Meta PyTorch

PyTorch is an open-source deep learning framework for researchers and developers building and deploying AI models. Its Python-native design and dynamic computation graphs support experimentation across language, vision, reinforcement learning and generative AI. Developers can train models across CPUs, GPUs and other accelerators, including distributed systems; ExecuTorch extends the ecosystem to on-device model execution.

Development began in 2016 among Torch contributors and Facebook’s AI research team, with Soumith Chintala, Adam Paszke and Sam Gross among its central creators. PyTorch combines imperative Python programming with accelerated execution, allowing users to inspect and debug models as ordinary programs. Its engineering contributions include parameter registration through attribute assignment and variable versioning to detect correctness problems with in-place operations during automatic differentiation. Following Meta’s 2022 transition announcement, stewardship moved to the independent PyTorch Foundation under the Linux Foundation, with Meta continuing to contribute and use the framework. Mark Collier became executive director in 2026, while Matt White became CTO.

The foundation expanded into a home for multiple projects in 2025. It reported 33 member organizations and more than 100 ecosystem projects in February 2026, reflecting organizational participation and software breadth. By July 2026, it hosted six projects: PyTorch, vLLM, DeepSpeed, Ray, Helion and Safetensors. Member commitments support shared infrastructure and community programs, while participating organizations also contribute engineering effort, hardware and operational expertise.