Torchvision operator torchvision::nms does not exist with NVIDIA PyTorch nv24.07 on Jetson Orin NX (JetPack 6.x)

Robotics & Edge Computing Jetson Systems Jetson Orin NX
1

Hello NVIDIA Team,

I am facing a torchvision ABI incompatibility issue on Jetson Orin NX 16GB while using NVIDIA-provided PyTorch (nv24.07).

System Details

  • Device: Jetson Orin NX 16GB

  • JetPack: 6.0 (CUDA 12.9)

  • OS: Ubuntu 22.04

  • Python: 3.10 (venv)

    image
    image512×154 46.2 KB

Installed Packages

pip show torch


Name: torch
Version: 2.4.0a0+3bcc3cddb5.nv24.7
Location: site-packages


pip show torchvision


Name: torchvision
Version: 0.19.0
Location: site-packages

Torch was installed from the official NVIDIA Jetson wheel, and torchvision was installed from the NVIDIA JetPack repo:

pip install torchvision==0.19.0 \
  --no-deps \
  --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v60

Issue Encountered

When importing torchvision ops, I consistently get:

RuntimeError: operator torchvision::nms does not exist

  • Clean virtual environment

  • Reinstall torch (NVIDIA nv24.07 wheel)

  • Force reinstall torchvision with:

    • --no-deps

    • --no-cache-dir

    • --force-reinstall

  • Manually deleting torchvision* directories

  • Clearing pip cache

The error persists, which suggests that:

  • The torchvision 0.19.0 wheel being installed is not the NVIDIA-compiled Jetson build, even though it comes from the JetPack repo URL.

Questions regarding my problem

  1. Is there an officially supported torchvision wheel that is ABI-compatible with PyTorch 2.4.0a0 nv24.07 on JetPack 6.x?

  2. If yes, could you please provide:

    • The exact torchvision version

    • Or a direct .whl download link

  3. If no:

    • Is building torchvision from source the recommended approach for JP6 + nv24.07?

Thank you for your time and support.
Looking forward to guidance on the official torch ↔ torchvision compatibility for JetPack 6.x.

Best regards,
Tanish Jain
Ai India innovation

4

Hi,

https://developer.download.nvidia.com/compute/redist/jp/v60
The link doesn’t have the torchvision so you will install the upstream package instead.

We don’t release TorchVision together with 2.4.0a0 nv24.07 so you will need to build it from the source.
But if newer JetPack is an option, please upgrade your device with JetPack 6.2.1 and install the packages from the following link:

Thanks.

1 Like
6

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.