The immense evolution in Large Language Models (LLMs) has underscored the importance of massive, heterogeneous, and high-quality data. A data recipe is a mixture of data from different sources for training LLMs, which plays a vital role in LLMs' performance. Existing open-source tools for LLM data processing are mostly tailored for specific data recipes. To continuously uncover the potential of LLMs, incorporate data from new sources, and improve LLMs' performance, we build a new system named Data-Juicer, with which we can efficiently generate diverse data recipes, explore different possibilities in forming data mixtures, and evaluate their effects on model performance. Different from traditional data-analytics pipelines, Data-Juicer faces some unique challenges. Firstly, the possible data sources for forming data recipes are truly heterogeneous and massive with various qualities. Secondly, it is extremely expensive to precisely evaluate data recipes' impact on LLMs' performance. Thirdly, the end users of Data-Juicer, model developers, need sufficient flexibility to configure and evaluate different data recipes. Data-Juicer features a fine-grained abstraction of pipelines for constructing data recipes, with over 50 built-in operators for easy composition and extension. By incorporating visualization and auto-evaluation capabilities, Data-Juicer enables a timely feedback loop for both LLM pre-training and fine-tuning. Further, Data-Juicer is optimized and integrated with ecosystems for LLM training, evaluation, and distributed computing. The data recipes derived with Data-Juicer gain notable improvements on state-of-the-art LLMs, by up to 7.45% increase in averaged score across 16 LLM benchmarks and 17.5% higher win rate in pair-wise GPT-4 evaluations. Our system, data recipes, and tutorials are released, calling for broader data-centric research on training and understanding LLMs.
We propose a distributed bundle adjustment (DBA) method using the exact Levenberg-Marquardt (LM) algorithm for super large-scale datasets. Most of the existing methods partition the global map to small ones and conduct bundle adjustment in the submaps. In order to fit the parallel framework, they use approximate solutions instead of the LM algorithm. However, those methods often give sub-optimal results. Different from them, we utilize the exact LM algorithm to conduct global bundle adjustment where the formation of the reduced camera system (RCS) is actually parallelized and executed in a distributed way. To store the large RCS, we compress it with a block-based sparse matrix compression format (BSMC), which fully exploits its block feature. The BSMC format also enables the distributed storage and updating of the global RCS. The proposed method is extensively evaluated and compared with the state-of-the-art pipelines using both synthetic and real datasets. Preliminary results demonstrate the efficient memory usage and vast scalability of the proposed method compared with the baselines. For the first time, we conducted parallel bundle adjustment using LM algorithm on a real datasets with 1.18 million images and a synthetic dataset with 10 million images (about 500 times that of the state-of-the-art LM-based BA) on a distributed computing system.
Visual retrieval tasks such as image retrieval and person re-identification (Re-ID) aim at effectively and thoroughly searching images with similar content or the same identity. After obtaining retrieved examples, re-ranking is a widely adopted post-processing step to reorder and improve the initial retrieval results by making use of the contextual information from semantically neighboring samples. Prevailing re-ranking approaches update distance metrics and mostly rely on inefficient crosscheck set comparison operations while computing expanded neighbors based distances. In this work, we present an efficient re-ranking method which refines initial retrieval results by updating features. Specifically, we reformulate re-ranking based on Graph Convolution Networks (GCN) and propose a novel Graph Convolution based Re-ranking (GCR) for visual retrieval tasks via feature propagation. To accelerate computation for large-scale retrieval, a decentralized and synchronous feature propagation algorithm which supports parallel or distributed computing is introduced. In particular, the plain GCR is extended for cross-camera retrieval and an improved feature propagation formulation is presented to leverage affinity relationships across different cameras. It is also extended for video-based retrieval, and Graph Convolution based Re-ranking for Video (GCRV) is proposed by mathematically deriving a novel profile vector generation method for the tracklet. Without bells and whistles, the proposed approaches achieve state-of-the-art performances on seven benchmark datasets from three different tasks, i.e., image retrieval, person Re-ID and video-based person Re-ID.
