Fast Transformer Decoding: One Write-Head Is All You Need (arxiv.org)

A more efficient way to infer with lower memory requirements.