nx_q8_0_to_f32.nx
buildroot/runtime/nx_q8_0_to_f32.nx
about
nx_q8_0_to_f32.nx -- bits-up Q8_0 dequantization to IEEE 754 binary32.
Q8_0 block layout (per ggml v3 spec; clean-room from public docs):
34 bytes per block of 32 values:
bytes [0..2] d (f16 scale)
bytes [2..34] qs (32 signed i8 quants)
Dequant: value_f32 = d_f32 * (i8)qs[i]
genealogy_id: ggml_q8_0_canon
lineage_id: substrate_q8_0_to_f32_v1
dependencies 5 imports · 6 importers
imports: nx_syscalls.nxnx_tier.nxnx_le.nxnx_f32.nxnx_f32_cvt.nx
imported by: nx_f32_lazy_weight.nxnx_f32_llama_layer_lazy_load.nxnx_gguf_load_f32.nxnx_q8_0_from_f32_kat.nxnx_q8_0_from_q5_0.nxnx_q8_0_simd_gate.nx
structs
| none |
consts
| 19 | const NX_Q8_0_OK: nx_int = 0 |
| 20 | const NX_Q8_0_VPB: i64 = 32 |
| 21 | const NX_Q8_0_BPB: i64 = 34 |
functions
| 23 | func _nx_q8_0_block_to_f32(buf: *u8, base_off: i64, take: i64, |
| 40 | func nx_q8_0_to_f32(buf: *u8, base_off: i64, n_values: i64, |