Quantum processing units (QPUs) are currently exclusively available from cloud vendors. However, with recent advancements, hosting QPUs is soon possible everywhere. Existing work has yet to draw from research in edge computing to explore systems exploiting mobile QPUs, or how hybrid applications can benefit from distributed heterogeneous resources. Hence, this work presents an architecture for Quantum Computing in the edge-cloud continuum. We discuss the necessity, challenges, and solution approaches for extending existing work on classical edge computing to integrate QPUs. We describe how warm-starting allows defining workflows that exploit the hierarchical resources spread across the continuum. Then, we introduce a distributed inference engine with hybrid classical-quantum neural networks (QNNs) to aid system designers in accommodating applications with complex requirements that incur the highest degree of heterogeneity. We propose solutions focusing on classical layer partitioning and quantum circuit cutting to demonstrate the potential of utilizing classical and quantum computation across the continuum. To evaluate the importance and feasibility of our vision, we provide a proof of concept that exemplifies how extending a classical partition method to integrate quantum circuits can improve the solution quality. Specifically, we implement a split neural network with optional hybrid QNN predictors. Our results show that extending classical methods with QNNs is viable and promising for future work.
Given the ubiquity of non-separable optimization problems in real worlds, in this paper we analyze and extend the large-scale version of the well-known cooperative coevolution (CC), a divide-and-conquer black-box optimization framework, on non-separable functions. First, we reveal empirical reasons of when decomposition-based methods are preferred or not in practice on some non-separable large-scale problems, which have not been clearly pointed out in many previous CC papers. Then, we formalize CC to a continuous-game model via simplification, but without losing its essential property. Different from previous evolutionary game theory for CC, our new model provides a much simpler but useful viewpoint to analyze its convergence, since only the pure Nash equilibrium concept is needed and more general fitness landscapes can be explicitly considered. Based on convergence analyses, we propose a hierarchical decomposition strategy for better generalization, as for any decomposition, there is a risk of getting trapped into a suboptimal Nash equilibrium. Finally, we use powerful distributed computing to accelerate it under the recent multi-level learning framework, which combines the fine-tuning ability from decomposition with the invariance property of CMA-ES. Experiments on a set of high-dimensional test functions validate both its search performance and scalability (w.r.t. CPU cores) on a clustering computing platform with 400 CPU cores.
We study the private $k$-median and $k$-means clustering problem in $d$ dimensional Euclidean space. By leveraging tree embeddings, we give an efficient and easy to implement algorithm, that is empirically competitive with state of the art non private methods. We prove that our method computes a solution with cost at most $O(d^{3/2}\log n)\cdot OPT + O(k d^2 \log^2 n / \epsilon^2)$, where $\epsilon$ is the privacy guarantee. (The dimension term, $d$, can be replaced with $O(\log k)$ using standard dimension reduction techniques.) Although the worst-case guarantee is worse than that of state of the art private clustering methods, the algorithm we propose is practical, runs in near-linear, $\tilde{O}(nkd)$, time and scales to tens of millions of points. We also show that our method is amenable to parallelization in large-scale distributed computing environments. In particular we show that our private algorithms can be implemented in logarithmic number of MPC rounds in the sublinear memory regime. Finally, we complement our theoretical analysis with an empirical evaluation demonstrating the algorithm's efficiency and accuracy in comparison to other privacy clustering baselines.
While modern internet services, such as chatbots, search engines, and online advertising, demand the use of large-scale deep neural networks (DNNs), distributed training and inference over heterogeneous computing systems are desired to facilitate these DNN models. Mixture-of-Experts (MoE) is one the most common strategies to lower the cost of training subject to the overall size of models/data through gating and parallelism in a divide-and-conquer fashion. While DeepSpeed has made efforts in carrying out large-scale MoE training over heterogeneous infrastructures, the efficiency of training and inference could be further improved from several system aspects, including load balancing, communication/computation efficiency, and memory footprint limits. In this work, we present a novel MoESys that boosts efficiency in both large-scale training and inference. Specifically, in the training procedure, the proposed MoESys adopts an Elastic MoE training strategy with 2D prefetch and Fusion communication over Hierarchical storage, so as to enjoy efficient parallelisms. For scalable inference in a single node, especially when the model size is larger than GPU memory, MoESys builds the CPU-GPU memory jointly into a ring of sections to load the model, and executes the computation tasks across the memory sections in a round-robin manner for efficient inference. We carried out extensive experiments to evaluate MoESys, where MoESys successfully trains a Unified Feature Optimization (UFO) model with a Sparsely-Gated Mixture-of-Experts model of 12B parameters in 8 days on 48 A100 GPU cards. The comparison against the state-of-the-art shows that MoESys outperformed DeepSpeed with 33% higher throughput (tokens per second) in training and 13% higher throughput in inference in general. Particularly, under unbalanced MoE Tasks, e.g., UFO, MoESys achieved 64% higher throughput with 18% lower memory footprints.
