EntroPack Documentation

General-purpose tensor compression for PyTorch, with lossless and adjustable lossy modes.

Usage

• Quick start
  • Installation
  • Direct tensor compression
  • Use Compressed Linear
• Compression configuration
  • Choose a configuration
  • Supported tensor dtypes
  • Parameter conventions
  • CompressionConfig
  • DFloat11Config
  • TileANSConfig
  • LatticeRANSConfig
  • Compression principles
• Tensor compression
  • Compress and restore
  • Inspect stored size
  • Move a compressed tensor
  • Save and load
  • Input requirements
• Compressed Linear usage
  • Replace an existing layer
  • Replace several layers in a model
  • Combine compression with FP8 or INT8 computation
  • Measure storage
  • Save and load a model

API reference

• API reference
  • Tensor encoding and decoding
  • CompressedTensor
  • CompressedLinear
  • CompressedFP8Linear and CompressedINT8Linear
  • Config classes

Compression principles

• DFloat11
  • Encoding
  • Decoding and storage
  • Usage
• Tile-ANS
  • Byte streams and probability tables
  • Tiled encoding and decoding
  • Usage
• EntroPack lattice compression
  • Lattice quantization and integer fields
  • Rate selection and refinement
  • Encoding and reconstruction
  • Usage