code wiki / q8
topic: q8
12 modules sharing the q8 name family (derived from the tree's prefix discipline).
The 'q8' topic family in the Nishi sovereign ecosystem focuses on quantization and dequantization operations for efficient tensor processing. Modules like nx_q8_0_dot handle fused integer dot products with dequantization, while nx_q8_0_from_f32_kat ensures roundtrip accuracy between F32 and Q8_0. nx_q8_0_to_f32 provides precise dequantization to IEEE 754 binary32, supporting verification and simulation across the ecosystem.
auto-narrated by the local model from this topic's module headers; links verified against the wiki index.
narrated overview -- maintained by the narration lane, module links verified against this wiki.
| module | description | lines | funcs |
|---|---|---|---|
| nx_q8_0_dot.nx | sovereign Q8_0 dequant + fused integer dot, verified on a REAL Z-Image DiT block. | 69 | 3 |
| nx_q8_0_from_f32.nx | quantize F32 -> Q8_0 blocks (the inverse of | 89 | 3 |
| nx_q8_0_from_f32_kat.nx | roundtrip KAT: F32 -> Q8_0 -> F32 within a quant | 37 | 1 |
| nx_q8_0_from_q5_0.nx | re-quantize a Q5_0 tensor to Q8_0 IN MEMORY. | 58 | 2 |
| nx_q8_0_linear_verify.nx | verify a full Q8_0 linear at REAL DiT scale vs a Python ggml golden. | 83 | 3 |
| nx_q8_0_simd_gate.nx | does the SIMD __f32_i8dot32 fused Q8_0 dequant-dot | 164 | 8 |
| nx_q8_0_to_f32.nx | bits-up Q8_0 dequantization to IEEE 754 binary32. | 54 | 2 |
| nx_q8_coldwarm.nx | decides the post-futex decode-matmul direction: | 96 | 5 |
| nx_q8_matmul_micro.nx | DECISIVE micro-probe of the REAL production Q8_0 | 146 | 5 |
| nx_q8_mix_ab.nx | same-process, same-load A/B of the REAL per-layer block | 128 | 4 |
| nx_q8_scaling.nx | pool SCALING probe: same-process ST vs MT(pool) cold Q8 | 104 | 4 |
| nx_q8_st_bw.nx | SINGLE-THREAD sovereign Q8 matmul bandwidth, to split the | 65 | 4 |