Pre-computing attention values for a batch of tokens before generating new tokens, often required when context changes.