code wiki / _hdl_build / nx_intfp_swiglu_gradcheck_gate.nx
nx_intfp_swiglu_gradcheck_gate.nx
buildroot/runtime/_hdl_build/nx_intfp_swiglu_gradcheck_gate.nx
about
nx_intfp_swiglu_gradcheck_gate.nx -- the OTHER transformer sublayer in integer: SwiGLU FFN, gradchecked.
y = (SiLU(x@Wg) (.) (x@Wu)) @ Wd , SiLU(z)=z*sigmoid(z)=z/(1+exp(-z)) -- reuses the proven fixed-point exp.
Full backward (dy->dWd, dm-> da/du -> dg via SiLU', -> dWg/dWu), ENTIRELY Q20 integer, gradchecked (integer
finite-diff, max-|grad|-relative metric) on all 3 weight matrices. SiLU'(z)=sig(z)(1+z(1-sig(z))). With this +
attention(48/48) both sublayers of a transformer layer are proven in integer. No float. license_tier: ORIGINAL
dependencies 1 imports · 0 importers
imports: nx_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
| 12 | const S: i64 = 1048576 // Q20 |
| 13 | const T: i64 = 3 |
| 14 | const DM: i64 = 4 |
| 15 | const HF: i64 = 6 |
functions
| 8 | func w(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 } |
| 9 | func wn(v: i64) -> i64 { if v==0 { sys_write(1,"0" as *u8,1); return 0 } var m: i64=v; if m<0{sys_write(1,"-" as *u8,1);m=0-m} let t: *u8=sys_mmap(24); var k: i64=0; while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1} let o: *u8=sys_mmap(24); var q: i64=k-1; var i: i64=0; while q>=0{o[i]=t[q];i=i+1;q=q-1} sys_write(1,o,i); return 0 } |
| 10 | func iabs(v: i64) -> i64 { if v<0 { return 0-v } return v } called by 1: gcheck |
| 17 | func fp_exp(xq: i64) -> i64 called by 1: sigmoid |
| 28 | func sigmoid(z: i64) -> i64 { let e: i64=fp_exp(0-z); return (S*S)/(S+e) } // 1/(1+exp(-z)), Q20 |
| 31 | func swiglu_fwd(x: *i64, Wg: *i64, Wu: *i64, Wd: *i64, g: *i64, a: *i64, u: *i64, m: *i64, y: *i64) -> i64 |
| 43 | func swiglu_bwd2(x: *i64, Wd: *i64, g: *i64, a: *i64, u: *i64, m: *i64, y: *i64, dWg: *i64, dWu: *i64, dWd: *i64) -> i64 |
| 58 | func gcheck(name: *u8, x: *i64, Wg: *i64, Wu: *i64, Wd: *i64, Wtgt: *i64, dW: *i64, ncell: i64, g: *i64, a: *i64, u: *i64, m: *i64, y: *i64) -> i64 |
| 76 | func main() -> i64 |