Update README.md

README.md

@@ -10,7 +10,7 @@ alt="\begin{align*}
 \end{align*}
 ">
 
-* Here R, K, V are generated by linear transforms of input, and W is parameter. Basically RWKV decomposes attention into R(target) * W(src, target) * K(src). So we can call R "receptance", and sigmoid means it's in 0~1 range.
+* The R, K, V are generated by linear transforms of input, and W is parameter. The idea of RWKV is to decompose attention into R(target) * W(src, target) * K(src). So we can call R "receptance", and sigmoid means it's in 0~1 range.
 
 * The Time-mix is similar to AFT (https://arxiv.org/abs/2105.14103). There are two differences.
 
@@ -26,6 +26,8 @@ alt="\text{softmax}_t(\text{K}_{u,c}) = \frac{\exp(\text{K}_{u,c})}{\sum_{v \leq
 "https://render.githubusercontent.com/render/math?math=%5Cdisplaystyle+W_%7Bt%2Cu%2Cc%7D%3Df_h%28t-u%29%5Ccdot+%5Calpha_h%28u%29+%5Ccdot+%5Cbeta_h%28t%29" 
 alt="W_{t,u,c}=f_h(t-u)\cdot \alpha_h(u) \cdot \beta_h(t)">
 
+(3) You don't need LayerNorm for Time-mix. In fact, the model converges faster when LayerNorm is removed.
+
 Moreover we multiply the final output of Time-mix layer by γ(t). The reason for the α β γ factors, is because the context size is smaller when t is small, and this can be compensated using the α β γ factors.
 
 * The Channel-mix is similar to GeGLU (https://arxiv.org/abs/2002.05202) with an extra R factor.