nx_f32_linear_attention.nx

nx_f32_linear_attention.nx -- sovereign O(n) linear attention (SANA-style), the DiT VRAM/speed lever. From DiT-IC/SANA (ingested 2026-07-01): replace softmax attention (O(n^2 d), materializes the n×n matrix) with LINEAR attention (O(n d^2), no n×n matrix): with a positive feature map phi(x)=relu(x)+1, S = Σ_j phi(K_j) ⊗ V_j (d_k × d_v) z = Σ_j phi(K_j) (d_k) out_i = ( phi(Q_i) @ S ) / ( phi(Q_i) · z ) For n_tokens >> head_dim this is a big compute/VRAM cut -> smaller-VRAM, faster, more-hardware Z-Image DiT. Gate: uniform K/Q -> out = mean(V) (matches softmax's uniform case); + a non-uniform hand-computed check. Q,K,V,out: flat *i64 f32 bits [n_tokens, head_dim]. license_tier: ORIGINAL