PyTorch and Torchvision version issue: RuntimeError: operator torchvision::nms does not exist (JetPack 6.2.1)

Robotics & Edge Computing Jetson Systems Jetson Nano
1

Hello!

Like a previous user, I am trying to get my code running correctly using CUDA on my Jetson AGX Orin.

I am running into the same error with torchvision:

RuntimeError: operator torchvision::nms does not exist

To reproduce the error, I just need to run code that imports torchvision:

import torch
import torchvision

Per the last discussion on this topic, this error results from incompatible PyTorch and Torchvision installations. Hence, I have reinstalled the two libraries multiple times in different ways, failing in different ways each time (like a previous user):

  1. Normal pip/pip3 install — pip3 defaults to cpu binaries for the packages, resulting in my code returning False for torch.cuda.is_available()

  2. Install torch as instructed here, from https://developer.download.nvidia.com/compute/redist/jp/v61/pytorch/torch-2.5.0a0+872d972e41.nv24.08.17622132-cp310-cp310-linux_aarch64.whl — results in torch and torchvision version discrepancy, BUT torch.cuda.is_available() returns True.

  3. Install torch and torchvision via jetson-containers — this results in the ‘nms does not exist’ error shown above.

Some information about my system:
Currently, it is running JetPack 6.2.1 with torch version “2.5.0a0+872d972e41.nv24.08.17622132-cp310-cp310-linux_aarch64” and torchvision version 0.20.0.

I would appreciate any help with this.

Thanks!

P.S. all the previous links to http://jetson.webredirect.org in the previous topics are dead

4

Hi,

Please try the legacy prebuilt wheels.

Thanks

5

Thanks! That seemed to do the trick!

6

Hi,

Thank you for your update.

Please mark the resolution to close the topic.

Thanks