I don’t understand why you don’t want to sum the logps. In any PyMC model the logps of every variable (and regardless of whether they are scalar, vectors or ndimensional variables) are always summed (and then summed together) for the purpose of sampling. For instance, you can only take gradients wrt to scalar quantities, so if we didn’t sum, NUTS couldn’t be used.
The only cases where you need elementwise probabilities (in PyMC) are for model comparison. If you really need them, you should create a new distribution or use DensityDist. Deterministic is not what you need from your description.