Distributed Training

Follow this topic Ask about this topic

Distributed training studies how to divide model computation and data across GPUs, computers, and sometimes geographically separated or unreliable workers. Common approaches include data, tensor, pipeline, and expert parallelism, alongside sharding, gradient compression, asynchronous updates, and communication-aware scheduling. Systems work measures throughput, memory use, and scaling on language models, graph neural networks, and scientific workloads, while fault-tolerant collectives, checkpointing, and robust aggregation address failures and malicious workers. A central challenge is balancing faster, larger-scale training against communication and synchronization costs without harming convergence, model quality, or privacy.

Latest papers

133 papers17 papers in the last four weeks, up 31% on the four weeks before. 0.1% of all new papers.

May 5, 2026

Resilient AI Supercomputer Networking using MRC and SRv6

Tail latency dominates the performance of synchronous pretraining jobs when running at very large scales. We describe a three-pronged approach: (1) a new RDMA-based transport protocol, MRC, sprays across many paths and actively load-balances between them, eliminating the issue of flow collisions (2) the use of multi-plane Clos topologies to get the benefits of high switch radix and redundancy, allowing training clusters well over 100K GPUs to be built as two-tier topologies while increasing physical redundancy, and (3) the use of static source-routing using SRv6 to allow MRC the freedom to bypass failures by itself. We describe our experiences running MRC and static SRv6 routing in production in OpenAI and Microsoft's largest training clusters, where it has been used to train the latest frontier models. We demonstrate how MRC allows AI training jobs to ride out many network failures that previously would have interrupted training.

Joao Araujo, Alex Chow, Mark Handley +47

Distributed Training

May 3, 2026

DBLP: Phase-Aware Bounded-Loss Transport for Burst-Resilient Distributed ML Training

Distributed machine learning (ML) training has become a necessity with the prevalence of billion to trillion-parameter-scale models. While prior work has improved training efficiency from the ML perspective at the application layer, it often fails to address transient congestion events at the network layer that introduce severe tail latency and training-time variability, thereby undermining the quality of service of distributed ML training systems. Existing network optimizations treat all gradients equally and thus fail to integrate sufficient model-training insights into communication protocol design. In this paper, we present Dynamic Bounded-Loss Protocol (DBLP), a burst-resilient, training-phase-aware, and hardware-agnostic transport protocol that incorporates model-level tolerance properties into gradient communication. By dynamically adjusting gradient loss tolerance across training phases, DBLP reduces overall training time and mitigates tail-latency collapse during transient high-loss events (i.e., microbursts). Compared to the current state-of-the-art solution (baseline), DBLP tolerates significantly higher loss while achieving comparable test accuracy, and reduces end-to-end training time by an average of 24.4% and a maximum of 33.9%. At microburst events, DBLP achieves up to 5.88x single-round communication latency speedups over the baseline, preventing burst-induced tail-latency spikes and maintaining stable training performance.

Zechen Ma, Zixi Qu, Jinyan Yi +2

Department of Computer Science, University of Toronto, Toronto, Canada

Communication-Efficient Distributed Training Distributed Training

May 3, 2026

Accelerating Long-Tail Generation in Synchronous RLHF Training via Adaptive Tensor Parallelism

Reinforcement Learning from Human Feedback (RLHF) has become a key post-training paradigm for improving model quality. However, the synchronous three-stage RLHF pipeline is often bottlenecked by the generation stage, where response-length skew causes the effective batch size to shrink rapidly during decoding, leaving GPUs underutilized while a few long responses remain unfinished. Mainstream frameworks employ a static tensor parallelism (TP) configuration that cannot adapt to changing batch characteristics, leaving substantial performance headroom unexplored. We propose PAT, an adaptive TP method that dynamically reconfigures TP during the generation stage of each RLHF iteration. PAT introduces two key techniques. First, a predictor-guided online reconfiguration method decides both the reconfiguration point and the target TP configuration based on offline profiling, triggering reconfiguration only when the predicted latency benefit outweighs the reconfiguration overhead. Second, a lightweight online reconfiguration mechanism updates only the states and layouts affected by TP changes: it adapts unfinished decoding states through a cost-model-based choice between KV-cache migration and recomputation, performs in-place weight resharding, and reuses cached communication groups. We implement PAT on top of SGLang and integrate it with the VeRL framework. Evaluations on LLaMA3.1-8B and Qwen3-14B using DeepScaleR show that PAT reduces generation latency by up to 34.6% and end-to-end RLHF training iteration latency by up to 27.2% compared to the original VeRL setup.

Long Zhao, Qinghe Wang, Jiaan Zhu +5

Anhui University · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center · University of Science and Technology of China +2

Tensor Parallelism LLM Inference Acceleration Distributed Training RL Post-Training

May 1, 2026

Adaptation of AI-accelerated CFD Simulations to the IPU platform

Intelligence Processing Units (IPU) have proven useful for many AI applications. In this paper, we evaluate them within the emerging field of AI for simulation, where traditional numerical simulations are supported by artificial intelligence approaches. We focus specifically on a program for training machine learning models supporting a computational fluid dynamics application. We use custom TensorFlow provided by the Poplar SDK to adapt the program for the IPU-POD16 platform and investigate its ease of use and performance scalability. Training a model on data from OpenFOAM simulations allows us to get accurate simulation state predictions in test time. We show how to utilize the popdist library to overcome a performance bottleneck in feeding training data to the IPU on the host side, achieving up to 34% speedup. Due to communication overheads, using data parallelism to utilize two IPUs instead of one does not improve the throughput. However, once the intra-IPU costs have been paid, the hardware capabilities for inter-IPU communication allow for good scalability. Increasing the number of IPUs from 2 to 16 improves the throughput from 560.8 to 2805.8 samples/s.

