nx_f32_transformer.nx

nx_f32_transformer.nx -- bits-up f32 Llama-class transformer block forward (single-token v1). L8 integration brick. Composes EVERY shipped primitive from this session arc: nx_f32_rmsnorm (mul + add + sqrt + div + cvt) nx_f32_matmul (mul + add) nx_f32_softmax (exp + add + div + max) nx_f32_silu (exp + mul + add + div) nx_f32_mul/add (scalar L4) All bits-up. No libm, no compiler-builtin float, no copied code. Single-token, single-head v1. Multi-token + multi-head + GQA + KV cache + RoPE are subsequent bricks (the structural scaffolding for real Llama-7B+ inference; this brick proves the math chain). Per-block formula (Llama-style, RMSNorm + SwiGLU FFN): attn_in = RMSNorm(x, gamma_attn) Q = matmul(attn_in, W_q) [1, head_dim] K = matmul(attn_in, W_k) V = matmul(attn_in, W_v) scores = matmul(Q, K^T) * attn_scale [1, 1] probs = softmax(scores) attn_out = matmul(probs, V) [1, head_dim] attn_proj = matmul(attn_out, W_o) [1, hidden_dim] x = x + attn_proj ffn_in = RMSNorm(x, gamma_ffn) gate = SiLU(matmul(ffn_in, W_gate)) [1, ffn_dim] up = matmul(ffn_in, W_up) [1, ffn_dim] hidden = gate * up [1, ffn_dim] ffn_proj = matmul(hidden, W_down) [1, hidden_dim] x = x + ffn_proj Weights bundle struct keeps the function signature under the NishiLang 16-arg limit while making call-sites self-documenting.