Recent years have seen an increase in the development of large deep learning (DL) models, which makes training efficiency crucial. Common practice is struggling with the trade-off between usability and performance. On one hand, DL frameworks such as PyTorch use dynamic graphs to facilitate model developers at a price of sub-optimal model training performance. On the other hand, practitioners propose various approaches to improving the training efficiency by sacrificing some of the flexibility, ranging from making the graph static for more thorough optimization (e.g., XLA) to customizing optimization towards large-scale distributed training (e.g., DeepSpeed and Megatron-LM). In this paper, we aim to address the tension between usability and training efficiency through separation of concerns. Inspired by DL compilers that decouple the platform-specific optimizations of a tensor-level operator from its arithmetic definition, this paper proposes a schedule language, Slapo, to decouple model execution from definition. Specifically, Slapo works on a PyTorch model and uses a set of schedule primitives to convert the model for common model training optimizations such as high-performance kernels, effective 3D parallelism, and efficient activation checkpointing. Compared to existing optimization solutions, Slapo progressively optimizes the model "as-needed" through high-level primitives, and thus preserving programmability and debuggability for users to a large extent. Our evaluation results show that by scheduling the existing hand-crafted optimizations in a systematic way using Slapo, we are able to improve training throughput by up to 2.92x on a single machine with 8 NVIDIA V100 GPUs, and by up to 1.41x on multiple machines with up to 64 GPUs, when compared to the out-of-the-box performance of DeepSpeed and Megatron-LM.

Multi-task model training has been adopted to enable a single deep neural network model (often a large language model) to handle multiple tasks (e.g., question answering and text summarization). Multi-task training commonly receives input sequences of highly different lengths due to the diverse contexts of different tasks. Padding (to the same sequence length) or packing (short examples into long sequences of the same length) is usually adopted to prepare input samples for model training, which is nonetheless not space or computation efficient. This paper proposes a dynamic micro-batching approach to tackle sequence length variation and enable efficient multi-task model training. We advocate pipelineparallel training of the large model with variable-length micro-batches, each of which potentially comprises a different number of samples. We optimize micro-batch construction using a dynamic programming-based approach, and handle micro-batch execution time variation through dynamic pipeline and communication scheduling, enabling highly efficient pipeline training. Extensive evaluation on the FLANv2 dataset demonstrates up to 4.39x higher training throughput when training T5, and 3.25x when training GPT, as compared with packing-based baselines. DynaPipe's source code is publicly available at https://github.com/awslabs/optimizing-multitask-training-through-dynamic-pipelines.

Multimodal model training takes multiple types of inputs to process with differently structured submodules, and aggregates outcomes from the submodules to learn the relationship among various types of inputs, e.g., correlating text to image for text-to-image generation. The differences of submodule architectures as well as their inputs lead to heterogeneity in terms of computation efficiency. Failing to account for such heterogeneity, existing distributed training systems treat all submodules as a monolithic entity and thus have sub-optimal performance. Moreover, the outcome aggregation phase introduces cross-sample dependencies with contrasting positive and negative sample pairs (i.e., contrastive loss). Such dependencies make the existing pipeline parallelism scheduling algorithms not applicable for multimodal training with contrastive loss. To address the limitations of existing solutions, we propose DISTMM. For a given multimodal model, DISTMM exploits the heterogeneity among submodules, applying different distributed parallelism strategies for each submodule, e.g., using Tensor Parallelism for a computation-intensive submodule, and Data Parallelism for a submodule with a small number of parameters. DISTMM balances the computation of parallelized submodules to reduce the computing resource idle time of waiting for the slowest submodule. DISTMM further optimizes the locality of submodules by leveraging the heterogeneous bandwidth of interconnections among accelerators. To address the limitation of existing pipeline execution schedules, we propose a new pipeline execution primitive, called batch-sync instruction, and a corresponding schedule, called DISTMM-Pipe. We build a prototype of DISTMM and evaluate it with existing solutions on models with various sizes ranging from 1.1 billion to 26 billion parameters and observe 1.32-3.27× speedup over Megatron-LM.