P. Rosciszewski, A. Krzywaniak, S. Iserte +2

Graphcore Poland · Dpt. of Construction and Mechanical Engineering, Universitat Jaume I, Spain · Barcelona Supercomputing Center, Spain +2

AI for Science PDE Surrogate Modeling Distributed Training

Apr 29, 2026

FloatSOM: GPU-Accelerated, Distributed, Topology-Flexible Self-Organizing Maps

GPU-accelerated Self-Organizing Map (SOM) implementations are among the most competitive options for large-scale SOM analysis, but growing dataset sizes increasingly challenge their practical use because workloads no longer fit cleanly within device-memory limits. We introduce FloatSOM, a SOM framework for scalable training and deployment that supports multi-GPU execution, out-of-memory disk-backed streaming, and novel topologies beyond regular lattices. We evaluate FloatSOM on 14 synthetic and real benchmark datasets together with controlled speed scaling benchmarks, and show that these improved topologies, combined with topology-aware hyperparameter fine-tuning, yield lower quantization error than current state-of-the-art SOM baselines. FloatSOM also sustains this performance at large scale with high-throughput distributed execution; in the largest benchmark, it trains a 1024-node SOM network on 1,000,000,000 samples with 50 features in 6.16 minutes on 8 GPUs across two separate high-performance-computing nodes.

Tony Xu, Sarah Klamt, Katherine Turner +3

John Curtin School of Medical Research, Australian National University · Department of Information Technology and Electrical Engineering, ETH Zürich · Mathematical Data Science Centre, Australian National University +2

GPU Acceleration High-Performance Computing Distributed Training Self-Organizing Maps

Apr 27, 2026

A Survey on Split Learning for LLM Fine-Tuning: Models, Systems, and Privacy Optimizations

Fine-tuning unlocks large language models (LLMs) for specialized applications, but its high computational cost often puts it out of reach for resource-constrained organizations. While cloud platforms could provide the needed resources, data privacy concerns make sharing sensitive information with third parties risky. A promising solution is split learning for LLM fine-tuning, which divides the model between clients and a server, allowing collaborative and secure training through exchanged intermediate data, thus enabling resource-constrained participants to adapt LLMs safely. % In light of this, a growing body of literature has emerged to advance this paradigm, introducing varied model methods, system optimizations, and privacy defense-attack techniques for split learning. To bring clarity and direction to the field, a comprehensive survey is needed to classify, compare, and critique these diverse approaches. This paper fills the gap by presenting the first extensive survey dedicated to split learning for LLM fine-tuning. We propose a unified, fine-grained training pipeline to pinpoint key operational components and conduct a systematic review of state-of-the-art work across three core dimensions: model-level optimization, system-level efficiency, and privacy preservation. Through this structured taxonomy, we establish a foundation for advancing scalable, robust, and secure collaborative LLM adaptation.

Zihan Liu, Yizhen Wang, Rui Wang +2

Zhejiang University, Ningbo, China

Privacy-Preserving Language Models LLM Fine-Tuning Distributed Training

Apr 27, 2026

TACO: Efficient Communication Compression of Intermediate Tensors for Scalable Tensor-Parallel LLM Training

Handling communication overhead in large-scale tensor-parallel training remains a critical challenge due to the dense, near-zero distributions of intermediate tensors, which exacerbate errors under frequent communication and introduce significant computational overhead during compression. To this end, we propose TACO (Tensor-parallel Adaptive COmmunication compression), a robust FP8-based framework for compressing TP intermediate tensors. First, we employ a data-driven reshaping strategy combined with an Adaptive Scale-Hadamard Transform to enable high-fidelity FP8 quantization, while its Dual-Scale Quantization mechanism ensures numerical stability throughout training. Second, we design a highly fused compression operator to reduce memory traffic and kernel launch overhead, allowing efficient overlap with communication. Finally, we integrate TACO with existing state-of-the-art methods for Data and Pipeline Parallelism to develop a compression-enabled 3D-parallel training framework. Detailed experiments on GPT models and Qwen model demonstrate up to 1.87X end-to-end throughput improvement while maintaining near-lossless accuracy, validating the effectiveness and efficiency of TACO in large-scale training.

