Diabetic retinopathy (DR) is a significant eye condition that can lead to vision loss and blindness in diabetic patients, emphasizing the importance of early detection and intervention. DR affects the blood vessels in the retina, often without initial symptoms, making regular comprehensive eye exams crucial for those with diabetes. This study addresses the challenge of DR detection by exploring various state-of-the-art deep learning models, including DeiT, MobileNet-V2 V2, and combinations of PiT architectures with MobileNet-V2. After evaluating these approaches, we focused on optimizing the PiT-Base-224 with MobileNet-V2 model, utilizing both Adam and Lamb optimizers to enhance its performance. By integrating the strengths of transformer-based architectures with pretrained CNNs, we developed a highly accurate and efficient diagnostic system. The results demonstrate the potential of this optimized model to enhance DR detection, achieving high accuracy and paving the way for accessible and reliable diagnostic tools in diabetic care with 99.43, 99.6, and 99.4 for accuracy, precision, and recall, respectively. This study highlights the effectiveness of our hybrid model and optimization strategy in accurately diagnosing diabetic retinopathy, emphasizing the importance of early detection and intervention in preventing vision loss.