Modified Llama-7B to 32k out of box (without finetuning) following the ntk-scaling recipe from this reddit thread.