The Mixture-of-Expert (MoE) technique plays a crucial role in expanding the size of DNN model parameters. However, it faces the challenge of extended all-to-all communication latency during the training process. Existing methods attempt to mitigate this issue by overlapping all-to-all with expert computation. Yet, these methods frequently fall short of achieving sufficient overlap, consequently restricting the potential for performance enhancements. In our study, we extend the scope of this challenge by considering overlap at the broader training graph level. During the forward pass, we enable non-MoE computations to overlap with all-to-all through careful partitioning and pipelining. In the backward pass, we achieve overlap with all-to-all by scheduling gradient weight computations. We implement these techniques in Lancet, a system using compiler-based optimization to automatically enhance MoE model training. Our extensive evaluation reveals that Lancet significantly reduces the time devoted to non-overlapping communication, by as much as 77 Moreover, it achieves a notable end-to-end speedup of up to 1.3 times when compared to the state-of-the-art solutions.

Multi-task model training has been adopted to enable a single deep neural network model (often a large language model) to handle multiple tasks (e.g., question answering and text summarization). Multi-task training commonly receives input sequences of highly different lengths due to the diverse contexts of different tasks. Padding (to the same sequence length) or packing (short examples into long sequences of the same length) is usually adopted to prepare input samples for model training, which is nonetheless not space or computation efficient. This paper proposes a dynamic micro-batching approach to tackle sequence length variation and enable efficient multi-task model training. We advocate pipeline-parallel training of the large model with variable-length micro-batches, each of which potentially comprises a different number of samples. We optimize micro-batch construction using a dynamic programming-based approach, and handle micro-batch execution time variation through dynamic pipeline and communication scheduling, enabling highly efficient pipeline training. Extensive evaluation on the FLANv2 dataset demonstrates up to 4.39x higher training throughput when training T5, and 3.25x when training GPT, as compared with packing-based baselines. DynaPipe's source code is publicly available at https://github.com/awslabs/optimizing-multitask-training-through-dynamic-pipelines.

Transformers are central in modern natural language processing and computer vision applications. Despite recent works devoted to reducing the quadratic cost of such models (as a function of the sequence length), dealing with ultra long sequences (e.g., with more than 16K tokens) remains challenging. Applications such as answering questions based on a book or summarizing a scientific article are inefficient or infeasible. Here, we propose to significantly improve the efficiency of Transformers for ultra long sequences, by compressing the sequence into a much smaller representation at each layer. Specifically, by exploiting the fact that in many tasks, only a small subset of special tokens (we call VIP-tokens) are most relevant to the final prediction, we propose a VIP-token centric compression (VCC) scheme which selectively compresses the sequence based on their impact on approximating the representation of the VIP-tokens. Compared with competitive baselines, our algorithm is not only efficient (achieving more than $3\times$ efficiency gain compared to baselines on 4K and 16K lengths), but also offers competitive/better performance on a large number of tasks. Further, we show that our algorithm scales to 128K tokens (or more) while consistently offering accuracy improvement.

Large deep learning models have recently garnered substantial attention from both academia and industry. Nonetheless, frequent failures are observed during large model training due to large-scale resources involved and extended training time. Existing solutions have significant failure recovery costs due to the severe restriction imposed by the bandwidth of remote storage in which they store checkpoints. This paper presents Gemini, a distributed training system that enables fast failure recovery for large model training by checkpointing to CPU memory of the host machines with much larger aggregated bandwidth. However, two challenges prevent naïvely checkpointing to CPU memory. First, the availability of checkpoints in CPU memory cannot be guaranteed when failures occur. Second, since the communication traffic for training and checkpointing share the same network, checkpoint traffic can interfere with training traffic and harm training throughput. To address these two challenges, this paper proposes: 1) a provably near-optimal checkpoint placement strategy to maximize the probability of failure recovery from checkpoints in CPU memory; and 2) a checkpoint traffic scheduling algorithm to minimize, if not eliminate, the interference of checkpoint traffic on model training. Our evaluation shows that overall Gemini achieves a faster failure recovery by more than 13× than existing solutions. Moreover, it achieves optimal checkpoint frequency, i.e., every iteration, and incurs no overhead on training throughput for large model training.

This work proposes POMP, a prompt pre-training method for vision-language models. Being memory and computation efficient, POMP enables the learned prompt to condense semantic information for a rich set of visual concepts with over twenty-thousand classes. Once pre-trained, the prompt with a strong transferable ability can be directly plugged into a variety of visual recognition tasks including image classification, semantic segmentation, and object detection, to boost recognition performances in a zero-shot manner. Empirical evaluation shows that POMP achieves state-of-the-art performances on 21 datasets, e.g., 67.0% average accuracy on 10 classification datasets (+3.1% compared to CoOp) and 84.4 hIoU on open-vocabulary Pascal VOC segmentation (+6.9 compared to ZSSeg). Our code is available at https://github.com/amazon-science/prompt-pretraining.

