Worked on the jeejeelee/vllm repository to address a critical bug in the KV offloading scheduler affecting MTP/Eagle models. Focused on backend development and distributed systems, the work involved implementing a consistent storable block calculation by separating prefill and decoding phases, which improved the correctness and reliability of KV offloading. Using Python and Pytest, the update prevented interior blocks at chunk boundaries from being dropped, thereby preserving stable prompt blocks and preventing permanent gaps in the KV cache. This fix reduced cache fragmentation, enabled reliable prefix reuse, and enhanced model throughput and reliability for production workloads involving large language models.