Build Pytorch 2.8 with CUDA, Numpy 2 on JetPack 6.2

Robotics & Edge Computing Jetson Systems Jetson AGX Orin
,
1

Hi. I need to run Pytorch 2.8 on Jetson AGX Orin with JetPack 6.2. The pre-built wheels from NVIDIA are great, but all seem to use Numpy 1, which causes issues once torch is imported.

I have followed the good instructions here: PyTorch environment and Zed SDK conflicts on Jetpack 6.2 - #5 by AastaLLL , ensuring that I install numpy 2 along with requirements.txt and clone torch 2.8 (not 2.7) for the build.

This successfully produced a wheel that I could import with numpy 2, but then torch.cuda.is_available() and torch.backends.cuda.is_built()both return False, while torch.version.cudareturns None.

What did I miss so that torch built without CUDA? Thanks for the help.

3

I fixed it! Here are tips that should generalize across Pytorch and JetPack versions, as well as specifics that worked in my case.

General Principles

  1. When building torch using setup.py, open the file to see which environment variables can be used and which are relevant to you. For example, though the post suggested USE_NATIVE_ARCH=1 for 2.7 builds, that environment variable doesn’t seem to be used when building torch 2.8.
  2. Install build dependencies in a virtual environment. This is usually my practice, but I originally skipped it for speed. Then I regretted it.
  3. Check the errors and CMake warnings during build. This will help you catch errors and kill the build early, rather than waiting 2 hours.

Specific Steps

  1. As suggested here,
git clone --depth=1 --recursive -b v<PYTORCH_VERSION> https://github.com/pytorch/pytorch
cd pytorch
  1. Create a virtual environment for the build packages (set that up yourself using venv, uv, etc.)
  2. Install packages outlined in original post into the virtual environment
  3. Install numpy>2 into the virtual environment
  4. Obtain your CUDA Arch Bin using jtop (mine was 8.7 for AGX Orin Dev Kit)
  5. Set environment variables and build as follows:
USE_CUDA=1 USE_CUDNN=1 USE_CUSPARSELT=1 USE_CUDSS=1 USE_CUFILE=1 TORCH_CUDA_ARCH_LIST="<CUDA Arch Bin>" USE_DISTRIBUTED=1 USE_FLASH_ATTENTION=1 USE_MEM_EFF_ATTENTION=1 USE_TENSORRT=0 python3 setup.py bdist_wheel

  1. During the build, check for Cmake warnings. For example, I encountered and fixed the following

    1. -- Could NOT find CUDA (missing: CUDA_TOOLKIT_ROOT_DIR CUDA_NVCC_EXECUTABLE CUDA_INCLUDE_DIRS CUDA_CUDART_LIBRARY) 
CMake Warning at cmake/public/cuda.cmake:31 (message):
  PyTorch: CUDA cannot be found. ...

      This happened because I didn’t include Jetson SDK Components when using the SDKManager to flash my Orin. I resolved it using sudo apt install nvidia-jetpack on the Orin

    2. -- Could NOT find CUSPARSELT (missing: CUSPARSELT_LIBRARY_PATH CUSPARSELT_INCLUDE_PATH) 
CMake Warning at cmake/public/cuda.cmake:276 (message):
  Cannot find cuSPARSELt library.  Turning the option off

      I downloaded cusparselt 0.6.3 from here: using the Linux → aarch64-jetson → Ubuntu 22.04 → deb (local) steps. The install_cusparselt.sh script mentioned here didn’t seem to work.

    3. -- Could NOT find CUDSS (missing: CUDSS_LIBRARY_PATH CUDSS_INCLUDE_PATH) 
CMake Warning at cmake/public/cuda.cmake:242 (message):
  Cannot find CUDSS library.  Turning the option off

      I downloaded cudss 0.7.1 from cuDSS 0.7.1 Downloads | NVIDIA Developer using the Linux → aarch64-jetson → Ubuntu 22.04 → deb (local) steps.

  2. Once your wheel is built in dist, you can move it to your desired deployment environment and install it there.

  3. Test your installed wheel in Python:

    import torch, numpy
print("torch: ", torch.__version__)
print("numpy: ", numpy.__version__)
print("torch.version.cuda: ", torch.version.cuda)
print("cuda built: ", torch.backends.cuda.is_built())
print("cuda available: ", torch.cuda.is_available())
print("cuda device count: ", torch.cuda.device_count())