Reproduce “MatMoE: Matryoshka Mixture-of-Experts with Dynamic Mixed-Precision Quantization for Efficient Inference” 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
Renjie Wei, Haochen Huang, Chuyu Qiu, Jinqi Wen, Dehao Xu, Meng Li
MatMoE allocates precision among activated experts at each decoding step according to routing scores, using a nested multi-precision model and floor-and-compensation to avoid an extra rounding-bitplane read at lower precision. On Qwen1.5-MoE-A2.7B, it scores 50.6% versus PMQ’s 32.2% on 8-shot GSM8K-CoT at target precision 2.25, a single-task result that does not guarantee gains elsewhere. It averages 1.5× speedup over llama.cpp on Raspberry Pi 5, an end-to-end system result.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.