Today we're officially moving Safetensors under the PyTorch Foundation, alongside PyTorch (of course), vLLM, DeepSpeed, Ray, and the recently-announced Helion. Concretely this means the trademark and repo are now held by the Linux Foundation rather than Hugging Face: neutral stewardship and open governance.
For local inference nothing changes today. Same format, same APIs, same Hub compatibility; we're working with the PyTorch team directly to see how to best integrate within PyTorch core.
What this does unlock is the ability to work more openly with the broader ecosystem on some further optimizations; more than a file format, there are some good opportunities for speedups across the board within the python/pytorch ecosystem: device-aware loading on different accelerators, tp/pp optimized loading, and new quantization/data types support.
We're currently refining our roadmap for the next few months/years and we'd be happy to work on it with you. Happy to answer questions about any of this, or the governance side.
interesting blog if not just for the interactive plots: getting to see how all the common ML architectures are linked together and their evolution has value.
I'm guessing that as we go deeper, we'll see fewer and fewer completely novel architectures that don't depend on the previous ones.
Today we're officially moving Safetensors under the PyTorch Foundation, alongside PyTorch (of course), vLLM, DeepSpeed, Ray, and the recently-announced Helion. Concretely this means the trademark and repo are now held by the Linux Foundation rather than Hugging Face: neutral stewardship and open governance.
For local inference nothing changes today. Same format, same APIs, same Hub compatibility; we're working with the PyTorch team directly to see how to best integrate within PyTorch core.
What this does unlock is the ability to work more openly with the broader ecosystem on some further optimizations; more than a file format, there are some good opportunities for speedups across the board within the python/pytorch ecosystem: device-aware loading on different accelerators, tp/pp optimized loading, and new quantization/data types support.
We're currently refining our roadmap for the next few months/years and we'd be happy to work on it with you. Happy to answer questions about any of this, or the governance side.