inference-methods.md

Multi-Query Attention: Fast Transformer Decoding: One Write-Head is All You Need(Nov 2019)，速度更快，更低的显存占用