EGRAPHS Community

Egglog and Equality Saturation in a Production Tensor Compiler

2026-09-17

Joe Fioti and Austin Glover of Luminal will share how equality saturation drives their production tensor compiler — what’s working, and where they see opportunities for the tools and techniques to grow.

Abstract

Modern machine learning applications are defined via directed acyclic graphs of tensor operations. Initially these programs were largely interpreted. Now, in production, these graphs are statically analyzed and transformed into optimized implementations. The space of possible transformations is large, and the performance implications are not modular. Luminal uses an equality saturation approach to separate the definition of legal transformations from the search for fast implementations. Applying equality saturation to this domain has yielded some interesting questions and challenges that we will discuss. Overall, we believe that equality saturation is a great fit for this problem and are excited to continue developing our compiler and growing our company.

Bio

Joe Fioti is the co-founder and CEO of Luminal, where he leads the company’s work building a search-based compiler that automatically discovers high-performance implementations of AI workloads on GPUs and emerging accelerators.

Austin Glover is Luminal’s founding compiler engineer. He works on the core compiler, including e-graph search, hardware-aware optimization, and code generation.

Luminal recently raised its Series A round and is growing its compiler team. Please inquire if you find this work interesting.