During this period, contributed to the alibaba/rtp-llm repository by implementing ROCm Prefill-Decode (pd separation) support for ROCm devices, focusing on improving synchronization and efficiency in attention mechanisms. Developed a new ROCm event lifecycle to create and manage synchronization events, enhancing task coordination on ROCm GPUs. Integrated cache storage into the context attention operation, which increased throughput and performance for attention workloads. The work leveraged C++ and device management skills, with an emphasis on performance optimization and ROCm-specific development. This feature addressed the need for more efficient attention processing, resulting in improved functionality and resource utilization on ROCm hardware.