HMC fails unexpectedly if data not scaled properly

Hi,
I have noticed that the HMC engine fails sometimes when data is not properly scaled. I was wondering if you could provide more clarity on this behaviour with a Hamiltonian MC theory interpretation. Since I can't reproduce this issue in BSTS, could this be related to a specific implementation in Tensorflow probability? In this tutorial: https://www.tensorflow.org/probability/examples/STS_approximate_inference_for_models_with_non_Gaussian_observations, it is mentioned indeed that scaling helps speed up the inference process, but in my case, no samples are accepted at all, as if the chain is not converging.
A corollary question: are there any useful tools in Tensorflow probability that help diagnose the HMC convergence, comparable to what we have in Stan for instance?
Thank you for your help.

Start with the linked TensorFlow Probability STS approximate-inference tutorial and the HMC engine behavior described in the report. No repository file or test is named; done would require a reproducible unscaled-data failure, an explanation of the rejection behavior, and identification of available convergence-diagnostic tools.