GradAlign for Training-free Model Performance Inference