[Huawei UCM Inference Memory Management Technology Officially Open-Sourced] Jinshi Data, November 5th: Huawei has officially announced the open-sourcing of its key technology for AI inference acceleration—UCM (Unified Cache Manager) inference memory data management. According to Huawei, UCM currently possesses four major capabilities: sparse attention, prefix caching, prefill offloading, and heterogeneous PD decoupling, achieving up to a 90% reduction in first token latency, a maximum 22-fold increase in system throughput, and a 10-fold expansion of the context window.