Fujitsu One Compression
A Python package that automatically compresses large language models, detecting available GPU memory, choosing a bit-width per layer, quantizing with error correction, and evaluating the result in one command.
OneComp (Fujitsu One Compression) runs model compression with a single command or a one-line function call: it detects the GPU's available memory, picks the best bit-width per layer, quantizes the model while propagating quantization error to later layers, evaluates the result, and saves it, fully automatically.
Its methods include Quantization Error Propagation for post-training accuracy correction, Layer-Projected Coordinate Descent for extending layer-wise quantization to arbitrary submodules, AutoBit for integer-linear-programming-based mixed-precision bit assignment under a memory budget, JointQ for jointly optimizing weight assignment and scale parameters, and a multi-envelope binary factorization quantizer for aggressive sub-1-bit compression. Quantized models can then be served through vLLM plugins, paired with Open WebUI for a chat interface.