Worked on performance optimization for the alibaba/rtp-llm repository, focusing on improving the efficiency of top log probability calculations within the CustomChatRenderer class. Replaced a full-array sort with a top-k selection approach in two code paths, reducing unnecessary computations and lowering CPU usage. This change, implemented in Python, targeted latency reduction for top-probability dependent rendering, resulting in faster response times and better throughput under high concurrency. Maintained code quality by ensuring all tests passed and updating documentation as needed. No new defects were introduced, and the optimization contributed to a more scalable and responsive LLM development workflow.