code wiki / _hdl_build / nx_qlayer_test.nx
nx_qlayer_test.nx
buildroot/runtime/_hdl_build/nx_qlayer_test.nx
about
nx_qlayer_test.nx -- the TEAM does a decode layer END TO END, every organ in its lane:
ENGINEER diagnoses the layer's bottleneck (roofline) -> MEMORY-bound (matvec has no reuse).
BUILDER picks the quality-balanced format under the Council floor -> Q4_K_M class (4-bit).
BUILDER builds the kernel: pre-quantize the matrix, run the hot streamed matvec.
ENGINEER verifies the WHOLE layer output 1:1 vs exact, max per-row error vs the floor.
Exit 0 only if: memory-bound diagnosed, Builder picks 4-bit, layer error within floor, >=3x less
data moved. This is the full session's stack (governor + roofline + quality balance + kernel)
composed into one autonomous decode layer. license_tier: ORIGINAL
dependencies 3 imports · 0 importers
imports: nx_qlayer.nxnx_builder_quant.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
| none |
functions
| 14 | func lt_puts(s: *u8) -> i64 { var n: i64 = 0; while s[n] != (0 as u8) { n = n + 1 } sys_write(1, s, n); return 0 } |
| 15 | func lt_num(v: i64) -> i64 { let bb: *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;k=1}; while m>0 {t[k]=48+(m%10); m=m/10; k=k+1}; var i: i64=0; while i<k {bb[i]=t[k-1-i]; i=i+1}; sys_write(1, bb, k); return 0 } |
| 17 | func main() -> i64 |