The ever-growing model size and scale of compute have attracted increasing interests in training deep learning models over multiple nodes. However, when it comes to training on cloud clusters, especially across remote clusters, huge challenges are faced. In this work, we introduce a general framework, Nebula-I, for collaboratively training deep learning models over remote heterogeneous clusters, the connections between which are low-bandwidth wide area networks (WANs). We took natural language processing (NLP) as an example to show how Nebula-I works in different training phases that include: a) pre-training a multilingual language model using two remote clusters; and b) fine-tuning a machine translation model using knowledge distilled from pre-trained models, which run through the most popular paradigm of recent deep learning. To balance the accuracy and communication efficiency, in Nebula-I, parameter-efficient training strategies, hybrid parallel computing methods and adaptive communication acceleration techniques are jointly applied. Meanwhile, security strategies are employed to guarantee the safety, reliability and privacy in intra-cluster computation and inter-cluster communication. Nebula-I is implemented with the PaddlePaddle deep learning framework, which can support collaborative training over heterogeneous hardware, e.g. GPU and NPU. Experiments demonstrate that the proposed framework could substantially maximize the training efficiency while preserving satisfactory NLP performance. By using Nebula-I, users can run large-scale training tasks over cloud clusters with minimum developments, and the utility of existed large pre-trained models could be further promoted. We also introduced new state-of-the-art results on cross-lingual natural language inference tasks, which are generated based upon a novel learning framework and Nebula-I.
Most AI projects start with a Python notebook running on a single laptop; however, one usually needs to go through a mountain of pains to scale it to handle larger dataset (for both experimentation and production deployment). These usually entail many manual and error-prone steps for the data scientists to fully take advantage of the available hardware resources (e.g., SIMD instructions, multi-processing, quantization, memory allocation optimization, data partitioning, distributed computing, etc.). To address this challenge, we have open sourced BigDL 2.0 at https://github.com/intel-analytics/BigDL/ under Apache 2.0 license (combining the original BigDL and Analytics Zoo projects); using BigDL 2.0, users can simply build conventional Python notebooks on their laptops (with possible AutoML support), which can then be transparently accelerated on a single node (with up-to 9.6x speedup in our experiments), and seamlessly scaled out to a large cluster (across several hundreds servers in real-world use cases). BigDL 2.0 has already been adopted by many real-world users (such as Mastercard, Burger King, Inspur, etc.) in production.
For intelligent home IoT services with sensors and machine learning, we need to upload IoT data to the cloud server which cannot share private data for training. A recent machine learning approach, called federated learning, keeps user data on the device in the distributed computing environment. Though federated learning is useful for protecting privacy, it experiences poor performance in terms of the end-to-end response time in home IoT services, because IoT devices are usually controlled by remote servers in the cloud. In addition, it is difficult to achieve the high accuracy of federated learning models due to insufficient data problems and model inversion attacks. In this paper, we propose a local IoT control method for a federated learning home service that recognizes the user behavior in the home network quickly and accurately. We present a federated learning client with transfer learning and differential privacy to solve data scarcity and data model inversion attack problems. From experiments, we show that the local control of home IoT devices for user authentication and control message transmission by the federated learning clients improves the response time to less than 1 second. Moreover, we demonstrate that federated learning with transfer learning achieves 97% of accuracy under 9,000 samples, which is only 2% of the difference from centralized learning.
