Struct ModelCompression 

pub struct ModelCompression { /* private fields */ }

Model compression

Implementations

impl ModelCompression

pub fn new() -> Self

pub fn initialize(&mut self) -> Result<(), MLError>

pub fn register_algorithm( &mut self, name: &str, algorithm: CompressionAlgorithm, )

Register a compression algorithm under the given name.

pub fn get_algorithm(&self, name: &str) -> Option<&CompressionAlgorithm>

Get a registered compression algorithm by name.

pub fn list_algorithms(&self) -> Vec<String>

List the names of all registered compression algorithms.

pub const fn pruning_mask_bytes(weight_count: usize) -> usize

Number of bytes needed for a packed one-bit-per-weight pruning mask.

pub fn mask_keeps(mask: &[u8], weight_index: usize) -> bool

Return whether weight_index is present in a packed pruning mask.

pub fn quantize_symmetric_int8_into( &mut self, weights: &[f64], out: &mut [i8], ) -> Result<QuantizationReport, MLError>

Run per-tensor symmetric signed-int8 post-training quantization.

out is caller-owned and receives the complete compressed payload. The returned scale is the only side metadata required to dequantize it.

pub fn dequantize_symmetric_int8_into( quantized: &[i8], parameters: QuantizationParameters, out: &mut [f64], ) -> Result<usize, MLError>

Dequantize a symmetric-int8 tensor into caller-owned floating-point storage.

pub fn prune_unstructured_into( &mut self, weights: &[f64], sparsity: f64, mask_out: &mut [u8], values_out: &mut [f64], scratch_indices: &mut [usize], ) -> Result<PruningReport, MLError>

Exact unstructured magnitude pruning.

The smallest-magnitude weights are removed. The result is a real sparse representation: a packed keep-mask and the retained values in original index order. scratch_indices is caller-provided sorting workspace.

pub fn prune_output_channels_into( &mut self, weights: &[f64], rows: usize, columns: usize, sparsity: f64, row_mask_out: &mut [u8], values_out: &mut [f64], score_scratch: &mut [f64], index_scratch: &mut [usize], ) -> Result<PruningReport, MLError>

Structured output-channel pruning for a row-major rows × columns matrix.

Entire rows with the smallest L2 norm are removed and packed contiguously. row_mask_out contains one keep bit per row.

pub fn unpack_pruned_weights_into( mask: &[u8], packed_values: &[f64], out: &mut [f64], ) -> Result<usize, MLError>

Reconstruct an unstructured sparse tensor into caller-owned dense storage.

pub fn distill_linear_student( &mut self, training_engine: &mut TrainingEngine, teacher: &Model, student: &mut Model, training_data: &[f64], hard_targets: Option<&[f64]>, distillation: DistillationConfig, training: &TrainingConfig, target_buffer: &mut [f64], ) -> Result<DistillationReport, MLError>

Distil any inference-supported teacher MLP into the existing single-linear SGD student. Teacher outputs are generated from real forward passes and optionally blended with hard targets in target_buffer.

pub fn record_compression( &mut self, algorithm_name: &str, original_size: usize, compressed_size: usize, accuracy_before: f64, accuracy_after: f64, ) -> Result<(), MLError>

Record the result of a compression operation and update the aggregate quality metrics.

pub fn get_quality_metrics(&self) -> &CompressionQualityMetrics

Access the aggregate compression quality metrics.

pub fn get_compression_statistics(&self) -> &CompressionStatistics

Access byte counts and the ratio from the most recent real compression.

pub fn compression_ratio(&self) -> f64

Return the overall compression ratio recorded so far.