The rapid growth of data generated by scientific computations and experimental facilities has created critical challenges in the storage and transmission of large-scale scientific data. Addressing these challenges requires efficient compression methods to reduce data size while maintaining acceptable levels of scientific integrity. Error-bounded lossy compression achieves high compression ratios while allowing flexible error control, making it ideal for scientific applications. GPU-based solutions further enhance this by leveraging parallelism for efficient in situ compression, reducing data movement and improving workflow efficiency. Currently, state-of-the-art GPU-accelerated lossy data compressors can be primarily categorized into transform-based methods (such as cuZFP) and prediction-based methods (such as cuSZ-i). The latter typically consists of a prediction-quantization module, which employs numerical methods to reduce data complexity, and a lossless compression module that further minimizes redundancy. We implemented a series of optimization strategies for the prediction module of cuSZ-i to maximize compression ratio, and extensively examined existing lossless compressors to integrate the most suitable lossless solution into cuSZ-i.