It is widely acknowledged that large models have the potential to deliver superior performance across a broad range of domains. Despite the remarkable progress made in the field of machine learning systems research, which has enabled the development and exploration of large models, such abilities remain confined to a small group of advanced users and industry leaders, resulting in an implicit technical barrier for the wider community to access and leverage these technologies. In this paper, we introduce PyTorch Fully Sharded Data Parallel (FSDP) as an industry-grade solution for large model training. FSDP has been closely co-designed with several key PyTorch core components including Tensor implementation, dispatcher system, and CUDA memory caching allocator, to provide non-intrusive user experiences and high training efficiency. Additionally, FSDP natively incorporates a range of techniques and settings to optimize resource utilization across a variety of hardware configurations. The experimental results demonstrate that FSDP is capable of achieving comparable performance to Distributed Data Parallel while providing support for significantly larger models with near-linear scalability in terms of TFLOPS.

In many practical applications of machine learning data arrives sequentially over time in large chunks. Practitioners have then to decide how to allocate their computational budget in order to obtain the best performance at any point in time. Online learning theory for convex optimization suggests that the best strategy is to use data as soon as it arrives. However, this might not be the best strategy when using deep non-linear networks, particularly when these perform multiple passes over each chunk of data rendering the overall distribution non i.i.d.. In this paper, we formalize this learning setting in the simplest scenario in which each data chunk is drawn from the same underlying distribution, and make a first attempt at empirically answering the following questions: How long should the learner wait before training on the newly arrived chunks? What architecture should the learner adopt? Should the learner increase capacity over time as more data is observed? We probe this learning setting using convolutional neural networks trained on classic computer vision benchmarks as well as a large transformer model trained on a large-scale language modeling task. Code is available at \url{www.github.com/facebookresearch/ALMA}.

This paper explores the environmental impact of the super-linear growth trends for AI from a holistic perspective, spanning Data, Algorithms, and System Hardware. We characterize the carbon footprint of AI computing by examining the model development cycle across industry-scale machine learning use cases and, at the same time, considering the life cycle of system hardware. Taking a step further, we capture the operational and manufacturing carbon footprint of AI computing and present an end-to-end analysis for what and how hardware-software design and at-scale optimization can help reduce the overall carbon footprint of AI. Based on the industry experience and lessons learned, we share the key challenges and chart out important development directions across the many dimensions of AI. We hope the key messages and insights presented in this paper can inspire the community to advance the field of AI in an environmentally-responsible manner.

Large-scale autoregressive language models such as GPT-3 are few-shot learners that can perform a wide range of language tasks without fine-tuning. While these models are known to be able to jointly represent many different languages, their training data is dominated by English, potentially limiting their cross-lingual generalization. In this work, we train multilingual autoregressive language models on a balanced corpus covering a diverse set of languages, and study their few- and zero-shot learning capabilities in a wide range of tasks. Our largest model with 7.5 billion parameters sets new state of the art in few-shot learning in more than 20 representative languages, outperforming GPT-3 of comparable size in multilingual commonsense reasoning (with +7.4% absolute accuracy improvement in 0-shot settings and +9.4% in 4-shot settings) and natural language inference (+5.4% in each of 0-shot and 4-shot settings). On the FLORES-101 machine translation benchmark, our model outperforms GPT-3 on 171 out of 182 translation directions with 32 training examples, while surpassing the official supervised baseline in 45 directions. We present a detailed analysis of where the model succeeds and fails, showing in particular that it enables cross-lingual in-context learning on some tasks, while there is still room for improvement on surface form robustness and adaptation to tasks that do not have a natural cloze form. Finally, we evaluate our models in social value tasks such as hate speech detection in five languages and find it has limitations similar to comparable sized GPT-3 models.

Mixture of Experts layers (MoEs) enable efficient scaling of language models through conditional computation. This paper presents a detailed empirical study of how autoregressive MoE language models scale in comparison with dense models in a wide range of settings: in- and out-of-domain language modeling, zero- and few-shot priming, and full fine-tuning. With the exception of fine-tuning, we find MoEs to be substantially more compute efficient. At more modest training budgets, MoEs can match the performance of dense models using $\sim$4 times less compute. This gap narrows at scale, but our largest MoE model (1.1T parameters) consistently outperforms a compute-equivalent dense model (6.7B parameters). Overall, this performance gap varies greatly across tasks and domains, suggesting that MoE and dense models generalize differently in ways that are worthy of future study. We make our code and models publicly available for research use.

