alexwengg's picture
Decoder: concatenate attention scores instead of K/V (no per-layer cache copy, ~17% faster decode, same output)
51046f0 verified