Light Just Cut KV Cache Memory Traffic to 1/16th

Light Just Cut KV Cache Memory Traffic to 1/16th The bottleneck in long-context LLM...