Pruning Foundation Models for High Accuracy without Retraining