RMSNorm
x b = RMSNorm ( x ) = x 1 n ∑ i = 1 n ( x i 2 ) + ϵ xb = \text{RMSNorm}(x) = \frac{x}{\sqrt{\frac{1}{n} \sum_{i=1}^{n} (x_i^2) + \epsilon}} xb=RMSNorm(x)=n1∑i=1n(xi2)+ϵ x
RoPE
对q,k进行PE
公式推导:
先给q(位置m),k(位置n)添加绝对位置信息: f(q, m), f(k, n)
相对位置信息:g(q, k, m-n)
即要构造出<f(q, m), f(k, n)> = g(q, k, m-n)
根据实部虚部推导就行,结果是 f ( q , m ) = q e i m θ f(q,m) = qe^{im\theta} f(q,m)=qeimθ, 即对q转 m θ m\theta mθ,即:
FFN
xb = RMSNorm(x, weight)
hb = xb @ w1, hb2 = xb @ w3
SwiGLU: hb * σ ( h b ) \sigma(hb) σ(hb) * hb2
xb = hb @ w2
残差处理: x += xb