Ouroboros: On Accelerating Training of Transformer-Based Language Models
This addresses the bottleneck of training large language models that require model parallelism due to size, offering a practical solution for researchers and practitioners in NLP.
The paper tackles the problem of accelerating training for large Transformer-based language models by proposing a model-parallel algorithm that overcomes backward locking and is applicable to language models, achieving faster speedup beyond data parallelism with comparable or better accuracy in experiments.
Language models are essential for natural language processing (NLP) tasks, such as machine translation and text summarization. Remarkable performance has been demonstrated recently across many NLP domains via a Transformer-based language model with over a billion parameters, verifying the benefits of model size. Model parallelism is required if a model is too large to fit in a single computing device. Current methods for model parallelism either suffer from backward locking in backpropagation or are not applicable to language models. We propose the first model-parallel algorithm that speeds the training of Transformer-based language models. We also prove that our proposed algorithm is guaranteed to converge to critical points for non-convex problems. Extensive experiments on Transformer and Transformer-XL language models demonstrate that the proposed algorithm obtains a much faster speedup beyond data parallelism, with comparable or better accuracy. Code to reproduce experiments is to be found at \url{https://github.com/LaraQianYang/Ouroboros}.