code wiki / (root) / nx_f32_qwen_layer.nx

nx_f32_qwen_layer.nx

buildroot/runtime/nx_f32_qwen_layer.nx

11916 B244 linesdepth 8pulls 12 transitivereach 1 importersview sourcekind tooltopic f32
docsdependenciesstructsconstsfunctions

about

nx_f32_qwen_layer.nx -- a FULL sovereign f32 Qwen transformer layer (the Z-Image text-encoder + companion chat model's repeating unit), assembled from the gated bricks. sd-server -> Nishi migration: the pre-norm GQA transformer layer (Llama/Qwen family). Per the real GGUF: RMSNorm -> [Wq/Wk/Wv] -> GQA attention (32 q / 8 kv heads) -> Wo -> residual -> RMSNorm -> SwiGLU (ffn_gate/ffn_up -> SiLU-gate -> ffn_down) -> residual. Composes nx_f32_rmsnorm + nx_f32_gqa_attention + nx_f32_silu + inline linears. Stack this x n_layers (a loop) = the full encoder forward -> real prompt embeddings. Same layer serves image-prompt-encoding AND companion chat (the shared-LLM reuse). x,out: flat *i64 f32 bits [n_tokens, hidden]. W{q}: [q_dim,hidden], W{k,v}: [kv_dim,hidden], Wo: [hidden,q_dim]; ffn Wg/Wu: [ffn_dim,hidden], Wd: [hidden,ffn_dim]; an_g/fn_g: [hidden]. Scratch is internal (fork-per-use). license_tier: ORIGINAL

dependencies 7 imports · 1 importers

nx_syscalls.nx nx_f32.nx nx_f32_div.nx nx_f32_cvt.nx nx_f32_rmsnorm.nx nx_f32_gqa_attention.nx nx_f32_activations.nx nx_f32_qwen_layer.nx nx_f32_qwen_encoder.nx

imports: nx_syscalls.nxnx_f32.nxnx_f32_div.nxnx_f32_cvt.nxnx_f32_rmsnorm.nxnx_f32_gqa_attention.nxnx_f32_activations.nx

imported by: nx_f32_qwen_encoder.nx

call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown

main sys_mmap nx_i32_to_f32 nx_f32_div nx_f32_classify nx_f32_exp_field nx_f32_mant_field nx_f32_sign nx_f32_mant_field ↻ nx_f32_exp_field ↻ nx_f32_qwen_layer nx_f32_qwen_layer_qk sys_mmap ↻ nx_i32_to_f32 ↻ nx_f32_rmsnorm nx_i32_to_f32 ↻ nx_f32_sqrt nx_f32_classify ↻ nx_f32_sign ↻ nx_f32_exp_field ↻ nx_f32_mant_field ↻ _f32_isqrt qw_linear nx_f32_add nx_f32_classify ↻ nx_f32_sign ↻ nx_f32_mant_field ↻ nx_f32_exp_field ↻ nx_f32_mul nx_f32_classify ↻ nx_f32_sign ↻ nx_f32_mant_field ↻ nx_f32_exp_field ↻ qw_head_norm sys_mmap ↻ nx_f32_rmsnorm ↻ nx_f32_gqa_attention sys_mmap ↻ nx_f32_attention nx_f32_add ↻

structs

none

consts

20const K_MAGIC_1000000: i64 = 1000000

functions

22func qw_linear(inp: *i64, n_tokens: i64, in_dim: i64, W: *i64, out_dim: i64, out: *i64) -> i64
38func qw_swiglu(inp: *i64, n_tokens: i64, hidden: i64, ffn_dim: i64, Wg: *i64, Wu: *i64, Wd: *i64, out: *i64, hb: *i64) -> i64
66func qw_head_norm(buf: *i64, n_tokens: i64, n_heads: i64, head_dim: i64, gain: *i64, eps: i64) -> i64
94func nx_f32_qwen_layer_qk(x: *i64, n_tokens: i64, hidden: i64, n_q_heads: i64, n_kv_heads: i64, head_dim: i64, ffn_dim: i64,
145func nx_f32_qwen_layer(x: *i64, n_tokens: i64, hidden: i64, n_q_heads: i64, n_kv_heads: i64, head_dim: i64, ffn_dim: i64,
155func main() -> i64