During pretraining, the Pre-LayerNorm transformer suffers from a gradient magnitude mismatch: gradients at early layers are much larger than at later layers. These issues can be alleviated by our proposed NormFormer architecture, which adds three normalization operations to each layer: a Layer Norm after self attention, head-wise scaling of self-attention outputs, and a Layer Norm after the first fully connected layer. The extra operations incur negligible compute cost ( 0.4% parameter increase), but improve pretraining perplexity and downstream task performance for both causal and masked language models ranging from 125 Million to 2.7 Billion parameters. For example, adding NormFormer on top of our strongest 1.3B parameter baseline can reach equal perplexity 24% faster, or converge 0.27 perplexity better in the same compute budget. This model reaches GPT3-Large (1.3B) zero shot performance 60% faster. For masked language modeling, NormFormer improves fine-tuned GLUE performance by 1.9% on average. Code to train NormFormer models is available in fairseq. The original transformer architecture (Vaswani et al., 2017) applies Layer Normalization (Ba et al., 2016) after each sublayer's residual connection ("Post-LN") in order to reduce the variance of the inputs to the following sublayer, i.e.: PostLN(x) LayerNorm(x Sublayer(x)), with Recent work has observed that Post-LN transformers tend to have larger magnitude gradients in later layers compared to earlier layers (Xiong et al., 2020) and has advocated moving the LayerNorm operation to the beginning of each sublayer ("Pre-LN"; see Figure 1, left), i.e.: PreLN(x) x Sublayer(LayerNorm(x)).

Current large-scale auto-regressive language models (Radford et al., 2019; Liu et al., 2018; Graves, 2013) display impressive fluency and can generate convincing text. In this work we start by asking the question: Can the generations of these models be reliably distinguished from real text by statistical discriminators? We find experimentally that the answer is affirmative when we have access to the training data for the model, and guardedly affirmative even if we do not. This suggests that the auto-regressive models can be improved by incorporating the (globally normalized) discriminators into the generative process. We give a formalism for this using the Energy-Based Model framework, and show that it indeed improves the results of the generative models, measured both in terms of perplexity and in terms of human evaluation.

In this work, we study how the large-scale pretrain-finetune framework changes the behavior of a neural language generator. We focus on the transformer encoder-decoder model for the open-domain dialogue response generation task. We find that after standard fine-tuning, the model forgets important language generation skills acquired during large-scale pre-training. We demonstrate the forgetting phenomenon through a detailed behavior analysis from the perspectives of context sensitivity and knowledge transfer. Adopting the concept of data mixing, we propose an intuitive fine-tuning strategy named "mix-review". We find that mix-review effectively regularize the fine-tuning process, and the forgetting problem is largely alleviated. Finally, we discuss interesting behavior of the resulting dialogue model and its implications.

Recent advances in generative modeling of text have demonstrated remarkable improvements in terms of fluency and coherency. In this work we investigate to which extent a machine can discriminate real from machine generated text. This is important in itself for automatic detection of computer generated stories, but can also serve as a tool for further improving text generation. We show that learning a dedicated scoring function to discriminate between real and fake text achieves higher precision than employing the likelihood of a generative model. The scoring functions generalize to other generators than those used for training as long as these generators have comparable model complexity and are trained on similar datasets.

User-generated comments in online social media have recently been gaining increasing attention as a viable source of general-purpose descriptive annotations for digital objects like photos or videos. Because users have different levels of expertise, however, the quality of their comments can vary from very useful to entirely useless. Our aim is to provide automated support for the curation of useful user-generated comments from public collections of digital objects. After constructing a crowd-sourced gold standard of useful and not useful comments, we use standard machine learning methods to develop a usefulness classifier, exploring the impact of surface-level, syntactic, semantic, and topic-based features in addition to extra-linguistic attributes of the author and his or her social media activity. We then adapt an existing model of prevalence detection that uses the learned classifier to investigate patterns in the commenting culture of two popular social media platforms. We find that the prevalence of useful comments is platform-specific and is further influenced by the entity type of the media object being commented on (person, place, event), its time period (e.g., year of an event), and the degree of polarization among commenters.