3 ms·Fast Transformer Decoding: One Write-Head Is All You Need4 points by hislaziness 3y agohislaziness 3y agoA more efficient way to infer with lower memory requirements.