/attention_sinks

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

Primary LanguagePythonApache License 2.0Apache-2.0

Stargazers

No one’s star this repository yet.