Safetensors is a model serialization format for deep learning models. It is [faster](https://huggingface.co/docs/safetensors/speed) and safer compared to other serialization formats like pickle (which is used under the hood in many deep learning libraries).
TGI depends on safetensors format mainly to enable [tensor parallelism sharding](./tensor_parallelism). For a given model repository during serving, TGI looks for safetensors weights. If there are no safetensors weights, TGI converts the PyTorch weights to safetensors format.