StreamingLLM

Attention-sink method enabling LLMs to stream infinite-length inputs without fine-tuning.