Reproduce “PAT: Pruning-Aware Tuning for Large Language Models” 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

Yijiang Liu, Huanrui Yang, Youxin Chen, Rongyu Zhang, Miao Wang, Yuan Du, Li Du

Pruning-Aware Tuning (PAT) integrates structural pruning directly into the fine-tuning of large language models, mitigating performance degradation typically associated with pruning. This method achieved up to 1.26% higher zero-shot accuracy compared to unpruned LoRA models on Llama2-7B at a 25% pruning ratio, alongside a 1.33x inference speedup and reduced VRAM usage.

Key insights

  • Integrating structural pruning directly into the fine-tuning process allows LLMs to adaptively learn optimal sparsity configurations for specific tasks, effectively mitigating performance loss and leveraging knowledge refinement.
  • The 'zero-preservation' property of RMSNorm allows for seamless propagation of pruned (zero-valued) channels through the network, enabling structural pruning without explicit modification of downstream layers.
  • Employing a trainable, globally unified sparsification mask and an efficient Hybrid-Identity-Operator (HIO) with orthogonality regularization (Identity Loss) enables robust and parameter-efficient simultaneous pruning and fine-tuning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.