On November 4th, the Store Computing Power China Tour Beijing event organized by the China Academy of Information and Communications Technology brought together nearly 20 industry representatives from China Mobile, Huawei, Silicon Base Liquid Technol

Currently, AI inference faces challenges such as upgraded KVCache storage requirements, multi-modal data collaboration, insufficient storage and computation bandwidth, load tidality, and cost control, said Zhou Yu, the chief project engineer of China

In this context, Huawei's open source UCM (UnifiedCacheManager) inferential memory data management technology is regarded as one of the key solutions to break the situation. This technology focuses on KVCache multi-level caching and inference memory

It is reported that UCM technology has recently been open-sourced on the Magic Engine community, with open-source content including four key capabilities: sparse attention, prefix caching, pre-fetch unloading, and heterogeneous PD decoupling. This ca

It is widely believed that the open source of UCM goes beyond mere technology sharing, enabling developers and enterprises to acquire industry-leading AI inference acceleration capabilities at lower cost and in a more convenient manner, thus promotin

Editor: Yang Jujuan

Proofread by Zhao Lin