JetPack 6.0 – Missing PyTorch + CUDA Support for Jetson Inference (Need Official Wheel or Build Instructions)

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin Nano
1

Hi team,

I’m working on a production-quality computer vision demo for a client using a Jetson Orin Nano (JetPack 6.0 / L4T R36.4.3). The project involves a YOLOv5 model running in real-time via webcam, and while the pipeline is fully functional, we’re currently stuck using CPU inference.

Despite building a clean Docker container and confirming that OpenCV, Torch, and my detection scripts are working — torch.cuda.is_available() returns False, and there’s no CUDA-enabled PyTorch wheel available for JetPack 6.0 via pip or the NVIDIA Python Index.

I’ve also attempted:

  • Installing torch==2.2.0+nv24.03 with --extra-index-url=https://pypi.nvidia.com (fails: no matching version)
  • Manually searching for .whl builds in the forums and NGC catalog
  • Reading JetPack 6 early access threads for hints on how to build from source (no confirmed recipe yet)

💬 My questions:

  1. Is NVIDIA planning to publish a CUDA-enabled PyTorch .whl for JetPack 6.0 soon?
  2. Is there a reliable workaround or unofficial wheel that supports GPU inference under JetPack 6.0?
  3. Should we consider downgrading to JetPack 5.1 if we require GPU support for YOLOv5 now?

As a developer using Jetson hardware specifically for its GPU acceleration capabilities, being limited to CPU inference at the moment is a major blocker.

Any help or updates would be appreciated!

Thanks,
Carelia Rojas

3

Hi,

Just want to double-confirm your environment.
Do you use JetPack 6.2? As r36.4.3 comes from JetPack 6.2.

Thanks.

4

Hi,

Yes, I’m using JetPack 6.2, which includes L4T r36.4.3.

Let me know if you need any more details about the setup.

Thanks!

5

Hi,

As TensorRT 10.3 removes the Caffe/uff support, some models and examples in jetson-inference won’t work.
It’s recommended to use JetPack 6.0/TensorRT 8 to run the jetson-inference.

But for PyTorch, you can find a prebuilt package for JetPack 6.2 below:

For YOLO, you can check if Ultralytics source can meet your requirements.
Their example works well on the JetPack 6.2 environment.

Thanks.

6

Hi,

Thanks for the guidance. I followed your instructions and installed PyTorch using the jp6/cu126 index from pypi.jetson-ai-lab.dev, as recommended. However, after installation, I confirmed that:

  • torch.cuda.is_available() returns False

  • torch.cuda.get_device_name(0) raises: AssertionError: Torch not compiled with CUDA enabled

This suggests that the installed torch package is still a CPU-only build, even though I’m running JetPack 6.2 on a Jetson Orin Nano.

Could you please confirm:

  1. Whether the jp6/cu126 index currently includes a CUDA-enabled PyTorch build for Jetson Orin Nano with JetPack 6.2?

  2. If there are specific version pins (e.g., torch==2.x.x) I should use to get GPU support?

  3. Any known issues or workarounds if the GPU version fails to install correctly?

Thanks in advance for your support.

Best regards,
Carelia