BoTorch is a PyTorch-based library for Bayesian Optimization, a technique for sequentially optimizing expensive-to-evaluate functions. It provides a modular and extensible interface for composing Bayesian optimization primitives including probabilistic models, acquisition functions, and optimizers. The library leverages PyTorch’s auto-differentiation and GPU acceleration, and supports Monte Carlo-based acquisition functions via the reparameterization trick for flexible model implementations.