I don't think this is intended for Llama 2? The Llama 3.1 and 3.2 series have very long context windows (128k tokens).