MicroGPT::TransformerBlock
Inherits MicroGPT::Layer < Reference < Object
Constructors
Instance methods
attn
Sourcebackward(grad : Mat) : Mat
Sourcebackward_accumulate(grad : Mat) : Mat
Sourceff
Sourceforward(x : Mat, mask : Mat | Nil = nil) : Mat
Pre-norm: x + attn(norm(x)), then x + ff(norm(x))
ln1
Sourceln2
Source