Efficiently Scale LLM Training Across a Large GPU Cluster with Alpa and Ray (developer.nvidia.com)