Federated learning is proposed by Google to safeguard data privacy through training models locally on users' devices. However, with deep learning models growing in size to achieve better results, it becomes increasingly difficult to accommodate the whole model on one single device. Thus, model parallelism is then used to divide the model weights among several devices. With this logic, the approach currently used evenly allocates weights among devices. However, in reality, a computation bottleneck may occur resulting from variant computing power of different users' devices. To address this problem, load balancing is needed to allocate the model weights based on the computational capability of the device. In this paper, we proposed Sky Computing, a load-balanced model parallelism framework to adaptively allocate the weights to devices. Sky Computing outperforms the baseline method by 55% in training time when training 160-layer BERT with 64 nodes. The source code can be found at https://github.com/hpcaitech/SkyComputing.
Most multi-agent reinforcement learning (MARL) methods are limited in the scale of problems they can handle. With increasing numbers of agents, the number of training iterations required to find the optimal behaviors increases exponentially due to the exponentially growing joint state and action spaces. This paper tackles this limitation by introducing a scalable MARL method called Distributed multi-Agent Reinforcement Learning with One-hop Neighbors (DARL1N). DARL1N is an off-policy actor-critic method that addresses the curse of dimensionality by restricting information exchanges among the agents to one-hop neighbors when representing value and policy functions. Each agent optimizes its value and policy functions over a one-hop neighborhood, significantly reducing the learning complexity, yet maintaining expressiveness by training with varying neighbor numbers and states. This structure allows us to formulate a distributed learning framework to further speed up the training procedure. Distributed computing systems, however, contain straggler compute nodes, which are slow or unresponsive due to communication bottlenecks, software or hardware problems. To mitigate the detrimental straggler effect, we introduce a novel coded distributed learning architecture, which leverages coding theory to improve the resilience of the learning system to stragglers. Comprehensive experiments show that DARL1N significantly reduces training time without sacrificing policy quality and is scalable as the number of agents increases. Moreover, the coded distributed learning architecture improves training efficiency in the presence of stragglers.
Offering incentives (e.g., coupons at Amazon, discounts at Uber and video bonuses at Tiktok) to user is a common strategy used by online platforms to increase user engagement and platform revenue. Despite its proven effectiveness, these marketing incentives incur an inevitable cost and might result in a low ROI (Return on Investment) if not used properly. On the other hand, different users respond differently to these incentives, for instance, some users never buy certain products without coupons, while others do anyway. Thus, how to select the right amount of incentives (i.e. treatment) to each user under budget constraints is an important research problem with great practical implications. In this paper, we call such problem as a budget-constrained treatment selection (BTS) problem. The challenge is how to efficiently solve BTS problem on a Large-Scale dataset and achieve improved results over the existing techniques. We propose a novel tree-based treatment selection technique under budget constraints, called Large-Scale Budget-Constrained Causal Forest (LBCF) algorithm, which is also an efficient treatment selection algorithm suitable for modern distributed computing systems. A novel offline evaluation method is also proposed to overcome an intrinsic challenge in assessing solutions' performance for BTS problem in randomized control trials (RCT) data. We deploy our approach in a real-world scenario on a large-scale video platform, where the platform gives away bonuses in order to increase users' campaign engagement duration. The simulation analysis, offline and online experiments all show that our method outperforms various tree-based state-of-the-art baselines. The proposed approach is currently serving over hundreds of millions of users on the platform and achieves one of the most tremendous improvements over these months.
Workflow scheduling is a long-studied problem in parallel and distributed computing (PDC), aiming to efficiently utilize compute resources to meet user's service requirements. Recently proposed scheduling methods leverage the low response times of edge computing platforms to optimize application Quality of Service (QoS). However, scheduling workflow applications in mobile edge-cloud systems is challenging due to computational heterogeneity, changing latencies of mobile devices and the volatile nature of workload resource requirements. To overcome these difficulties, it is essential, but at the same time challenging, to develop a long-sighted optimization scheme that efficiently models the QoS objectives. In this work, we propose MCDS: Monte Carlo Learning using Deep Surrogate Models to efficiently schedule workflow applications in mobile edge-cloud computing systems. MCDS is an Artificial Intelligence (AI) based scheduling approach that uses a tree-based search strategy and a deep neural network-based surrogate model to estimate the long-term QoS impact of immediate actions for robust optimization of scheduling decisions. Experiments on physical and simulated edge-cloud testbeds show that MCDS can improve over the state-of-the-art methods in terms of energy consumption, response time, SLA violations and cost by at least 6.13, 4.56, 45.09 and 30.71 percent respectively.
