Fast Transformer Decoding: One Write-Head is All You Need

Abstract

Multi-head attention layers, as used in the Transformer neural sequence model, are a powerful alternative to RNNs for moving information across and between sequences. While training these layers is generally fast and simple, due to parallelizability across the length of the sequence, incremental inference (where such paralleization is impossible) is often slow, due to the memory-bandwidth cost of repeatedly loading the large "keys" and "values" tensors. We propose a variant called multi-query attention, where the keys and values are shared across all of the different attention "heads", greatly reducing the size of these tensors and hence the memory bandwidth requirements of incremental decoding. We verify experimentally that the resulting models can indeed be much faster to decode, and incur only minor quality degradation from the baseline.

Cited by in corpus (10)

• GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding
• A Comprehensive Overview and Comparative Analysis on Deep Learning Models: CNN, RNN, LSTM, GRU
• Single Headed Attention RNN: Stop Thinking With Your Head
• LambdaNetworks: Modeling Long-Range Interactions Without Attention
• Data Movement Is All You Need: A Case Study on Optimizing Transformers
• Faster Transformer Decoding: N-gram Masked Self-Attention
• Time-based Sequence Model for Personalization and Recommendation Systems
• Fast Interleaved Bidirectional Sequence Generation
• Improved Transformer for High-Resolution GANs
• EL-Attention: Memory Efficient Lossless Attention for Generation