Summary of Compact: Compressed Activations For Memory-efficient Llm Training, by Yara Shamshoum et al.
CompAct: Compressed Activations for Memory-Efficient LLM Training
by Yara Shamshoum, Nitzan Hodos, Yuval Sieradzki, Assaf Schuster
First submitted to arxiv on: 20 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computation and Language (cs.CL)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A novel technique called CompAct reduces peak memory utilization on Graphics Processing Units (GPUs) by 25-30% during pretraining and 50% during fine-tuning of Large Language Models (LLMs). This is achieved by compressing low-rank activations to be used in the backward pass, unlike previous methods that only reduce optimizer overheads or the number of trained parameters. CompAct uses random projection matrices to avoid additional memory overheads. Compared to existing techniques for either pretraining or fine-tuning, CompAct substantially improves compute-performance tradeoffs. The technique’s savings are expected to scale even higher for larger models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary CompAct is a new way to make Large Language Models (LLMs) use less memory when training on Graphics Processing Units (GPUs). Right now, peak device memory can be a big problem when training LLMs. CompAct fixes this by storing low-rank, compressed activations that are used in the backward pass. This helps reduce the required memory and makes it possible to train larger models. |
Keywords
» Artificial intelligence » Fine tuning » Pretraining
