Any insight into why this occurs during adaptation? Is it an issue of parameter initialization, either for the model or for NUTS?
During the adaptation phase, NUTS tries step sizes based on an exponential schedule. When sites are constrained and we use a large step size, it is easy for the integrator to generate samples which result in a NaN when transformed back to the constrained support of the original sample site. This isn’t a problem as it will try out a smaller value for the step size when this happens, though I agree that the warning can be confusing. Thanks for reporting this - I will look into silencing these warnings for the adaptation period.