Man Liu, Xingchen Liu, Xingjian Tian +8

Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences · Institute of Computing Technology, Chinese Academy of Sciences · University of Science and Technology of China

Communication-Efficient Distributed Training Tensor Parallelism Distributed Training

Apr 27, 2026

FreeScale: Distributed Training for Sequence Recommendation Models with Minimal Scaling Cost

Modern industrial Deep Learning Recommendation Models typically extract user preferences through the analysis of sequential interaction histories, subsequently generating predictions based on these derived interests. The inherent heterogeneity in data characteristics frequently result in substantial under-utilization of computational resources during large-scale training, primarily due to computational bubbles caused by severe stragglers and slow blocking communications. This paper introduces FreeScale, a solution designed to (1) mitigate the straggler problem through meticulously load balanced input samples (2) minimize the blocking communication by overlapping prioritized embedding communications with computations (3) resolve the GPU resource competition during computation and communication overlapping by communicating through SM-Free techniques. Empirical evaluation demonstrates that FreeScale achieves up to 90.3% reduction in computational bubbles when applied to real-world workloads running on 256 H100 GPUs.

Chenhao Feng, Haoli Zhang, Shakhzod Ali-Zade +17

1Meta Inc.

Sequential Recommendation Communication-Efficient Distributed Training Distributed Optimization Distributed Training

Apr 27, 2026

CommFuse: Hiding Tail Latency via Communication Decomposition and Fusion for Distributed LLM Training

The rapid growth in the size of large language models has necessitated the partitioning of computational workloads across accelerators such as GPUs, TPUs, and NPUs. However, these parallelization strategies incur substantial data communication overhead significantly hindering computational efficiency. While communication-computation overlap presents a promising direction, existing data slicing based solutions suffer from tail latency. To overcome this limitation, this research introduces a novel communication-computation overlap technique to eliminate this tail latency in state of the art overlap methods for distributed LLM training. The aim of this technique is to effectively mitigate communication bottleneck of tensor parallelism and data parallelism for distributed training and inference. In particular, we propose a novel method termed CommFuse that replaces conventional collective operations of reduce-scatter and all-gather with decomposed peer-to-peer (P2P) communication and schedules partitioned computations to enable fine-grained overlap. Our method provides an exact algorithm for reducing communication overhead that eliminates tail latency. Moreover, it presents a versatile solution compatible with data-parallel training and various tensor-level parallelism strategies, including TPSP and UP. Experimental evaluations demonstrate that our technique consistently achieves lower latency, superior Model FLOPS Utilization (MFU), and high throughput.

Rezaul Karim, Austin Wen, Wang Zongzuo +3

Toronto Ascend Team

Communication-Efficient Distributed Training Distributed Optimization Tensor Parallelism Distributed Training LLM Training

Apr 23, 2026

Decoupled DiLoCo for Resilient Distributed Pre-training

Modern large-scale language model pre-training relies heavily on the single program multiple data (SPMD) paradigm, which requires tight coupling across accelerators. Due to this coupling, transient slowdowns, hardware failures, and synchronization overhead stall the entire computation, wasting significant compute time at scale. While recent distributed methods like DiLoCo reduced communication bandwidth, they remained fundamentally synchronous and vulnerable to these system stalls. To address this, we introduce Decoupled DiLoCo, an evolution of the DiLoCo framework designed to break the lock-step synchronization barrier and go beyond SPMD to maximize training goodput. Decoupled DiLoCo partitions compute across multiple independent

learners'' that execute local inner optimization steps. These learners asynchronously communicate parameter fragments to a central synchronizer, which circumvents failed or straggling learners by aggregating updates using a minimum quorum, an adaptive grace window, and dynamic token-weighted merging. Inspired by

chaos engineering'', we achieve significantly improved training efficiency in failure-prone environments with millions of simulated chips with strictly zero global downtime, while maintaining competitive model performance across text and vision tasks, for both dense and mixture-of-expert architectures.

Arthur Douillard, Keith Rush, Yani Donchev +14

1Google DeepMind · 2Google Research

Language Model Pretraining Distributed Optimization High-Performance Computing Fault-Tolerant Distributed Training Distributed Training

Apr 21, 2026

Towards Energy Impact on AI-Powered 6G IoT Networks: Centralized vs. Decentralized

The emergence of sixth-generation (6G) technologies has introduced new challenges and opportunities for machine learning (ML) applications in Internet of Things (IoT) networks, particularly concerning energy efficiency. As model training and data transmission contribute significantly to energy consumption, optimizing these processes has become critical for sustainable system design. This study first conduct analysis on the energy consumption model for both centralized and decentralized architecture and then presents a testbed deployed within the German railway infrastructure, leveraging sensor data for ML-based predictive maintenance. A comparative analysis of distributed versus Centralized Learning (CL) architectures reveals that distributed models maintain competitive predictive accuracy (~90%) while reducing overall electricity consumption by up to 70%. These findings underscore the potential of distributed ML to improve energy efficiency in real-world IoT deployments, particularly by mitigating transmission-related energy costs.

