I think this is a bug with pm.sample_posterior_predictive. Would you mind filing an issue in the PyMC github repo? In the short term, you can subsample trace, and then pass your smaller trace to pm.sample_posterior_predictive, and leave out both the samples=10, keep_size=False arguements.
Before it errors, the reason it’s slow is that you have 1500 points in X_new. GP predictions also scale poorly with data set size. Reducing this will give a big speed up.
EDIT: See @OriolAbril answer below, that’s the correct one.