Struct BitLinear 

pub struct BitLinear { /* private fields */ }

BitLinear layer with ternary weights and INT8 activations.

This is a drop-in replacement for candle_nn::Linear that uses:

• Ternary weights {-1, 0, +1} with per-group scales
• INT8 activation quantization with per-token scales

Example

use bitnet_rs::{BitLinear, BitNetConfig};
use candle_core::{Device, Tensor};

let device = Device::Cpu;
let config = BitNetConfig::default();

// Create from existing weights
let weight = Tensor::randn(0.0f32, 1.0, (512, 256), &device)?;
let layer = BitLinear::from_weight(&weight, None, &config)?;

// Forward pass
let input = Tensor::randn(0.0f32, 1.0, (4, 256), &device)?;
let output = layer.forward(&input)?;

Implementations

impl BitLinear

pub fn from_weight( weight: &Tensor, bias: Option<&Tensor>, config: &BitNetConfig, ) -> Result<Self>

Create a new BitLinear layer from a weight tensor.

Arguments

• weight - Weight tensor [out_features, in_features]
• bias - Optional bias tensor [out_features]
• config - BitNet configuration

Errors

Returns error if weight quantization fails.

pub fn from_quantized( weight: TernaryWeight, bias: Option<Tensor>, config: BitNetConfig, device: Device, ) -> Self

Create a new BitLinear layer from pre-quantized weights.

Arguments

• weight - Pre-quantized ternary weight
• bias - Optional bias tensor
• config - BitNet configuration
• device - Device for operations

pub fn in_features(&self) -> usize

Get the input features dimension.

pub fn out_features(&self) -> usize

Get the output features dimension.

pub const fn quantized_weight(&self) -> &TernaryWeight

Get reference to the quantized weights.

pub const fn bias(&self) -> Option<&Tensor>

Get reference to the bias.

pub const fn config(&self) -> &BitNetConfig

Get reference to the configuration.

pub const fn device(&self) -> &Device

Get the device.

pub fn sparsity(&self) -> f32

Get the weight sparsity.

pub fn compression_ratio(&self) -> f32

Get the compression ratio.

pub fn forward_quantized(&self, input: &Tensor) -> Result<Tensor>

Forward pass with explicit activation quantization.

This method:

1. Quantizes input activations to INT8
2. Dequantizes weights for matmul (or uses optimized kernel)
3. Performs the linear transformation
4. Adds bias if present

Arguments

• input - Input tensor [batch, …, in_features]

Errors

Returns error if forward pass fails.

Trait Implementations

impl Debug for BitLinear

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Module for BitLinear

fn forward(&self, input: &Tensor) -> Result<Tensor>