Infinite Context Windows? LLMs for Streaming Applications with Attention Sinks

LLMs trained with a finite attention window can be extended to infinite sequence lengths without any fine-tuning.

applications attention context context windows fine-tuning llms plain english papers streaming windows