Reproduce “Activation-Aware Weight Tensorization: A Calibration-Time Preconditioner for Tensor-Network LLM Compression” 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
Alessandro Beatini, Marco Maronese, Emanuele Rodolà
Activation-aware Weight Tensorization (AWT) improves tensor-network compression of language-model layers by using unlabeled calibration inputs to estimate channel importance, rescale weight columns before an unchanged Tensor Train or Tree Tensor Network decomposition, and compensate at inference. For Llama 3.1 8B suffix replacement, it closes 59.9% of the dense-to-vanilla perplexity gap at 6× compression; this is not full-model compression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.