nx_quant_block.nx

nx_quant_block.nx -- 4-bit block quantization on i64 substrate. Algo-led memory win per the min-hardware-floor cardinal. Modern LLMs and diffusion models DO NOT fit on small hardware at f32 precision; q4_K-style quantization is what makes Llama-3 70B run on a 24 GB consumer GPU and what makes Z-Image fit on a 16 GB card. On a $50 SBC with 4-8 GB total RAM, quantization is the difference between "model loads" and "model OOMs." v1 shape: q4_0-style symmetric quantization on i64 substrate. * Block size: 32 values per block (standard) * Per-block scale: one i64 (Q-format multiplier) * Values: 4-bit signed nibbles in [-7, 7]; packed two per byte * Storage per block: 8 bytes (scale) + 16 bytes (nibbles) = 24 bytes * Dense storage: 32 i64 = 256 bytes * Compression: 256 / 24 ~ 10.67x When the compiler-I2 f-types land, this slots in q4_K (k-means per- group scales) + q5_K + q8_0 without API change. The block layout is the contract; the per-block calibration is the dtype. Algorithmic precision note (the audit-honest answer): * Per-value max rounding error = scale/2 = max_abs/14 * Relative error per value ≤ 1/14 ≈ 7.1% in the worst case * Mean error across a uniform-distribution block is ~1/28 ≈ 3.6% * Oracle round-trip uses eps_q10 = 102 (10%) -- well above mean error, below worst-case, so failures here are real corruption not normal quantization noise. genealogy_id: q4_0_ggml_2023 + q4_K_ggml_2024 + lloyd_max_1957 + jacob_kligys_2018_8bit_quant lineage_id: substrate_quant_block_v1