Fetching the paper…

GEAR: An Efficient KV Cache Compression Recipe for Near-Lossless Generative Inference of LLM · Around