Reproduce “DictLLM: Post-training Compression of Large Language Model with Dictionary Kernels” 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
Jinho Park, Se Young Chun, Mingoo Seok
DictLLM compresses language models after training by representing related layers with a shared dictionary and sparse, layer-specific coefficients, then refitting the dictionary and quantizing both factors. For LLaMA-7B, its NVFP4 representation labeled 5× has WikiText-2 perplexity 8.48, versus 5.68 originally, and a 5.08× compression gain. This matrix-size estimate excludes embeddings and the language-model head; it does not establish complete-model deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.