Developed a Connectionist Temporal Classification (CTC) loss operator for the FlagOpen/FlagGems repository, focusing on accelerating sequence-to-sequence model training. The implementation included both forward and backward computation paths, supporting a variety of input shapes and configurations to enhance flexibility in deep learning workflows. Comprehensive unit and integration tests were added to ensure numerical correctness and stability, addressing edge cases commonly encountered in sequence modeling. Leveraging Python, CUDA, and PyTorch, the work enabled more reliable and potentially faster training for machine learning models that rely on CTC loss, contributing to improved robustness in sequence-to-sequence applications within the repository.