Can I install TRT 8.5 on Jetpack4.6.3?

Robotics & Edge Computing Jetson Systems Jetson Nano
1

Be default, there is TRT8.2 on Jetpack4.6.3. But our model is not good with TRT8.2, our model is good with (JP5.1 and TRT8.5). Can I install TRT 8.5 on Jetpack4.6.3?

2

Hi,

This looks like a Jetson issue. Please refer to the below samples in case useful.

For any further assistance, we will move this post to to Jetson related forum.

Thanks!

4

@spolisetty
Please support me. It was 5 days from my first post.

I tried to use docker image deepstream-l4t:6.2-triton (TRT8.5.2, CUDA11.4) in (JP4.6.3 + DS6.0 + TRT8.2.1, CUDA 10.2). In this container I can not convert qat onnx to qat engine with this error

Cuda failure: initialization error
autorun.sh: line 6:    16 Aborted                 (core dumped)

although I have set --runtime nvidia when running container.

I can convert qat onnx to qat engine in docker container deepstream-l4t:6.0-triton (TRT8.2, CUDA 10.2) and outside of container (directly convert on host JP463, DS6.0, TRT8.2, CUDA10.2).

I want to add some infomation. My qat onnx is generated in Docker container in PC (TRT8.2). In PC, I can convert qat onnx to qat engine in docker container with TRT8.2 or TRT8.5.2.

5

Hi,

We are moving this post to the Jetson related forum to get better help.

Thank you.

8

@spolisetty
May be missmatch version of CUDA (11.4 in container l4t) and version of CUDA in Jetson host (10.2).

I tried to install CUDA 10.2 in container l4t (with TRT8.5.2) but, it is only for AMD64, not for ARM64.

A also tried install TRT8.5.2 on container nvcr.io/nvidia/l4t-cuda:10.2.460-runtime, but I could not install with the followning error

The following packages have unmet dependencies:
 tensorrt : Depends: libnvinfer8 (= 8.5.2-1+cuda11.8) but it is not going to be installed
            Depends: libnvinfer-plugin8 (= 8.5.2-1+cuda11.8) but it is not going to be installed
            Depends: libnvparsers8 (= 8.5.2-1+cuda11.8) but it is not going to be installed
            Depends: libnvonnxparsers8 (= 8.5.2-1+cuda11.8) but it is not going to be installed
            Depends: libnvinfer-bin (= 8.5.2-1+cuda11.8) but it is not going to be installed
            Depends: libnvinfer-dev (= 8.5.2-1+cuda11.8) but it is not going to be installed
            Depends: libnvinfer-plugin-dev (= 8.5.2-1+cuda11.8) but it is not going to be installed
            Depends: libnvparsers-dev (= 8.5.2-1+cuda11.8) but it is not going to be installed
            Depends: libnvonnxparsers-dev (= 8.5.2-1+cuda11.8) but it is not going to be installed
            Depends: libnvinfer-samples (= 8.5.2-1+cuda11.8) but it is not going to be installed

Download file is for TRT8.5.2+cuda11.8, there is not TRT8.5.2+cuda10.2. So I could not install it.

9

Hi,

Do you use Jetson Nano? If yes, the latest TensorRT version is 8.2.

There are some underlying dependencies between the GPU driver and the CUDA version.
So you can only use CUDA 10.2 on r32 (JetPack 4.x).

Is your model only work on TensorRT 8.5?
How do you know that? Is there another topic that tracks the model issue?

Thanks.

10

I use Jetson Xavier NX.

My Jetpack is 4.6.3 with CUDA 10.2 + TRT8.2.1. It means that we can not use Docker container deepstream-l4t:6.2-triton (with TRT8.5.2 and CUDA11.4) on JP4.6.3 with CUDA 10.2 + TRT8.2.1. Is that right?

My model can work with TRT8.5 if I install JP5.1 on Jetson. But I want to check my model with TRT8.5 on Jetson with Jetpack4.6.3.

11

Hi,

Sorry, you will need to use Deepstream 6.0.1 for JetPack 4.
Thanks.

12

@AastaLLL

I don’t need to Deepstream. Now I only want to test speed of engine model by using trtexec

13

Hi,

You can try nvcr.io/nvidia/l4t-tensorrt:r8.2.1-runtime:
NVIDIA L4T TensorRT | NVIDIA NGC

But on JetPack 4, the container mounts libraries from the Jetson native.
So you will need to install TensorRT first on the Nano.

Thanks.