quantize::quadtree

More...

Classes

	Name
class	quantize::quadtree::QuadtreeEncoder

Detailed Description

Quadtree adaptive block-size encoder for SGFP4 v2.

Per D-01: Full quadtree implementation. Encode tries largest block first
(64x64), measures Laplacian-weighted error, splits into 4 children if error
exceeds configurable threshold, recurses down to min_block_size (default 4x4).

The encoder is designed to be consumed by FP4Exporter. It accepts callable
hooks for FP4_AFFINE and T158_AFFINE fitting, keeping the quadtree logic
independent of the specific encode implementation.

Dual-mode selection per D-04: prefer T158 when t158_error <= (1.0 + delta) * fp4_error.
Per-weight max error guard per RESEARCH.md Pitfall 4: if any individual weight
reconstruction error exceeds 5 * scale, reject T158 and force FP4_AFFINE.

Hysteresis per RESEARCH.md Pitfall 1: if parent block was accepted, require child
error to be <= threshold * 0.8 (20% improvement) before splitting. Allow 10% slack
(accept if error <= threshold * 1.1) when min_block_size not yet reached.

Max recursion depth = 4 levels (64->32->16->8->4). Raises ValueError if exceeded.

---

Updated on 2026-07-25 at 22:56:57 +0000