code wiki / _hdl_build / nx_nofloat_train.nx
nx_nofloat_train.nx
buildroot/runtime/_hdl_build/nx_nofloat_train.nx
about
nx_nofloat_train.nx -- SOVEREIGN NO-FLOAT TRAINING CORE (ledger #1, the biggest gap). Integer Q16 fixed-point
gradient descent: forward (pred=w*x+b), loss (sum err^2), backprop (dw=err*x, db=err), SGD update -- 100% integer,
no float. Learning is DEMONSTRATED: loss drops + weights converge to the true params. The training analog of the
no-float inference moat.
EXCEED vs Unsloth (eff_unsloth.raw) / MLPerf (bt_mlperf.raw): training is DETERMINISTIC/bit-exact reproducible
(same weights every run, on any hardware) -- float/CUDA training is NOT (nondeterministic kernels, hardware drift).
SOVEREIGN (no CUDA/PyTorch/GPU), runs on any CPU. BEHIND: raw speed (no GPU kernels) + scale (tiny vs billions).
T1 loss DECREASED (learning happened). T2 weights CONVERGED to truth (w~2.0, b~1.0 in Q16).
T3 (EXCEED) DETERMINISM: two training runs -> BIT-IDENTICAL weights. T4 final loss is small.
expect_exit: 0 Sovereign: nx_syscalls.
dependencies 2 imports · 0 importers
imports: nx_syscalls.nxnx_g_puts_lib.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
| 18 | const Q: i64 = 16 |
| 19 | const ONE: i64 = 65536 |
functions
| 14 | func g_pn(v: i64) -> i64 { let b: *u8=sys_mmap(28); var x: i64=v; if x<0{b[0]=45;sys_write(1,b,1);x=0-x} if x==0{b[0]=48;sys_write(1,b,1);return 0} var d: i64=0; var y: i64=x; while y>0{d=d+1;y=y/10} var i: i64=d-1; y=x; while i>=0{b[i]=(48+(y%10)) as u8;y=y/10;i=i-1} sys_write(1,b,d); return 0 } |
| 15 | func ck(name: *u8, c: i64) -> i64 { if c==1 { g_puts(" PASS " as *u8) } else { g_puts(" FAIL " as *u8) } g_puts(name); g_puts("\n" as *u8); return c } |
| 16 | func have(path: *u8) -> i64 { let fd: i64=sys_openat_rd(path); if fd<0 { return 0 } sys_close(fd); return 1 } |
| 20 | func fxmul(a: i64, b: i64) -> i64 { return (a*b)>>Q } // Q16 fixed-point multiply called by 1: train |
| 21 | func iabs(v: i64) -> i64 { if v<0 { return 0-v } return v } called by 1: main |
| 24 | func train(wout: *i64, epochs: i64, lr: i64) -> i64 |
| 45 | func main() -> i64 |