Install Older Version of TensorRT for Jetpack

Hi, I have a model (TensorRT engine) that was made to run in Jetpack 4.4 (CUDA 10 and TRT 7), however upon upgrading to Jetpack 5, it doesn’t work, as Jetpack 5 uses CUDA 11 and TRT 7, so I’m trying to build a docker container that contains the appropriate versions. I was able to find the CUDA 10 container for jetson however the TRT container only has v8.

Where can I find the appropriate packages for TensorRT 7 for Jetpack?

Hi @davidru35bs, I’m not sure that downgrading CUDA like that across major JetPack versions would work or is supported - what you could do is try adding to your container the NVIDIA apt server from JetPack 4.4 (find this in your under your /etc/apt/sources.list.d on your JetPack 4.4 system). Then install CUDA Toolkit, cuDNN, and TensorRT into your container from apt (I would use l4t-base:r35.1.0 as your base image)

Again, I’m unsure if this would cleanly install/run or not, and what I’d actually recommend is to rebuild your TensorRT engine in JetPack 5. TensorRT engines aren’t portable across TensorRT versions and they are intended to be rebuilt for new versions of TensorRT that comes out. So this issue for you will only become more exacerbated as time goes on. Are you able to rebuild your model against TRT 8?

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