nx_genarch.nx

nx_genarch.nx -- derive a DiT's architecture FROM THE MODEL FILE, so parts are hot-swappable. Operator, 2026-08-07: *"make sure that they are hot swappable with other models till we can pay to make our own model"*. Z-Image Turbo is a MIX of parts (DiT + Qwen text encoder + VAE); the right goal is not to rebuild each part but to be able to drop a different one in. A hardcoded architecture constant is the thing that prevents that. This lane has already been burned four times by assumed constants -- layer count (30, not the 36 I assumed), norm_eps, adaLN chunk order, SDPA scale -- and each was silently wrong rather than loudly wrong. ★★★★★ AN ARCHITECTURE CONSTANT IN THE CODE IS A MODEL THAT CANNOT BE SWAPPED. Everything below is read off tensor SHAPES, because this GGUF carries zero metadata KV pairs (measured: `kv 0`) -- so the shapes are the only self-description the file has. SLOTS 0 hidden D from attention.qkv.weight dim0 1 head_dim from attention.q_norm.weight length 2 n_heads qkv_width / 3 / head_dim (q,k,v equal-width; verified, not assumed) 3 qkv_width from attention.qkv.weight dim1 4 ffn_dim from feed_forward.w1.weight dim1 5 n_layers probed: layers.N until absent 6 n_refiners probed: context_refiner.N until absent 7 adaln_embed from adaLN_modulation.0.weight dim0 8 n_chunks adaLN_modulation.0.weight dim1 / D 9 out_dim from final_layer.linear.weight dim1 (= patch*patch*latent_channels) 10 x_embed_in from x_embedder.weight dim0 license_tier: ORIGINAL