Parallel posterior evaluation

I am using the CSIS class to learn amortized posterior hypotheses for a vision problem, and running inference on a new observation becomes too slow when num_inference_samples increases. Basically, running my generative model requires rendering a scene. There is any way of evaluating posterior hypotheses in parallel? Because the _traces() method that computes posterior hypotheses and the correspondent log_weight runs with for i in range(self.num_samples):. Thanks in advance!