nx_f32_attention.nx

nx_f32_attention.nx -- software-f32 multi-token scaled-dot-product attention (single head), the first brick of the DiT diffusion-backbone half of the sd-server -> Nishi migration. `nx_f32_mha.nx` is SINGLE-TOKEN (autoregressive decode, n_tokens=1, degenerate softmax). The DiT needs MULTI-TOKEN / bidirectional self-attention over all latent (+text) tokens -- exactly the follow-on its header flags ("adds K^T transpose + multi-row softmax"). This is the standard prefill attention: scores[i][j] = dot(Q[i], K[j]) * scale (Q@K^T, scaled) probs[i] = softmax(scores[i]) (over the j axis, numerically stable) out[i] = sum_j probs[i][j] * V[j] (probs@V) Composes the gated primitives nx_f32_add/mul + nx_f32_softmax. Single-head; multi-head = caller loops heads over the head_dim slices of Q/K/V (RoPE + QK-RMSNorm wrap this in the DiT block, next rung). Q,K,V,out: flat *i64 f32 bits, [n_tokens, head_dim] row-major. score_row/prob_row: caller scratch [n_tokens]. license_tier: ORIGINAL