Reproduce “UnitaryWaveLM: Orthonormal Quadrature Fourier Wave Language Model With O(1) Pre-RoPE Cavity (112.02 MiB) and Zero nn.Linear or Softmax Layers on 8 GB Apple M1 CPU” 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
Jun Hyuk Lim (Wick-Lim)
UnitaryWaveLM enables token-by-token generation with a 1.7B-parameter model on an 8 GB Apple M1 CPU. It computes projections in Fourier coordinates, approximates Softmax normalization with an eight-stage repeated-squaring rule, and uses a fixed 512-slot KV cache. In the hybrid-tiering setup, memory use is 2.15 GB RSS with no swap; this bounded-memory result does not establish language-quality parity or unlimited context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.