I am surprised that discrete switch point works well here, still, I suggest you to use a continuous step function instead of switch, more details see: https://stackoverflow.com/questions/49144144/convert-numpy-function-to-theano/49152694#49152694
Back to your original question, in pm.sample each sampler has some target that it would try to reach during tuning. For example, in Metropolis, it adjusts the scale of the transition kernel so that the acceptance probability is around 50%
In NUTS, tuning is much more complex as it optimizes the step size for leapfrog integrator, the mass matrix of the Hamiltonian transition kernel. The best reference for this is the Stan User Manuel http://mc-stan.org/users/documentation/ chapter 34.2, although we dont use the excat tuning algorithm as Stan.