News

Meta Cedes Open Source PyTorch AI Framework to New Foundation

• By David Ramel
• 09/12/2022

Meta is ceding control of its wildly popular PyTorch machine learning framework to a new organization set up under the big open source umbrella of the Linux Foundation.

The new PyTorch Foundation will now control the project via a new governing body initially comprising representatives from AMD, Amazon Web Services (AWS), Google Cloud, Meta, Microsoft Azure and Nvidia.

The move was announced in a Sept. 12 Facebook post from Meta (and Facebook) co-founder Mark Zuckerberg, which in just about five hours has already amassed more than 20,000 reactions and some 4.6 thousand comments. Speaking to the aforementioned popularity, he noted that PyTorch has become one of the leading artificial intelligence (AI) platforms, boasting more than 150,000 projects on GitHub that were built on the framework.

Meta emphasized the move will democratize access to state-of-the-art AI tools, libraries and other components to accelerate progress in AI, as it plans to continue investing in PyTorch and using it as its primary framework for AI research and production applications at the company.

"PyTorch was built with an open source, community-first philosophy, and that won't change with the transition to the Foundation," the company said in a Sept. 12 announcement. "When researchers and developers open-source their code, others around the world can share their work, learn from each other's advances and then contribute back to the AI community.

"Going forward, the framework's contributors will benefit from the robust governance, diverse leadership and additional investments provided by the new PyTorch Foundation partners. The Foundation will strive to adhere to four principles: remaining open, maintaining neutral branding, staying fair and forging a strong technical identity. One of its main priorities will be to keep a clear separation between the business and technical governance of PyTorch."

Soumith Chintala, a PyTorch maintainer, noted that PyTorch celebrated its fifth-year anniversary in January, growing far beyond the team's initial vision.

"The business governance of PyTorch was fairly unstructured for quite some time since launch -- we operated like a scrappy startup," she said in a Sept. 12 post. "Team members at Meta spent the time and energy to structure this properly and organize PyTorch into an organizationally more healthy entity. Meta helped PyTorch with introducing many structures, such as Contributor License Agreements, Branding Guidelines, and Trademark registration. Keeping PyTorch's organizational health up to check is essential and beneficial for the community. The next stage of our organizational progress is to support the interests of multiple stakeholders, hence moving to a foundation is good. We chose the Linux Foundation as it has vast organization experience hosting large multi-stakeholder open-source projects with the right balance of organizational structure and finding specific solutions for these projects.

"Simultaneously, the technical governance of PyTorch has been a loosely structured community model of open-source development -- a set of people maintaining PyTorch by area with their responsibility often tied to their individual identity rather than their employment. While we kept a codified list at the PyTorch - Maintainers page, the technical governance was not formalized nor codified. As PyTorch scales as a community, the next step is to structure and codify. The PyTorch Technical Governance now supports a hierarchical maintainer structure and clear outlining of processes around day to day work and escalations. This doesn't change how we run things, but it does add discipline and openness that at our scale feels essential and timely."