Tag
This paper identifies a fundamental limitation (shrinkage bias) in non-uniform FP4 quantization formats for LLM pretraining and proposes UFP4, a uniform 4-bit training recipe that outperforms existing E2M1-based methods.