Accurate KV Cache Eviction via Anchor Direction Projection for Efficient LLM Inference

Open in new window