:
I also tried installing from the Jetson-AI-Lab custom index at:
https://pypi.jetson-ai-lab.io/jp6/cu126
using the command:
pip3 install torch torchvision torchaudio --extra-index-url https://pypi.jetson-ai-lab.io/jp6/cu126
However, this installed a CPU-only version (torch==2.9.0+cpu) — and torch.cuda.is_available() returned False. This confirms that the index is not serving Jetson-native, CUDA-enabled builds for JetPack 6.1 / Python 3.10 / CUDA 12.6 (even though it’s supposed to).
I also attempted to use .whl files shared in past forum posts (e.g., from nvidia.box.com), but many of those links are now dead (404 errors). This makes it difficult to retrieve any verified torch-2.3.0-cp310-cp310-linux_aarch64.whl or matching torchvision builds that are known to work for Jetson AGX Orin running JetPack 6.1.