Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation
This addresses a fundamental limitation in transformer models for tasks requiring long sequences, such as language modeling, by providing an efficient extrapolation method.
The paper tackles the problem of enabling transformer models to extrapolate to longer input sequences than seen during training by introducing Attention with Linear Biases (ALiBi), a position representation method that biases attention scores based on distance. It shows that a 1.3 billion parameter model trained on sequences of length 1024 achieves the same perplexity on sequences of length 2048 as a model trained on length 2048, while training 11% faster and using 11% less memory.
Since the introduction of the transformer model by Vaswani et al. (2017), a fundamental question has yet to be answered: how does a model achieve extrapolation at inference time for sequences that are longer than it saw during training? We first show that extrapolation can be enabled by simply changing the position representation method, though we find that current methods do not allow for efficient extrapolation. We therefore introduce a simpler and more efficient position method, Attention with Linear Biases (ALiBi). ALiBi does not add positional embeddings to word embeddings; instead, it biases query-key attention scores with a penalty that is proportional to their distance. We show that this method trains a 1.3 billion parameter model on input sequences of length 1024 that extrapolates to input sequences of length 2048, achieving the same perplexity as a sinusoidal position embedding model trained on inputs of length 2048 but training 11% faster and using 11% less memory. ALiBi's inductive bias towards recency also leads it to outperform multiple strong position methods on the WikiText-103 benchmark.