Prompt caching is a performance feature that stores an LLM’s internal computations (key-value tensors) for stable prefix tokens - View it on GitHub