Tensormesh uses a form of kv caching to make inference more than ten times more efficient. Source link
