key-value loading

difference-aware sparsity (1) stripe granularity (1) anchor computation (1) extended context (1) large language model (1) sparse attention (1) inference optimization (1) efficient computing (1) transformer architecture (1)