Build a modular drop-in replacement for standard MLP blocks that uses tree-structured routing. This provides a path for developers to reduce inference compute costs while maintaining performance without a dedicated router network.