Reproduce “ApexQuant: Data-Free Elastic Quantization by Residual Re-Isotropization” 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
Aksel Fristrup, Sumit Pandey, Ankit Kariryaa
ApexQuant compresses weights without calibration data by repeatedly rotating residual weight error, quantizing it with the same stage, and subtracting the reconstruction. Fresh Haar rotations make each residual direction uniform, yielding predictable geometric decay in weight-space error. On Qwen2.5-7B, its two-bit trellis reached 0.477 nats per token versus 9.985 for data-free SINQ at near-matched storage, but calibrated quantization performed better.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.