Taco: Cross platform LLM based compressor, eliminates floating point (github.com)

Taco, a novel lossless compressor utilizing a language model (GPT-2), has been introduced as a proof-of-concept tool that leverages integer-only arithmetic for cross-platform compatibility. This means that files compressed on one machine can be reliably decompressed on another, ensuring that every CPU and GPU produces identical results. The compression technique involves predicting data sequences, storing only the unpredicted elements with a range coder. While Taco’s current speed is relatively slow—compressing text at about 85-129 KB/s on GPU and 0.3-2 KB/s on CPU—it offers a unique promise: consistent outputs across varied hardware, eliminating discrepancies often found with floating-point calculations. This development is significant for the AI/ML community, as it brings together advancements in machine learning and compression techniques, ensuring that files retain integrity regardless of the device or architecture used. Taco's integer-based model conversion ensures that no floating-point arithmetic affects the results, a crucial factor for applications demanding precision. Furthermore, it stands out against other existing models, such as ts_zip, by separating its design principles, providing well-defined rules for integer operations, and offering multiple independent implementations that yield the same compressed data. This combination of features positions Taco as a unique player in the landscape of compression tools, paving the way for future enhancements and broader applicability.