code wiki / _hdl_build / nx_nofloat_k3interleave_gate.nx
nx_nofloat_k3interleave_gate.nx
buildroot/runtime/_hdl_build/nx_nofloat_k3interleave_gate.nx
about
nx_nofloat_k3interleave_gate.nx -- gate: the 3:1 KDA:MLA interleaved K3 block structure, assembled from the
proven KDA + MLA organs, deterministic at depth.
T1 the 4-layer interleave [KDA,KDA,KDA,MLA] runs end-to-end + transforms the input
T2 BOTH attention types are LIVE: perturbing ONLY an MLA weight changes the output, AND perturbing ONLY a
KDA weight changes the output (neither path is dead / optimized away)
T3 DEPTH-STABLE: output bounded through the interleaved stack (RMSNorm holds)
T4 DETERMINISTIC bit-identical at depth (float interleaved stacks drift + compound)
T5 the schedule is mechanically 3:1: layer_type over 8 layers == [0,0,0,1,0,0,0,1] (6 KDA : 2 MLA)
license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0
dependencies 3 imports · 0 importers
imports: nx_nofloat_k3interleave.nxnx_gate_verdict.nxnx_syscalls.nx
imported by: nobody (leaf or entry point)
call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown
structs
| none |
consts
| 14 | const IG_T: i64 = 4 |
| 15 | const IG_D: i64 = 4 |
| 16 | const IG_DFF: i64 = 8 |
| 17 | const IG_NH: i64 = 2 |
| 18 | const IG_HD: i64 = 2 |
| 19 | const IG_LC: i64 = 2 |
| 20 | const IG_L: i64 = 4 |
| 21 | const IG_Q: i64 = 65536 |
| 22 | const IG_SCALE: i64 = 46341 // 1/sqrt(hd=2) in Q16 |
functions
| 24 | func ig_fill(m: *i64, n: i64, seed: i64) -> i64 called by 1: main |
| 30 | func main() -> i64 |