Skip to main content

Module layer

bitnet_quantize

Module layer

Expand description

Neural network layers for BitNet.

This module provides:

BitLinear: Drop-in replacement for nn::Linear with ternary weights
Straight-Through Estimator for training

Structs§

BitLinear: BitLinear layer with ternary weights and INT8 activations.

Functions§

int8_ste: Apply INT8 STE (quantize to [-127, 127]).
ste_backward: Compute STE backward pass (identity gradient).
ste_forward: Apply STE forward pass (quantize then dequantize).
ternary_ste: Apply ternary STE (quantize to {-1, 0, +1}).