Chapter 10.
Frameworks

In the previous chapters, we examined the low-level building blocks of machine learning systems, from hardware accelerators and memory hierarchies to custom GPU kernels. To build and run models on top of these low-level components, developers use high-level frameworks that translate the mathematical operations of a model into efficient execution on hardware.

PyTorch is the dominant framework in research and production today, alongside major alternatives like TensorFlow, Keras, and JAX. In this chapter, we examine the role of a framework in the overall machine learning system stack and the core mechanisms that make it work.