KeDiff: Key Similarity-Based KV Cache Eviction for Long-Context LLM Inference in Resource-Constrained Environments Paper • 2504.15364 • Published Apr 21, 2025 • 3