Struct MinMaxQuantizer 

pub struct MinMaxQuantizer { /* private fields */ }

Calibrates quantization parameters using tensor min/max statistics.

Implementations

impl MinMaxQuantizer

pub fn new( bits: u32, scheme: QuantScheme, granularity: QuantGranularity, ) -> Self

Create a new quantizer.

Panics

Panics if bits is 0 or > 16.

pub fn int8_symmetric() -> Self

Standard INT8 symmetric per-tensor quantizer.

pub fn int4_per_group(group_size: usize) -> Self

Standard INT4 symmetric per-group quantizer (group = 128, as in GGML).

pub fn calibrate(&self, tensor: &[f32]) -> QuantResult<QuantParams>

Calibrate parameters from a flat tensor.

For PerChannel, the tensor is assumed to be in row-major layout with n_channels rows of length tensor.len() / n_channels.

Errors

• QuantError::EmptyInput — if tensor is empty.
• QuantError::GroupSizeMismatch — if PerGroup size does not divide.
• QuantError::DimensionMismatch — if PerChannel axis is inconsistent.

pub fn calibrate_2d( &self, tensor: &[f32], rows: usize, cols: usize, ) -> QuantResult<QuantParams>

Calibrate from a 2-D tensor (rows = channels).

Returns one (scale, zp) per row.

Errors

• QuantError::EmptyInput if rows == 0.
• QuantError::DimensionMismatch if cols == 0.

pub fn quantize( &self, tensor: &[f32], params: &QuantParams, ) -> QuantResult<Vec<i32>>

Quantize a flat tensor given pre-computed params (PerTensor mode).

Returns Vec<i32> of integer codes.

Errors

• QuantError::InvalidScale if params.scales[0] <= 0.

pub fn quantize_grouped( &self, tensor: &[f32], params: &QuantParams, group_size: usize, ) -> QuantResult<Vec<i32>>

Quantize using per-group params.

Errors

• QuantError::GroupSizeMismatch if tensor size is not divisible by group_size.

pub fn dequantize(&self, codes: &[i32], params: &QuantParams) -> Vec<f32>

Dequantize integer codes back to f32.

pub fn dequantize_grouped( &self, codes: &[i32], params: &QuantParams, group_size: usize, ) -> Vec<f32>

Dequantize per-group codes.

Trait Implementations

impl Clone for MinMaxQuantizer

fn clone(&self) -> MinMaxQuantizer

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for MinMaxQuantizer

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

Formats the value using the given formatter.