Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism

Recent work in unsupervised language modeling demonstrates that training large neural language models advances the state of the art in Natural Language Processing applications. However, for very large models, memory constraints limit the size of models that can be practically trained. Model parallelism allows us to train larger models, because the parameters can be split across multiple processors. In this work, we implement a simple, efficient intra-layer model parallel approach that enables training state of the art transformer language models with billions of parameters. Our approach does not require a new compiler or library changes, is orthogonal and complimentary to pipeline model parallelism, and can be fully implemented with the insertion of a few communication operations in native PyTorch. We illustrate this approach by converging an 8.3 billion parameter transformer language model using 512 GPUs, making it the largest transformer model ever trained at 24x times the size of BERT and 5.6x times the size of GPT-2. We sustain up to 15.1 PetaFLOPs per second across the entire application with 76% scaling efficiency, compared to a strong single processor baseline that sustains 39 TeraFLOPs per second, which is 30% of peak FLOPs. The model is trained on 174GB of text, requiring 12 ZettaFLOPs over 9.2 days to converge. Transferring this language model achieves state of the art (SOTA) results on the WikiText103 (10.8 compared to SOTA perplexity of 16.4) and LAMBADA (66.5% compared to SOTA accuracy of 63.2%) datasets. We release training and evaluation code, as well as the weights of our smaller portable model, for reproducibility.