Hi, have you tested the performance on quantized KV Cache? Is it possible to keep a resonable high performance under int8 quantization?