Reproduce “QSVD: Efficient Low-rank Approximation for Unified Query-Key-Value Weight Compression in Low-Precision Vision-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yutong Wang, Haiyu Wang, Sai Qian Zhang
QSVD introduces a unified framework for compressing Vision-Language Models by applying singular value decomposition to joint Query-Key-Value weight matrices and integrating an adaptive post-training quantization scheme. This approach substantially reduces computational cost, memory footprint, and KV cache size while maintaining high accuracy and achieving significant inference speedups.
Key insights
- Applying SVD to joint QKV weights is more efficient than individual decompositions, reducing weight parameters, computational cost, and KV cache size by a larger margin.
- Low-rank approximation can act as a regularization technique, sometimes improving VLM accuracy (e.g., on VizWiz) and reducing hallucination compared to full-precision models.
- Dynamically allocating SVD ranks across layers using importance scores is crucial for preserving model accuracy under aggressive compression ratios.
- Learning an optimal `β` parameter for each layer helps smooth activation outliers in the SVD-compressed representations, enabling more robust low-precision quantization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.