Worked on the kvcache-ai/ktransformers repository to optimize the CPUInfer backend initialization process, focusing on backend development and performance optimization using Python. Refactored the backend to implement lazy initialization, ensuring that the backend is only instantiated when the requested thread count increases. This approach reduced unnecessary overhead by preventing repeated backend instantiation, which in turn lowered startup latency and improved throughput for concurrent CPU inference workloads. Lifecycle guards were added to align backend state with thread count, minimizing potential race conditions. The work targeted scalable, low-latency CPU inference and enhanced resource utilization in multi-threaded environments, contributing to more efficient backend operations.