The Dexterous Grasp Transformer (DGTR) is a novel discriminative framework for generating diverse and high-quality dexterous grasp poses from a complete object point cloud. DGTR formulates dexterous grasp generation as a set prediction task and employs a transformer-based model inspired by Detection Transformers. The model uses a transformer decoder and learnable queries to predict a diverse set of grasp poses in one forward pass. To address optimization challenges, DGTR introduces progressive strategies for training and testing. During training, a dynamic-static matching training (DSMT) strategy is used to enhance optimization stability. During testing, an adversarial-balanced test-time adaptation (AB-TTA) strategy is introduced to refine grasp poses directly in the parameter space of the dexterous hand. These strategies significantly improve grasp quality and diversity. Experimental results on the DexGraspNet dataset show that DGTR outperforms previous methods in multiple metrics without data preprocessing. DGTR achieves high-quality and diverse grasp poses with one forward pass, making it the first work to introduce set prediction formulation into dexterous grasp generation. The framework demonstrates strong performance in robotic dexterous grasping scenarios.The Dexterous Grasp Transformer (DGTR) is a novel discriminative framework for generating diverse and high-quality dexterous grasp poses from a complete object point cloud. DGTR formulates dexterous grasp generation as a set prediction task and employs a transformer-based model inspired by Detection Transformers. The model uses a transformer decoder and learnable queries to predict a diverse set of grasp poses in one forward pass. To address optimization challenges, DGTR introduces progressive strategies for training and testing. During training, a dynamic-static matching training (DSMT) strategy is used to enhance optimization stability. During testing, an adversarial-balanced test-time adaptation (AB-TTA) strategy is introduced to refine grasp poses directly in the parameter space of the dexterous hand. These strategies significantly improve grasp quality and diversity. Experimental results on the DexGraspNet dataset show that DGTR outperforms previous methods in multiple metrics without data preprocessing. DGTR achieves high-quality and diverse grasp poses with one forward pass, making it the first work to introduce set prediction formulation into dexterous grasp generation. The framework demonstrates strong performance in robotic dexterous grasping scenarios.