Is this the same as the prompt caching that other API's (Anthropc, OpenAI, etc) have had, just open source and for vLLM?