BPDQ is a post-training quantization method presented in the paper BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models. It constructs a variable quantization grid via bit-plane decomposition and scalar coefficients, allowing for high-fidelity performance even in low-bit regimes (2–3 bits) where conventional PTQ typically degrades.

Resources

Paper: BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models
GitHub: KingdalfGoodman/BPDQ

Note on size

The figure reported on Hugging Face (24.2 GB for Qwen2.5-72B-BPD2-G256) is the checkpoint file size on disk, measured in GB (10⁹ bytes). The 22.69 GB value cited in our paper refers to the peak VRAM during inference on a single RTX 3090 — a different measurement. Note also that Linux tools such as du -h display sizes in GiB (2³⁰ bytes), so 24.2 GB ≈ 22.54 GiB on disk; this is a unit convention, not a discrepancy.

Citation

If you find BPDQ useful in your research, please cite:

bibtex

@article{chen2026bpdq,
  title={BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models},
  author={Chen, Junyu and Li, Jungang and Xiong, Jing and Wang, Wenjie and Yang, Qingyao and Xiao, He and Li, Zhen and Wu, Taiqiang and Chen, Mengzhao and Peng, Zhen and others},
  journal={arXiv preprint arXiv:2602.04163},
  year={2026}
}

goodman20241017/Qwen2.5-72B-BPD2-G256

Qwen2.5-72B-BPD2-G256

Resources

Note on size

Citation

goodman20241017/
Qwen2.5-72B-BPD2-G256