QSVD: Efficient Low-Rank Approximation for Unified Query-Key-Value Weight Compression in Low-Precision Vision-Language Models

April 29, 2026 · View on GitHub

This repository provides the official implementation of QSVD, a method for efficient low-rank approximation that unifies Query-Key-Value (QKV) weight compression in low-precision Vision-Language Models (VLMs).

QSVD Overview

🎉 Our work has been accepted to NeurIPS 2025 as a Spotlight paper.

🌟 Highlights

🧩 Joint QKV Decomposition:
QSVD performs a joint singular value decomposition on the concatenated query–key–value weight matrices $[W_q, W_k, W_v]$ , sharing a common down-projection $W_{qkv}^{d}$ .
→ Reduces parameters, low-rank KV-cache, and FLOPs compared to per-matrix SVD.
📊 Cross-Layer Rank Allocation:
Introduces an adaptive cross-layer rank allocation strategy to allocate ranks across all layers based on each singular value’s contribution to model loss, enabling fine-grained, gradient-guided truncation.
→ Preserves critical components while truncating redundant ones across all layers.
🎯 Post-Training Quantization for Low-Rank VLMs:
Combines dual orthogonal rotations $(H_1, \ H_2)$ to smooth channel-wise outliers in both activations and latent buffers, together with an adaptive exponent β that rescales singular values to balance channel distributions and reduce quantization error.
→ Jointly suppresses activation variance and outlier amplification, enabling stable low-precision inference with minimal degradation.

🔧 Requirements

This implementation utilizes the myllava repository, adapted from the original LLaVA repo. Please follow the steps below to set up the environment:

git submodule update --init --recursive
conda create -n QSVD python=3.10 -y
conda activate QSVD
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu126
pip install --no-build-isolation -r requirements.txt

📊 Evaluation

To evaluate QSVD and reproduce our results, follow the steps below.

📁 Dataset Preparation

Follow the LLaVA evaluation guide to prepare the following datasets:

ScienceQA (Train) LLaVA ScienceQA train
VizWiz (Test) Update the paths in eval_*.py and data_utils.py accordingly.

🛠 Evaluation Toolkit

We use third_party/VLMEvalKit for evaluation. Please follow its Quickstart for environment setup and usage.

▶️ Running Evaluations

We provide pre-computed calibration cache files to directly reproduce the main QSVD results without rerunning the whitening data and gradient collection. All pre-computed cache files are organized under the cache_file directory.

Each cache package includes:

Activation-aware whitening data for ASVD-style preprocessing, computed from 256 calibration samples.
Gradient-square expectations of all singular values, estimated on the same dataset.
Final importance scores used for cross-layer rank allocation in the joint QKV SVD.

Currently released model cache:

SmolVLM-Instruct 2B
LLaVA-Next 7B
LLaVA-Next 13B

To reproduce our main results on ScienceQA with the provided cache files, you can directly run the following example for LLaVA-Next 7B:

export HF_HOME='your_hf_home'
cd path_to_QSVD/fake_quant
conda activate QSVD
bash path_to_QSVD/scripts/fp16_cache_llavanext.sh 0.9

This script will automatically:

Load the pre-computed calibration cache from cache_file/llava-next-7b.
Apply the joint QKV SVD using the stored whitening data, and adaptively truncate singular values based on the pre-computed importance scores and rank budgets.
Evaluate the compressed model under the default R₁ = 60 % / R₂ = 22.5 % configuration.

💡 Note:
The argument 0.9 in the command specifies the rank budget ratio used for joint QKV SVD.
After decomposition, the overall preserved rank is approximately 0.9 / 2 = 0.45, meaning that about 45 % of the original rank is retained (dividing by 2 is a legacy setting in the current script).
This configuration corresponds to R₁ = 60 % / R₂ = 22.5 % in our paper.

Similarly, using 0.8 results in an effective ratio of 0.8 / 2 = 0.4, i.e., retaining 40 % of the rank, which corresponds to R₁ = 53.33 % / R₂ = 20.0 %.

The mapping between rank ratio → (R₁, R₂) follows directly from the definition of joint QKV SVD in QSVD.

You can dynamically adjust this parameter (0.9, 0.8, etc.) to control the effective rank budget, balancing the trade-off between compression ratio and accuracy.

