Struct SmoothQuantMigrator 

pub struct SmoothQuantMigrator {
    pub config: SmoothQuantConfig,
}

Applies per-channel scaling to balance quantization difficulty.

The migrator operates on linear layers:

• Activations X — shape [n_tokens, n_channels]
• Weights W — shape [n_out, n_channels] (transposed: Y = X W^T)

Fields

config: SmoothQuantConfig

Migration configuration.

Implementations

impl SmoothQuantMigrator

pub fn new(alpha: f32) -> Self

Create a migrator with the given alpha.

pub fn compute_migration_scales( &self, act_max: &[f32], weight_max: &[f32], ) -> QuantResult<Vec<f32>>

Compute per-channel migration scales from pre-aggregated statistics.

Parameters

• act_max — per-channel max absolute value of activations (length n_ch).
• weight_max — per-channel (column) max absolute value of weights (length n_ch).

Returns

Scale vector s of length n_ch where s[j] = act_max[j]^alpha / weight_max[j]^(1−alpha).

Errors

• QuantError::DimensionMismatch — act_max and weight_max differ in length.
• QuantError::EmptyInput — either slice is empty.

pub fn compute_act_stats( acts: &[f32], n_tokens: usize, n_channels: usize, ) -> QuantResult<Vec<f32>>

Compute per-channel max absolute values from an activation tensor.

Parameters

• acts — row-major activation matrix [n_tokens, n_channels].
• n_tokens — number of tokens (rows).
• n_channels — hidden dimension (columns).

Errors

• QuantError::DimensionMismatch — slice length ≠ n_tokens × n_channels.
• QuantError::EmptyInput — either dimension is 0.

pub fn compute_weight_stats( weights: &[f32], n_out: usize, n_channels: usize, ) -> QuantResult<Vec<f32>>

Compute per-column (input-channel) max absolute values from a weight matrix.

Parameters

• weights — row-major weight matrix [n_out, n_channels].
• n_out — number of output features (rows).
• n_channels — number of input features (columns).

Errors

• QuantError::DimensionMismatch — slice length ≠ n_out × n_channels.
• QuantError::EmptyInput — either dimension is 0.

pub fn smooth_activations( acts: &mut [f32], scales: &[f32], n_tokens: usize, n_channels: usize, ) -> QuantResult<()>

Divide each activation channel j by scales[j] in-place.

Errors

• QuantError::DimensionMismatch — inconsistent lengths.

pub fn smooth_weights( weights: &mut [f32], scales: &[f32], n_out: usize, n_channels: usize, ) -> QuantResult<()>

Multiply each weight column j (input channel) by scales[j] in-place.

Weights are assumed to have shape [n_out, n_channels].

Errors

• QuantError::DimensionMismatch — inconsistent lengths.

pub fn smooth_layer( &self, acts: &mut [f32], weights: &mut [f32], n_tokens: usize, n_channels: usize, n_out: usize, ) -> QuantResult<Vec<f32>>

Smooth a complete linear layer: compute scales, apply to activations and weights.

Parameters

• acts — mutable activation matrix [n_tokens, n_channels].
• weights — mutable weight matrix [n_out, n_channels].
• n_tokens — token (batch) dimension.
• n_channels — input feature dimension.
• n_out — output feature dimension.

Returns

The per-channel migration scales used (length n_channels).

Errors

Propagates all dimension and empty-input errors from sub-operations.

Trait Implementations

impl Clone for SmoothQuantMigrator

fn clone(&self) -> SmoothQuantMigrator

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for SmoothQuantMigrator

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

Formats the value using the given formatter.

impl Copy for SmoothQuantMigrator