My cuda version is 8.0 and cudnn is 5.1.5
There is a how-to somewhere. I came across it and it looked like a lot of work. Another alternative is to configure bazel on your host to cross-compile for the Jetson.
My use case is 100% inference (no on-line training) so I have decided to learn how to optimize and apply my models using TensorRT instead. So my host machine will handle the training, and the Jetson will be specifically for running inference using the model. Maybe you can use this approach too.
First of all you may like update your OS via Jetpack 3.2 re-flash.
You will get cuda 9 then you may either use ready whl, or repeat installation steps from an instruction.