Scalable Efficient Training of Large Language Models with Low-dimensional Projected Attention
This work addresses the problem of scalable and efficient training for large language models, offering a method that enhances performance without compromising efficiency, though it appears incremental as it builds on existing low-rank and attention mechanisms.
The paper tackles the challenge of improving both effectiveness and efficiency in large language model training by introducing Low-dimensional Projected Attention (LPA), which applies low-rank modules only to attention layers, resulting in up to 12.4% time savings and approximately 5% improvement in test perplexity and downstream tasks compared to vanilla Transformers.
Improving the effectiveness and efficiency of large language models (LLMs) simultaneously is a critical yet challenging research goal. In this paper, we find that low-rank pre-training, normally considered as efficient methods that will compromise performance, can be scalably effective when reduced parameters are precisely targeted. Specifically, applying the low-dimensional module only to the attention layer -- resolves this issue and enhances both effectiveness and efficiency. We refer to this structure as Low-dimensional Projected Attention (LPA) and provide an explanatory analysis. Through extensive experimentation at parameter scales of 130M, 370M, and scaling up to 3B, we have validated the effectiveness and scalability of LPA. Our results show that LPA model can save up to 12.4% in time while achieving an approximate 5% improvement in test perplexity (ppl) and on downstream tasks compared with the vanilla Transformer.