Recent years have seen an increase in the development of large deep learning (DL) models, which makes training effi-ciency crucial. Common practice is struggling with the trade-off between usability and performance. On one hand, DL frameworks such as PyTorch use dynamic graphs to facilitate model developers at a price of sub-optimal model training performance. On the other hand, practitioners propose various approaches to improving the training efficiency by sacrificing some of the flexibility, ranging from making the graph static for more thorough optimization (e.g., XLA) to customizing optimization towards large-scale distributed training (e.g., DeepSpeed and Megatron-LM). In this paper, we aim to address the tension between usability and training efficiency through separation of concerns. Inspired by DL compilers that decouple the platform-specific optimizations of a tensor-level operator from its arithmetic definition, this paper proposes a schedule language, Slapo, to decouple model execution from definition. Specifically, Slapo works on a PyTorch model and uses a set of schedule primitives to convert the model for common model training optimizations such as high-performance kernels, effective 3D parallelism, and efficient activation checkpointing. Compared to existing optimization solutions, Slapo progressively optimizes the model “as-needed” through high-level primitives, and thus preserving programmability and debuggability for users to a large extent. Our evaluation results show that by scheduling the existing hand-crafted optimizations in a systematic way using Slapo, we are able

Pre-trained large language models can efficiently interpolate human-written prompts in a natural way. Multitask prompted learning can help generalization through a diverse set of tasks at once, thus enhancing the potential for more effective downstream fine-tuning. To perform efficient multitask-inference in the same batch, parameter-efficient fine-tuning methods such as prompt tuning have been proposed. However, the existing prompt tuning methods may lack generalization. We propose SPT, a semi-parametric prompt tuning method for multitask prompted learning. The novel component of SPT is a memory bank from where memory prompts are retrieved based on discrete prompts. Extensive experiments, such as (i) fine-tuning a full language model with SPT on 31 different tasks from 8 different domains and evaluating zero-shot generalization on 9 heldout datasets under 5 NLP task categories and (ii) pretraining SPT on the GLUE datasets and evaluating fine-tuning on the SuperGLUE datasets, demonstrate effectiveness of SPT.

The mixture of Expert (MoE) parallelism is a recent advancement that scales up the model size with constant computational cost. MoE selects different sets of parameters (i.e., experts) for each incoming token, resulting in a sparsely-activated model. Despite several successful applications of MoE, its training efficiency degrades significantly as the number of experts increases. The routing stage in MoE relies on the efficiency of the All2All communication collective, which suffers from network congestion and has poor scalability. To mitigate these issues, we introduce SMILE, which exploits heterogeneous network bandwidth and splits a single-step routing into bi-level routing. Our experimental results show that the proposed method obtains a 2.5x speedup over Switch Transformer in terms of pretraining throughput on the Colossal Clean Crawled Corpus without losing any convergence speed.

Existing out-of-distribution (OOD) detection methods are typically benchmarked on training sets with balanced class distributions. However, in real-world applications, it is common for the training sets to have long-tailed distributions. In this work, we first demonstrate that existing OOD detection methods commonly suffer from significant performance degradation when the training set is long-tail distributed. Through analysis, we posit that this is because the models struggle to distinguish the minority tail-class in-distribution samples, from the true OOD samples, making the tail classes more prone to be falsely detected as OOD. To solve this problem, we propose Partial and Asymmetric Supervised Contrastive Learning (PASCL), which explicitly encourages the model to distinguish between tail-class in-distribution samples and OOD samples. To further boost in-distribution classification accuracy, we propose Auxiliary Branch Finetuning, which uses two separate branches of BN and classification layers for anomaly detection and in-distribution classification, respectively. The intuition is that in-distribution and OOD anomaly data have different underlying distributions. Our method outperforms previous state-of-the-art method by 1.29%, 1.45%, 0.69% anomaly detection false positive rate (FPR) and 3.24%, 4.06%, 7.89% in-distribution classification accuracy on CIFAR10-LT, CIFAR100-LT, and ImageNet-LT, respectively. Code and pre-trained models are available at https://github.com/amazon-research/long-tailed-ood-detection.

