Multi-head = attention done several times in parallel. The token vector is split into equal slices (heads); each head runs its own attention independently, so different heads can focus on different things; then the pieces are concatenated and mixed by a final projection Wₒ. Same math, just sliced and re-joined.
One generation step, in three phases: 1 the latest token's Q is multiplied with every cached K → softmax weights; 2 those weights blend the cached Vs into an output that predicts the next token; 3 that new token's own K, V slide into the cache — ready for the next step. Past K/V are reused, never recomputed. Single head.