IntraSlice: Towards High-Performance Structural Pruning with Block-Intra PCA for LLMs
This addresses deployment challenges for LLMs by improving structured pruning methods to reduce model size while maintaining performance, though it appears incremental as an enhancement to existing PCA-based approaches.
The paper tackles performance degradation in structured pruning of Large Language Models by proposing IntraSlice, a block-wise module-intra PCA compression pruning framework that fuses transformation matrices without extra parameters. Experimental results on Llama2, Llama3, and Phi models show superior compression performance compared to baselines at the same compression ratio or inference speed.
Large Language Models (LLMs) achieve strong performance across diverse tasks but face deployment challenges due to their massive size. Structured pruning offers acceleration benefits but leads to significant performance degradation. Recent PCA-based pruning methods have alleviated this issue by retaining key activation components, but are only applied between modules in order to fuse the transformation matrix, which introduces extra parameters and severely disrupts activation distributions due to residual connections. To address these issues, we propose IntraSlice, a framework that applies block-wise module-intra PCA compression pruning. By leveraging the structural characteristics of Transformer modules, we design an approximate PCA method whose transformation matrices can be fully fused into the model without additional parameters. We also introduce a PCA-based global pruning ratio estimator that further considers the distribution of compressed activations, building on conventional module importance. We validate our method on Llama2, Llama3, and Phi series across various language benchmarks. Experimental results demonstrate that our approach achieves superior compression performance compared to recent baselines at the same compression ratio or inference speed.