Where the KV cache lives
model
layer
attention
heads
▶ zoom in