PyTorch Foundation
The Linux Foundation’s home for PyTorch and, since May 2025, five more AI projects: vLLM, DeepSpeed, Ray, Helion and Safetensors.
What it is
The PyTorch Foundation is part of the non-profit Linux Foundation. Meta set it up on September 12, 2022 to take over PyTorch, the deep-learning framework Facebook had released in 2016, with AMD, Amazon Web Services, Google Cloud, Meta, Microsoft Azure and NVIDIA on its first governing board. On May 7, 2025 it became an umbrella foundation and accepted two more projects, vLLM and DeepSpeed; Ray followed in October 2025, and Helion and Safetensors in April 2026. In October 2026 it hosted those six projects, and its PyTorch Conference in San Jose drew more than 3,500 people.
For someone choosing a technology, the badge says who holds the project. A hosted project has won a two-thirds vote of the Technical Advisory Council (TAC), and the Linux Foundation’s lawyers have taken over its assets, so the company that wrote it no longer controls it alone. Applicants must show production use, a strong and diverse contributor base, compatibility with recent PyTorch releases and the OpenSSF Best Practices silver badge, and each hosted project has a representative on the TAC.
It is not a maturity ladder. Unlike the CNCF, the foundation has no sandbox or graduated stages: the TAC files each hosted project as Platform, if it spans several stages of the AI lifecycle such as training and inference, or Vertical, if it serves one field such as protein folding, and reviews every project once a year, with retirement the only other status. The much larger PyTorch Ecosystem, more than 100 projects in February 2026 including SGLang, verl and Unsloth, is a curated list of independently governed projects, not part of the foundation.
Its projects in this catalog
RayOpen sourceA Python framework that spreads functions (tasks) and stateful objects (actors) across a cluster, with libraries on top for data processing, training, tuning and model serving.
Ray ServeOpen sourcePart of RayRay’s library for online inference: Python classes become deployments whose replicas run as Ray actors across a cluster, with request batching, model composition, autoscaling to zero and an OpenAI-compatible LLM server built on vLLM.In depth
The six hosted projects#
As listed on pytorch.org in October 2026. ExecuTorch, Meta’s runtime for phones and microcontrollers, became part of PyTorch core in April 2026 rather than a separate hosted project.
| Project | Joined | Contributed by | What it is |
|---|---|---|---|
| PyTorch | September 12, 2022, the founding project | Meta | The deep-learning framework; its CI costs more than $1.5 million a month, paid mostly by Meta and AWS |
| vLLM | May 7, 2025 | Started at UC Berkeley | An inference and serving engine for large language models, built around PagedAttention |
| DeepSpeed | May 7, 2025 | Microsoft | A distributed training library: ZeRO, 3D parallelism and inference acceleration |
| Ray | October 22, 2025 | Anyscale; started at UC Berkeley | Distributed compute for data processing, training and inference; more than 237 million downloads when it joined |
| Helion | April 7, 2026 | Meta | A Python-embedded language for machine-learning kernels that compiles to Triton and TileIR and autotunes them |
| Safetensors | April 8, 2026 | Hugging Face | A file format for model weights that can’t run arbitrary code when it is loaded |
Ecosystem first, then hosting#
A project reaches the foundation in two steps. It first applies to the PyTorch Ecosystem, run by the TAC’s Ecosystem Working Group; ecosystem projects keep their own governance and appear on the PyTorch landscape, and 14 joined in 2025. Only an approved ecosystem project can then apply to be foundation-hosted, through an issue in the pytorch-fdn/foundation-hosted repository, with a voting TAC member as its sponsor.
The TAC reviews the application at one of its monthly meetings and needs a two-thirds majority to accept it. A declined project stays in the ecosystem and may reapply after three months if it has made tangible progress. After a yes, the Linux Foundation’s legal team completes the contribution and the transfer of the project’s assets, and foundation staff set up its branding, website, marketing and CI. The TAC reviews every hosted project each year and can retire one that has gone inactive or stopped meeting the requirements.
Who decides#
Business and technical decisions are split. The Governing Board runs the foundation under its charter: each top-tier member appoints a representative, the other paying members elect up to three between them, and the TAC has a seat. The foundation updated its membership tiers by February 2026, when it had 33 member organisations, and nine more joined at the October 2026 conference, with Nebius, Nscale and Shopify as Platinum members.
The TAC is the technical body. In October 2026 its 21 members included a representative of each hosted project, among them Meta for PyTorch, Inferact for vLLM, Snowflake for DeepSpeed, Anyscale for Ray and Hugging Face for Safetensors, alongside engineers from members such as AMD, Arm, Google, Huawei, Intel, Microsoft and NVIDIA; Thomas Viehmann of Lightning AI chairs it. It runs five working groups: CI infrastructure, multi-cloud, ecosystem, accelerators and security. Each project keeps its own maintainers and technical governance; PyTorch’s is written down by its maintainers.
