I have written a description of how I think things should go forward in this open PR. This refactors lpn, lpmn, and clpmn to behave in a consistent way that supports the many different behaviours we want to support. I see this setting out the model we want to apply going forward.
So everyone knows, my choices are guided by my experience of, and the difficulty in, implementing these changes in a way that is elegant at the Python level, flexible and efficient at the C++ level, and supportable on CUDA.