cache-fit means that all of the model's weights reside permanently in L3 cache, so that each forward pass reads the weights from cache, thereby bypassing DRAM's bandwidth bottleneck and access latency.