nx_gen_ditfull.nx

nx_gen_ditfull.nx -- the COMPLETE sovereign Z-Image DiT forward, weights from the GGUF. txt_embed --[context_refiner x2, UNMODULATED]--> txt img_embed --[noise_refiner x2, modulated ]--> img concat(txt, img) --[layers x30, modulated]--> h --[final layer]--> [64, L] Everything after the text encoder and the latent embedding. No oracle tensor in the loop: weights are read from the model file, adaLN is computed per block, and every activation is the previous stage's own output. Usage: nx_gen_ditfull <model_id> <gguf_path> [workers] [reference] THE FINAL LAYER IS NOT A BLOCK and does not reuse br_block: scale = W_ada1 . silu(t_emb) + b <- silu HERE, unlike the blocks' adaLN x = LayerNorm(x) <- LayerNorm (mean-centred), NOT RMSNorm, no affine x = x * (1 + scale) x = W_lin . x + b Both differences were read out of FinalLayer::forward, not assumed. Reusing the block's adaLN path here would apply no silu and subtract no mean, and would still produce a finite tensor. ⚠TOKEN PADDING IS NOT IMPLEMENTED: this shape needs none (512 text and 256 image tokens are both multiples of SEQ_MULTI_OF=32, so n_pad=0). A shape that needs padding must append the learned cap_pad_token / x_pad_token first; this organ REFUSES rather than silently skipping it. license_tier: ORIGINAL