ParentFull threadxiphias2·Llama.cpp already moves 1.5bit to the cores, they are converted in SRAM which has much higher bandwidth.FP4 only helps batched inferencing.View on HN