src/agiws_neural_quant/nf4/__init__.py

"""NF4 (NormalFloat 4-bit) quantization primitives — QLoRA paper format.

Pure primitive subpackage: tensor-level quantization functions only. No
nn.Module wrapper classes — the unified QuantizedModule (base.py) is the
single place for quantized layer wrappers.

Exports:
  NF4_LUT                   — 16 fixed non-uniform levels (const tensor)
  quantize_nf4              — quantize weight tensor to NF4 indices + per-group scale
  dequantize_nf4            — reconstruct weight from indices + scale
  pack_nf4 / unpack_nf4     — pack two 4-bit indices per byte
  double_quantize_scales_2d — quantize scales to int8 + per-block fp32 scale
  dequantize_scales_2d       — reconstruct scales from int8 + block_scale
"""

from agiws_neural_quant.nf4.nf4 import (
    NF4_LUT, quantize_nf4, dequantize_nf4, pack_nf4, unpack_nf4,
)
from agiws_neural_quant.nf4.double_quant import (
    double_quantize_scales_2d, dequantize_scales_2d,
)

__all__ = [
    "NF4_LUT",
    "quantize_nf4",
    "dequantize_nf4",
    "pack_nf4",
    "unpack_nf4",
    "double_quantize_scales_2d",
    "dequantize_scales_2d",
]