On-device memory concerns in distributed deep learning have become severe due to (i) the growth of model size in multi-GPU training, and (ii) the wide adoption of deep neural networks for federated learning on IoT devices which have limited storage. In such settings, communication efficient optimization methods are attractive alternatives, however they still struggle with memory issues. To tackle these challenges, we propose an communication efficient method called contractive error feedback (ConEF). As opposed to SGD with error-feedback (EFSGD) that inefficiently manages memory, ConEF obtains the sweet spot of convergence and memory usage, and achieves communication efficiency by leveraging biased and all-reducable gradient compression. We empirically validate ConEF on various learning tasks that include image classification, language modeling, and machine translation and observe that ConEF saves 80\% - 90\% of the extra memory in EFSGD with almost no loss on test performance, while also achieving 1.3x - 5x speedup of SGD. Through our work, we also demonstrate the feasibility and convergence of ConEF to clear up the theoretical barrier of integrating ConEF to popular memory efficient frameworks such as ZeRO-3.

In this paper we present a Click-through-rate (CTR) prediction model for product advertisement at Amazon. CTR prediction is challenging because the model needs to a) learn from text and numeric features, b) maintain low-latency at inference time, and c) adapt to a temporal advertisement distribution shift. Our proposed model is DCAF-BERT, a novel lightweight cache-friendly factorized model that consists of twin-structured BERT-like encoders for text with a mechanism for late fusion for tabular and numeric features. The factorization of the model allows for compartmentalised retraining which enables the model to easily adapt to distribution shifts. The twin encoders are carefully trained to leverage historical CTR data, using a large pre-trained language model and cross-architecture knowledge distillation (KD). We empirically find the right combination of pretraining, distillation and fine-tuning strategies for teacher and student which leads to a 1.7% ROC-AUC lift over the previous best model offline. In an online experiment we show that our compartmentalised refresh strategy boosts the CTR of DCAF-BERT by 3.6% on average over the baseline model consistently across a month.

Existing general purpose frameworks for gigantic model training, i.e., dense models with billions of parameters, cannot scale efficiently on cloud environment with various networking conditions due to large communication overheads. In this paper, we propose MiCS, which Minimizes the Communication Scale to bring down communication overhead. Specifically, by decreasing the number of participants in a communication collective, MiCS can utilize heterogeneous network bandwidth, reduce network traffic over slower links, reduce the latency of communications for maintaining high network bandwidth utilization, and amortize expensive global gradient synchronization overhead. Our evaluation on AWS shows that the system throughput of MiCS is up to 2.89× that of the state-of-the-art large model training systems. MiCS achieves near-linear scaling efficiency, which is up to 1.27× that of DeepSpeed. MiCS allows us to train a proprietary model with 100 billion parameters on 512 GPUs with 99.4% weak-scaling efficiency, and it is able to saturate over 54.5% theoretical computation power of each GPU on a public cloud with less GPU memory and more restricted networks than DGX-A100 clusters.

Adversarial training (AT) defends deep neural networks against adversarial attacks. One challenge that limits its practical application is the performance degradation on clean samples. A major bottleneck identified by previous works is the widely used batch normalization (BN), which struggles to model the different statistics of clean and adversarial training samples in AT. Although the dominant approach is to extend BN to capture this mixture of distribution, we propose to completely eliminate this bottleneck by removing all BN layers in AT. Our normalizer-free robust training (NoFrost) method extends recent advances in normalizer-free networks to AT for its unexplored advantage on handling the mixture distribution challenge. We show that NoFrost achieves adversarial robustness with only a minor sacrifice on clean sample accuracy. On ImageNet with ResNet50, NoFrost achieves 74.06% clean accuracy, which drops merely 2.00% from standard training. In contrast, BN-based AT obtains 59.28% clean accuracy, suffering a significant 16.78% drop from standard training. In addition, NoFrost achieves a 23.56% adversarial robustness against PGD attack, which improves the 13.57% robustness in BN-based AT. We observe better model smoothness and larger decision margins from NoFrost, which make the models less sensitive to input perturbations and thus more robust. Moreover, when incorporating more data augmentations into NoFrost, it achieves comprehensive robustness against multiple distribution shifts. Code and pre-trained models are public at https://github.com/amazon-research/normalizer-free-robust-training.

