New trick makes 4-bit AI attention faster on Nvidia GPUs

Original: Hardware-Aware FP4 FlashAttention-4

Source: arxiv.org ↗