[ICLR 2025] COAT: Compressing Optimizer States and Activation for Memory-Efficient FP8 Training - View it on GitHub