Compiling TensorRT in ARM docker container

I’m attempting to construct a docker image to build our code using a suitable LT4 base image ( This docker container runs in AWS on a native ARM instance. I have installed nvidia-tensorrt-dev , nvidia-cudnn8-dev and nvidia-cuda-dev from the ubuntu repo provided in the image. My TRT application fails when trying to link against and, which do not exist. Where am I expected to find these files or is ARM native compiling not supported outside of Jetson devices?

Hi @tabbott, some drivers are mounted into the containers by the NVIDIA Container Runtime, including those:

$ cat /etc/nvidia-container-runtime/host-files-for-container.d/l4t.csv | grep libnvdla
lib, /usr/lib/aarch64-linux-gnu/tegra/
lib, /usr/lib/aarch64-linux-gnu/tegra/

Typically, one would set their default docker-runtime to nvidia (like here), so that CUDA/ect was runnable during docker build operations. However on AWS, you wouldn’t have that, and I wouldn’t know what happens if you try copying them in from a real Jetson.

Thanks @dusty_nv. Where would I be able to download the 2 missing files if I wanted to get them from a real Jetson? I don’t have access to any Jetson hardware.

@tabbott you might be able to extract them from the L4T Driver Package (Jetson Linux 35.4.1 | NVIDIA Developer) but I would highly recommend getting real Jetson hardware to test/build on sooner rather than later.

Thanks again. For completeness, I was able to find the files using the link you provided in the Driver Package (BSP). Inside of the archive there is a /Linux_for_Tegra/nv_tegra/nvidia_drivers.tbz2. In that the files are located within the directory /usr/lib/aarch64-linux-gnu/tegra/. The complete list of required additional libs are as follows:

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