SpeedLoader: An I/O efficient scheme for heterogeneous and distributed LLM operation

With the surging growth of model parameters, foundation models pose unprecedented challenges to traditional computational infrastructures.