Unable to use TensorRT 7.2.3 for Ubuntu 18.04 and CUDA 11.2

TensorRT download page includes TensorRT 7.2.3 for Ubuntu 18.04 and CUDA 11.1 & 11.2 TAR package.
This package cannot be used to compile TensorRT samples for CUDA 11.2 due to missing libraries (searches for CUDA 11.1).

Environment

TensorRT Version: 7.2.3.4
GPU Type: /
Nvidia Driver Version: 460.32.03
CUDA Version: 11.2
CUDNN Version: 8.1.1.33
Operating System + Version: Ubuntu 18.04
Python Version (if applicable): /
TensorFlow Version (if applicable): /
PyTorch Version (if applicable): /
Baremetal or Container (if container which image + tag): Baremetal

Steps To Reproduce

cd /my/custom/path/to/tensorrt/samples/
make

outputs:

…/…/lib/libnvinfer.so: undefined reference to nvrtcDestroyProgram@libnvrtc.so.11.1' ../../lib/libmyelin.so: undefined reference to nvrtcAddNameExpression@libnvrtc.so.11.1’
…/…/lib/libnvinfer.so: undefined reference to nvrtcCreateProgram@libnvrtc.so.11.1' ../../lib/libnvinfer.so: undefined reference to nvrtcCompileProgram@libnvrtc.so.11.1’
…/…/lib/libnvinfer.so: undefined reference to nvrtcGetPTX@libnvrtc.so.11.1' ../../lib/libnvinfer.so: undefined reference to nvrtcVersion@libnvrtc.so.11.1’
…/…/lib/libnvinfer.so: undefined reference to nvrtcGetProgramLog@libnvrtc.so.11.1' ../../lib/libmyelin.so: undefined reference to nvrtcGetLoweredName@libnvrtc.so.11.1’
…/…/lib/libnvinfer.so: undefined reference to nvrtcGetErrorString@libnvrtc.so.11.1' ../../lib/libnvinfer.so: undefined reference to nvrtcGetProgramLogSize@libnvrtc.so.11.1’
…/…/lib/libnvinfer.so: undefined reference to `nvrtcGetPTXSize@libnvrtc.so.11.1’

ldd /my/custom/path/to/tensorrt/targets/x86_64-linux-gnu/lib/libnvinfer.so
linux-vdso.so.1 (0x00007ffff7ffa000)
libcudnn.so.8 => /cm/shared/apps/cudnn8.1-cuda11.2/8.1.1.33/lib64/libcudnn.so.8 (0x00007fffd2147000)
libmyelin.so.1 => /cm/shared/apps/tensorrt-cuda11.2/7.2.3.4/lib/libmyelin.so.1 (0x00007fffd18c7000)
libnvrtc.so.11.1 => not found

Could you please provide some suggestions to workaround this issue?

Thanks a lot.

Kind regards,
GT

Hi @gianvito.taneburgo,
It seems to be some setup relate issue. Could you please try to run the same thing using TRT NGC container?
https://docs.nvidia.com/deeplearning/tensorrt/container-release-notes/running.html
NVIDIA NGC

Please let us know if you are getting the same issue even on NGC container.

Thanks

Hi SunilJB and thanks for your response.
As weird as it may sound, I am not ultimately interested in running TensorRT, but only in packaging it.
In other words, a working TensorRT NGC container wouldn’t solve the issue, because the container setup is different from the baremetal environment I am using/building. And I am not interested in extending the TensorRT container.

Hi @gianvito.taneburgo ,

We request you to please validate TensorRT installation steps you’ve followed. Please refer installation guide for correct steps
https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html
Also as mentioned above, please make sure you have required system dependencies.
https://docs.nvidia.com/deeplearning/tensorrt/support-matrix/index.html

Thank you.

Hi spolisetty and thanks for your response.

There is no TensorRT installation step to validate, since the problem I am raising happens while building TensorRT (with make) from the source code (i.e. before installation).

I am using CUDA 11.2 and cuDNN 8.1.1, that are supported both in the support matrix (Support Matrix :: NVIDIA Deep Learning TensorRT Documentation) and in the download page (leading to TensorRT-7.2.3.4.Ubuntu-18.04.x86_64-gnu.cuda-11.1.cudnn8.1.tar.gz).

It looks like TensorRT is simply looking for a CUDA 11.1 library, while CUDA 11.2 is installed…

Please, let me know how shall I proceed.
Thanks a lot.

Kind regards,
GT

Hi @gianvito.taneburgo,

For this build we need to install CUDA 11.1. In later versions backward/forward compatibility is improved.

Thank you.

Ok, thank you. So, I will just wait for an update to fix the issue.

Kind regards,
GT

1 Like