quantize::quadtree::QuadtreeEncoder¶

Public Functions¶

	Name
	init(self self, Dict] thresholds[int, Dict[str, float], float ternary_delta, Callable fit_fp4, Callable fit_t158, laplacian laplacian, int min_block_size =4)
List[dict]	encode(self self, np.ndarray superblock_64x64)

Protected Functions¶

	Name
List[dict]	_try_block(self self, np.ndarray superblock, int y, int x, int size, bool parent_accepted, int depth =0)
np.ndarray	_reconstruct(np.ndarray region, dict result, CodeMode mode =CodeMode.FP4_AFFINE)
bool	_t158_has_outlier(np.ndarray region, dict t158_result)

Protected Attributes¶

	Name
	_thresholds
	_ternary_delta
	_fit_fp4
	_fit_t158
	_laplacian
	_min_block_size

Detailed Description¶

class quantize::quadtree::QuadtreeEncoder;

Encode a 64x64 superblock into variable-sized blocks using quadtree recursion.

Constructor args:
    thresholds: Dict mapping block_size (int) -> {"max_mse": float, "max_relative": float}.
                Thresholds per block size for split decisions.
    ternary_delta: D-04 delta value for T158 preference:
                   prefer T158 when t158_err <= (1.0 + delta) * fp4_err.
    min_block_size: Minimum block edge size. Must be in {4, 8, 16, 32, 64}.
                    Default: 4.
    fit_fp4: Callable(region: np.ndarray) -> dict.
             Must return {scale, bias, l2_error, payload, n_weights}.
    fit_t158: Callable(region: np.ndarray) -> dict.
              Must return {scale, bias, l2_error, payload, n_weights}.
    laplacian: LaplacianWeightedError instance for error computation.

Public Functions Documentation¶

function init¶

__init__(
    self self,
    Dict] thresholds[int, Dict[str, float],
    float ternary_delta,
    Callable fit_fp4,
    Callable fit_t158,
    laplacian laplacian,
    int min_block_size =4
)

function encode¶

List[dict] encode(
    self self,
    np.ndarray superblock_64x64
)

Encode a 64x64 superblock into a list of block dicts.

Each dict contains: {y, x, size, mode, payload, header, scale, bias, error}.

Args:
    superblock_64x64: 2D numpy array of shape (64, 64), float32.

Returns:
    List of dict, one per leaf block. Blocks cover the full 64x64 area
    without overlap or gaps.

Protected Functions Documentation¶

function _try_block¶

List[dict] _try_block(
    self self,
    np.ndarray superblock,
    int y,
    int x,
    int size,
    bool parent_accepted,
    int depth =0
)

Recursive quadtree encode. Returns list of block dicts.

Args:
    superblock: The full 64x64 superblock array.
    y: Top-left row of this block.
    x: Top-left column of this block.
    size: Edge size of this block (power of 2).
    parent_accepted: Whether the parent block was accepted
                     (used for hysteresis).
    depth: Current recursion depth.

Returns:
    List of dict, one per leaf block.

function _reconstruct¶

static np.ndarray _reconstruct(
    np.ndarray region,
    dict result,
    CodeMode mode =CodeMode.FP4_AFFINE
)

Reconstruct a region from encode result for error computation.

Args:
    region: 2D array of original weights.
    result: Encode result dict with scale/bias.
    mode: CodeMode.FP4_AFFINE (round-to-nearest nibble codes) or
          CodeMode.T158_AFFINE (ternary threshold codes). Using the
          wrong mode's codebook here systematically distorts the
          dual-mode selection error comparison.

function _t158_has_outlier¶

static bool _t158_has_outlier(
    np.ndarray region,
    dict t158_result
)

Check if any individual weight error exceeds kT158MaxPerWeightErrorScale * scale.```


## Protected Attributes Documentation

### variable _thresholds

```python
_thresholds;

variable _ternary_delta¶

_ternary_delta;

variable _fit_fp4¶

_fit_fp4;

variable _fit_t158¶

_fit_t158;

variable _laplacian¶

_laplacian;

variable _min_block_size¶

_min_block_size;

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