[Feat]: Using Safetensors instead of Pickle Libraries for security purposes

[cogspot]

Description
There are inherent risks of locally hosting models especially from China since CCP requires, by law, that software must be backdoored. PyTorch models are loaded using a serialization library called Pickle, which creates a virtual machine very briefly during the process of loading model weights. It is possible to hide malicious code inside the pickle library, which then executes upon the VM being created. SafeTensors allows you to mitigate this security risk.

Use Case
Securely offload and load locally hosted models

[a-ghorbani]

Could you point out where we mentioned that we are using Pickle?

PocketPal AI uses the GGUF format: https://huggingface.co/docs/hub/en/gguf