How to dynamically add a stop gradient in TensorFlow for truncated backprop through time?

Question

Suppose I have an RNN, and I would like to do truncated backpropagation through time. Typically you can do truncated backprop at every K interval, but instead I want to back