Anjie Qiu, Donglin Wang, Sanket Partani +2

*Institute for Wireless Communication and Navigation (WiCoN), University of Kaiserslautern-Landau (RPTU) · +Intelligent Networks Research Group (IN), German Research Center for Artificial Intelligence (DFKI)

Energy-Efficient ML Internet of Things Decentralized Learning Distributed Training

Apr 17, 2026

Scalable and Adaptive Parallel Training of Graph Transformer on Large Graphs

Graph foundation models have demonstrated remarkable adaptability across diverse downstream tasks through large-scale pretraining on graphs. However, existing implementations of the backbone model, graph transformers, are typically limited to single-GPU systems, leading to long training times or out-of-memory issues on large graphs. Moreover, parallelizing graph transformer training over the full graph is challenging, as efficiency depends heavily on both the graph structure and system characteristics, such as bandwidth and memory capacity. In this work, we introduce a distributed training framework for graph transformers, which automatically selects and optimizes parallelization strategies based on the graph structure and hardware configuration. With our implementation of distributed sparse operations, we accelerate sparse graph attention by up to 3.8x and reduce memory consumption by 78% compared to state-of-the-art frameworks. On large graph benchmarks, our proposed framework achieves up to 6x speedup with system scaling up to 8 GPUs. These results demonstrate that the proposed framework improves the scalability of graph transformers, bringing them closer to serving as practical graph foundation models.

Jun-Liang Lin, Kamesh Madduri, Mahmut Taylan Kandemir

The Pennsylvania State University University Park, PA, USA

Distributed Training Graph Transformer

Apr 17, 2026

Training Time Prediction for Mixed Precision-based Distributed Training

Accurate prediction of training time in distributed deep learning is crucial for resource allocation, cost estimation, and job scheduling. We observe that the floating-point precision setting is a key determinant of training time, leading to training time variations of ~2.4x over its minimum. However, existing studies on distributed training time prediction rely on static model computation graphs that do not capture precision variations, including mixed precision. According to our experiments, training time prediction without considering precision results in significant prediction errors - reaching up to 147.85% in mean absolute percentage error (MAPE). To address this issue, we propose a precision-aware distributed training time predictor that achieves robust accuracy across diverse precision settings, including mixed precision, with 9.8% MAPE.

Minchul Kang, Changyong Shin, Jinwoo Jeong +5

Korea University · KT Corporation

Deep Learning Optimization Distributed Training

Apr 17, 2026

PINNACLE: An Open-Source Computational Framework for Classical and Quantum PINNs

We present PINNACLE, an open-source computational framework for physics-informed neural networks (PINNs) that integrates modern training strategies, multi-GPU acceleration, and hybrid quantum-classical architectures within a unified modular workflow. The framework enables systematic evaluation of PINN performance across benchmark problems including 1D hyperbolic conservation laws, incompressible flows, and electromagnetic wave propagation. It supports a range of architectural and training enhancements, including Fourier feature embeddings, random weight factorization, strict boundary condition enforcement, adaptive loss balancing, curriculum training, and second-order optimization strategies, with extensibility to additional methods. We provide a comprehensive benchmark study quantifying the impact of these methods on convergence, accuracy, and computational cost, and analyze distributed data parallel scaling in terms of runtime and memory efficiency. In addition, we extend the framework to hybrid quantum-classical PINNs and derive a formal estimate for circuit-evaluation complexity under parameter-shift differentiation. Results highlight the sensitivity of PINNs to architectural and training choices, confirm their high computational cost relative to classical solvers, and identify regimes where hybrid quantum models offer improved parameter efficiency. PINNACLE provides a foundation for benchmarking physics-informed learning methods and guiding future developments through quantitative assessment of their trade-offs.

Shimon Pisnoy, Hemanth Chandravamsi, Ziv Chen +4

†Faculty of Mechanical Engineering, Technion Israel Institute of Technology, Haifa, Israel. · ‡Corresponding author. · §The Andrew and Erna Viterbi Faculty of Electrical & Computer Engineering, Technion Israel Institute of Technology, Haifa, Israel. +2

Neural PDE Solvers PDE Surrogate Modeling Hybrid Quantum-Classical ML Distributed Training Physics-Informed ML

Apr 15, 2026

SparseBalance: Load-Balanced Long Context Training with Dynamic Sparse Attention

While sparse attention mitigates the computational bottleneck of long-context LLM training, its distributed training process exhibits extreme heterogeneity in both 1) sequence length and 2) sparsity sensitivity, leading to a severe imbalance problem and sub-optimal model accuracy. Existing algorithms and training frameworks typically focus on single issue, failing to systematically co-optimize these two problems. Therefore, we propose SparseBalance, a novel algorithm-system co-design framework, which exploits the sparsity and sequence heterogeneity to optimize model accuracy and system efficiency jointly. First, we propose workload-aware dynamic sparsity tuning, which employs a bidirectional sparsity adjustment to eliminate stragglers and exploit inherent bubbles for free accuracy. Second, we propose a sparsity-aware batching strategy to achieve coarse-grained balance, which complements dynamic sparsity tuning. Experimental results demonstrate that SparseBalance achieves up to a 1.33

\times

end-to-end speedup while still improving the long-context capability by 0.46% on the LongBench benchmark.

