nx_q6_k_to_f32.nx

nx_q6_k_to_f32.nx -- bits-up Q6_K dequantization to IEEE 754 binary32. Q6_K super-block layout (per ggml v3 spec; clean-room). 210 bytes per super-block of 256 values: bytes [0..128] ql 128 bytes, 4-bit packed (low/high nibble) bytes [128..192] qh 64 bytes, 2-bit packed (4 pairs per byte) bytes [192..208] scales 16 i8 sub-block scales bytes [208..210] d f16 super-scale Dequant (per ggml dequantize_row_q6_K): Process super-block in 2 chunks of 128 values each (c in 0..1). Each chunk consumes 64 bytes of ql, 32 bytes of qh, 8 scales. Inside a chunk, l loops 0..31, producing 4 output positions per l: y[c*128 + l ] = d * sc[c*8 + (l/16) + 0] * q1 y[c*128 + l + 32] = d * sc[c*8 + (l/16) + 2] * q2 y[c*128 + l + 64] = d * sc[c*8 + (l/16) + 4] * q3 y[c*128 + l + 96] = d * sc[c*8 + (l/16) + 6] * q4 where: q1 = (ql[c*64 + l ] & 0x0F) | ((qh[c*32 + l] >> 0) & 3) << 4 - 32 q2 = (ql[c*64 + l + 32] & 0x0F) | ((qh[c*32 + l] >> 2) & 3) << 4 - 32 q3 = (ql[c*64 + l ] >> 4 ) | ((qh[c*32 + l] >> 4) & 3) << 4 - 32 q4 = (ql[c*64 + l + 32] >> 4 ) | ((qh[c*32 + l] >> 6) & 3) << 4 - 32 (Rewrite 2026-05-21 after live-fire on real Qwen2.5 showed the naive per-position algorithm doesn't match ggml's interleaving.) genealogy_id: ggml_q6_k_canon lineage_id: substrate_q6_k_to_f32_v1