2 ms·Dynamic Memory Compression: Retrofitting LLMs for Accelerated Inference3 points by veryluckyxyz 3y ago