For LLaVA-Next 13B, modify the script scripts/fp16_cache_llavanext.sh:

Set the cache path to:

cache_file="../cache_file/llava-next-13b"

Change the model argument to:

--model llava-hf/llava-v1.6-vicuna-13b-hf

Then run the same command:

bash path_to_QSVD/scripts/fp16_cache_llavanext.sh 0.9

For SmolVLM-Instruct 2B, use the corresponding script:

bash path_to_QSVD/scripts/fp16_cache_smolvlm.sh 1.5

Similarly, the argument 1.5 here specifies a rank budget ratio corresponding to an effective 75 % retained rank, which maps to R₁ = 100 % / R₂ = 50 % in our joint-SVD configuration.

🔎 More Details

For more usage and custom evaluations, explore the instructions and scripts in fake_quant and scripts. Currently, we only support SmolVLM, LLaVA-v1.5, LLaVA-Next models. You can simply run the mainsmolvlm.py, mainllava.py, or mainllavanext.py accordingly to reproduce the results in the paper. The most important arguments are:

--model: Model name (or path to the weights)
--seed: Control the random seed
--nsamples: Number of samples for SVD calibration
--rotate: Whether we want to rotate the model (apply quarot)
--tasks: Tasks for LM-Eval
--cal_dataset: Calibration dataset for GPTQ quantization/SVD calibration (currently supports ScienceQA_Train)
--eval_dataset: Evaluation dataset (currently support ScienceQA_TEST and VizWiz)
--a_bits: Number of bits for activation quantization
--w_bits: Number of bits for weight quantization
--v_bits: Number of bits for value quantization (deprecated if using SVD)
--k_bits: Number of bits for key quantization (deprecated if using SVD)
--w_clip: Whether we want to clip the weights
--a_clip_ratio: The ratio of clipping for activation
--vita_clip_ratio: Override the ratio of clipping for vit activation
--lma_clip_ratio: Override the ratio of clipping for language model activation
--k_clip_ratio: The ratio of clipping for key (deprecated if using SVD)
--v_clip_ratio: The ratio of clipping for value (deprecated if using SVD)
--w_asym: Whether we want to use asymmetric quantization for weights
--a_asym: Whether we want to use asymmetric quantization for activation
--v_asym: Whether we want to use asymmetric quantization for value
--k_asym: Whether we want to use asymmetric quantization for key
--a_groupsize: The group size for activation quantization
--w_groupsize: The group size for weight quantization
--v_groupsize: The group size for value quantization
--k_groupsize: The group size for key quantization
--svd_mode: Choose how sigma is fused in SVD weights
--qkv_fuse: Whether we concatenate QKV for the joint SVD proposed in our paper
--calib_method: Choose SVD whitening method (abs_max and abs_mean for ASVD-style)
--rank_ratio: 2 * SVD rank ratio (the factor of 2 is a legacy setting)
--act_aware: Whether to use activation-aware SVD
--had_rank: Whether to add rotation (H₂ in our paper) in SVD latent activation
--svd_lm: Whether we apply SVD
--act_alpha: Activation-aware SVD related hyperparameter of ASVD
--vit_module: Whether we apply quantization in ViT
--grad_info: Whether we use the cross-layer rank allocation proposed in our paper
--beta_then_svd: Whether we apply SVD after ViT quantization
--cache_file: Path to pre-computed calibration cache file folder
--basepath: Path to the parent folder of myllava (where we store ScienceQA and VizWiz dataset)

For example, to run the ScienceQA evaluation of llava-v1.5-7b model with quantizing all weights and activations, you can run the following command:

cd QSVD/fake_quant
python mainllava.py --model liuhaotian/llava-v1.5-7b  \
                --a_bits 4 \
                --w_bits 4 \
                --cal_dataset ScienceQA_Train \
                --eval_dataset ScienceQA_TEST \
                --w_rtn \
                --w_clip \
                --lma_clip_ratio 0.9 \
                --nsamples 256 \
                --seed 0 \
                --svd_mode "U" \
                --qkv_fuse \
                --calib_method 'abs_mean' \
                --rank_ratio 1.5 \
                --act_aware \
                --had_rank \
                --svd_lm \
                --act_alpha 0.5 \
                --setting "/sqa/online_then_qkvlm_svd/seed0" \
                --rotate \
                --vit_module \
                --vit_online \
                --grad_info \
                --beta_then_svd

📚 Citation

If QSVD helps your research or applications, please cite our paper:

@article{wang2025qsvd,
  title={QSVD: Efficient Low-rank Approximation for Unified Query-Key-Value Weight Compression in Low-Precision Vision-Language Models},
  author={Wang, Yutong and Wang, Haiyu and Zhang, Sai Qian},
  journal={arXiv preprint arXiv:2510.16292},
  year={2025}
}

🤝 Contributing

This project builds upon the excellent work of:

We thank these projects for their contributions to the community.