Deep neural networks (DNNs) exploit many layers and a large number of parameters to achieve excellent performance. The training process of DNN models generally handles large-scale input data with many sparse features, which incurs high Input/Output (IO) cost, while some layers are compute-intensive. The training process generally exploits distributed computing resources to reduce training time. In addition, heterogeneous computing resources, e.g., CPUs, GPUs of multiple types, are available for the distributed training process. Thus, the scheduling of multiple layers to diverse computing resources is critical for the training process. To efficiently train a DNN model using the heterogeneous computing resources, we propose a distributed framework, i.e., Paddle-Heterogeneous Parameter Server (Paddle-HeterPS), composed of a distributed architecture and a Reinforcement Learning (RL)-based scheduling method. The advantages of Paddle-HeterPS are three-fold compared with existing frameworks. First, Paddle-HeterPS enables efficient training process of diverse workloads with heterogeneous computing resources. Second, Paddle-HeterPS exploits an RL-based method to efficiently schedule the workload of each layer to appropriate computing resources to minimize the cost while satisfying throughput constraints. Third, Paddle-HeterPS manages data storage and data communication among distributed computing resources. We carry out extensive experiments to show that Paddle-HeterPS significantly outperforms state-of-the-art approaches in terms of throughput (14.5 times higher) and monetary cost (312.3% smaller). The codes of the framework are publicly available at: https://github.com/PaddlePaddle/Paddle.
Classic and deep generalized canonical correlation analysis (GCCA) algorithms seek low-dimensional common representations of data entities from multiple ``views'' (e.g., audio and image) using linear transformations and neural networks, respectively. When the views are acquired and stored at different computing agents (e.g., organizations and edge devices) and data sharing is undesired due to privacy or communication cost considerations, federated learning-based GCCA is well-motivated. In federated learning, the views are kept locally at the agents and only derived, limited information exchange with a central server is allowed. However, applying existing GCCA algorithms onto such federated learning settings may incur prohibitively high communication overhead. This work puts forth a communication-efficient federated learning framework for both linear and deep GCCA under the maximum variance (MAX-VAR) formulation. The overhead issue is addressed by aggressively compressing (via quantization) the exchanging information between the computing agents and a central controller. Compared to the unquantized version, our empirical study shows that the proposed algorithm enjoys a substantial reduction of communication overheads with virtually no loss in accuracy and convergence speed. Rigorous convergence analyses are also presented, which is a nontrivial effort. Generic federated optimization results do not cover the special problem structure of GCCA. Our result shows that the proposed algorithms for both linear and deep GCCA converge to critical points at a sublinear rate, even under heavy quantization and stochastic approximations. In addition, in the linear MAX-VAR case, the quantized algorithm approaches a global optimum in a geometric rate under reasonable conditions. Synthetic and real-data experiments are used to showcase the effectiveness of the proposed approach.
Data augmentation (DA) algorithms are widely used for Bayesian inference due to their simplicity. In massive data settings, however, DA algorithms are prohibitively slow because they pass through the full data in any iteration, imposing serious restrictions on their usage despite the advantages. Addressing this problem, we develop a framework for extending any DA that exploits asynchronous and distributed computing. The extended DA algorithm is indexed by a parameter $r \in (0, 1)$ and is called Asynchronous and Distributed (AD) DA with the original DA as its parent. Any ADDA starts by dividing the full data into $k$ smaller disjoint subsets and storing them on $k$ processes, which could be machines or processors. Every iteration of ADDA augments only an $r$-fraction of the $k$ data subsets with some positive probability and leaves the remaining $(1-r)$-fraction of the augmented data unchanged. The parameter draws are obtained using the $r$-fraction of new and $(1-r)$-fraction of old augmented data. For many choices of $k$ and $r$, the fractional updates of ADDA lead to a significant speed-up over the parent DA in massive data settings, and it reduces to the distributed version of its parent DA when $r=1$. We show that the ADDA Markov chain is Harris ergodic with the desired stationary distribution under mild conditions on the parent DA algorithm. We demonstrate the numerical advantages of the ADDA in three representative examples corresponding to different kinds of massive data settings encountered in applications. In all these examples, our DA generalization is significantly faster than its parent DA algorithm for all the choices of $k$ and $r$. We also establish geometric ergodicity of the ADDA Markov chain for all three examples, which in turn yields asymptotically valid standard errors for estimates of desired posterior quantities.
