Worked on the alibaba/MNN repository to enhance caching reliability for large language model workloads by addressing a critical issue in the prefix disk cache. Using C++ and backend development skills, implemented a targeted bug fix that ensured the prefix disk cache loaded correctly after the first response, resolving a persistent loading problem. Introduced a verification mechanism for sync files and enforced the creation and validation of prefix cache files, improving cache management and file handling under high-load conditions. These changes led to more consistent response times and reduced cache-related failures, strengthening the system’s resilience and reliability in demanding production environments.