---
license: llama2
base_model: meta-llama/Llama-2-70b-hf
tags:
- quantization
- harp
- 2-bit
- extreme-quantization
library_name: transformers
---

# Llama-2-70b-QuIP-HARP-2Bit

2-bit quantized version of [meta-llama/Llama-2-70b-hf](https://huggingface.co/meta-llama/Llama-2-70b-hf), produced with **HARP** (Hadamard-Preconditioned Adaptive Rotations), a learnable structured orthogonal incoherence processor for extreme LLM quantization.

- **Paper:** [HARP: Hadamard-Preconditioned Adaptive Rotation Processor for Extreme LLM Quantization](https://arxiv.org/abs/2605.29843) (EMNLP 2026)
- **Code:** [brain-lab-research/HARP](https://github.com/brain-lab-research/HARP)
- **Backend:** QuIP# (`harp_quip/`), HARP as a drop-in replacement for the fixed randomized Hadamard/RHT incoherence-processing stage
- **Collection:** [brain-lab/harp-quantized-models](https://huggingface.co/collections/brain-lab/harp-quantized-models)

## What is HARP?

Ultra-low-bit PTQ methods such as QuIP# rely on a randomized Hadamard transform (RHT) to make weights more incoherent before quantization. HARP replaces this fixed transform with a **learned** structured orthogonal processor. It starts from a Hadamard-style initialization and is fit only on calibration data, so no model retraining is needed. Stored in int8, it adds negligible bits-per-parameter overhead. Everything else in the QuIP# pipeline (codebooks, LDLQ solver, HF export, inference) stays unchanged.

## Results

WikiText2 / C4 perplexity at ~2 bits, context length 4096 (BPP includes HARP processor storage):

| Method | BPP↓ | Wiki2↓ | C4↓ |
|---|---|---|---|
| QuIP# (fixed RHT) | 2.00 | 4.16 | 6.01 |
| **QuIP# + HARP (this model)** | **2.02** | **4.01** | **5.82** |

## Usage

This checkpoint uses QuIP#-style packed weights and requires the `harp_quip` inference code (a fork of QuIP#) rather than plain `transformers.AutoModelForCausalLM`:

```bash
git clone https://github.com/brain-lab-research/HARP
cd HARP/harp_quip
pip install -r requirements.txt
cd quiptools && python setup.py install && cd ..
```

```python
from lib.utils.unsafe_import import model_from_hf_path

model, model_str = model_from_hf_path("brain-lab/Llama-2-70b-QuIP-HARP-2Bit")
```

See `harp_quip/eval/eval_ppl.py` and `harp_quip/eval/eval_zeroshot.py` for perplexity / zero-shot evaluation scripts, and the main [repo README](https://github.com/brain-lab-research/HARP) for the full quick-start guide.

## License

Governed by the [Llama 2 license](https://ai.meta.com/llama/license/). This is a derivative of Meta's Llama 2 70B; use is subject to Meta's Llama 2 Community License Agreement.

## Citation

```bibtex
@article{zagitov2026harp,
  title={HARP: Hadamard-Preconditioned Adaptive Rotation Processor for Extreme LLM Quantization},
  author={Artur Zagitov and Gleb Molodtsov and Aleksandr Beznosikov},
  year={2026},
  eprint={2605.29843},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2605.29843},
}
```