I doubt there is any algorithm or inference method can guarantee give you a global optimum except in simple low dimension problems.
What you can try as a close approximate maybe: sample from the posterior, compute the logp, and do a small grid search around the region of the maximum of the logp