Compiling machine learning models efficiently demands representations distinct from those found in general-purpose compilers. While standard Intermediate Representations (IRs) serve well for typical software, they often struggle to represent and optimize the high-level operations and graph structures inherent in ML workloads. This gap necessitates specialized, multi-level IRs designed specifically for the abstractions and transformations common in machine learning.

This chapter focuses on these advanced IRs. We will analyze why conventional compiler IRs are often insufficient for expressing and optimizing ML computations effectively. You will study the design principles underpinning multi-level representations, which allow optimizations at various abstraction levels, from entire graphs down to low-level instructions. We will take a detailed look at MLIR (Multi-Level Intermediate Representation) as a prominent example, examining its core structure, including dialects and operations. You'll learn how MLIR represents high-level ML framework graphs, the process of progressively lowering these representations towards hardware targets, and how its extensible nature allows for custom dialects supporting new hardware or domains. This chapter provides the foundation for understanding how modern ML compilers manage complexity and enable powerful optimizations through sophisticated intermediate representations.