Safetensors joins PyTorch Foundation for community-driven AI safety

At a glance:

• Safetensors, a secure format for sharing ML models, joins the PyTorch Foundation
• The move ensures vendor-neutral governance and community-driven development
• Upcoming enhancements include device-aware loading, parallel APIs, and quantization support

Safetensors' journey to the PyTorch Foundation

Safetensors, a Hugging Face project, was created to address the need for a secure way to store and share machine learning model weights without the risk of executing arbitrary code. The format, consisting of a JSON header and raw tensor data, has been widely adopted by the ML community and has become the preferred method for sharing models on the Hugging Face Hub and other platforms.

Why the PyTorch Foundation?

The decision to join the PyTorch Foundation, hosted by the Linux Foundation, ensures that Safetensors has a vendor-neutral home and belongs to the community. This move allows for more companies and contributors to participate in the project's governance, reflecting the diverse needs of the community building on top of it. The core maintainers from Hugging Face, Luc and Daniel, will continue to lead the project, but the trademark, repository, and governance will now sit with the Linux Foundation.

Impact on users and contributors

For most users, the transition to the PyTorch Foundation will have no immediate impact. The format, APIs, and Hub integration will remain unchanged, ensuring a seamless experience. For contributors, the path to becoming a maintainer is now formally documented and open to anyone in the community. The project's governance is outlined in the GOVERNANCE.md and MAINTAINERS.md files in the repository, providing a stable, long-term foundation for organizations building on top of Safetensors.

Future developments

Safetensors is working with the PyTorch team to integrate the format as a serialization system for PyTorch models. The project's roadmap includes device-aware loading and saving, first-class APIs for Tensor Parallel and Pipeline Parallel loading, and support for various quantization formats such as FP8, GPTQ, AWQ, and sub-byte integer types. Being part of the PyTorch Foundation allows Safetensors to collaborate with other hosted projects in solving these ecosystem-wide challenges.

Getting involved

Safetensors is open source and welcomes contributions at every level, from bug reports and documentation to new features and governance participation. Developers, researchers, and organizations that build on Safetensors are encouraged to get involved in shaping its direction by opening issues, starting discussions, or reaching out to the maintainers directly.

Conclusion

Safetensors' move to the PyTorch Foundation marks a significant milestone in the project's journey, ensuring a community-driven approach to AI safety. With a vendor-neutral home and a clear governance structure, Safetensors is well-positioned to continue its growth and address the evolving needs of the ML community.