Sparse is Enough in Scaling Transformers
This addresses the scalability and efficiency bottleneck for researchers and practitioners using large Transformer models, though it builds incrementally on prior sparsity approaches.
The authors tackled the problem of Transformer models being expensive to train and slow at decoding by leveraging sparsity across all layers, proposing Scaling Transformers that achieve the same perplexity as standard Transformers with the same parameters while enabling much faster unbatched decoding and competitive performance on long text summarization.
Large Transformer models yield impressive results on many tasks, but are expensive to train, or even fine-tune, and so slow at decoding that their use and study becomes out of reach. We address this problem by leveraging sparsity. We study sparse variants for all layers in the Transformer and propose Scaling Transformers, a family of next generation Transformer models that use sparse layers to scale efficiently and perform unbatched decoding much faster than the standard Transformer as we scale up the model size. Surprisingly, the sparse layers are enough to obtain the same perplexity as the standard Transformer with the same number of parameters. We also integrate with prior sparsity approaches to attention and enable fast inference on long sequences even with limited memory. This results in performance competitive to the state-of-the-art on long text summarization.