论文检索

Lifang Chen, Zhichao Chen, Liqi Yan, Yanyu Cheng, Fangli Guan, Pan Li

The computational power required for training deep learning models has been skyrocketing in the past decade as they scale with big data, and has become a very expensive and scarce resource. Therefore, distributed training, which can leverage distributed available computational power, is vital for efficient large-scale model training. However, most previous distributed training frameworks like DDP and DeepSpeed are primarily designed for co-located clusters under homogeneous computing and communication conditions, and hence cannot account for geo-distributed clusters with both computing and communication heterogeneity. To address this challenge, we develop a new data parallel based distributed training framework called Co-Adaptive Data Parallelism (C-ADP). First, we consider a data owner and parameter server that distributes data to and coordinates the collaborative learning across all the computing devices. We employ local training and delayed parameter synchronization to reduce communication costs. Second, we formulate a data parallel scheduling optimization problem to minimize the training time by optimizing data distribution. Third, we devise an efficient algorithm to solve this scheduling problem, and formally prove that the obtained solution is optimal in the asymptotic sense. Experiments on the ImageNet100 dataset demonstrate that C-ADP achieves fast convergence in heterogeneous distributed training environments. Compared to Distributed Data Parallel (DDP) and DeepSpeed, C-ADP achieves 21.6 times and 26.3 times improvements in FLOPS, respectively, and a reduction in training time of about 72% and 47%, respectively.

WenZheng Zhang, Yang Hu, Jing Shi, Xiaoying Bai

Scaling Deep Neural Networks (DNNs) requires significant computational resources in terms of GPU quantity and compute capacity. In practice, there usually exists a large number of heterogeneous GPU devices due to the rapid release cycle of GPU products. It is highly needed to efficiently and economically harness the power of heterogeneous GPUs, so that it can meet the requirements of DNN research and development. The paper introduces Poplar, a distributed training system that extends Zero Redundancy Optimizer (ZeRO) with heterogeneous-aware capabilities. We explore a broader spectrum of GPU heterogeneity, including compute capability, memory capacity, quantity and a combination of them. In order to achieve high computational efficiency across all heterogeneous conditions, Poplar conducts fine-grained measurements of GPUs in each ZeRO stage. We propose a novel batch allocation method and a search algorithm to optimize the utilization of heterogeneous GPUs clusters. Furthermore, Poplar implements fully automated parallelism, eliminating the need for deploying heterogeneous hardware and finding suitable batch size. Extensive experiments on three heterogeneous clusters, comprising six different types of GPUs, demonstrate that Poplar achieves a training throughput improvement of 1.02-3.92x over current state-of-the-art heterogeneous training systems.

Shuyu Dong, Michele Sebag, Kento Uemura, Akito Fujii, Shuang Chang, Yusuke Koyanagi, Koji Maruhashi

Causal learning tackles the computationally demanding task of estimating causal graphs. This paper introduces a new divide-and-conquer approach for causal graph learning, called DCILP. In the divide phase, the Markov blanket MB(Xi) of each variable Xi is identified, and causal learning subproblems associated with each MB(Xi) are independently addressed in parallel. This approach benefits from a more favorable ratio between the number of data samples and the number of variables considered. In counterpart, it can be adversely affected by the presence of hidden confounders, as variables external to MB(Xi) might influence those within it. The reconciliation of the local causal graphs generated during the divide phase is a challenging combinatorial optimization problem, especially in large-scale applications. The main novelty of DCILP is an original formulation of this reconciliation as an integer linear programming (ILP) problem, which can be delegated and efficiently handled by an ILP solver. Through experiments on medium to large scale graphs, and comparisons with state-of-the-art methods, DCILP demonstrates significant improvements in terms of computational complexity, while preserving the learning accuracy on real-world problem and suffering at most a slight loss of accuracy on synthetic problems.

