IDEA Prune: An Integrated Enlarge-and-Prune Pipeline in Generative Language Model Pretraining
Apple Machine Learning Research
The paper proposes an integrated enlarge-and-prune pipeline for pretraining generative language models before structured pruning. It frames the work around efficiency under limited inference budgets and emphasizes token efficiency versus training target-size models from scratch. As a result, pruning methods explicitly include enlarged model pretraining and aim to determine when this step is worthwhile even if the enlarged model is not deployed.
Why it matters
Recent advancements in large language models have intensified the need for efficient and deployable models within limited inference budgets. Structured pruning pipelines have shown promise in token efficiency compared to training target-size models from scratch. In this paper, we advocate incorporating enlarged model pretraining, which is often ignored in previous works, into pruning. We study the enlarge-and-prune pipeline as an integrated system to address two critical questions: whether it is worth pretraining an enlarged model even when the model is never deployed, and how to optimize the…
