Glaucoma is a leading cause of irreversible blindness worldwide. While deep learning approaches using fundus images have largely improved early diagnosis of glaucoma, variations in images from different devices and locations (known as domain shifts) challenge the use of pre-trained models in real-world settings. To address this, we propose a novel Graph-guided Test-Time Adaptation (GTTA) framework to generalize glaucoma diagnosis models to unseen test environments. GTTA integrates the topological information of fundus images into the model training, enhancing the model's transferability and reducing the risk of learning spurious correlation. During inference, GTTA introduces a novel test-time training objective to make the source-trained classifier progressively adapt to target patterns with reliable class conditional estimation and consistency regularization. Experiments on cross-domain glaucoma diagnosis benchmarks demonstrate the superiority of the overall framework and individual components under different backbone networks.