Infer_discrete using trained model

Hi I am a beginner and am trying to draw posterior samples of discrete latent variables. My question is how to get a “trained_model”, after running mcmc, to use in infer_discrete(trained_model,…)() so that posterior distribution for the continuous latent variables will be used. I found an example for the svi case using poutine.replay, but not for the mcmc case. appreciate it!