PyTorch, the popular open source machine learning framework, has announced new developments at the PyTorch Conference. These include the release of PyTorch 2.1 and the introduction of PyTorch Edge and ExecuTorch, which enable AI inference on mobile and edge devices.
ExecuTorch, developed by Meta Platforms (formerly Facebook), is already being used to power Ray-Ban smart glasses and the Quest 3 VR headset. The technology allows for on-device AI inference and is now open source as part of the PyTorch project.
During the conference, Ibrahim Haddad, executive director of the PyTorch Foundation, highlighted the progress made by the organization over the past year. He emphasized the benefits of hosting projects in a neutral open home, such as the Linux Foundation, to facilitate growth.
PyTorch 2.1 expands the capabilities for inference, which historically hasn’t been widely used compared to training AI models. IBM has contributed to PyTorch 2.1 to improve inference for server deployments. The update also includes performance enhancements and support for automatic dynamic shapes and accelerated numerical calculations.
In a keynote session at the conference, Mergen Nachin from Meta explained the importance of ExecuTorch. It provides an end-to-end solution for deploying AI on mobile and edge devices. ExecuTorch converts PyTorch models into optimized native programs, allowing them to run on a wide range of devices. It offers portability and consistency across different targets, improving developer productivity.
ExecuTorch has been successfully used in Meta’s Ray-Ban smart glasses and is now open source to encourage collaboration and address fragmentation in deploying AI models on edge devices.
By introducing these advancements, PyTorch aims to enable a new era of on-device AI inference capabilities. This will empower organizations to take full advantage of AI technology in various industries and applications.
VentureBeat supports technical decision-makers by providing knowledge about transformative enterprise technology. For more information, subscribe to VentureBeat’s Briefings.
