Model Architecture & Training MethodsNoam Shazeer
Fast Transformer Decoding
Fast Transformer Decoding — Multi-Query Attention
Why it matters
Reduced the K/V heads to one, redefining the KV cache as the decoding bottleneck
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/1911.02150Added in Week 35, 2026 · See that week