onecomp 1.0.0

Python package for LLM compression

Fujitsu One Compression

Fujitsu One Compression (OneComp) is a Python package for LLM compression.

📖 Documentation

Full documentation is available at https://FujitsuResearch.github.io/OneCompression/.

📦 Features

• Quantization Error Propagation (QEP): A post-training quantization method that corrects quantization errors by propagating them to subsequent layers, improving the accuracy of quantized LLMs. See Arai & Ichikawa, NeurIPS 2025 for details. The original reference implementation is available at FujitsuResearch/qep.
• vLLM Plugin Integration: Serve OneComp-quantized models with vLLM via built-in plugins for DBF and Mixed-GPTQ quantization methods.
• AutoBit: Mixed-precision quantization with ILP-based bitwidth assignment. Automatically estimates the target bitwidth from available VRAM and assigns per-layer bitwidths to minimize quantization error under the memory budget.
• JointQ: Joint quantization method that optimizes weight assignments and scale parameters simultaneously for improved quantization accuracy. Supports group-wise quantization (e.g., 4-bit, groupsize=128).
• LoRA SFT Post-Process: Fine-tune quantized models with LoRA adapters for accuracy recovery or domain-specific knowledge injection. Supports SFT loss, teacher distillation, and intermediate block alignment.
• Rotation Preprocessing: SpinQuant/OstQuant-based rotation preprocessing that reduces quantization error by learning optimal rotation matrices before quantization. Rotation/scaling matrices are absorbed into model weights, with online Hadamard hooks automatically registered at load time. Supports Llama and Qwen3 architectures.
• (TBD)

🤖 Supported Models

OneComp has been verified with the following model architectures. Other Hugging Face-compatible models may work but are currently untested.

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

Note: Support for additional architectures is planned. Contributions and test reports are welcome.

🔧 Installation

for users (pip)

1. Install PyTorch

Please install the appropriate version of PyTorch.

✅ CPU-only

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

✅ CUDA-enabled

Choose the appropriate CUDA version for your system:

CUDA Version	Installation Command
CUDA 11.8	pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
CUDA 12.1	pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
CUDA 12.4	pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
CUDA 12.6	pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
CUDA 12.8	pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

Check your CUDA version:

nvcc --version

or

nvidia-smi

Verify PyTorch GPU support:

import torch
print(torch.cuda.is_available())

2. Install onecomp

Once PyTorch is installed, you can install onecomp:

pip install onecomp

for developers (uv : recommended)

Install uv

uv is a fast Python package and project manager written in Rust. It offers a drop-in replacement for pip and pip-tools while also managing virtual environments and Python installations. With its Rust-based dependency resolver and the uv.lock lockfile, uv provides deterministic and reproducible environments across development machines and CI pipelines.

# install uv (for macOS or Linux)
curl -LsSf https://astral.sh/uv/install.sh | sh

git clone <git repository URL>
cd OneCompression
uv sync --extra cu128 --extra dev

The uv sync command creates a Python virtual environment and installs all dependent libraries.

The --extra cu128 option installs the CUDA-enabled version of PyTorch (along with torchvision from the same CUDA index). Replace cu128 with the appropriate variant for your environment: cpu, cu118, cu121, cu124, cu126, or cu128. PyTorch will be automatically downloaded by uv, so you do not need to install it beforehand.

Adding --extra dev installs additional packages for development.

To use vLLM for serving quantized models, add --extra vllm:

uv sync --extra cu128 --extra dev --extra vllm

Note: --extra vllm may take a long time on the first run if a pre-built xformers wheel is not available for your Python/CUDA combination (e.g. Python 3.13). Using Python 3.12 typically avoids this.

Running commands (uv environment)

In the environment created by uv sync, you can run commands in two ways:

Option 1: Use uv run (no activation needed)

uv run pytest tests/ -v
uv run python example/example1.py
uv run black --check onecomp/

Option 2: Activate the virtual environment (traditional approach)

source .venv/bin/activate
pytest tests/ -v
python example/example1.py
black --check onecomp/

for developers (pip)

git clone <git repository URL>
cd OneCompression

# First, install PyTorch with CUDA support for your environment
pip install torch --index-url https://download.pytorch.org/whl/cu128
# Then install onecomp with development dependencies
pip install -e ".[dev]"

Replace cu128 with the appropriate variant for your environment: cpu, cu118, cu121, cu124, cu126, or cu128.

Building Documentation Locally

uv sync --extra cu128 --extra dev --extra docs
uv run mkdocs serve

Then open http://127.0.0.1:8000 in your browser.

🚀 Examples

Category	Script	Description
Quantization	example_gptq.py	GPTQ quantization
	example_qep_gptq.py	GPTQ + QEP (error propagation)
	example_jointq.py	JointQ quantization
	example_autobit.py	AutoBit mixed-precision quantization
	example_auto_run.py	AutoBit with automatic VRAM estimation
Save / Load	example_save_load.py	Save and load quantized models
Rotation Preprocessing	example_llama_preprocess_rtn.py	Rotation preprocessing + RTN (TinyLlama)
	example_preprocess_save_load.py	Save and load rotation-preprocessed quantized models
Post-Process	example_lora_sft.py	LoRA SFT post-quantization fine-tuning
	example_lora_sft_knowledge.py	LoRA SFT knowledge injection
vLLM	example_gptq_vllm_inference.py	GPTQ + QEP quantization and vLLM inference
	example_autobit_vllm_inference.py	AutoBit quantization and vLLM inference

🔌 vLLM Inference

OneComp-quantized models can be served with vLLM via built-in plugins (DBF, Mixed-GPTQ).

# uv users
uv sync --extra cu128 --extra vllm

# pip users
pip install vllm

See the vLLM Inference guide for details.

📄 License

See LICENSE for more details.

Citation

OneComp technical report (coming soon on ArXiv):

@misc{onecomp2026,
  title={TBD},
  author={TBD},
  year={2026},
  note={arXiv preprint coming soon}
}

QEP (Quantization Error Propagation):

@inproceedings{
arai2025quantization,
title={Quantization Error Propagation: Revisiting Layer-Wise Post-Training Quantization},
author={Yamato Arai and Yuma Ichikawa},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=a3l3K9khbL}
}