Reproduce “Revisiting Error Feedback for Model 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
Wai-Yiu Keung, Chi Cui, Wing-Kin Ma, Jesús Cid-Sueiro, Hei Victor Cheng
The paper redesigns quantization feedback to account for earlier rounding errors and downstream sensitivity. It analyzes joint rotation-feedback under fixed error-statistics assumptions. On Llama-2-7B at M=8, objective-aware EFQ+RHT lowers perplexity from 9.120 to 8.704, raises accuracy from 53.15% to 53.73%, and reduces final-logit MSE from 2.077 to 1.739. The theorem optimizes a trace surrogate, not finite-level performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.