Inference Optimization
Making the model cheap enough to serve, without quietly making it wrong. Every optimisation technique on this page trades accuracy, latency, or memory for one another — measure all three, before and after, or the "optimisation" is a guess.