Ant Open Source Trillion Parameter Reinforcement Learning High Performance Weight Exchange Framework Awex

On November 20th, Ant Group announced the open-source of the trillion-parameter reinforcement learning high-performance weight exchange framework Awex. It is reported that in October of this year, Ant Group open-sourced two trillion-parameter flagship models, including the non-thinking Ling-1T and the thinking model Ring-1T. Among them, Ring-1T, based on its self-developed high-performance weight exchange framework Awex, achieves weight synchronization at the trillion-parameter level in 5-10 seconds on a cluster of several thousand GPU cards.