Developed and delivered Attention Sinks support in the AITer Flash Attention backend for the jeejeelee/vllm repository, focusing on enabling variable-length sequence handling and speculative decoding for deep learning workloads. The work centered on backend development using Python and PyTorch, with careful integration of ROCm compatibility to broaden deployment options. Emphasis was placed on maintainability and clear commit traceability, ensuring robust feature delivery without introducing major bugs. This addition improved the versatility and performance of user workloads in ROCm-enabled environments, addressing the need for flexible sequence processing in machine learning applications while maintaining a clean and well-documented codebase.