Reproduce “SubRot: Signed Gradient Subspace Calibration for VLM Rotation Quantization” 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
Zhenhao Shang, Haizhao Jing, Haokui Zhang, Guoting Wei, Rong Xiao, Jianqing Gao, Peng Wang
SubRot guides low-bit VLM quantization by using activation gradients to identify sensitive channel directions, steering quantization errors with stable signed effects while restraining errors in other sensitive directions and retaining reconstruction loss. In a five-benchmark average for LLaVA-NeXT-7B at W4A4, it scores 48.8 versus FlatQuant’s 47.4. Evaluation quantizes the language-model component, and averages do not imply gains on every benchmark.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.