Hongtao Xu, Jianchao Tan, Yuxuan Hu +8

School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, Beijing, China · Meituan, Beijing, China · University of Chinese Academy of Sciences, Beijing, China

Sparse Attention Distributed Training Long-Context Language Modeling

Apr 8, 2026

SL-FAC: A Communication-Efficient Split Learning Framework with Frequency-Aware Compression

The growing complexity of neural networks hinders the deployment of distributed machine learning on resource-constrained devices. Split learning (SL) offers a promising solution by partitioning the large model and offloading the primary training workload from edge devices to an edge server. However, the increasing number of participating devices and model complexity leads to significant communication overhead from the transmission of smashed data (e.g., activations and gradients), which constitutes a critical bottleneck for SL. To tackle this challenge, we propose SL-FAC, a communication-efficient SL framework comprising two key components: adaptive frequency decomposition (AFD) and frequency-based quantization compression (FQC). AFD first transforms the smashed data into the frequency domain and decomposes it into spectral components with distinct information. FQC then applies customized quantization bit widths to each component based on its spectral energy distribution. This collaborative approach enables SL-FAC to achieve significant communication reduction while strategically preserving the information most crucial for model convergence. Extensive experiments confirm the superior performance of SL-FAC for improving the training efficiency.

Zehang Lin, Miao Yang, Haihan Zhu +9

School of Computer and Information Engineering, Xiamen University of Technology, Xiamen, 361000, China · School of Artificial Intelligence, Hebei University of Technology · Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong +5

Activation Compression Communication-Efficient Distributed Training Distributed Training Activation Quantization Neural Network Compression

Apr 8, 2026

NestPipe: Large-Scale Recommendation Training on 1,500+ Accelerators via Nested Pipelining

Modern recommendation models have increased to trillions of parameters. As cluster scales expand to O(1k), distributed training bottlenecks shift from computation and memory to data movement, especially lookup and communication latency associated with embeddings. Existing solutions either optimize only one bottleneck or improve throughput by sacrificing training consistency. This paper presents NestPipe, a large-scale decentralized embedding training framework that tackles both bottlenecks while preserving synchronous training semantics. NestPipe exploits two hierarchical sparse parallelism opportunities through nested pipelining. At the inter-batch level, Dual-Buffer Pipelining (DBP) constructs a staleness-free five-stage pipeline through dual-buffer synchronization, mitigating lookup bottlenecks without embedding staleness. At the intra-batch level, we identify the embedding freezing phenomenon, which inspires Frozen-Window Pipelining (FWP) to overlap All2All communication with dense computation via coordinated stream scheduling and key-centric sample clustering. Experiments on production GPU and NPU clusters with 1,536 workers demonstrate that NestPipe achieves up to 3.06x speedup and 94.07% scaling efficiency.

Zhida Jiang, Zhaolong Xing, Huichao Chai +12

Huawei

Distributed Training Pipeline Parallelism Communication-Efficient Distributed Training Recommender Systems

Apr 3, 2026

Communication-free Sampling and 4D Hybrid Parallelism for Scalable Mini-batch GNN Training

Graph neural networks (GNNs) are widely used for learning on graph datasets derived from various real-world scenarios. Learning from extremely large graphs requires distributed training, and mini-batching with sampling is a popular approach for parallelizing GNN training. Existing distributed mini-batch approaches have significant performance bottlenecks due to expensive sampling methods and limited scaling when using data parallelism. In this work, we present ScaleGNN, a 4D parallel framework for scalable mini-batch GNN training that combines communication-free distributed sampling, 3D parallel matrix multiplication (PMM), and data parallelism. ScaleGNN introduces a uniform vertex sampling algorithm, enabling each process (GPU device) to construct its local mini-batch, i.e., subgraph partitions without any inter-process communication. 3D PMM enables scaling mini-batch training to much larger GPU counts than vanilla data parallelism with significantly lower communication overheads. We also present additional optimizations to overlap sampling with training, reduce communication overhead by sending data in lower precision, kernel fusion, and communication-computation overlap. We evaluate ScaleGNN on five graph datasets and demonstrate strong scaling up to 2048 GPUs on Perlmutter, 2048 GCDs on Frontier, and 1024 GPUs on Tuolumne. On Perlmutter, ScaleGNN achieves 3.5x end-to-end training speedup over the SOTA baseline on ogbn-products.

Cunyang Wei, Siddharth Singh, Aishwarya Sarkar +7

University of Maryland, College Park, USA · NVIDIA, Inc. · Iowa State University, USA +3

Graph Neural Networks Distributed Training Communication-Efficient Distributed Training Distributed Optimization

Apr 2, 2026

ModTrans: Translating Real-world Models for Distributed Training Simulator

Large-scale distributed training has been a research hot spot in machine learning systems for industry and academia in recent years. However, conducting experiments without physical machines and corresponding resources is difficult. One solution is to leverage distributed training simulators, but current ones like ASTRA-sim do not support importing real-world developed models, which poses challenges for ML researchers seeking to use them. Based on this challenge, we developed ModTrans, a translator supporting format translation from any real-world model to the ASTRA-sim simulator's input, removing the barrier between machine learning experts and machine learning system researchers. The experiment results show that ModTrans's cost is negligible.

