@gaddamanil16 did you ever get this to work with non-toy datasets, with MCMC instead of ADVI?
I’m having a similar issue, even on very small datasets I can’t get MCMC to work. The reason I want to use MCMC is to find the ground truth of the posterior distribution, and use this to benchmark my own codes that employ variational inference - perhaps LDA is a case where MCMC is prohibitively computationally expensive.
I would recommend using Scikit learns LDA model rather than PYMC3 if you want to just the latent topics present in a collection of documents as it’s very fast and works well, the issue being that it won’t tell you much about the underlying Bayesian model.