Compiler 4.0 - Agentic ML Compilers for the Edge

Explore the evolution of compiler technology in this conference talk that traces the development from early assembly translators to the cutting-edge Compiler 4.0 paradigm. Learn how traditional optimizing compilers like GCC and LLVM automated code transformation for CPUs, and discover how ML compilers such as XLA, TVM, TorchInductor, and MLIR-based stacks revolutionized compilation by shifting focus from programs to computational graphs. Understand the limitations of current Compiler 3.0 systems when dealing with the growing diversity of model architectures and hardware targets, which creates complex combinatorial optimization problems that static heuristics cannot efficiently solve. Examine yasp.compile, an innovative agentic ML compiler that represents the Compiler 4.0 paradigm, designed to reason explicitly about hardware constraints and generate low-level implementations specifically tailored to individual model-hardware pairs. Discover how this approach combines hardware-aware graph optimization with learned code generation and cost modeling to reduce manual kernel engineering requirements and improve adaptability across heterogeneous edge accelerators, addressing the scalability challenges faced by current compilation systems.