Yi Lyu

University of Wisconsin-Madison Madison, WI, USA

Distributed Training

Apr 1, 2026

EmbedPart: Embedding-Driven Graph Partitioning for Scalable Graph Neural Network Training

Graph Neural Networks (GNNs) are widely used for learning on graph-structured data, but scaling GNN training to massive graphs remains challenging. To enable scalable distributed training, graphs are divided into smaller partitions that are distributed across multiple machines such that inter-machine communication is minimized and computational load is balanced. In practice, existing partitioning approaches face a fundamental trade-off between partitioning overhead and partitioning quality. We propose EmbedPart, an embedding-driven partitioning approach that achieves both speed and quality. Instead of operating directly on irregular graph structures, EmbedPart leverages node embeddings produced during the actual GNN training workload and clusters these dense embeddings to derive a partitioning. EmbedPart achieves more than 100x speedup over Metis while maintaining competitive partitioning quality and accelerating distributed GNN training. Moreover, EmbedPart naturally supports graph updates and fast repartitioning, and can be applied to graph reordering to improve data locality and accelerate single-machine GNN training. By shifting partitioning from irregular graph structures to dense embeddings, EmbedPart enables scalable and high-quality graph data optimization.

Nikolai Merkel, Ruben Mayer, Volker Markl +1

Technical University of Munich (TUM) · University of Bayreuth · TU Berlin, BIFOLD, DFKI +1

Graph Partitioning Graph Neural Networks Distributed Training Graph Clustering

Mar 30, 2026

OptINC: Optical In-Network-Computing for Scalable Distributed Learning

Distributed learning is widely used for training large models on large datasets by distributing parts of the model or dataset across multiple devices and aggregating the computed results for subsequent computations or parameter updates. Existing communication algorithms for distributed learning such as ring all-reduce result in heavy communication overhead between servers. Since communication in large-scale systems uses optical fibers, we propose an Optical In-Network-Computing (OptINC) architecture to offload the computation in servers onto the optical interconnects. To execute gradient averaging and quantization in the optical domain, we incorporate optical devices such as Mach-Zehnder-Interferometers (MZIs) into the interconnects. Such a de facto optical neural network (ONN) can effectively reduce the communication overhead in existing distributed training solutions. To reduce dataset complexity for training this neural network, a preprocessing algorithm implemented in the optical domain is also proposed. Hardware cost is lowered by approximating the weight matrices of the optical neural network with unitary and diagonal matrices, while the accuracy is maintained by a proposed hardware-aware training algorithm. The proposed solution was evaluated on real distributed learning tasks, including ResNet50 on CIFAR-100, and a LLaMA-based network on Wikipedia-1B. In both cases, the proposed framework can achieve comparable training accuracy to the ring all-reduce baseline, while eliminating communication overhead.

Sijie Fei, Grace Li Zhang, Bing Li +1

Technical University of Munich, Munich, Germany · Technical University of Darmstadt, Darmstadt, Germany · Technical University of Ilmenau, Ilmenau, Germany

Gradient Quantization Communication-Efficient Distributed Training Distributed Training Photonic Computing

Mar 30, 2026

Gradient Manipulation in Distributed Stochastic Gradient Descent with Strategic Agents: Truthful Incentives with Convergence Guarantees

Distributed learning has gained significant attention due to its advantages in scalability, privacy, and fault tolerance.In this paradigm, multiple agents collaboratively train a global model by exchanging parameters only with their neighbors. However, a key vulnerability of existing distributed learning approaches is their implicit assumption that all agents behave honestly during gradient updates. In real-world scenarios, this assumption often breaks down, as selfish or strategic agents may be incentivized to manipulate gradients for personal gain, ultimately compromising the final learning outcome. In this work, we propose a fully distributed payment mechanism that, for the first time, guarantees both truthful behaviors and accurate convergence in distributed stochastic gradient descent. This represents a significant advancement, as it overcomes two major limitations of existing truthfulness mechanisms for collaborative learning:(1) reliance on a centralized server for payment collection, and (2) sacrificing convergence accuracy to guarantee truthfulness. In addition to characterizing the convergence rate under general convex and strongly convex conditions, we also prove that our approach guarantees the cumulative gain that an agent can obtain through strategic behavior remains finite, even as the number of iterations approaches infinity--a property unattainable by most existing truthfulness mechanisms. Our experimental results on standard machine learning tasks, evaluated on benchmark datasets, confirm the effectiveness of the proposed approach.

Ziqin Chen, Yongqiang Wang

Department of Electrical and Computer Engineering, Clemson University, Clemson, SC 29634 USA

Distributed Training Decentralized Learning Distributed Optimization Stochastic Optimization Convergence

Mar 18, 2026

Tula: Optimizing Time, Cost, and Generalization in Distributed Large-Batch Training

Distributed training increases the number of batches processed per iteration either by scaling-out (adding more nodes) or scaling-up (increasing the batch-size). However, the largest configuration does not necessarily yield the best performance. Horizontal scaling introduces additional communication overhead, while vertical scaling is constrained by computation cost and device memory limits. Thus, simply increasing the batch-size leads to diminishing returns: training time and cost decrease initially but eventually plateaus, creating a knee-point in the time/cost versus batch-size pareto curve. The optimal batch-size therefore depends on the underlying model, data and available compute resources. Large batches also suffer from worse model quality due to the well-known generalization gap. In this paper, we present Tula, an online service that automatically optimizes time, cost, and convergence quality for large-batch training of convolutional models. It combines parallel-systems modeling with statistical performance prediction to identify the optimal batch-size. Tula predicts training time and cost within 7.5-14% error across multiple models, and achieves up to 20x overall speedup and improves test accuracy by 9% on average over standard large-batch training on various vision tasks, thus successfully mitigating the generalization gap and accelerating training at the same time.

