To understand why a GPU prefill path is often significantly faster than single-token decode, we have to look at how Large