“The ScaleSearchAttention algorithm improves Wikitext-2 perplexity (PPL) by up to 0.77 points for the Llama 3.1 70B model.”