code wiki / (root) / nx_block_quant_h2h_gate.nx

nx_block_quant_h2h_gate.nx

buildroot/runtime/nx_block_quant_h2h_gate.nx

6006 B87 linesdepth 3pulls 3 transitivereach 0 importersview sourcekind gate/proof
docsdependenciesstructsconstsfunctions

about

nx_block_quant_h2h_gate.nx -- work the BEHIND scorecard line (quantization) with a real technique: BLOCK (per-group) quantization + FRACTIONAL scales (what GGUF Q4_K/Q8 do). A separate scale per small block of B weights adapts to local magnitude; fractional scale (round(w*127/max), deq=q*max/127) avoids the coarse integer-scale error that hurt the earlier per-channel try. Measured on REAL BF16 weights: per-tensor vs block-64 vs block-16 weight relative-L2 error. HONEST: report how much it closes the 0.8%->0.04% gap AND what remains (production 0.04% still needs QAT/4-bit-aware). ours measured. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL

dependencies 2 imports · 0 importers

nx_syscalls.nx nx_gate_verdict.nx nx_block_quant_h2h_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 bq_puts sys_write sys_mmap sys_openat_rd sys_read sys_lseek sys_close bf16_to_q16 sys_exit blockerr absq qdeq rdiv isqrt bq_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 ↻

structs

none

consts

none

functions

11func bq_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
12func bq_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
13func absq(x: i64) -> i64 { if x<0 { return 0-x } return x }
called by 1: blockerr
14func rdiv(a: i64, b: i64) -> i64 { if b==0 { return 0 } if a>=0 { return (a+(b>>1))/b } return 0 - (((0-a)+(b>>1))/b) }
called by 1: qdeq
15func isqrt(v: i64) -> i64 { if v<=0 { return 0 } if v<4 { return 1 } var x: i64=v; var y: i64=(x+1)>>1; var go: i64=1; while go==1 { if y<x { x=y; y=(x+v/x)>>1 } else { go=0 } } return x }
called by 1: blockerr
16func bf16_to_q16(bf: i64) -> i64 { let sign: i64=(bf>>15)&1; let exp: i64=(bf>>7)&255; let mant: i64=bf&127; if exp==0 { return 0 } if exp==255 { if sign==1 { return 0-2147483647 } return 2147483647 } let m: i64=128+mant; let e: i64=exp-118; var v: i64=0; if e>=0 { v=m<<e } else { v=m>>(0-e) } if sign==1 { v=0-v } return v }
called by 1: main
18func qdeq(w: i64, mx: i64) -> i64 { if mx<=0 { return 0 } var q: i64=rdiv(w*127, mx); if q>127 { q=127 } if q<0-127 { q=0-127 } return rdiv(q*mx, 127) }
called by 1: blockerr calls 1: rdiv
21func blockerr(W: *i64, N: i64, B: i64) -> i64
called by 1: main calls 3: absqqdeqisqrt
39func main() -> i64