[AUTOML24] Compressing Large Language Models via Neural Architecture Search