AI BriefingFujitsu One Compression (3 min read)
Key point
Fujitsu has released OneComp, an open-source Python library that improves the quantization accuracy of LLMs.
Details
Fujitsu One Compression (OneComp) is an open-source Python library for Post-Training Quantization (PTQ) of large language models (LLMs). It supports state-of-the-art algorithms such as GPTQ, DBF, and RTN, and in particular implements Quantization Error Propagation (QEP), a new method presented at NeurIPS 2025.
Key features are as follows:
- QEP (Quantization Error Propagation): Propagates quantization error to subsequent layers to correct the model's accuracy.
- vLLM plugin integration: Enables DBF and Mixed-GPTQ quantized models to be served immediately on vLLM.
- AutoBit: Provides mixed-precision quantization that automatically allocates per-layer bit widths based on available VRAM.
- JointQ: Jointly optimizes weight allocation and scale parameters to improve quantization accuracy.
- LoRA SFT post-processing: Uses LoRA adapters to recover the accuracy of quantized models or inject domain-specific knowledge.
- Rotation preprocessing: Learns rotation matrices based on SpinQuant/OstQuant to minimize quantization error.
It has currently been validated on Llama (TinyLlama, Llama-2, Llama-3) and Qwen3 (0.6B–32B) architectures. With just a single line of code, users can automate the entire process from model quantization to evaluation (Perplexity and Accuracy) and model saving.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.