I have a model that I need to run in real time scenarios using a Jetson. In my computer it runs very fast with JIT and now is time to move on a Jetson for production.
I wonder if running this model through using a Docker Image will be as fast as running it from source, installing all the libs from scratch.
I think the speed difference is small enough for the main question to be which gets you better support / a better workflow for your life-cycle, so likely you want to go with what NVidia recommends unless you have some other support arrangement.
I’m saying this as someone who always builds his own PyTorch from source and does offer support for bespoke installations on a commercial basis.