From PyTorch to Hardware: An Introduction to LLM Compilers

When we write a model in PyTorch, how does a high-level tensor operation eventually become something that runs efficiently on a GPU, TPU, or NPU?

In this talk, we will follow the journey from PyTorch to hardware and explore the role that ML/LLM compilers play in between. Starting from computation graphs and tensor operations, we will look at how a compiler transforms high-level model computations into efficient programs for increasingly specialized hardware.

We will compare this process with traditional programming-language compilers, examine what makes compiling machine-learning workloads different, and use a small interactive example to see how graph transformations and optimizations affect execution.

Finally, we will discuss why ML compilers have become increasingly important for LLMs, where performance depends not only on computation, but also on memory movement, kernel execution, data layout, and the characteristics of the target hardware.

The goal of this talk is not to teach model training or mathematical optimization, but to provide software engineers with a practical mental model of what happens between writing a model and actually running it efficiently on hardware.

Slides: https://speakerdeck.com/tommyhonym/pytorch-ai-compiler