sparsity ratio
Týr-the-Pruner: Structural Pruning LLMs via Global Sparsity Distribution Optimization
Structural pruning enhances hardware-agnostic inference efficiency for large language models (LLMs) yet often fails to maintain comparable performance. Local pruning performs efficient layer-by-layer compression but ignores global topology. Although global pruning aims to identify an optimal sparse model, intuitive methods typically adopt a two-stage paradigm that first evaluates substructure saliency and then applies global pruning, which ignores inter-structure dependencies and fails to achieve end-to-end optimization. To address these limitations, we propose Týr-the-Pruner, an efficient end-to-end search-based global structural pruning framework.
DiscoveringSparsityAllocationforLayer-wise PruningofLargeLanguageModels
In this paper, we present DSA, the first automated framework for discovering sparsity allocation schemes for layer-wise pruning in Large Language Models (LLMs). LLMs have become increasingly powerful, but their large parameter counts make them computationally expensive. Existing pruning methods for compressing LLMs primarily focus on evaluating redundancies and removing element-wise weights. However, these methods fail to allocate adaptive layerwise sparsities, leading to performance degradation in challenging tasks.
Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment Wei Li
Large Language Models (LLMs) have revolutionized the field of natural language processing with their impressive capabilities.
The Elastic Lottery Ticket Hypothesis
Work was done when the author interned at Microsoft.
Appendix RepresentationLearningProcess
Here we provide more experimental results. Specifically, we evaluate the representational similarity using the CKA value of the same layer from the model (ResNet-32) with different sparsity at each epoch and compare them with the final model. In our work, we evaluate four different types of freezing schemes (Sec. Inthis case, we can keep the single-shot & resume has the same FLOPs reduction asthe single-shot scheme, and the entire network can be fine-tuned at the end of training with a small learningrate. For the periodically freezing scheme, we let the selected layers freeze periodically with a given frequency so that all the layers/blocks are able to be updated at different stages of the training process.
