code wiki / (root) / nx_nofloat_attention_gate.nx

nx_nofloat_attention_gate.nx

buildroot/runtime/nx_nofloat_attention_gate.nx

7690 B164 linesdepth 3pulls 3 transitivereach 0 importersview sourcekind gate/prooftopic nofloat
docsdependenciesstructsconstsfunctions

about

nx_nofloat_attention_gate.nx -- ASSEMBLE the trio into one DETERMINISTIC transformer attention layer. Composes CAP-NOFLOAT-LAYERNORM + CAP-NOFLOAT-GEMM (matmul) + CAP-NOFLOAT-SOFTMAX, all integer Q16: LayerNorm(X) -> Q,K,V = Xn*Wq/Wk/Wv -> scores = QK^T * 1/sqrt(D) -> softmax per row -> out = attn*V + X. Every op integer -> the WHOLE LAYER is bit-exact DETERMINISTIC. Proven by running the full forward with the score-dot and output-weighted-sum in FORWARD then REVERSE order: the output must be bit-identical (float attention CANNOT promise this). This is the unit a coding-LLM is stacked from -- now sovereign + reproducible. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL

dependencies 2 imports · 0 importers

nx_syscalls.nx nx_gate_verdict.nx nx_nofloat_attention_gate.nx

imports: nx_syscalls.nxnx_gate_verdict.nx

imported by: nobody (leaf or entry point)

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

main at_puts sys_write sys_mmap attn_fwd ln_row isqrt proj qmul qmul ↻ softmax_row fx_exp absd at_num sys_mmap ↻ sys_write ↻ gv_ctr sys_mmap ↻ gv_verdict gv_puts sys_write ↻ gv_num sys_mmap ↻ sys_write ↻ sys_munmap gv_journal sys_openat_append sys_mmap ↻ gv_catn sys_mmap ↻ sys_munmap ↻ sys_now_realtime_sec sys_mmap ↻ sys_clock_gettime_real gv_cat sys_write ↻ sys_close sys_munmap ↻ sys_exit

structs

none

consts

11const Q16: i64 = 65536
12const LOG2E: i64 = 94548
13const PC0: i64 = 65536
14const PC1: i64 = 45426
15const PC2: i64 = 15743
16const PC3: i64 = 4367

functions

18func at_puts(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 }
called by 1: main calls 1: sys_write
19func at_num(v: i64) -> i64 { let b: *u8=sys_mmap(28); var m: i64=v; if m<0{m=0-m;sys_write(1,"-" as *u8,1)} let t: *u8=sys_mmap(28); var k: i64=0; if m==0{t[0]=48 as u8;k=1} while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1} var i: i64=0; while i<k{b[i]=t[k-1-i];i=i+1} sys_write(1,b,k); return 0 }
called by 1: main calls 2: sys_mmapsys_write
20func qmul(a: i64, b: i64) -> i64 { return (a*b) >> 16 }
called by 2: projattn_fwd
21func absd(a: i64, b: i64) -> i64 { if a>b { return a-b } return b-a }
called by 1: main
23func isqrt(v: i64) -> i64
called by 1: ln_row
32func fx_exp(x: i64) -> i64
called by 1: softmax_row
47func ln_row(x: *i64, off: i64, D: i64) -> i64
called by 1: attn_fwd calls 1: isqrt
59func proj(src: *i64, W: *i64, dst: *i64, T: i64, D: i64) -> i64
called by 1: attn_fwd calls 1: qmul
68func softmax_row(sc: *i64, at: *i64, row: i64, T: i64) -> i64
called by 1: attn_fwd calls 1: fx_exp
77func attn_fwd(X: *i64, Wq: *i64, Wk: *i64, Wv: *i64, Xn: *i64, Q: *i64, K: *i64, V: *i64, sc: *i64, at: *i64, out: *i64, T: i64, D: i64, scale: i64, rev: i64) -> i64
called by 1: main calls 4: ln_rowprojqmulsoftmax_row
106func main() -> i64