Struct VSAGradientCompressor 

pub struct VSAGradientCompressor { /* private fields */ }

Compress gradients using Vector Symbolic Architecture.

This compressor uses proper VSA operations:

1. Project each gradient to hyperdimensional space
2. Bind each projected gradient with a unique random key
3. Bundle all bound vectors into a single superposition
4. Unbind during decompression to extract individual gradients

The bundled representation achieves significant compression while the bind/unbind operations enable accurate reconstruction due to the quasi-orthogonality of random keys in high dimensions.

Example

use vsa_optim_rs::vsa::VSAGradientCompressor;
use vsa_optim_rs::VSAConfig;

let compressor = VSAGradientCompressor::new(1_000_000, VSAConfig::default());

// After computing gradients
let (compressed, metadata) = compressor.compress(&gradients)?;
let reconstructed = compressor.decompress(&compressed, &metadata)?;

Implementations

impl VSAGradientCompressor

pub fn new(param_count: usize, config: VSAConfig) -> VSAGradientCompressor

Create a new VSA gradient compressor.

Arguments

• param_count - Total number of model parameters
• config - VSA configuration

pub const fn compressed_dim(&self) -> usize

Get the hypervector dimension.

pub fn compress( &mut self, gradients: &HashMap<String, Tensor>, ) -> Result<(PackedTritVec, HashMap<String, GradientMetadata>), OptimError>

Compress gradients to bundled hyperdimensional representation.

Algorithm

1. Project each gradient to hypervector space
2. Quantize to ternary {-1, 0, +1}
3. Bind with unique random key
4. Bundle all bound vectors via element-wise sum

Arguments

• gradients - Map of parameter names to gradient tensors

Returns

Tuple of (bundled hypervector, metadata for reconstruction).

pub fn decompress( &mut self, bundled: &PackedTritVec, metadata: &HashMap<String, GradientMetadata>, ) -> Result<HashMap<String, Tensor>, OptimError>

Decompress gradients from bundled hypervector.

Algorithm

For each gradient:

1. Unbind with the gradient’s key to extract from bundle
2. Inverse project back to gradient space
3. Apply stored scale factor

Arguments

• bundled - Bundled hypervector from compress
• metadata - Metadata from compression

Returns

Map of reconstructed gradients.

pub fn get_compression_stats(&self) -> CompressionStats

Get compression statistics.

pub fn clear_cache(&mut self)

Clear caches to free memory.