Module quantization 

Quantization modules for BitNet.

This module provides:

• Weight quantization: AbsMean -> {-1, 0, +1}
• Activation quantization: Per-token AbsMax -> INT8

Structs

QuantizedActivations

Quantized activations with per-token scales.

TernaryWeight

Ternary weight representation with per-group scales.

Functions

dequantize_activations

Dequantize INT8 activations back to float tensor.

dequantize_weights

Dequantize ternary weights back to float tensor.

quantize_activations

Quantize activations using per-token AbsMax scaling to INT8.

quantize_ste

Apply quantization in a differentiable way using Straight-Through Estimator.

quantize_weights

Quantize a weight tensor to ternary using AbsMean quantization.