Safetensors conceptual guide (#905)

IDK what else to add in this guide, I looked for relevant code in TGI
codebase and saw that it's used in quantization as well (maybe I could
add that?)

docs/source/_toctree.yml

@@ -21,6 +21,8 @@
 - sections:
   - local: conceptual/streaming
     title: Streaming
+  - local: conceptual/safetensors
+    title: Safetensors
   - local: conceptual/flash_attention
     title: Flash Attention
   title: Conceptual Guides

docs/source/conceptual/safetensors.md

+# Safetensors
+
+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. 
+
+You can learn more about safetensors by reading the [safetensors documentation](https://huggingface.co/docs/safetensors/index).
\ No newline at end of file