An intermediate representation that breaks down ML operations into minimal, recursively-defined building blocks for performance modeling.