nx_gen_qknorm_verify.nx

nx_gen_qknorm_verify.nx -- SOVEREIGN per-head QK-RMSNorm over a PACKED qkv, verified vs oracle. out[t][h][d] = qkv[t][(head_base+h)*head_dim + d] / sqrt(mean_d(.^2) + eps) * w[d] QK-norm normalizes each attention head's query (or key) vector independently over head_dim, reading from the fused qkv projection where q, k and v are packed head-major in one row. Z-Image, Qwen-Image and most recent DiTs do this; it is the first stage of the attention core. Usage: nx_gen_qknorm_verify <model> <qkv> <norm_w> <out> <head_dim> <n_heads> <head_base> [rows] [eps_recip] head_base 0 for q; n_heads for k (q, k, v are packed in that order) WHY THIS IS A SEPARATE ORGAN AND NOT THE RMSNORM ONE: the two tensors have DIFFERENT row strides. The packed qkv row is (n_q + n_k + n_v) * head_dim wide, while the output row is n_heads * head_dim. A generic normalizer that assumed one stride would read the wrong head and still produce finite, plausible numbers -- the failure mode this lane keeps meeting. Hardware __f32_* intrinsics in the hot loop, never the nx_f32_* software twins. license_tier: ORIGINAL