Disclaimer: I am not an expert here, but my two cents. Why is 30 s disappointing? Are you trying to scale to larger dataset? I read in the pymc docs that often pymc is equal to or slower than other packages bayesian packages when the data is small, as there is a lot of overhead compiling. But as you fit larger data the compiling overhead becomes a smaller percentage of compute time and thus pymc becomes faster than some of the other bayesian software packages. So if you need to scale the model it might not perform as badly as you think.
1 Like