Hamiltonian Monte Carlo is not guaranteed to always work, which is why the implementation of dynamic Hamiltonian Monte Carlo in Stan is heavily instrumented to detect problems. For more information see https://arxiv.org/abs/1701.02434.
Presumably your implementation does not (at least you have not informed us of any diagnostics), in which case it’s impossible for us to tell what could be going on. But in all likelihood your model is pathological (it might have a reasonable mode, but the shape around that mode is not well-behaved) and even our best implementation of Hamiltonian Monte Carlo can’t fit it. Indeed as @bgoodri notes the kind of model that you are trying to fit is notoriously nasty.