As we scale to more massive machine learning models, the frequent synchronization demands inherent in data-parallel approaches create significant slowdowns, posing a critical challenge to further scaling. Recent work [11, 24] develops and analyzes an approach (DiLoCo) that relaxes synchronization demands via periodic synchronization. However, these works do not carefully analyze how DiLoCo's behavior changes with model size. In this work, we study the scaling law behavior of DiLoCo when training LLMs under a fixed compute budget. We focus on how algorithmic factors, including number of model replicas, hyperparameters, and token budget affect training in ways that can be accurately predicted via scaling laws. We find that DiLoCo scales both predictably and robustly with model size. When well-tuned, DiLoCo scales better than data-parallel training with model size, and can outperform data-parallel training even at small model sizes. Our results showcase a more general set of benefits of DiLoCo than previously documented, including increased optimal batch sizes, improved downstream generalization with scale, and improved evaluation loss for a fixed token budget.

As we scale to more massive machine learning models, the frequent synchronization demands inherent in data-parallel approaches create significant slowdowns, posing a critical challenge to further scaling. Recent work develops an approach (DiLoCo) that relaxes synchronization demands without compromising model quality. However, these works do not carefully analyze how DiLoCo's behavior changes with model size. In this work, we study the scaling law behavior of DiLoCo when training LLMs under a fixed compute budget. We focus on how algorithmic factors, including number of model replicas, hyperparameters, and token budget affect training in ways that can be accurately predicted via scaling laws. We find that DiLoCo scales both predictably and robustly with model size. When well-tuned, DiLoCo scales better than data-parallel training with model size, and can outperform data-parallel training even at small model sizes. Our results showcase a more general set of benefits of DiLoCo than previously documented, including increased optimal batch sizes, improved downstream generalization with scale, and improved evaluation loss for a fixed token budget.

On High-Performance Computing (HPC) systems, several hyperparameter configurations can be evaluated in parallel to speed up the Hyperparameter Optimization (HPO) process. State-of-the-art HPO methods follow a bandit-based approach and build on top of successive halving, where the final performance of a combination is estimated based on a lower than fully trained fidelity performance metric and more promising combinations are assigned more resources over time. Frequently, the number of epochs is treated as a resource, letting more promising combinations train longer. Another option is to use the number of workers as a resource and directly allocate more workers to more promising configurations via data-parallel training. This article proposes a novel Resource-Adaptive Successive Doubling Algorithm (RASDA), which combines a resource-adaptive successive doubling scheme with the plain Asynchronous Successive Halving Algorithm (ASHA). Scalability of this approach is shown on up to 1,024 Graphics Processing Units (GPUs) on modern HPC systems. It is applied to different types of Neural Networks (NNs) and trained on large datasets from the Computer Vision (CV), Computational Fluid Dynamics (CFD), and Additive Manufacturing (AM) domains, where performing more than one full training run is usually infeasible. Empirical results show that RASDA outperforms ASHA by a factor of up to 1.9 with respect to the runtime. At the same time, the solution quality of final ASHA models is maintained or even surpassed by the implicit batch size scheduling of RASDA. With RASDA, systematic HPO is applied to a terabyte-scale scientific dataset for the first time in the literature, enabling efficient optimization of complex models on massive scientific data. The implementation of RASDA is available on https://github.com/olympiquemarcel/rasda

Graph Neural Networks (GNNs) have emerged as powerful tools for supervised machine learning over graph-structured data, while sampling-based node representation learning is widely utilized in unsupervised learning. However, scalability remains a major challenge in both supervised and unsupervised learning for large graphs (e.g., those with over 1 billion nodes). The scalability bottleneck largely stems from the mini-batch sampling phase in GNNs and the random walk sampling phase in unsupervised methods. These processes often require storing features or embeddings in memory. In the context of distributed training, they require frequent, inefficient random access to data stored across different workers. Such repeated inter-worker communication for each mini-batch leads to high communication overhead and computational inefficiency. We propose GraphScale, a unified framework for both supervised and unsupervised learning to store and process large graph data distributedly. The key insight in our design is the separation of workers who store data and those who perform the training. This separation allows us to decouple computing and storage in graph training, thus effectively building a pipeline where data fetching and data computation can overlap asynchronously. Our experiments show that GraphScale outperforms state-of-the-art methods for distributed training of both GNNs and node embeddings. We evaluate GraphScale both on public and proprietary graph datasets and observe a reduction of at least 40% in end-to-end training times compared to popular distributed frameworks, without any loss in performance. While most existing methods don't support billion-node graphs for training node embeddings, GraphScale is currently deployed in production at TikTok enabling efficient learning over such large graphs.