Xiaohan Zhang, Zhenyu Sun, Yukui Qiu, Junyan Su, Qi Liu

Currently, 3D rendering for large-scale free camera trajectories, namely, arbitrary input camera trajectories, poses significant challenges: 1) The distribution and observation angles of the cameras are irregular, and various types of scenes are included in the free trajectories; 2) Processing the entire point cloud and all images at once for large-scale scenes requires a substantial amount of GPU memory. This paper presents a Toy-GS method for accurately rendering large-scale free camera trajectories. Specifically, we propose an adaptive spatial division approach for free trajectories to divide cameras and the sparse point cloud of the entire scene into various regions according to camera poses. Training each local Gaussian in parallel for each area enables us to concentrate on texture details and minimize GPU memory usage. Next, we use the multi-view constraint and position-aware point adaptive control (PPAC) to improve the rendering quality of texture details. In addition, our regional fusion approach combines local and global Gaussians to enhance rendering quality with an increasing number of divided areas. Extensive experiments have been carried out to confirm the effectiveness and efficiency of Toy-GS, leading to state-of-the-art results on two public large-scale datasets as well as our SCUTic dataset. Our proposal demonstrates an enhancement of 1.19 dB in PSNR and conserves 7 G of GPU memory when compared to various benchmarks.

Linyang He, Qiaolin Wang, Xilin Jiang, Nima Mesgarani

Transformer-based speech language models (SLMs) have significantly improved neural speech recognition and understanding. While existing research has examined how well SLMs encode shallow acoustic and phonetic features, the extent to which SLMs encode nuanced syntactic and conceptual features remains unclear. By drawing parallels with linguistic competence assessments for large language models, this study is the first to systematically evaluate the presence of contextual syntactic and semantic features across SLMs for self-supervised learning (S3M), automatic speech recognition (ASR), speech compression (codec), and as the encoder for auditory large language models (AudioLLMs). Through minimal pair designs and diagnostic feature analysis across 71 tasks spanning diverse linguistic levels, our layer-wise and time-resolved analysis uncovers that 1) all speech encode grammatical features more robustly than conceptual ones. 2) Despite never seeing text, S3M match or surpass ASR encoders on every linguistic level, demonstrating that rich grammatical and even conceptual knowledge can arise purely from audio. 3) S3M representations peak mid-network and then crash in the final layers, whereas ASR and AudioLLM encoders maintain or improve, reflecting how pre-training objectives reshape late-layer content. 4) Temporal probing further shows that S3Ms encode grammatical cues 500 ms before a word begins, whereas AudioLLMs distribute evidence more evenly—indicating that objectives shape not only where but also when linguistic information is most salient. Together, these findings establish the first large-scale map of contextual syntax and semantics in speech models and highlight both the promise and the limits of current SLM training paradigms.

Optimization · Large Scale, Parallel and Distributed

Jiarui Fang, Jinzhe Pan, Aoyu Li, Xibo Sun, WANG Jiannan

This paper presents PipeFusion, an innovative parallel methodology to tackle the high latency issues associated with generating high-resolution images using diffusion transformers (DiTs) models. PipeFusion partitions images into patches and the model layers across multiple GPUs. It employs a patch-level pipeline parallel strategy to orchestrate communication and computation efficiently. By capitalizing on the high similarity between inputs from successive diffusion steps, PipeFusion reuses one-step stale feature maps to provide context for the current pipeline step. This approach notably reduces communication costs compared to existing DiTs inference parallelism, including tensor parallel, sequence parallel and DistriFusion. PipeFusion enhances memory efficiency through parameter distribution across devices, ideal for large DiTs like Flux.1. Experimental results demonstrate that PipeFusion achieves state-of-the-art performance on 8$\times$L40 PCIe GPUs for Pixart, Stable-Diffusion 3, and Flux.1 models. Our Source code is available at \url{https://github.com/xdit-project/xDiT}.

Optimization · Large Scale, Parallel and Distributed

Yongqiang Yao, Jingru Tan, Kaihuan Liang, Feizhao Zhang, Jiahao Hu, Shuo Wu, Yazhe Niu, Ruihao Gong, Dahua Lin, Ningyi Xu

