← Efficient Large-Scale Language Model Training on GPU ...