nx_gen_taedec.nx

nx_gen_taedec.nx -- SOVEREIGN TAESD (tiny VAE) decoder: latent -> RGB pixels. The last stage before an image. TAESD is 9.4MB of plain f32 3x3 convolutions where the full VAE is 320MB with attention, so it is the cheapest honest route from a sovereign latent to a sovereign picture. h = tanh(z/3)*3 (the caller's fixture already starts here) conv3x3 16->64 (+bias), relu [TAEBlock x3] upsample2 conv3x3 64->64 (no bias) [TAEBlock x3] upsample2 conv3x3 64->64 (no bias) [TAEBlock x3] upsample2 conv3x3 64->64 (no bias) [TAEBlock x1] conv3x3 64->3 (+bias) TAEBlock: conv0 -> relu -> conv2 -> relu -> conv4, then + input, then relu Usage: nx_gen_taedec <model_id> <taesd.safetensors> [workers] [in] [ref] ⚠WHY THIS DOES NOT CALL nx_f32_conv2d_forward That organ's inner loop uses nx_f32_add / nx_f32_mul -- the SOFTWARE IEEE-754 pair -- not the hardware __f32_* intrinsics beside them. At this decoder's ~20 G MACs that is the difference between ~25 minutes and seconds. Same defect class this lane has now hit four times, here in a pre-existing shared organ (filed separately; fixing it there is a fleet change, not a local one). ★★★★★ WHEN A SOVEREIGN KERNEL IS 10x UNDER PEAK, SUSPECT AN nx_-PREFIXED HELPER. Layout is planar packed f32 [C][H][W], W contiguous -- the same order ggml uses ([W,H,C,N]) and the same order safetensors stores conv weights ([C_out][C_in][KH][KW]). license_tier: ORIGINAL