KV Cache

During generation, the model recomputes attention for all past tokens at each step. The KV cache stores intermediate results to avoid this. What does it store? Flag format: CONGRESS{[what_is_stored]} Example: CONGRESS{embedding_vectors}