| """ | |
| BitLinear: Ultra-Low-Precision Linear Layers for PyTorch | |
| A PyTorch extension implementing 1.58-bit ternary linear layers for extreme | |
| compression in neural networks, particularly Transformers. | |
| """ | |
| __version__ = "0.1.0" | |
| from .layers import BitLinear, MultiTernaryLinear, convert_linear_to_bitlinear | |
| from .functional import bitlinear_python, greedy_ternary_decomposition | |
| from .quantization import ( | |
| ternary_quantize, | |
| absmax_scale, | |
| weight_to_ternary, | |
| ) | |
| from .packing import ( | |
| pack_ternary_base3, | |
| unpack_ternary_base3, | |
| estimate_memory_savings, | |
| ) | |
| __all__ = [ | |
| "BitLinear", | |
| "MultiTernaryLinear", | |
| "convert_linear_to_bitlinear", | |
| "bitlinear_python", | |
| "greedy_ternary_decomposition", | |
| "ternary_quantize", | |
| "absmax_scale", | |
| "weight_to_ternary", | |
| "pack_ternary_base3", | |
| "unpack_ternary_base3", | |
| "estimate_memory_savings", | |
| ] | |