Best way to install LibTorch?

[SkyDroneM]

Hi,

What is the best/easiest way to install LibTorch so that I can use it in C++ on a Jetson AGX Xavier? I am currently using the CPU version from pytorch website so torch::cuda::is_available() returns false at the moment. (CUDA 10.2 is installed if I check in the terminal). The CUDA version on pytorch.org is unfortunately not built for aarch64 so I need to build it on my device I suppose? Or do you provide a pre-built aarch64 version of libtorch somewhere that I can just drag and drop onto my jetson?

Thankful for any help I can get.

[AastaLLL]

Hi,

We do have the prebuilt PyTorch package for the Jetson user.
You can find the package here or a container here.

Thanks.

[SkyDroneM]

Hi thanks for the response,

Unfortunately I run into this error after following the installation instructions and verify using python:

python3 -c 'import torch'
...
OSError: libmpi_cxx.so.20: cannot open shared object file: No such file or directory

sudo apt-get install libopenblas-base libopenmpi-dev libomp-dev
...
libomp-dev is already the newest version (1:10.0-50~exp1).
libopenmpi-dev is already the newest version (4.0.3-0ubuntu1).
libopenblas-base is already the newest version (0.3.8+ds-1ubuntu0.20.04.1).

EDIT:

Also tried to run the docker image (Using jetpack 4.6.1) and no luck:

sudo docker run -it --rm --runtime nvidia --network host nvcr.io/nvidia/l4t-pytorch:r32.6.1-pth1.9-py3
...
docker: Error response from daemon: failed to create shim: OCI runtime create failed: container_linux.go:380: starting container process caused: error adding seccomp filter rule for syscall clone3: permission denied: unknown.

[dusty_nv]

@SkyDroneM

docker: Error response from daemon: failed to create shim: OCI runtime create failed: container_linux.go:380: starting container process caused: error adding seccomp filter rule for syscall clone3: permission denied: unknown.

Can you try following these steps to restore your docker install? Docker fails to create container after upgrading docker on Jetpack 4.9 · Issue #108 · dusty-nv/jetson-containers · GitHub

@SkyDroneM

libomp-dev is already the newest version (1:10.0-50~exp1).
libopenmpi-dev is already the newest version (4.0.3-0ubuntu1).
libopenblas-base is already the newest version (0.3.8+ds-1ubuntu0.20.04.1).

Is this on JetPack 5.0 DP? Because those package versions are for focal (20.04) not bionic (18.04)

On JetPack 5.0, I have the same versions installed and don’t get the MPI library error. You might want to try installing my PyTorch 1.11 wheel instead and see if that helps.