[Megatron](https://arxiv.org/pdf/1909.08053.pdf) is a large, powerful transformer developed by the Applied Deep Learning Research team at NVIDIA. This repository is for ongoing research on training large transformer language models at scale. We developed efficient, model-parallel, and multinode training of [GPT-2](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) and [BERT](https://arxiv.org/pdf/1810.04805.pdf) using mixed precision.
Our codebase is capable of efficiently training a 72-layer, 8.3 billion parameter GPT-2 language model with 8-way model and 64-way data parallelism across 512 GPUs. We sustain 15.1 PetaFLOPs across the entire application with 76% scaling efficiency when compared to a strong single GPU baseline that sustains 39 TeraFLOPs, which is 30% of peak FLOPs. We find that bigger language models are able to surpass current GPT-2-1.5B WikiText-103 perplexities in as little as 5 epochs of training. Using our GPT-2 model we achieve SOTA results on the WikiText-103 (10.8 compared to SOTA perplexity of 15.8) and LAMBADA (66.5% compared to SOTA accuracy of 63.2%) datasets.
Our codebase is capable of efficiently training a 72-layer, 8.3 billion parameter GPT-2 language model with 8-way model and 64-way data parallelism across 512 GPUs. We sustain 15.1 PetaFLOPs across the entire application with 76% scaling efficiency when compared to a strong single GPU baseline that sustains 39 TeraFLOPs, which is 30% of peak theoritical FLOPs. Using our GPT-2 model we achieve SOTA results on the WikiText-103 (10.8 compared to SOTA perplexity of 15.8) and LAMBADA (66.5% compared to SOTA accuracy of 63.2%) datasets.
For BERT training, we swapped the position of the layer normalization and the residual connection in the model architecture, which allowed the models to continue to improve as they were scaled up. Our code trains BERT Large with 336 million parameters on 64 V100 GPUs in 3 days, achieving a language model loss of 1.58, SQuAD 2.0 F1-score of 88.1, and RACE accuracy of 83.0. On an ensemble of BERT-like models with 3.9 billion parameters, we reached a loss of 1.16, SQuAD 2.0 F1-score of 91.7, and RACE accuracy of 90.9.
For BERT training, we swapped the position of the layer normalization and the residual connection in the model architecture (similar to GPT-2 architucture), which allowed the models to continue to improve as they were scaled up. Our BERT models with 3.9 billion parameters reaches a loss of 1.16, SQuAD 2.0 F1-score of 91.7, and RACE accuracy of 90.9%.
