Currently hosted modules work best only on the first layer of model depth (i.e. host_max_depth=1). Stability breaks when the host tries to load an offloaded module group.
This feature would be built alongside a modular loading mechanism that can be shared between distributed models, validators, and workers.