Install and use more than one TensorRT versions in the same x64 host

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Hello,
I have a jetson TX2 and a Xavier boards.

I need to work with TensorRT 4.x for the TX2 and with TensorRT 5.x for the Xavier.

I developed my CNN with TF and till now i used the TF to TRT conversion tools locally install in my x64 Linux host which were part of the TensorRT 4.x package.

Now i need to install the 5.x package for the Xavier.

Is it possible to install these two TensorRT versions without performing the upgrade from 4.x version to 5.x version?

If not, is there any other way to have two TensorRT versions and each time to select the required one?

Maybe a TensorRT Docker? How?
Or maybe can i install the 5.x package with the conversion tools on the Xavier itself without and usage of the x64 host?

Regard,

2

On the x86 host, you can use TensorRT NVIDIA GPU Cloud (NGC) Container. start a 4.x container, then a separate 5.x container, both can point to the same volume (same source, model).

Another method is to use the .tar package instead of the .deb package. With the .tar package, you can you just change the environment LD_LIBRARY_PATH to the vesion you want.

3

Is the .tar option possible if I want to use an edge device’s TRT version (e.g. Jetson TX-2 or Xavier)? so I’ll compile a model for an edge device whilst on the host?

Thanks