There are a number of tricks for stabilizing the sampling - you can try an alternate representation with triangular matrices or possibly using an ordered transform on the variances. This notebook from the PyMC3 docs has some examples of how to do this, but you may need to tweak some of the code to make it run.
1 Like