<p>The impact of tomato leaf diseases on tomato production worldwide is one of the most significant risks, as the diseases cause massive yield losses and high management costs. Modern precision agriculture systems, therefore, need to be accurate, timely, and automated in the diagnosis of diseases. The paper proposes a vision transformer with Cascaded Group Attention (ViT-CGA) model to classify tomato leaf diseases efficiently and reliably. The proposed architecture integrates patch-based token embeddings with multi-scale grouped attention and cascaded refinement to capture small-scale disease symptoms and long-range contextual dependencies that traditional convolutional neural networks tend to miss. The model is trained and tested on a massive, mixed dataset of several tomato leaf disease classes under different environmental conditions. Occlusion, illumination variation, and leaf overlap are situations that require extensive data augmentation to address class imbalance and improve robustness to these conditions. The experimental findings indicate that ViT-CGA is consistently qualitatively better than the baseline CNN, YOLOv7, and SFMS models, as it is more accurate across all evaluation scenarios. The classification accuracy of the proposed method is between 95.6% and 96.9%, the precision is up to 97.2%, the recall is up to 97.5%, and the false classification rate is considerably less. The ability to remain robust even in stressful environmental conditions is over 90%, and real-time inference can be attained at latency rates of 31 ms on CPU and 44 ms on edge devices. These results affirm the efficacy, efficiency, and practicality of ViT-CGA for the real-time use of the agricultural decision-support and sustainable crop management systems.</p>