KeyDiff: Key Similarity-Based KV Cache Eviction for Long-Context LLM Inference in Resource-Constrained Environments

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