PyTorch

PyTorch anchors much of modern deep learning in both research and production. Originally developed at Meta AI and now governed by the PyTorch Foundation, it provides GPU-accelerated tensor computation similar to NumPy and deep neural networks built on a tape-based autograd system, so models execute eagerly and can change their behavior at runtime. Core components include torch for tensors, torch.autograd for reverse-mode automatic differentiation, torch.nn for network modules, and torch.jit for TorchScript compilation. The framework is written in Python over a C++ core and runs on CPU as well as NVIDIA CUDA, AMD ROCm, and Intel GPUs, drawing on cuDNN, NCCL, and MKL for performance. Its Python-first philosophy keeps abstractions minimal, makes debugging work with standard tools, and allows extensions in Python or C++. A large ecosystem of libraries and pretrained models has grown around it, and it serves as the default framework for many researchers, data scientists, and engineers shipping neural networks. Open source under a BSD-style license.