QERA: an Analytical Framework for Quantization Error Reconstruction

This is the official implementation of the ICLR'25 paper "QERA: an Analytical Framework for Quantization Error Reconstruction".

Env Setup

git clone git@github.com:ChengZhang-98/QERA.git
cd QERA
git submodule update --init
conda env create -f environment.yml
conda activate qera
pip install -r requirements.txt
pip install -e .

Entry Points

In the source code and scripts, we use the following abbreviations for the low-rank term types:

• If --disable-qera is set, no low-rank terms are used, i.e., weight-only quantization.
• Else:
  • identity: Truncated SVD on the quantized weight matrix, i.e., ZeroQuant-V2
  • lqer: The heuristic method proposed in LQER paper
  • diag: QERA-approx in our paper.
  • exact: QERA-exact in our paper.

Post-Training Quantization

• ptq_bf16_baseline.py evaluates BF16 baseline.
• ptq_q_baseline.py evaluates PTQ baseline.
• ptq_pipeline.py runs data calibration (if needed), computes low-rank terms, and evaluates the quantized model.
• ptq_pipeline_chunked.py runs data calibration (if needed), and computes low-rank terms for a chunk of layers. This is useful for large models. If all chunks (layers) are computed, this script also triggers the evaluation of the quantized model.
  • chunk_checker.py checks the completion of the chunks (optional).

Quantized LoRA Fine-Tuning

• adapt_and_save.py run data calibration, quantizes the model, computes the initial value of the low-rank terms, and saves the quantized model + low-rank terms.
• glue_train.py fine-tunes the qLoRA-adapted model with low-rank terms on GLUE tasks.
• clm_train.py fine-tunes the qLoRA-adapted model with low-rank terms on WikiText2.
• gsm8k_train.py fine-tunes the qLoRA-adapted model with low-rank terms on GSM8K.

Experiment Scripts

See experiments/ptq and experiments/qpeft for PTQ and qLoRA fine-tuning experiments, respectively.

Citation

@article{zhang2024qera,
  title={QERA: an Analytical Framework for Quantization Error Reconstruction},
  author={Zhang, Cheng and Wong, Jeffrey TH and Xiao, Can and Constantinides, George A and Zhao, Yiren},
  journal={arXiv preprint arXiv:2410.06040},
  year={2024}
}