tomaarsen/attention_sinks

Extend existing LLMs way beyond the original training length with constant memory usage, without retraining