STAT: Shrinking Transformers After Training

TL;DR

STAT is a fast, fine-tuning-free pruning algorithm for transformer models that effectively reduces model size while maintaining accuracy, applicable to various architectures and benchmarks.

Contribution

We introduce a novel, gradient-free pruning method that compresses transformers post-training using principled matrix factorizations, without fine-tuning.

Findings

Preserves accuracy after pruning with minimal data.

Compresses large models in hours on a single GPU.

Outperforms existing gradient-free pruning methods.

Abstract

We present STAT: a simple algorithm to prune transformer models without any fine-tuning. STAT eliminates both attention heads and neurons from the network, while preserving accuracy by calculating a correction to the weights of the next layer. Each layer block in the network is compressed using a series of principled matrix factorizations that preserve the network structure. Our entire algorithm takes minutes to compress BERT, and less than three hours to compress models with 7B parameters using a single GPU. Using only several hundred data examples, STAT preserves the output of the network and improves upon existing gradient-free pruning methods. It is even competitive with methods that include significant fine-tuning. We demonstrate our method on both encoder and decoder architectures, including BERT, DistilBERT, and Llama-2 using benchmarks such as GLUE, Squad, WikiText2.