PyTorch

PyTorch is an open-source machine learning framework developed by Meta's AI Research lab, distinguished from TensorFlow by its commitment to eager execution and dynamic computation graphs. Where TensorFlow originally required users to define a static graph before execution, PyTorch allows tensors to flow through operations imperatively — each line of Python executes immediately, gradients accumulate dynamically via autograd, and debugging requires nothing more than a standard debugger. This design choice made PyTorch the dominant framework for research prototyping, while TensorFlow retained advantages in production deployment.

The tension between PyTorch and TensorFlow is not merely technical preference. It reflects a philosophical disagreement about the nature of machine learning code: should it be a declarative specification optimized by a compiler (TensorFlow's graph model) or an imperative program that happens to be differentiable (PyTorch's dynamic model)? PyTorch 2.0's introduction of TorchScript and graph compilation via TorchDynamo represents a partial convergence — even the champions of dynamism recognize that performance requires the very abstractions they once rejected.

PyTorch's ascendancy in academic research created a pipeline effect: papers were written with PyTorch, reference implementations were released in PyTorch, and industrial labs adopted PyTorch to reproduce academic results. This is path dependence at the framework level — not necessarily the best technology winning, but the technology that captured the research community first.