Struct QuantizedLinear 

pub struct QuantizedLinear { /* private fields */ }

A weight matrix quantized to i8, with per-channel or per-tensor scale/zero_point.

Forward pass: dequantize weights on the fly, then f64 matmul.

Layout

weight_q has shape (out_features, in_features) — the same convention as LinearExpert::weights (see moe/expert.rs).

Implementations

impl QuantizedLinear

pub fn from_fp( weight: &Array2<f64>, params: &QuantizationParams, ) -> Result<Self, QuantizationError>

Quantize an existing f64 weight matrix using the provided params.

Only Int8 quantization type is supported. Use crate::quantization::calibrate_linear to produce params.

Errors

• QuantizationError::InvalidParams if qtype != Int8.
• QuantizationError::ShapeMismatch if the weight is not 2-D or the scale/zero_point vectors are the wrong length for PerChannel.

pub fn with_bias(self, bias: Array1<f64>) -> Result<Self, QuantizationError>

Attach a bias vector of length out_features.

Errors

QuantizationError::ShapeMismatch if bias.len() != out_features.

pub fn forward(&self, x: &Array2<f64>) -> Array2<f64>

Dequantize and run matmul.

Input x must have shape [batch, in_features]. Output has shape [batch, out_features].

fp = (q - zero_point[c]) * scale[c] per element, where c is the output channel (row) index when granularity == PerChannel, or 0 for PerTensor.

pub fn dequantize(&self) -> Array2<f64>

Dequantize the stored i8 weights back to f64.

For PerTensor: all elements use scale[0] / zero_point[0]. For PerChannel: each row c uses scale[c] / zero_point[c].

pub fn out_features(&self) -> usize

Return the output feature dimension.

pub fn in_features(&self) -> usize

Return the input feature dimension.

pub fn granularity(&self) -> QuantizationGranularity

Return the quantization granularity.

pub fn scales(&self) -> &[f64]

Return the scale factors (length 1 for PerTensor, out_features for PerChannel).