code wiki / (root) / nx_q4k_q24_gate.nx

nx_q4k_q24_gate.nx

buildroot/runtime/nx_q4k_q24_gate.nx

1454 B28 linesdepth 6pulls 8 transitivereach 0 importersview sourcekind gate/prooftopic q4k
docsdependenciesstructsconstsfunctions

about

nx_q4k_q24_gate.nx -- prove the exact Q24 super-scale decode on the REAL Q4_K super-scales. Guards Fix A+B of the Q4_K dequant tech-debt: `_gguf_f16_to_q24` must decode the real blk.0.attn_q super-scales (d_raw=0x062e, dmin_raw=0x1078) with ZERO loss and reproduce the ggml-correct element0. Golden (from the ggml forensic oracle): d->1582, dmin->9152, element0 = d*sc0*q0 - dmin*m0 = 1582*33*7 - 9152*37 = 26818 (Q24) = +0.0015985. Q10 gave 0 (underflow), Q14 gave -65 (sign flip). license_tier: ORIGINAL

dependencies 4 imports · 0 importers

nx_syscalls.nx nx_tier.nx nx_le.nx nx_gguf_load.nx nx_q4k_q24_gate.nx

imports: nx_syscalls.nxnx_tier.nxnx_le.nxnx_gguf_load.nx

imported by: nobody (leaf or entry point)

call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown

main _gguf_f16_to_q24

structs

none

consts

none

functions

13func main() -> i64