Sahil Tyagi, Feiyi Wang

Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA

Distributed Training Hyperparameter Optimization Neural Network Generalization

Mar 17, 2026

Byzantine-Robust and Communication-Efficient Distributed Training: Compressive and Cyclic Gradient Coding

In this paper, we study the problem of distributed training (DT) under Byzantine attacks with communication constraints. While prior work has developed various robust aggregation rules at the server to enhance robustness to Byzantine attacks, the existing methods suffer from a critical limitation in that the solution error does not diminish when the local gradients sent by different devices vary considerably, as a result of data heterogeneity among the subsets held by different devices. To overcome this limitation, we propose a novel DT method, cyclic gradient coding-based DT (LAD). In LAD, the server allocates the entire training dataset to the devices before training begins. In each iteration, it assigns computational tasks redundantly to the devices using cyclic gradient coding. Each honest device then computes local gradients on a fixed number of data subsets and encodes the local gradients before transmitting to the server. The server aggregates the coded vectors from the honest devices and the potentially incorrect messages from Byzantine devices using a robust aggregation rule. Leveraging the redundancy of computation across devices, the convergence performance of LAD is analytically characterized, demonstrating improved robustness against Byzantine attacks and significantly lower solution error. Furthermore, we extend LAD to a communication-efficient variant, compressive and cyclic gradient coding-based DT (Com-LAD), which further reduces communication overhead under constrained settings. Numerical results validate the effectiveness of the proposed methods in enhancing both Byzantine resilience and communication efficiency.

Chengxi Li, Youssef Allouah, Rachid Guerraoui +2

Division of Information Science and Engineering, School of Electrical Engineering and Computer Science, KTH Royal Institute of Technology, 10044 Stockholm, Sweden · Stanford University, CA 94305, United States · School of Computer and Communication Sciences, the Swiss Federal Institute of Technology in Lausanne (EPFL), 1015 Lausanne, Switzerland

Distributed Training Communication-Efficient Distributed Training

Mar 11, 2026

Thousand-GPU Large-Scale Training and Optimization Recipe for AI-Native Cloud Embodied Intelligence Infrastructure

Embodied intelligence is a key step towards Artificial General Intelligence (AGI), yet its development faces multiple challenges including data, frameworks, infrastructure, and evaluation systems. To address these issues, we have, for the first time in the industry, launched a cloud-based, thousand-GPU distributed training platform for embodied intelligence, built upon the widely adopted LeRobot framework, and have systematically overcome bottlenecks across the entire pipeline. At the data layer, we have restructured the data pipeline to optimize the flow of embodied training data. In terms of training, for the GR00T-N1.5 model, utilizing thousand-GPU clusters and data at the scale of hundreds of millions, the single-round training time has been reduced from 15 hours to just 22 minutes, achieving a 40-fold speedup. At the model layer, by combining variable-length FlashAttention and Data Packing, we have moved from sample redundancy to sequence integration, resulting in a 188% speed increase; π-0.5 attention optimization has accelerated training by 165%; and FP8 quantization has delivered a 140% speedup. On the infrastructure side, relying on high-performance storage, a 3.2T RDMA network, and a Ray-driven elastic AI data lake, we have achieved deep synergy among data, storage, communication, and computation. We have also built an end-to-end evaluation system, creating a closed loop from training to simulation to assessment. This framework has already been fully validated on thousand-GPU clusters, laying a crucial technical foundation for the development and application of next-generation autonomous intelligent robots, and is expected to accelerate the arrival of the era of human-machine integration.

Yongjian Guo, Yunxuan Ma, Haoran Sun +22

Distributed Training Embodied AI Robot Foundation Models

Mar 10, 2026

Exploiting Adaptive Channel Pruning for Communication-Efficient Split Learning

Split learning (SL) transfers most of the training workload to the server, which alleviates computational burden on client devices. However, the transmission of intermediate feature representations, referred to as smashed data, incurs significant communication overhead, particularly when a large number of client devices are involved. To address this challenge, we propose an adaptive channel pruning-aided SL (ACP-SL) scheme. In ACP-SL, a label-aware channel importance scoring (LCIS) module is designed to generate channel importance scores, distinguishing important channels from less important ones. Based on these scores, an adaptive channel pruning (ACP) module is developed to prune less important channels, thereby compressing the corresponding smashed data and reducing the communication overhead. Experimental results show that ACP-SL consistently outperforms benchmark schemes in test accuracy. Furthermore, it reaches a target test accuracy in fewer training rounds, thereby reducing communication overhead.

