Stable and Efficient Reinforcement Learning for Trillion-Parameter LLMs - View it on GitHub