code wiki / _hdl_build / nx_f32_moe_train_gate.nx

nx_f32_moe_train_gate.nx

buildroot/runtime/_hdl_build/nx_f32_moe_train_gate.nx

20505 B257 linesdepth 4pulls 5 transitivereach 0 importersview sourcekind gate/prooftopic f32

docs dependencies structs consts functions

dependencies 4 imports · 0 importers

imports: nx_gate_gn.nxnx_gate_base.nxnx_f32_hw.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

16	const VMAX: i64 = 40
17	const D: i64 = 12
18	const HF: i64 = 16
19	const NE: i64 = 4 // experts
20	const TOPK: i64 = 2 // active experts per token (sparse)

functions

22	func grow(name: u8, ok: i64) -> i64 { if ok==1 { gw(" PASS " as u8) } else { gw(" FAIL " as *u8) } gw(name); gw(" calls 1: gw
24	func gm(x: i64) -> i64 { return gn(f32_int(f32_mul(x, f32_of(1000)))) } called by 1: main calls 4: gn f32_int f32_mul f32_of
25	func gu(x: i64) -> i64 { return gn(f32_int(f32_mul(x, f32_of(1000000)))) } // micro (x1e6) -- resolves tiny grads called by 6: f32_exp f32_sigmoid fwd argmax_absgrad grad_ok main calls 1: f32_sub
26	func f32_le(x: i64, y: i64) -> i64 { let d: i64=f32_sub(x,y) & 0xFFFFFFFF; if ((d>>31)&1)==1 { return 1 } if (d & 0x7FFFFFFF)==0 { return 1 } return 0 } called by 1: adam1 calls 3: f32_of f32_div f32_add
27	func f32_sqrt(x: i64) -> i64 { if (x & 0x7FFFFFFF)==0 { return f32_of(0) } var y: i64=x; var i: i64=0; while i<16 { y=f32_div(f32_add(y, f32_div(x,y)), f32_of(2)); i=i+1 } return y } called by 2: f32_sigmoid fwd calls 7: f32_div f32_of f32_mul f32_le f32_int f32_add+1
28	func f32_exp(x: i64) -> i64
35	func f32_log(x: i64) -> i64 { let b: i64=x & 0xFFFFFFFF; let e: i64=((b>>23)&0xFF)-127; let m: i64=(b & 0x7FFFFF)\|0x3F800000; let u: i64=f32_div(f32_sub(m,f32_of(1)),f32_add(m,f32_of(1))); let u2: i64=f32_mul(u,u); var t: i64=u; var s: i64=u; var k: i64=1; while k<=7 { t=f32_mul(t,u2); s=f32_add(s,f32_div(t,f32_of((2*k)+1))); k=k+1 } let ln2: i64=f32_div(f32_of(693147),f32_of(1000000)); return f32_add(f32_mul(f32_of(e),ln2),f32_mul(f32_of(2),s)) } called by 2: f32_silu f32_silu_d calls 6: f32_le f32_of f32_div f32_add f32_exp f32_neg
36	func f32_sigmoid(z: i64) -> i64 { if f32_le(f32_of(0),z)==1 { return f32_div(f32_of(1), f32_add(f32_of(1), f32_exp(f32_neg(z)))) } let e: i64=f32_exp(z); return f32_div(e, f32_add(f32_of(1), e)) } called by 1: fwd calls 2: f32_mul f32_sigmoid
37	func f32_silu(z: i64) -> i64 { return f32_mul(z, f32_sigmoid(z)) } called by 1: bwd calls 5: f32_sigmoid f32_mul f32_add f32_of f32_sub
38	func f32_silu_d(z: i64) -> i64 { let s: i64=f32_sigmoid(z); return f32_mul(s, f32_add(f32_of(1), f32_mul(z, f32_sub(f32_of(1), s)))) }
41	func MWg(M: i64, e: i64) -> i64 { return M[3+e] as *i64 } called by 2: fwd bwd
42	func MWu(M: i64, e: i64) -> i64 { return M[3+NE+e] as *i64 } called by 2: fwd bwd
43	func MWd(M: i64, e: i64) -> i64 { return M[3+2NE+e] as i64 }
49	func fwd(M: i64, tok: i64, tgt: i64, vv: i64, S: i64) -> i64
86	func loss_only(M: i64, tok: i64, tgt: i64, vv: i64, S: i64) -> i64 { return fwd(M, tok, tgt, vv, S) }
89	func bwd(M: i64, tok: i64, tgt: i64, vv: i64, S: i64, G: *i64) -> i64
131	func det_fill(dst: i64, n: i64, seed: i64) -> i64 { var i: i64=0; while i<n { let s: i64=(((i2654435761)+seed) % 13) - 6; dst[i]=f32_div(f32_of(s), f32_of(24)); i=i+1 } return 0 }
134	func argmax_absgrad(Gd: *i64, cnt: i64) -> i64 { var bi: i64=0; var bv: i64=Gd[0] & 0x7FFFFFFF; var i: i64=1; while i<cnt { let av: i64=Gd[i] & 0x7FFFFFFF; if f32_le(bv,av)==1 { bv=av; bi=i } i=i+1 } return bi }
137	func grad_ok(ana: i64, fd: i64) -> i64
146	func zero(P: *i64, n: i64) -> i64 { var i: i64=0; while i<n { P[i]=f32_of(0); i=i+1 } return 0 } called by 1: main calls 6: f32_of f32_add f32_mul f32_sub f32_div f32_sqrt
147	func adam1(P: i64, Gd: i64, Mo: i64, Vo: i64, cnt: i64, lr: i64, b1: i64, b2: i64, bc1: i64, bc2: i64, aeps: i64) -> i64
153	func main() -> i64