Simplifying 4-bit AI training with a wider-range scaling trick

Original: UE5M3 FP4 Block Scaling for Stable Language Model Pretraining

Source: arxiv.org ↗