Did you really mean all those normal(1, 1) priors rather than normal(0, 1)?
It would be more straightforward to a 14-element vector mu and set its elements rather than define 14 independent variables (m1 to m14) and then stack them.
You set init_log_vars to 0, then call exp() on it and multiply. This is just a big no-op that yields sample = zs[i] + init_means.
In my experience, nan values on model evaluation usually means arguments that are out of support.