One possible is that in v3.0, after the ADVI init the mass matrix is fixed and overestimated. So it just taking large step size and completely ignore the funnel. And on master currently we have adapt_diag, which sometimes (depending on the starting point) might adapt to a very narrow local geometry.