Standard execution modes in machine learning frameworks, like eager execution or Ahead-of-Time (AOT) compilation, involve specific trade-offs between flexibility and optimization potential. Just-In-Time (JIT) compilation operates differently, deferring the final code generation until runtime. This strategy enables optimizations informed by dynamic runtime context, such as observed tensor shapes (e.g., $B \times C \times H \times W$) or values.

This chapter covers JIT compilation methods relevant to ML workloads. We will analyze graph acquisition techniques like tracing and scripting, examine requirements for JIT-specific intermediate representations, and study runtime specialization and adaptive compilation strategies. The architecture and function of significant JIT systems, including TensorFlow XLA and PyTorch JIT (TorchScript), serve as practical examples.