Appearance
TransformerBlock
Source: models/TransformerBlock.luau
Pre-norm transformer block: h = x + SA(LN(x)); h = h + FFN(LN(h)). Optional residual scaling: initType "rezero" (alpha=0) or "deepnorm" (alpha = (2L)^(1/4) as in DeepNorm, 2203.00555).