Tuesday, September 27

Meta Splits From PyTorch Foundation to Make AI Framework Vendor Neutral

The PyTorch logo.
Enlarge / The PyTorch logo on a fancy gradient.

PyTorch Foundation

Meta AI today announced that governance of PyTorch, a popular open-source deep learning framework, has been transferred to an independent organization called the PyTorch Foundation. It will operate as part of the nonprofit Linux Foundation, and its board includes representatives from Nvidia, Meta, Google, Microsoft, Amazon, and AMD.

The move comes as adoption of the PyTorch framework has become widely used in the deep learning industry to power many natural language and computer vision projects, including Tesla Autopilot. Meta cites over 2,400 contributors and 150,000 projects built on the framework. Running PyTorch in its own foundation avoids potential conflicts of interest that could arise if PyTorch were only controlled by Meta, which created the framework.

“The PyTorch Foundation will strive to adhere to four principles,” Meta wrote in a blog post announcing the news. “Stay open, maintain a neutral brand image, stay fair, and forge a strong technical identity. One of the foundation’s top priorities is to maintain a clear separation between the business and technical governance of PyTorch.”

The PyTorch project started in 2016 as an implementation of the Torch library in Python, and it gained fame for tensor computation and its tape-based autograd, which records operations like a tape recorder and then plays them back to the back to calculate the gradients.

If that sounds complicated, that’s because it is, but PyTorch makes these functions easier by putting them together in a building block of code that can be used in more complex machine learning projects without needing each new project reinvents the fundamentals every time. Layers and layers of libraries and frameworks make up most of today’s modern software applications, and this is especially true in the field of machine learning.

According to Meta, the transition to the PyTorch Foundation will not affect any existing PyTorch code, and Meta remains committed to investing in the now independent framework in the future.