hf_text-generation-inference/docs/source/conceptual/safetensors.md

8 lines
680 B
Markdown
Raw Normal View History

# Safetensors
2024-02-16 03:58:58 -07:00
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).
2024-02-16 03:58:58 -07:00
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.
2024-02-16 03:58:58 -07:00
You can learn more about safetensors by reading the [safetensors documentation](https://huggingface.co/docs/safetensors/index).