TensorRT installation on Drive AGX

Hardware Platform: DRIVE AGX Xavier™ Developer Kit
Software Version: DRIVE Software 10.0
Host Machine Version: native Ubuntu 18.04
SDK Manager Version:

I have Drive Software 10.0 installed with the SDK Manager. I expect that TensorRT Version 5.1.4 should automatically be installed on the target as part of this package.
How do I verify that TensorRT has correctly been installed on the Drive AGX?
I can see libnvinfer.so.5.1.4 in /usr/local/driveworks/lib/.
However, i do not get any succesful response with:
dpkg -l | grep nvinfer

Is this normal or do I need further steps to install TensorRT?
When trying to import NvInfer.h, I receive an error that the file does not exist. Is this just a linking error or is the package not correctly installed?

Also, is the TensorRT Python API still not supported for the Drive AGX?

Hi @philipp.heidenreichdo0x6,

Please upload the file for our analysis. Thanks!


Dear @philipp.heidenreichdo0x6,
In DRIVE SW 10.0, we have removed all header filers and TensorRT samples on target to save filesystem space. The TensorRT libs are expected to present at driveworks lib location. So, you are expected to cross compile TensorRT sample on host and link the libs on target for execution.
TensorRT Python API is not supported on DRIVE AGX platform.

OK, we do not need the samples, so this is not the issue.
To clarify:
Is it expected that “dpkg -l | grep nvinfer” lists my TensorRT Version when installed correctly on the Nvidia Drive AGX? Or is it sufficient when libnvinfer.so.5.1.4 is inside the /usr/local/driveworks/lib/ directory?

Please provide full logs (all ~/.nvsdkm) for our investigation. Thanks!

.nvsdkm.zip (1.5 MB)

My main issue is that I can’t find the header files needed to compile a minimum working example of TensorRT (ideally with ONNX Parser). I do find the shared libraries of libnvinfer.so.5, libnvparsers.so and libnvonnxparser.so.0 inside /usr/local/driveworks/lib/ and inside /usr/local/driveworks/targets/aarch64_Linux/lib/. But I do not find any header files like NvInfer.h. Where should they be located?
When i try:
#include NvInfer.h
I get: “No such file or directory” even though I did:
target_link_libraries(my_project nvinfer nvparsers nvonnxparser)
in my CMakeLists.txt.

Our team is investigating your issue.

Before we have any update, please try to extract deb packages from ~/Downloads/nvidia/sdkm_downloads/drive-t186ref-linux- and install on your target machine.

BTW, NvInfer.h should be in libnvinfer-dev_5.1.4-1+cuda10.2_arm64.deb in the run file.

Are there any news regarding this issue?

We still did not get a minimum working example to run.

We haven’t got any progress due to the holiday season.

Have you manually installed the deb packages in ~/Downloads/nvidia/sdkm_downloads/drive-t186ref-linux- on you target system? All packages you need should be there.

Please ignore my previous suggestions about manually installing packages on the target system.

You should cross compile samples on host system. That’s why there is no those files installed on target. Thanks!