nx_quant_calibration_gate.nx
buildroot/runtime/nx_quant_calibration_gate.nx
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
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
structs
| none |
consts
| none |
functions
| 12 | func 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 } |
| 13 | func 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 } |
| 14 | func absq(x: i64) -> i64 { if x<0 { return 0-x } return x } |
| 15 | func rdiv(a: i64, b: i64) -> i64 { if a>=0 { return (a+(b>>1))/b } return 0 - (((0-a)+(b>>1))/b) } |
| 16 | func clampq(x: i64, lo: i64, hi: i64) -> i64 { if x<lo { return lo } if x>hi { return hi } return x } called by 1: measure |
| 19 | func measure(val: *i64, isout: *i64, N: i64, scale: i64, res: *i64) -> i64 |
| 33 | func main() -> i64 |