Training Long-Context Large Language Models (LLMs) is challenging, as hybrid training with long-context and short-context data often leads to workload imbalances. Existing works mainly use data packing to alleviate this issue, but fail to consider imbalanced attention computation and wasted communication overhead. This paper proposes Hierarchical Balance Packing (HBP), which designs a novel batch-construction method and training recipe to address those inefficiencies. In particular, the HBP constructs multi-level data packing groups, each optimized with a distinct packing length. It assigns training samples to their optimal groups and configures each group with the most effective settings, including sequential parallelism degree and gradient checkpointing configuration. To effectively utilize multi-level groups of data, we design a dynamic training pipeline specifically tailored to HBP, including curriculum learning, adaptive sequential parallelism, and stable loss. Our extensive experiments demonstrate that our method significantly reduces training time over multiple datasets and open-source models while maintaining strong performance. For the largest DeepSeek-V2 (236B) MoE model, our method speeds up the training by 2.4$\times$ with competitive performance. Codes will be released at https://github.com/ModelTC/HBP.

Optimization · Large Scale, Parallel and Distributed

Jianze Li, Jiezhang Cao, Zichen Zou, Xiongfei Su, Xin Yuan, Yulun Zhang, Yong Guo, Xiaokang Yang

Diffusion models have demonstrated excellent performance for real-world image super-resolution (Real-ISR), albeit at high computational costs. Most existing methods are trying to derive one-step diffusion models from multi-step counterparts through knowledge distillation (KD) or variational score distillation (VSD). However, these methods are limited by the capabilities of the teacher model, especially if the teacher model itself is not sufficiently strong. To tackle these issues, we propose a new One-Step \textbf{D}iffusion model with a larger-scale \textbf{D}iffusion \textbf{D}iscriminator for SR, called D$^3$SR. Our discriminator is able to distill noisy features from any time step of diffusion models in the latent space. In this way, our diffusion discriminator breaks through the potential limitations imposed by the presence of a teacher model. Additionally, we improve the perceptual loss with edge-aware DISTS (EA-DISTS) to enhance the model's ability to generate fine details. Our experiments demonstrate that, compared with previous diffusion-based methods requiring dozens or even hundreds of steps, our D$^3$SR attains comparable or even superior results in both quantitative metrics and qualitative evaluations. Moreover, compared with other methods, D$^3$SR achieves at least $3\times$ faster inference speed and reduces parameters by at least 30\%.

Optimization · Large Scale, Parallel and Distributed

Jia Gu, Songxi Chen

This paper considers decentralized Federated Learning (FL) under het- erogeneous distributions among distributed clients or data blocks for the M- estimation. The mean squared error and consensus error across the estima- tors from different clients via the decentralized stochastic gradient descent algorithm are derived. The asymptotic normality of the Polyak–Ruppert (PR) averaged estimator in the decentralized distributed setting is attained, which shows that its statistical efficiency comes at a cost as it is more restrictive on the number of clients than that in the distributed M-estimation. To overcome the restriction, a one-step estimator is proposed which permits a much larger number of clients while still achieving the same efficiency as the original PR-averaged estimator in the nondistributed setting. The confidence regions based on both the PR-averaged estimator and the proposed one-step estimator are constructed to facilitate statistical inference for decentralized FL.

Optimization · Large Scale, Parallel and Distributed

Guoliang He, Youhe Jiang, Wencong Xiao, Jiang Kaihua, Shuguang Wang, Jun Wang, Du Zixian, Zhuo Jiang, Xinlei Zhang, Binhang Yuan 等

The scaling law for large language models (LLMs) depicts that the path towards machine intelligence necessitates training at large scale. Thus, companies continuously build large-scale GPU clusters, and launch training jobs that span over thousands of computing nodes. However, LLM pre-training presents unique challenges due to its complex communication patterns, where GPUs exchange data in sparse yet high-volume bursts within specific groups. Inefficient resource scheduling exacerbates bandwidth contention, leading to suboptimal training performance. This paper presents Arnold, a scheduling system summarizing our experience to effectively align LLM communication patterns to data center topology at scale. In-depth characteristic study is performed to identify the impact of physical network topology to LLM pre-training jobs. Based on the insights, we develop a scheduling algorithm to effectively align communication patterns to physical network topology in data centers. Through simulation experiments, we show the effectiveness of our algorithm in reducing the maximum spread of communication groups by up to $1.67$x. In production training, our scheduling system improves the end-to-end performance by $10.6\%$ when training with more than $9600$ Hopper GPUs, a significant improvement for our training pipeline.

