code wiki / (root) / nx_quant_calibration_gate.nx

nx_quant_calibration_gate.nx

buildroot/runtime/nx_quant_calibration_gate.nx

5628 B84 linesdepth 3pulls 3 transitivereach 0 importersview sourcekind gate/prooftopic quant
docsdependenciesstructsconstsfunctions

about

nx_quant_calibration_gate.nx -- the UNGLAMOROUS production engineering an integer stack actually needs: dynamic-range / outlier handling, MEASURED honestly (not a toy that rigs the alternative). The real obstacle to integer inference is OUTLIERS: a few huge values force a coarse per-tensor scale that crushes the common values into a handful of levels (huge error on the bulk). Production (SmoothQuant/AWQ/ per-channel) fixes this. Here we measure it directly in Q16: quantize the SAME data to INT8 with NAIVE per-tensor min/max vs CLIPPED calibration, and report the bulk error AND -- honestly -- the outlier error that clipping does NOT fix for free. Evidence, with the tradeoff stated, not "we're uniquely smart". No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL

dependencies 2 imports · 0 importers

nx_syscalls.nx nx_gate_verdict.nx nx_quant_calibration_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 qc_puts sys_write sys_mmap absq rdiv measure clampq rdiv ↻ absq ↻ qc_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

none

functions

12func qc_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
13func qc_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
14func absq(x: i64) -> i64 { if x<0 { return 0-x } return x }
called by 2: measuremain
15func rdiv(a: i64, b: i64) -> i64 { if a>=0 { return (a+(b>>1))/b } return 0 - (((0-a)+(b>>1))/b) }
called by 2: measuremain
16func clampq(x: i64, lo: i64, hi: i64) -> i64 { if x<lo { return lo } if x>hi { return hi } return x }
called by 1: measure
19func measure(val: *i64, isout: *i64, N: i64, scale: i64, res: *i64) -> i64
called by 1: main calls 3: clampqrdivabsq
33func main() -> i64