AI Compiler - A system that generates optimized kernel code for AI model inference

AI compilers are commonly understood as systems that generate optimized kernel code for AI inference.

Most existing approaches frames this problem using traditional language compiler abstractions and optimization techniques.

In contrast, an alternative viewpoint is to treat AI models as algebraic computations systmes, where compilation focuses on the transformation and composition of matamtical operations rather than program semantics. 

 

Compiler-based optimization approaches applied to AI inference workloads frame AI models as programs and focus on decision-making problems such as scheduling, kernel selection, and pass ordering.
While effective in practice, this framing inherently limits optimization to execution strategies rather than mathematical transformation of operations.
As a result, fundamental issues of incompleteness are addressed empirically rather than explicitly, leaving the optimization problem statistically effective but theoretically unresolved.