Population-based multi-agent reinforcement learning (PB-MARL) refers to the series of methods nested with reinforcement learning (RL) algorithms, which produces a self-generated sequence of tasks arising from the coupled population dynamics. By leveraging auto-curricula to induce a population of distinct emergent strategies, PB-MARL has achieved impressive success in tackling multi-agent tasks. Despite remarkable prior arts of distributed RL frameworks, PB-MARL poses new challenges for parallelizing the training frameworks due to the additional complexity of multiple nested workloads between sampling, training and evaluation involved with heterogeneous policy interactions. To solve these problems, we present MALib, a scalable and efficient computing framework for PB-MARL. Our framework is comprised of three key components: (1) a centralized task dispatching model, which supports the self-generated tasks and scalable training with heterogeneous policy combinations; (2) a programming architecture named Actor-Evaluator-Learner, which achieves high parallelism for both training and sampling, and meets the evaluation requirement of auto-curriculum learning; (3) a higher-level abstraction of MARL training paradigms, which enables efficient code reuse and flexible deployments on different distributed computing paradigms. Experiments on a series of complex tasks such as multi-agent Atari Games show that MALib achieves throughput higher than 40K FPS on a single machine with $32$ CPU cores; 5x speedup than RLlib and at least 3x speedup than OpenSpiel in multi-agent training tasks. MALib is publicly available at https://github.com/sjtu-marl/malib.
We study optimization methods to train local (or personalized) models for decentralized collections of local datasets with an intrinsic network structure. This network structure arises from domain-specific notions of similarity between local datasets. Examples for such notions include spatio-temporal proximity, statistical dependencies or functional relations. Our main conceptual contribution is to formulate federated learning as generalized total variation (GTV) minimization. This formulation unifies and considerably extends existing federated learning methods. It is highly flexible and can be combined with a broad range of parametric models, including generalized linear models or deep neural networks. Our main algorithmic contribution is a fully decentralized federated learning algorithm. This algorithm is obtained by applying an established primal-dual method to solve GTV minimization. It can be implemented as message passing and is robust against inexact computations that arise from limited computational resources including processing time or bandwidth. Our main analytic contribution is an upper bound on the deviation between the local model parameters learnt by our algorithm and an oracle-based clustered federated learning method. This upper bound reveals conditions on the local models and the network structure of local datasets such that GTV minimization is able to pool (nearly) homogeneous local datasets.
Virtual Reality (VR) is becoming ubiquitous with the rise of consumer displays and commercial VR platforms. Such displays require low latency and high quality rendering of synthetic imagery with reduced compute overheads. Recent advances in neural rendering showed promise of unlocking new possibilities in 3D computer graphics via image-based representations of virtual or physical environments. Specifically, the neural radiance fields (NeRF) demonstrated that photo-realistic quality and continuous view changes of 3D scenes can be achieved without loss of view-dependent effects. While NeRF can significantly benefit rendering for VR applications, it faces unique challenges posed by high field-of-view, high resolution, and stereoscopic/egocentric viewing, typically causing low quality and high latency of the rendered images. In VR, this not only harms the interaction experience but may also cause sickness. To tackle these problems toward six-degrees-of-freedom, egocentric, and stereo NeRF in VR, we present the first gaze-contingent 3D neural representation and view synthesis method. We incorporate the human psychophysics of visual- and stereo-acuity into an egocentric neural representation of 3D scenery. We then jointly optimize the latency/performance and visual quality while mutually bridging human perception and neural scene synthesis to achieve perceptually high-quality immersive interaction. We conducted both objective analysis and subjective studies to evaluate the effectiveness of our approach. We find that our method significantly reduces latency (up to 99% time reduction compared with NeRF) without loss of high-fidelity rendering (perceptually identical to full-resolution ground truth). The presented approach may serve as the first step toward future VR/AR systems that capture, teleport, and visualize remote environments in real-time.