AGX Orin: nvcc missing & torch.cuda.is_available() False after JetPack 6.x Installs

Robotics & Edge Computing Jetson Systems Jetson AGX Xavier
,
1

Hi all,

I’m having ongoing issues with my AGX Orin after installing JetPack 6.2 (and previously 6.1). The JetPack installs appear successful, but CUDA and PyTorch are not working correctly.

Here’s a breakdown:

  1. nvcc --version Not Found: This is a primary problem. I’ve tried reinstalling the CUDA toolkit, but nvcc remains unrecognized.
  2. PyTorch CUDA Issues:
  • I’m using official PyTorch wheels for Jetson.
  • Even when nvidia-smi once showed GPU activity, torch.cuda.is_available() is always False.
  1. Multiple Reinstalls: I’ve reflashed and reinstalled JetPack 6.2 and 6.1 multiple times without success; the errors persist.

It seems CUDA isn’t being installed/recognized correctly, preventing PyTorch from using the GPU.

Could anyone offer guidance or point to resources for a proper JetPack 6.x (6.2 or 6.1) setup on AGX Orin, ensuring CUDA works and is accessible to PyTorch?

I’m looking for tips on:

  • Ensuring nvcc is installed and in the system PATH.
  • Verifying CUDA toolkit installation independently.
  • The correct method for installing compatible PyTorch wheels for the Jetson CUDA version.
  • Troubleshooting nvidia-smi showing GPU details while torch.cuda.is_available() is false (and confirming if nvidia-smi is appropriate for Jetson, or if I should use jtop/tegrastats).
3

Hi,

You will need to install the package that is built with CUDA support.
For example, you can find our PyTorch prebuilt below:

Thanks.