When working with deep learning frameworks, you define models using high-level Python abstractions. However, to generate efficient machine code for a GPU or accelerator, the compiler requires a more structured and explicit view of the program. This internal format is known as the Intermediate Representation (IR). The IR serves as the common language between the framework frontend and the hardware backend, allowing the compiler to analyze the code structure without relying on the quirks of the source language.

In this chapter, we examine how machine learning models are translated into computation graphs. Unlike general-purpose IRs used in standard software compilers, ML compiler IRs are domain-specific. They treat tensor operations as first-class citizens. You will look at how a model is represented as a directed acyclic graph (DAG) where nodes represent operators, such as convolution or matrix multiplication, and edges represent the flow of data.

We will cover the following core concepts:

• Dataflow Graphs: How dependencies between operations dictate execution order and enable parallelism.
• Tensor Metadata: How the compiler manages shape information and data types (dtypes). For example, verifying that a matrix multiplication between a tensor of shape $(M, K)$ and $(K, N)$ produces an output of shape $(M, N)$.
• Shape Compilation: The distinction between static shapes, where dimensions are known at compile time, and dynamic shapes, where dimensions depend on runtime input.

Finally, we will practice inspecting the IR generated by tools like TVM or MLIR. Being able to read this representation is a necessary skill for debugging graph transformations and understanding why certain optimizations succeed or fail. By the end of this section, you will have a clear mental model of how the compiler views your neural network before it begins the optimization process.