FlexRank: Nested Low-Rank Knowledge Decomposition for Adaptive Model Deployment
This addresses the problem of rigid and expensive deployment of large models like LLMs and ViTs for practitioners, offering an incremental improvement in efficiency.
The paper tackles the high cost of deploying large neural networks by proposing FlexRank, a method that extracts nested, importance-ordered submodels from pretrained models to enable adaptive deployment across different computational budgets, achieving a graceful trade-off between cost and performance without retraining.
The growing scale of deep neural networks, encompassing large language models (LLMs) and vision transformers (ViTs), has made training from scratch prohibitively expensive and deployment increasingly costly. These models are often used as computational monoliths with fixed cost, a rigidity that does not leverage overparametrized architectures and largely hinders adaptive deployment across different cost budgets. We argue that importance-ordered nested components can be extracted from pretrained models, and selectively activated on the available computational budget. To this end, our proposed FlexRank method leverages low-rank weight decomposition with nested, importance-based consolidation to extract submodels of increasing capabilities. Our approach enables a "train-once, deploy-everywhere" paradigm that offers a graceful trade-off between cost and performance without training from scratch for each budget - advancing practical deployment of large models.