Optimization · Large Scale, Parallel and Distributed

Jinda Jia, Cong Xie, Hanlin Lu, Fanjiang Ye, Hao Feng, Daoce Wang, Haibin Lin, Zhi Zhang, Xin Liu

Distributed training often suffers from high communication overhead due to large-scale gradient synchronization. Although gradient compression—particularly at 4-bit or even lower precision—significantly reduces transfer volume, it typically results in sacrifice in precision and degradation of the final model accuracy. In this work, we introduce DUO, a distributed training framework designed to mitigate accuracy degradation incurred by gradient compression without involving additional overhead. DUO achieves this by inserting an additional high-precision gradient synchronization step into a previously computation-only phase, so that its communication is fully hidden by computation. We provide a comprehensive theoretical proof of convergence for DUO and validate its effectiveness through extensive pre-training experiments on GPT models. Our results indicate that DUO effectively restores accuracy when using 4-bit gradient compression, achieving performance comparable to uncompressed training. Remarkably, DUO maintains minimal accuracy degradation even under extreme compression scenarios, including 1-bit gradients or complete omission of the low-precision gradient communication step (0-bit transmission).

Optimization · Large Scale, Parallel and Distributed

Sikai Bai, Haoxi Li, Jie ZHANG, Zicong Hong, Song Guo

Despite the significant breakthrough of Mixture-of-Experts (MoE), the increasing scale of these MoE models presents huge memory and storage challenges. Existing MoE pruning methods, which involve reducing parameter size with a uniform sparsity across all layers, often lead to suboptimal outcomes and performance degradation due to varying expert redundancy in different MoE layers. To address this, we propose a non-uniform pruning strategy, dubbed Differentiable Expert Pruning (DiEP), which adaptively adjusts pruning rates at the layer level while jointly learning inter-layer importance, effectively capturing the varying redundancy across different MoE layers. By transforming the global discrete search space into a continuous one, our method handles exponentially growing non-uniform expert combinations, enabling adaptive gradient-based pruning. Extensive experiments on five advanced MoE models demonstrate the efficacy of our method across various NLP tasks. Notably, \textbf{DiEP} retains around 92\% of original performance on Mixtral 8$\times$7B with only half the experts, outperforming other pruning methods by up to 7.1% on the challenging MMLU dataset.

Optimization · Large Scale, Parallel and Distributed

Ke Li, Yan Ding, Zhiqin Zhu, Shenhai Zheng

Heterogeneous federated learning enables collaborative training across clients under dual heterogeneity of models and data, posing challenges for effective knowledge transfer. Federated mutual learning employs proxy models to bridge cross-model knowledge exchange; however, existing methods remain limited to direct alignment between the outputs of private and proxy models, ignoring the deep discrepancies in representation and decision spaces between them. Such cognitive biases cause knowledge to be transferred only at shallow levels and trigger performance bottlenecks. To address this, this paper proposes FedKWAZ to identify and exploit Knowledge Weak-Aware Zones (KWAZ)—spatial zones of deep knowledge misalignment between private and proxy models, further refined into Semantic Weak-Aware Zones and Decision Weak-Aware Zones, which characterize cognitive misalignments in representation and decision spaces as focal targets for enhanced bidirectional distillation. FedKWAZ designs a Hierarchical Adaptive Patch Mixing (HAPM) mechanism to generate multiple mixed samples and employs a Knowledge Discrepancy Perceptron (KDP) to select the samples exhibiting the largest representation and decision discrepancies, thereby mining critical KWAZ. These modules are integrated into a two-stage mutual learning framework, achieving global class-level representation-decision consistency alignment and local KWAZ-guided refinement, structurally bridging cognitive biases across heterogeneous mutual learning models. Experimental results on multiple datasets and model configurations demonstrate the superior performance of FedKWAZ.

