Unable to generate TRT Engine

[MrOCW]

Attempting to convert models like yolov5s and centerpoint to an engine on a Jetson Xavier AGX (JP 5.0.2) results in the following error:

[tensorrt_yolo_node-1] 1: [executionResources.cpp::setTacticSources::152] Error Code 1: Cudnn (Could not initialize cudnn, please check cudnn installation.)
[tensorrt_yolo_node-1] 2: [builder.cpp::buildSerializedNetwork::609] Error Code 2: Internal Error (Assertion enginePtr != nullptr failed. )
[tensorrt_yolo_node-1] Fail to create serialized network

Running dpkg --list | grep libcudnn:

hi  libcudnn8                                                   8.2.4.15-1+cuda11.4                  arm64        cuDNN runtime libraries
hi  libcudnn8-dev                                               8.2.4.15-1+cuda11.4                  arm64        cuDNN development libraries and headers

Running dpkg --list | grep libnvinfer

hi  libnvinfer-dev                                              8.2.4-1+cuda11.4                     arm64        TensorRT development libraries and headers
hi  libnvinfer-plugin-dev                                       8.2.4-1+cuda11.4                     arm64        TensorRT plugin libraries
hi  libnvinfer-plugin8                                          8.2.4-1+cuda11.4                     arm64        TensorRT plugin libraries
hi  libnvinfer8                                                 8.2.4-1+cuda11.4                     arm64        TensorRT runtime libraries

[maxime-clem]

Can you share the output of nvidia-smi and nvcc -v ?

[MrOCW]

Here is the output of ls -l /usr/lib/aarch64-linux-gnu/ symlink.txt

I run the container with
docker run -it --name="autoware" --net=host --runtime nvidia --privileged --env="NVIDIA_DRIVER_CAPABILITIES=all"
--env="DISPLAY"
--env="QT_X11_NO_MITSHM=1"
--volume="/tmp/.X11-unix:/tmp/.X11-unix:rw"
--volume=some other volumes autoware

[manato]

Thank you for sharing. Hmm..., I could not find obvious problems from the information you shared.
Does your image based on one of here ? This repo's images make it easy to work with Jetson and docker, so you might want to use them. You can identify the L4T version of your JP by running cat /etc/nv_tegra_release , and output will be like the following:

# R34 (release), REVISION: 1.1, GCID: 30414990, BOARD: t186ref, EABI: aarch64, DATE: Tue May 17 04:20:55 UTC 2022

this suggests the L4T version is 34.1.1

[MrOCW]

Yes I am using one of the images there

[MrOCW]

I have used the default trt & cuDNN versions since the documentation was oudated & it works now. Thanks

[liuypcoding]

@MrOCW I have used the default trt & cuDNN versions since the documentation was oudated & it works now. Thanks

which trt & cuDNN versions are using, Thanks.

[maxime-clem]

The issue was caused by incorrect version of cudnn and tensorrt.
This was caused by outdated instructions which have now been updated: #2885