TokenFormer treats model parameters as tokens, enabling efficient scaling without full retraining, to cut training costs by 90% 🤯