Model evaluation by predicting generalization error on unlabelled graph under distribution shifts is critical for reliable deployment in practice. Previous methods measure network parameter discrepancy between a well-trained model and its retrained version to evaluate model performance, where the retrained model is supervised by the well-trained model with test graph on the retraining process. However, we reveal that not all parameters are equal correlated with generalization error, which could negative influence on evaluating performance. To address this gap, we develop a score named GradNorm to predict generalization error. Specifically, GradNorm quantifies the learning discrepancy via the dot product of the flattened model parameters and gradients accumulated during retraining, weighting the parameters by gradient history. Extensive experiments demonstrate the effectiveness of GradNorm on node classification tasks under graph distribution shifts across various Graph Neural Network (GNN) models.