Production inference framework for LLMs with RadixAttention caching. Powers 400k+ GPUs generating trillions of tokens daily.