Fujitsu One Compression (OneComp)

A Python package for LLM compression.
Full documentation: https://FujitsuResearch.github.io/OneCompression/

Features

Quantization Error Propagation (QEP): Post-training quantization with error propagation to subsequent layers (Arai & Ichikawa, NeurIPS 2025)
vLLM Plugin Integration: Serve quantized models with vLLM via built-in DBF and Mixed-GPTQ plugins
AutoBit: Mixed-precision quantization with ILP-based bitwidth assignment and automatic VRAM estimation
JointQ: Joint quantization optimizing weight assignments and scale parameters simultaneously
LoRA SFT Post-Process: Fine-tune quantized models with LoRA for accuracy recovery or knowledge injection
Rotation Preprocessing: SpinQuant/OstQuant-based rotation preprocessing (Llama, Qwen3)

Architecture	Verified Models	Status
Llama	TinyLlama, Llama-2, Llama-3	Verified
Qwen3	Qwen3-0.6B ~ 32B	Verified

pip install onecomp