nx_f32_adaln.nx

nx_f32_adaln.nx -- software-f32 adaptive LayerNorm (adaLN / adaLN-Zero) modulation, the DiT-specific brick. sd-server -> Nishi migration: this is what turns a plain transformer block into a DIFFUSION transformer (DiT) block (Peebles & Xie 2023 "Scalable Diffusion Models with Transformers"; Z-Image uses it). A small MLP maps the conditioning vector (timestep + pooled text) to per-block modulation params {shift, scale, gate} for each sublayer; the block then applies: modulate(norm(x), shift, scale) = norm(x) * (1 + scale) + shift (broadcast over tokens) x = x + gate * sublayer(modulate(...)) (gated residual; adaLN-Zero starts gate=0) This file owns the two elementwise ops (modulate + gated-residual). The conditioning MLP that produces shift/scale/gate is a plain linear (nx_f32_matmul) the caller supplies. Composes nx_f32_add/mul + cvt. x, out, res, h: flat *i64 f32 bits [n_tokens, D]. scale, shift, gate: [D] f32 (per-feature, token-broadcast). license_tier: ORIGINAL