It seems to be an issue with compatibility? There is a brief explanation on the HuggingFace SafeTensors page which seems to suggest that shared tensors is a torch-specific feature that isn’t compatible with HuggingFace.
A similar problem was resolved by setting safe_serialization=False but I couldn’t find any other work around and the SafeTensors page seems to suggest that unless HuggingFace addresses this, there really isn’t any other approach.