We present results from a large-scale experiment on pretraining encoders with non-embedding parameter counts ranging from 700M to 9.3B, their subsequent distillation into smaller models ranging from 17M-170M parameters, and their application to the Natural Language Understanding (NLU) component of a virtual assistant system. Though we train using 70% spoken-form data, our teacher models perform comparably to XLM-R and mT5 when evaluated on the written-form Cross-lingual Natural Language Inference (XNLI) corpus. We perform a second stage of pretraining on our teacher models using in-domain data from our system, improving error rates by 3.86% relative for intent classification and 7.01% relative for slot filling. We find that even a 170M-parameter model distilled from our Stage 2 teacher model has 2.88% better intent classification and 7.69% better slot filling error rates when compared to the 2.3B-parameter teacher trained only on public data (Stage 1), emphasizing the importance of in-domain data for pretraining. When evaluated offline using labeled NLU data, our 17M-parameter Stage 2 distilled model outperforms both XLM-R Base (85M params) and DistillBERT (42M params) by 4.23% to 6.14%, respectively. Finally, we present results from a full virtual assistant experimentation platform, where we find that models trained using our pretraining and distillation pipeline outperform models distilled from 85M-parameter teachers by 3.74%-4.91% on an automatic measurement of full-system user dissatisfaction.

We present results from a large-scale experiment on pretraining encoders with non-embedding parameter counts ranging from 700M to 9.3B, their subsequent distillation into smaller models ranging from 17M-170M parameters, and their application to the Natural Language Understanding (NLU) component of a virtual assistant system. Though we train using 70% spoken-form data, our teacher models perform comparably to XLM-R and mT5 when evaluated on the written-form Cross-lingual Natural Language Inference (XNLI) corpus. We perform a second stage of pretraining on our teacher models using in-domain data from our system, improving error rates by 3.86% relative for intent classification and 7.01% relative for slot filling. We find that even a 170M-parameter model distilled from our Stage 2 teacher model has 2.88% better intent classification and 7.69% better slot filling error rates when compared to the 2.3B-parameter teacher trained only on public data (Stage 1), emphasizing the importance of in-domain data for pretraining. When evaluated offline using labeled NLU data, our 17M-parameter Stage 2 distilled model outperforms both XLM-R Base (85M params) and DistillBERT (42M params) by 4.23% to 6.14%, respectively. Finally, we present results from a full virtual assistant experimentation platform, where we find that models trained using our pretraining and distillation pipeline outperform models distilled from 85M-parameter teachers by 3.74%-4.91% on an automatic measurement of full-system user dissatisfaction.

Communication overhead severely hinders the scalability of distributed machine learning systems. Recently, there has been a growing interest in using gradient compression to reduce the communication overhead of the distributed training. However, there is little understanding of applying gradient compression to adaptive gradient methods. Moreover, its performance benefits are often limited by the non-negligible compression overhead. In this paper, we first introduce a novel adaptive gradient method with gradient compression. We show that the proposed method has a convergence rate of $\mathcal{O}(1/\sqrt{T})$ for non-convex problems. In addition, we develop a scalable system called BytePS-Compress for two-way compression, where the gradients are compressed in both directions between workers and parameter servers. BytePS-Compress pipelines the compression and decompression on CPUs and achieves a high degree of parallelism. Empirical evaluations show that we improve the training time of ResNet50, VGG16, and BERT-base by 5.0%, 58.1%, 23.3%, respectively, without any accuracy loss with 25 Gb/s networking. Furthermore, for training the BERT models, we achieve a compression rate of 333x compared to the mixed-precision training.

Time series data have grown at an explosive rate in numerous domains and have stimulated a surge of time series modeling research. A comprehensive comparison of different time series models, for a considered data analytics task, provides useful guidance on model selection for data analytics practitioners. Data scarcity is a universal issue that occurs in a vast range of data analytics problems, due to the high costs associated with collecting, generating, and labeling data as well as some data quality issues such as missing data. In this paper, we focus on the temporal classification/regression problem that attempts to build a mathematical mapping from multivariate time series inputs to a discrete class label or a real-valued response variable. For this specific problem, we identify two types of scarce data: scarce data with small samples and scarce data with sparsely and irregularly observed time series covariates. Observing that all existing works are incapable of utilizing the sparse time series inputs for proper modeling building, we propose a model called sparse functional multilayer perceptron (SFMLP) for handling the sparsity in the time series covariates. The effectiveness of the proposed SFMLP under each of the two types of data scarcity, in comparison with the conventional deep sequential learning models (e.g., Recurrent Neural Network, and Long Short-Term Memory), is investigated through mathematical arguments and numerical experiments.