Safetensors has joined the PyTorch Foundation as a new project. The format, developed by Hugging Face, aims to make the distribution of AI models more secure by eliminating the risks of arbitrary code execution.
Safetensors is joining the PyTorch Foundation. This news comes a day after Helion also joined the foundation, and is also being announced during the first European PyTorch conference in Paris.
With this addition, the PyTorch Foundation is accelerating its ambition to grow. “It is not only important to expand the capabilities of existing projects under the Foundation’s wing, but we must also add entirely new projects,” says Executive Director Mark Collier. Safetensors joins PyTorch, DeepSpeed, Ray, vLLM, and Helion in the foundation.
A more secure format for LLMs
Safetensors is a serialization format for tensors and metadata in machine learning models. The project aims to address a known risk in older pickle-based formats, where model files can be exploited to execute arbitrary, untrusted code as soon as developed by Hugging Face they are loaded.
“In some cases, old formats allow code to be injected into open weight models,” Collier explains. “SafeTensors is a new format for publishing models, ensuring it is clear that they are what they claim to be.”
New standard
“SafeTensors is thus lending itself to becoming a standard,” he continues. “The entire industry wants this problem solved. It doesn’t need to happen a thousand times: once is enough. A solution under the wings of a non-profit like the PyTorch Foundation can ensure broad adoption. Moreover, many contributions only make the standard more secure.”
Safetensors was developed and is maintained by Hugging Face. In the open-source ML ecosystem, it has since grown into a widely used format for the distribution of models and associated metadata.
More than PyTorch
With its inclusion in the PyTorch Foundation, the project gains a more formal place within the open-source AI landscape. For its part, the Foundation further cements itself as an organization that supports not only the PyTorch framework, but a growing ecosystem of open tools that make open AI possible.