PATCH: Learnable Tile-level Hybrid Sparsity for LLMs
This addresses the deployment cost problem for LLM users by offering a more flexible and efficient pruning method, though it is incremental in improving existing sparsity techniques.
The paper tackles the trade-off between accuracy and hardware efficiency in pruning large language models by introducing PATCH, a hybrid sparsity framework that allows continuous sparsity ratios and achieves 1.18x-1.38x speedup with 0.37%-2.96% accuracy improvement over state-of-the-art methods on models like LLaMA-2 7B.
Large language models (LLMs) deliver impressive performance but incur prohibitive memory and compute costs at deployment. Model pruning is an effective way to reduce these overheads, yet existing approaches face challenges: unstructured sparsity, where nonzeros can appear anywhere, preserves accuracy but yields irregular access patterns that prevent GPU acceleration, while semi-structured 2:4 sparsity is hardware-friendly but enforces a rigid 50% pattern that degrades model quality. To bridge this gap, we introduce PATCH, a hybrid sparsity framework that enables a continuous sparsity ratio between 0% and 50%. PATCH partitions weight matrices into tiles, assigning each tile to be either dense or 2:4 sparse via a learnable mask selection mechanism. This design provides fine-grained control over accuracy-acceleration tradeoffs and supports non-uniform sparsity across layers, leading to superior overall quality. Across models from 0.5B to 8B parameters, PATCH consistently narrows the gap to dense accuracy while delivering practical speedups. For instance, on LLaMA-2 7B with an A6000 GPU, PATCH achieves 1.18x-1.38x end-to-end speedup over dense baselines while improving accuracy by 0.37%-2.96% compared to the state-of-the-art 2:4 pruning method, MaskLLM.