Fast Transformer Decoding: One Write-Head Is All You Need (opens in new tab)

A more efficient way to infer with lower memory requirements.