Paid staff run the rest. In February 2026 Matt White, executive director for nearly two years, became the foundation’s CTO, and Mark Collier, a co-founder of OpenStack, took over as executive director, in charge of finances, events, staff and membership.
Where it came from2016–2026
Facebook released PyTorch in 2016, and by 2022 it had more than 2,400 contributors. On September 12, 2022 Meta moved it to a new foundation at the Linux Foundation, with the companies whose chips and clouds it ran on, AMD, AWS, Google Cloud, Microsoft Azure and NVIDIA, on the board beside Meta. For two and a half years the foundation hosted only PyTorch, while a list of ecosystem projects built on it grew to about 120.
On May 7, 2025, at PyTorch Day France in Paris, it announced that it would become an umbrella foundation, with vLLM and DeepSpeed as its first new hosted projects; it then had more than 30 member companies. Ray joined at the PyTorch Conference in San Francisco on October 22, 2025, putting model development, inference and distributed execution under one foundation. Sixteen companies joined as members that year.
In February 2026 Mark Collier replaced Matt White as executive director. At the first PyTorch Conference Europe, in Paris on April 7 and 8, 2026, Meta contributed Helion and Hugging Face contributed Safetensors, and ExecuTorch became part of PyTorch core. In October 2026 Nscale, which plans to buy Ray’s developer Anyscale, joined as a Platinum member.
Why it’s built this way
- A Linux Foundation project, not a Meta one
- PyTorch ran on chips and clouds sold by Meta’s partners and competitors, so its first board seated AMD, AWS, Google Cloud, Microsoft Azure and NVIDIA beside Meta, with the Linux Foundation holding the project.
- An umbrella for the AI stack
- Projects such as vLLM and DeepSpeed were already built on PyTorch, and the foundation wanted a vendor-neutral home for projects across the AI lifecycle, from training and inference to agents, not only for the framework.
- Ecosystem projects and hosted projects
- Ecosystem projects keep their own governance and only get listed; hosted projects adopt the foundation’s governance and get full operational support. Requiring ecosystem status first means the TAC already knows a project before it takes on its assets.
- Categories by scope, not by maturity
- The umbrella was set up for Platform projects that span several stages of the AI lifecycle and Vertical projects for fields such as bioinformatics, geospatial intelligence and protein folding, so the TAC classifies projects by what they cover rather than ranking them.
Timeline
- 2016
Facebook releases PyTorch
- 2022
Meta moves PyTorch to the new PyTorch Foundation on September 12
- 2025
Becomes an umbrella foundation on May 7 with vLLM and DeepSpeed; Ray joins on October 22
- 2026
Mark Collier becomes executive director in February; Helion and Safetensors join in April
- Linux Foundation: Meta transitions PyTorch to the Linux Foundation (September 12, 2022)
- PyTorch Foundation: Expands to umbrella foundation, welcomes vLLM and DeepSpeed (May 7, 2025)
- Linux Foundation: PyTorch Foundation welcomes Ray (October 22, 2025)
- PyTorch Foundation in 2025: A year in review (January 2026)
- PyTorch Foundation: The next chapter, together (leadership transition, February 2026)
- PyTorch Conference Europe 2026 recap (Helion, Safetensors, ExecuTorch; April 2026)
- PyTorch Conference North America 2026 recap (new members, Nscale and Anyscale; October 2026)
Sources · facts checked 2026-10-28
- PyTorch Foundation (hosted projects, governance)
- Linux Foundation: Meta transitions PyTorch to the Linux Foundation (founding board, contributors; September 12, 2022)
- PyTorch Foundation: Expands to umbrella foundation, welcomes vLLM and DeepSpeed (ecosystem vs hosted, Platform and Vertical; May 7, 2025)
- Linux Foundation: PyTorch Foundation welcomes Ray (237 million downloads; October 22, 2025)
- PyTorch Foundation welcomes Helion as a foundation-hosted project (ExecuTorch joins PyTorch core; April 7, 2026)
- Linux Foundation: PyTorch Foundation announces Safetensors as newest contributed project (April 8, 2026)
- pytorch-fdn/foundation-hosted: How to apply (TAC sponsor, two-thirds vote, reapplication)
- pytorch-fdn/foundation-hosted: Foundation-hosted project intake process (eligibility, onboarding, retirement)
- pytorch-fdn/tac: PyTorch Technical Advisory Council (monthly meetings, working groups)
- PyTorch Foundation: Technical Advisory Council (members and project representatives)
- PyTorch Foundation charter (membership classes, Governing Board, TAC)
- PyTorch Foundation: The next chapter, together (33 members, more than 100 ecosystem projects, working groups, leadership; February 2026)
- PyTorch Foundation in 2025: A year in review (ecosystem projects, new members, CI costs)
- PyTorch Conference Europe 2026 recap (April 7–8, 2026)
- PyTorch Conference North America 2026 recap (3,500+ attendees, nine new members; October 2026)