Use tfd.Independent for multivariate response distributions

[dirmeier]

... i think we can use that instead of looping over q univariate distributions for multivariate y.