In general, there is no good way to inference unknown number of mixture component in pymc3 (I would say even in general, as the problem is really difficult). My suggestion for you is to estimate a truncated mixture model, where you always have more mixture component than you need. And if the true latent component distributions contains only 2 components, the rest of the component would just end up with very small weight.
You can have a look at https://docs.pymc.io/notebooks/dp_mix.html