Speeding up the large-scale distributed training is challenging in that it requires improving various components of training including load balancing, communication, optimizers, etc. We present novel approaches for fast large-scale training of BERT model which individually ameliorates each component thereby leading to a new level of BERT training performance. Load balancing is imperative in distributed BERT training since its training datasets are characterized by samples with various lengths. Communication cost, which is proportional to the scale of distributed training, needs to be hidden by useful computation. In addition, the optimizers, e.g., ADAM, LAMB, etc., need to be carefully re-evaluated in the context of large-scale distributed training. We propose two new ideas, (1) local presorting based on dataset stratification for load balancing and (2) bucket-wise gradient clipping before allreduce which allows us to benefit from the overlap of gradient computation and synchronization as well as the fast training of gradient clipping before allreduce. We also re-evaluate existing optimizers via hyperparameter optimization and utilize ADAM, which also contributes to fast training via larger batches than existing methods. Our proposed methods, all combined, give the fastest MLPerf BERT training of 25.1 (22.3) seconds on 1,024 NVIDIA A100 GPUs, which is 1.33x (1.13x) and 1.57x faster than the other top two (one) submissions to MLPerf v1.1 (v2.0). Our implementation and evaluation results are available at MLPerf v1.1~v2.1.

Here, message passing semantics is leveraged, which helps any process to communicate with other processes through the messages. Different communication backends are used, and the communication doesn't need to be from the same machine. But, of course, multiple processes should be started in parallel, and coordination tools in the cluster must be enabled. There are three main components in the torch. First, distributed as distributed data-parallel training, RPC-based distributed training, and collective communication.

Developing high-performing predictive models for large tabular data sets is a challenging task. The state-of-the-art methods are based on expert-developed model ensembles from different supervised learning methods. Recently, automated machine learning (AutoML) is emerging as a promising approach to automate predictive model development. Neural architecture search (NAS) is an AutoML approach that generates and evaluates multiple neural network architectures concurrently and improves the accuracy of the generated models iteratively. A key issue in NAS, particularly for large data sets, is the large computation time required to evaluate each generated architecture. While data-parallel training is a promising approach that can address this issue, its use within NAS is difficult. For different data sets, the data-parallel training settings such as the number of parallel processes, learning rate, and batch size need to be adapted to achieve high accuracy and reduction in training time. To that end, we have developed AgEBO-Tabular, an approach to combine aging evolution (AgE), a parallel NAS method that searches over neural architecture space, and an asynchronous Bayesian optimization method for tuning the hyperparameters of the data-parallel training simultaneously. We demonstrate the efficacy of the proposed method to generate high-performing neural network models for large tabular benchmark data sets. Furthermore, we demonstrate that the automatically discovered neural network models using our method outperform the state-of-the-art AutoML ensemble models in inference speed by two orders of magnitude while reaching similar accuracy values.

The enormous amount of data and computation required to train DNNs have led to the rise of various parallelization strategies. Broadly, there are two strategies: 1) Data-Parallelism -- replicating the DNN on multiple processes and training on different training samples, and 2) Model-Parallelism -- dividing elements of the DNN itself into partitions across different processes. While data-parallelism has been extensively studied and developed, model-parallelism has received less attention as it is non-trivial to split the model across processes. In this paper, we propose HyPar-Flow: a framework for scalable and user-transparent parallel training of very large DNNs (up to 5,000 layers). We exploit TensorFlow's Eager Execution features and Keras APIs for model definition and distribution. HyPar-Flow exposes a simple API to offer data, model, and hybrid (model + data) parallel training for models defined using the Keras API. Under the hood, we introduce MPI communication primitives like send and recv on layer boundaries for data exchange between model-partitions and allreduce for gradient exchange across model-replicas. Our proposed designs in HyPar-Flow offer up to 3.1x speedup over sequential training for ResNet-110 and up to 1.6x speedup over Horovod-based data-parallel training for ResNet-1001; a model that has 1,001 layers and 30 million parameters. We provide an in-depth performance characterization of the HyPar-Flow framework on multiple HPC systems with diverse CPU architectures including Intel Xeon(s) and AMD EPYC. HyPar-Flow provides 110x speed up on 128 nodes of the Stampede2 cluster at TACC for hybrid-parallel training of ResNet-1001.

In this paper, we present BigDL, a distributed deep learning framework for Big Data platforms and workflows. It is implemented on top of Apache Spark, and allows users to write their deep learning applications as standard Spark programs (running directly on large-scale big data clusters in a distributed fashion). It provides an expressive, "data-analytics integrated" deep learning programming model, so that users can easily build the end-to-end analytics + AI pipelines under a unified programming paradigm; by implementing an AllReduce like operation using existing primitives in Spark (e.g., shuffle, broadcast, and in-memory data persistence), it also provides a highly efficient "parameter server" style architecture, so as to achieve highly scalable, data-parallel distributed training. Since its initial open source release, BigDL users have built many analytics and deep learning applications (e.g., object detection, sequence-to-sequence generation, visual similarity, neural recommendations, fraud detection, etc.) on Spark.