Worked on enhancing the robustness of the llama.cpp server cache by addressing a critical issue in token cache integrity. Focused on correcting the cache reuse logic when llama_kv_cache_seq_rm is invoked, this work improved session stability and ensured token validity throughout generation workflows. The solution, implemented in C++ with a focus on server development best practices, reduced the risk of token loss and contributed to more reliable long-running sessions. All changes were thoroughly regression-tested and reviewed before integration into the ggerganov/llama.cpp repository, demonstrating careful debugging, disciplined version control, and a commitment to maintainability and reliability in production environments.