Optimization · Large Scale, Parallel and Distributed

HongXin Xu, Tianyu Guo, Xianwei Zhang

To accelerate large language model (LLM) inference, pipeline parallelism partitions model layers into sequential stages, each assigned to a different device for concurrent execution. However, this method often suffers from pipeline bubbles caused by imbalanced computation in the tail stage. While upstream stages focus solely on layer-forward operations, the final stage must also handle post-processing tasks like sampling, introducing significant latency. This uneven workload leads to pipeline misalignment, forcing upstream stages to idle and degrading overall performance. Existing frameworks typically distribute layers evenly across stages without accounting for computational load differences. To address this, we propose DynaPipe, a dynamic layer redistribution scheme that adaptively balances computation by predicting execution latency in real time. Moreover, we introduce an asynchronous key-value (KV) cache migration coordinator to enable non-blocking layer redistribution during inference. Experiments on representative LLMs demonstrate that DynaPipe reduces average end-to-end request latency by 8% to 49% across diverse workloads, outperforming state-of-the-art pipeline parallelism systems.

Optimization · Large Scale, Parallel and Distributed

SARIT KHIRIRAT, Abdurakhmon Sadiev, Artem Riabinin, Eduard Gorbunov, Peter Richtarik

We provide the first proof of convergence for normalized error feedback algorithms across a wide range of machine learning problems. Despite their popularity and efficiency in training deep neural networks, traditional analyses of error feedback algorithms rely on the smoothness assumption that does not capture the properties of objective functions in these problems. Rather, these problems have recently been shown to satisfy generalized smoothness assumptions, and the theoretical understanding of error feedback algorithms under these assumptions remains largely unexplored. Moreover, to the best of our knowledge, all existing analyses under generalized smoothness either i) focus on single-node settings or ii) make unrealistically strong assumptions for distributed settings, such as requiring data heterogeneity, and almost surely bounded stochastic gradient noise variance. In this paper, we propose distributed error feedback algorithms that utilize normalization to achieve the $\mathcal{O}(1/\sqrt{K})$ convergence rate for nonconvex problems under generalized smoothness. Our analyses apply for distributed settings without data heterogeneity conditions, and enable stepsize tuning that is independent of problem parameters. Additionally, we provide strong convergence guarantees of normalized error feedback algorithms for stochastic settings. Finally, we show that due to their larger allowable stepsizes, our new normalized error feedback algorithms outperform their non-normalized counterparts on various tasks, including the minimization of polynomial functions, logistic regression, and ResNet-20 training.

Optimization · Large Scale, Parallel and Distributed

Xinyu Wang, Jonas M. Kübler, Kailash Budhathoki, Yida Wang, Matthäus Kleindessner

When serving a single base LLM with several different LoRA adapters simultaneously, the adapters cannot simply be merged with the base model’s weights as the adapter swapping would create overhead and requests using different adapters could not be batched. Rather, the LoRA computations have to be separated from the base LLM computations, and in a multi-device setup the LoRA adapters can be sharded in a way that is well aligned with the base model’s tensor parallel execution, as proposed in S-LoRA. However, the S-LoRA sharding strategy encounters some communication overhead, which may be small in theory, but can be large in practice. In this paper, we propose to constrain certain LoRA factors to be block-diagonal, which allows for an alternative way of sharding LoRA adapters that does not require any additional communication for the LoRA computations. We demonstrate in extensive experiments that our block-diagonal LoRA approach is similarly parameter efficient as standard LoRA (i.e., for a similar number of parameters it achieves similar downstream performance) and that it leads to significant end-to-end speed-up over S-LoRA. For example, when serving on eight A100 GPUs, we observe up to 1.79x (1.23x) end-to-end speed-up with 0.87x (1.74x) the number of adapter parameters for Llama-3.1-70B, and up to 1.63x (1.3x) end-to-end speed-up with 0.86x (1.73x) the number of adapter parameters for Llama-3.1-8B.

Optimization · Large Scale, Parallel and Distributed

Kenta Niwa, Yuki Takezawa, Guoqiang Zhang, W. Kleijn

For communication-efficient decentralized learning, it is essential to employ dynamic graphs designed to improve the expected spectral gap by reducing deviations from global averaging. The $1$-peer exponential graph demonstrates its finite-time convergence property--achieved by maximizing the expected spectral gap--but only when the number of nodes $n$ is a power of two. However, its efficiency across any $n$ and the commutativity of mixing matrices remain unexplored. We delve into the principles underlying the $1$-peer exponential graph to explain its efficiency across any $n$ and leverage them to develop new dynamic graphs. We propose two new dynamic graphs: the $k$-peer exponential graph and the null-cascade graph. Notably, the null-cascade graph achieves finite-time convergence for any $n$ while ensuring commutativity. Our experiments confirm the effectiveness of these new graphs, particularly the null-cascade graph, in most test settings.

