Prefix Cache

📝 VLCACHE: Computing 2% Vision Tokens and Reusing 98% for Vision–Language Inference