nx_q5_k_ggml_kat.nx
buildroot/runtime/nx_q5_k_ggml_kat.nx
about
nx_q5_k_ggml_kat.nx -- permanent golden KAT for nx_q5_k_to_f32 vs a ggml Q5_K oracle.
Embeds the real 176-byte super-block 0 of blk.0.attn_v (Q5_K) from the Huihui Q4_K_M model (file offset
1050331744) and asserts nx_q5k_block_to_f32 reproduces the ggml golden at POSITION-SENSITIVE points
(element0/32/64 + a non-uniform dot) -- so a wrong nibble/sub-block layout (the Q4_K failure mode) is
caught. Golden from a hand-rolled ggml dequantize_row_q5_K (scratchpad/q5k_golden.py):
element0=-0.004717 (bits 3147470080), element32=-0.049673 (3175839104), element64=-0.022964 (3166445568),
dot(sum y[i]*((i%5)+1))=-2.357033 (bits 3222722976).
license_tier: ORIGINAL
dependencies 7 imports · 0 importers
imports: nx_syscalls.nxnx_tier.nxnx_le.nxnx_f32.nxnx_f32_cvt.nxnx_f32_div.nxnx_q5_k_to_f32.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
| 18 | func kat_hexval(c: i64) -> i64 called by 1: main |
| 24 | func main() -> i64 |