A Sparsity Predicting Approach for Large Language Models via Activation Pattern Clustering