kwaai-compression

Compression utilities for KwaaiNet distributed ML operations.

This crate provides compression algorithms for efficient gradient and tensor transfer, including:

• Blockwise 8-bit Quantization: ~4x compression with minimal accuracy loss
• Sparse Gradient Compression: Top-K selection for bandwidth efficiency
• Delta Encoding: Only transfer changes

Example

use kwaai_compression::{BlockwiseQuantizer, CompressedData, Compressor};
use candle_core::Tensor;

let quantizer = BlockwiseQuantizer::new(64);
let tensor = Tensor::zeros(&[1024], candle_core::DType::F32, &candle_core::Device::Cpu)?;

let compressed = quantizer.compress(&tensor)?;
println!("Compression ratio: {:.2}x", compressed.compression_ratio());

let decompressed = quantizer.decompress(&compressed)?;
# let _ = decompressed;
# Ok::<(), Box<dyn std::error::Error>>(())