Optimization · Large Scale, Parallel and Distributed

Haizhou Du, Yiran Xiang, Yiwen Cai, Xiufeng Liu, Zonghan Wu, Huan Huo, Guodong Long

Heterogeneous Federated Learning (HtFL) enables collaborative learning across clients with diverse model architectures and non-IID data distributions, which are prevalent in real-world edge computing applications. Existing HtFL approaches typically employ proxy datasets to facilitate knowledge sharing or implement coarse-grained model-level knowledge transfer. However, such approaches not only elevate risks of user privacy leakage but also lead to the loss of fine-grained model-specific knowledge, ultimately creating barriers to effective knowledge sharing. To address these challenges, we propose FedFree, a novel data-free and model-free HtFL framework featuring two key innovations. First, FedFree introduces a reverse layer-wise knowledge transfer mechanism that aggregates heterogeneous client models into a global model solely using Gaussian-based pseudo data, eliminating reliance on proxy datasets. Second, it leverages Knowledge Gain Entropy (KGE) to guide targeted layer-wise knowledge alignment, ensuring that each client receives the most relevant global updates tailored to its specific architecture. We provide rigorous theoretical convergence guarantees for FedFree and conduct extensive experiments on CIFAR-10 and CIFAR-100. Results demonstrate that FedFree achieves substantial performance gains, with relative accuracy improving up to 46.3% over state-of-the-art baselines. The framework consistently excels under highly heterogeneous model/data distributions and in large scale settings.

Optimization · Large Scale, Parallel and Distributed

Hong Huang, Jinhai Yang, Yuan Chen, Jiaxun Ye, Dapeng Wu

Federated Learning (FL) enables collaborative model training across distributed clients without data sharing, but its high computational and communication demands strain resource-constrained devices. While existing methods use dynamic pruning to improve efficiency by periodically adjusting sparse model topologies while maintaining sparsity, these approaches suffer from issues such as **greedy adjustments**, **unstable topologies**, and **communication inefficiency**, resulting in less robust models and suboptimal performance under data heterogeneity and partial client availability. To address these challenges, we propose **Fed**erated **R**obust pruning via combinatorial **T**hompson **S**ampling (FedRTS),a novel framework designed to develop robust sparse models. FedRTS enhances robustness and performance through its Thompson Sampling-based Adjustment (TSAdj) mechanism, which uses probabilistic decisions informed by stable and farsighted information, instead of deterministic decisions reliant on unstable and myopic information in previous methods. Extensive experiments demonstrate that FedRTS achieves state-of-the-art performance in computer vision and natural language processing tasks while reducing communication costs, particularly excelling in scenarios with heterogeneous data distributions and partial client participation. Our codes are available at: https://github.com/Little0o0/FedRTS.

Optimization · Large Scale, Parallel and Distributed

Ziming Liu, Shaoyu Wang, Shenggan Cheng, Zhongkai Zhao, Kai Wang, Xuanlei Zhao, James Demmel, Yang You

Training Transformer models on long sequences in a distributed setting poses significant challenges in terms of efficiency and scalability. Current methods are either constrained by the number of attention heads or excessive communication overheads. To address this problem, we propose StarTrail, a multi-dimensional concentric distributed training system for long sequences, fostering an efficient communication paradigm and providing additional tuning flexibility for communication arrangements. Specifically, StarTrail introduces an extra parallel dimension and divides the peer-to-peer communication into sub-rings to substantially reduce communication volume and avoid bandwidth bottlenecks. Through comprehensive experiments across diverse hardware environments and on both Natural Language Processing (NLP) and Computer Vision (CV) tasks, we demonstrate that our approach significantly surpasses state-of-the-art methods that support Long sequence lengths, achieving performance improvements of up to 77.12% on GPT-style models and up to 114.33% on DiT (Diffusion Transformer) models without affecting the computations results.