Deep universal probabilistic programming with Python and PyTorch (github.com)

Pyro is a mature deep probabilistic programming library built on PyTorch that emphasizes universality, scalability, minimalism and flexibility. Designed originally at Uber AI and now maintained by community contributors (including a Broad Institute team) under the Linux Foundation, Pyro positions itself as a “universal” PPL capable of expressing any computable probability distribution while scaling to large datasets with minimal overhead versus hand‑written code. That combination makes it attractive for researchers and engineers who need expressive probabilistic models integrated into modern deep‑learning stacks. Technically, Pyro provides a small core of composable abstractions to express generative models and inference procedures, offering high‑level automatisms for common workflows and direct hooks for expert customization. It supports amortized inference patterns common in deep probabilistic models and leverages PyTorch tensors and autograd for scalable stochastic variational inference and MCMC. Installation is straightforward (pip install pyro-ppl or pip install git+https://github.com/pyro-ppl/pyro.git for the latest) and examples/tutorials require optional extras. If you use Pyro in published work, cite the JMLR paper: “Pyro: Deep Universal Probabilistic Programming” (Bingham et al., 2019).