Jialei Tan, Zheng Lin, Xiangming Cai +4

School of Physics and Information Engineering, Fuzhou University, Fuzhou 350108, China · Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong · School of Electrical Engineering and Telecommunications, University of New South Wales Sydney, Sydney, NSW 2052, Australia +3

Distributed Training Neural Network Pruning

Mar 9, 2026

Split Federated Learning Architectures for High-Accuracy and Low-Delay Model Training

Can we find a network architecture for ML model training so as to optimize training loss (and thus, accuracy) in Split Federated Learning (SFL)? And can this architecture also reduce training delay and communication overhead? While accuracy is not influenced by how we split the model in ordinary, state-of-the-art SFL, in this work we answer the questions above in the affirmative. Recent Hierarchical SFL (HSFL) architectures adopt a three-tier training structure consisting of clients, (local) aggregators, and a central server. In this architecture, the model is partitioned at two partitioning layers into three sub-models, which are executed across the three tiers. Despite their merits, HSFL architectures overlook the impact of the partitioning layers and client-to-aggregator assignments on accuracy, delay, and overhead. This work explicitly captures the impact of the partitioning layers and client-to-aggregator assignments on accuracy, delay and overhead by formulating a joint optimization problem. We prove that the problem is NP-hard and propose the first accuracy-aware heuristic algorithm that explicitly accounts for model accuracy, while remaining delay-efficient. Simulation results on public datasets show that our approach can improve accuracy by 3%, while reducing delay by 20% and overhead by 50%, compared to state-of-the-art SFL and HSFL schemes.

Yiannis Papageorgiou, Yannis Thomas, Ramin Khalili +1

Department of Informatics, Athens University of Economics and Business, Athens, Greece · Huawei Technologies Duesseldorf Gmbh, Munich, Germany

Split Federated Learning Federated Learning Distributed Training Hierarchical Federated Learning Communication-Efficient Distributed Training

Mar 9, 2026

Covenant-72B: Pre-Training a 72B LLM with Trustless Peers Over-the-Internet

Recently, there has been increased interest in globally distributed training, which has the promise to both reduce training costs and democratize participation in building large-scale foundation models. However, existing models trained in a globally distributed manner are relatively small in scale and have only been trained with whitelisted participants. Therefore, they do not yet realize the full promise of democratized participation. In this report, we describe Covenant-72B, an LLM produced by the largest collaborative globally distributed pre-training run (in terms of both compute and model scale), which simultaneously allowed open, permissionless participation supported by a live blockchain protocol. We utilized a state-of-the-art communication-efficient optimizer, SparseLoCo, supporting dynamic participation with peers joining and leaving freely. Our model, pre-trained on approximately 1.1T tokens, performs competitively with fully centralized models pre-trained on similar or higher compute budgets, demonstrating that fully democratized, non-whitelisted participation is not only feasible, but can be achieved at unprecedented scale for a globally distributed pre-training run.

Joel Lidin, Amir Sarfi, Erfan Miahi +6

Covenant AI · Mila

Language Model Pretraining Distributed Training Decentralized Learning

Mar 7, 2026

A Distributed Gaussian Process Model for Multi-Robot Mapping

We propose DistGP: a multi-robot learning method for collaborative learning of a global function using only local experience and computation. We utilise a sparse Gaussian process (GP) model with a factorisation that mirrors the multi-robot structure of the task, and admits distributed training via Gaussian belief propagation (GBP). Our loopy model outperforms Tree-Structured GPs \cite{bui2014tree} and can be trained online and in settings with dynamic connectivity. We show that such distributed, asynchronous training can reach the same performance as a centralised, batch-trained model, albeit with slower convergence. Last, we compare to DiNNO \cite{yu2022dinno}, a distributed neural network (NN) optimiser, and find DistGP achieves superior accuracy, is more robust to sparse communication and is better able to learn continually.

Seth Nabarro, Mark van der Wilk, Andrew J. Davison

Dyson Robotics Lab, Imperial College London · Department of Computer Science, University of Oxford

Sparse Gaussian Processes Gaussian Belief Propagation Gaussian Processes Multi-Robot Systems Distributed Training Bayesian Inference

Mar 6, 2026

NEST: Network- and Memory-Aware Device Placement For Distributed Deep Learning

The growing scale of deep learning demands distributed training frameworks that jointly reason about parallelism, memory, and network topology. Prior works often rely on heuristic or topology-agnostic search, handling communication and memory separately. Without per-device memory awareness, these methods typically ensure feasibility post hoc by sharding parameters and activations across many devices, increasing synchronization, inflating communication, and underutilizing compute-limiting scalability and efficiency on real datacenter networks. We present NEST, a network-, compute-, and memory-aware device placement framework that unifies model parallelism, topology modeling, and memory feasibility via structured dynamic programming. NEST's DP operates on operator graphs with tensor and expert parallel configurations, explicit allreduce latencies across hierarchical or arbitrary networks, and memory/compute profiles. By factoring parallelism across tensor, pipeline, data, and expert dimensions, NEST defines a principled search space for hybrid strategies while jointly optimizing co-location, network latency, and memory feasibility. Evaluations across diverse hardware and networks show NEST achieves up to 2.43 times higher throughput, better memory efficiency, and improved scalability over state-of-the-art baselines, providing a foundation for co-designing parallelization strategies and datacenter interconnects for next-generation AI infrastructure. The source code of NEST is available at: https://github.com/scai-tech/Nest

Irene Wang, Vishnu Varma Venkata, Arvind Krishnamurthy +1

Georgia Institute of Technology, Atlanta, GA, USA · University of Washington, Seattle, WA, USA

Distributed Training Communication-Efficient Distributed Training