Marco Canini

Aditya K Kamath, Arvind Krishnamurthy, Marco Canini et al.

Machine learning (ML) training and inference often process data sets far exceeding GPU memory capacity, forcing them to rely on PCIe for on-demand tensor transfers, causing critical transfer bottlenecks. Lossy compression has been proposed to relieve bottlenecks but introduces workload-dependent accuracy loss, making it complex or even prohibitive to use in existing ML deployments. We explore lossless compression as an alternative that avoids this deployment complexity. We identify where lossless compression can be integrated into ML pipelines while minimizing interference with GPU execution. Based on our findings, we introduce Invariant Bit Packing (IBP), a novel lossless compression algorithm designed to minimize data transfer time for ML. IBP identifies and eliminates invariant bits across groups of tensors, improving throughput through GPU-optimized decompression that leverages warp parallelism, low-overhead bit operations, and asynchronous PCIe transfers. We provide easy-to-use APIs, showcasing them by adding IBP support to GNN training, as well as DLRM and LLM inference frameworks. IBP achieves, on average, 74% faster GNN training, 180% faster DLRM embedding lookup, and 24% faster LLM inference.

Fares Fourati, Salma Kharrat, Vaneet Aggarwal et al.

Federated learning, an emerging machine learning paradigm, enables clients to collaboratively train a model without exchanging local data. Clients participating in the training process significantly impact the convergence rate, learning efficiency, and model generalization. We propose a novel approach, client filtering, to improve model generalization and optimize client participation and training. The proposed method periodically filters available clients to identify a subset that maximizes a combinatorial objective function with an efficient greedy filtering algorithm. Thus, the clients are assessed as a combination rather than individually. We theoretically analyze the convergence of federated learning with client filtering in heterogeneous settings and evaluate its performance across diverse vision and language tasks, including realistic scenarios with time-varying client availability. Our empirical results demonstrate several benefits of our approach, including improved learning efficiency, faster convergence, and up to 10% higher test accuracy than training without client filtering.

Saverio Pasqualoni, Tommaso Bonato, Lorenzo Piarulli et al.

Collective operations are cornerstones of both HPC applications and large-scale AI training and inference, yet benchmarking them in a systematic and reproducible way remains difficult on modern systems due to the complexity of their hardware and software stacks. Existing suites primarily report end-to-end timings and offer limited support for controlled algorithm and configuration selection, fine-grained profiling, and capturing the runtime environment. We present PICO (Performance Insights for Collective Operations), an open-source framework that decouples portable experiment setup from platform execution, provides a backend-adaptive parameter selection interface across MPI and NCCL, supplies plain-MPI reference collective implementations, optionally instrumentable, and records the system configuration for reproducible comparisons. Evaluated on three major supercomputers, PICO shows that default collective algorithms and transport settings can be up to $5\times$ slower than the best available choice. It provides diagnostic evidence by isolating topology sensitive algorithmic choices and, through instrumentation, reveals detailed algorithmic breakdowns. To assess end-to-end effects of benchmark-informed tuning and evaluate application-level impacts, we replay open-source LLM training traces in ATLAHS simulator with optimized collective profiles identified by PICO, achieving reductions in training times of up to $44\%$.

Tie Ma, Yixi Chen, Vaastav Anand et al.

We present MAESTRO, an evaluation suite for the testing, reliability, and observability of LLM-based MAS. MAESTRO standardizes MAS configuration and execution through a unified interface, supports integrating both native and third-party MAS via a repository of examples and lightweight adapters, and exports framework-agnostic execution traces together with system-level signals (e.g., latency, cost, and failures). We instantiate MAESTRO with 12 representative MAS spanning popular agentic frameworks and interaction patterns, and conduct controlled experiments across repeated runs, backend models, and tool configurations. Our case studies show that MAS executions can be structurally stable yet temporally variable, leading to substantial run-to-run variance in performance and reliability. We further find that MAS architecture is the dominant driver of resource profiles, reproducibility, and cost-latency-accuracy trade-off, often outweighing changes in backend models or tool settings. Overall, MAESTRO enables systematic evaluation and provides empirical guidance for designing and optimizing agentic systems.

Kaihua Liang, Xin Tan, An Zhong et al.

Diffusion Large Language Models (DLLMs) offer a compelling alternative to Auto-Regressive models, but their deployment is constrained by high decoding cost. In this work, we identify a key inefficiency in DLLM decoding: while computation is parallelized over token blocks, only a small subset of tokens is decodable at each diffusion step, causing most compute to be wasted on non-decodable tokens. We further observe a strong correlation between attention-derived token importance and token-wise decoding probability. Based on this insight, we propose FOCUS -- an inference system designed for DLLMs. By dynamically focusing computation on decodable tokens and evicting non-decodable ones on-the-fly, FOCUS increases the effective batch size, alleviating compute limitations and enabling scalable throughput. Empirical evaluations demonstrate that FOCUS achieves up to 3.52$\times$ throughput improvement over the production-grade engine LMDeploy, while preserving or improving generation quality across multiple benchmarks. The FOCUS system is publicly available on GitHub: https://github.com/sands-lab/FOCUS.

Janez Božič, Amândio R. Faustino, Boris Radovič et al.

Progressing beyond centralized AI is of paramount importance, yet, distributed AI solutions, in particular various federated learning (FL) algorithms, are often not comprehensively assessed, which prevents the research community from identifying the most promising approaches and practitioners from being convinced that a certain solution is deployment-ready. The largest hurdle towards FL algorithm evaluation is the difficulty of conducting real-world experiments over a variety of FL client devices and different platforms, with different datasets and data distribution, all while assessing various dimensions of algorithm performance, such as inference accuracy, energy consumption, and time to convergence, to name a few. In this paper, we present CoLExT, a real-world testbed for FL research. CoLExT is designed to streamline experimentation with custom FL algorithms in a rich testbed configuration space, with a large number of heterogeneous edge devices, ranging from single-board computers to smartphones, and provides real-time collection and visualization of a variety of metrics through automatic instrumentation. According to our evaluation, porting FL algorithms to CoLExT requires minimal involvement from the developer, and the instrumentation introduces minimal resource usage overhead. Furthermore, through an initial investigation involving popular FL algorithms running on CoLExT, we reveal previously unknown trade-offs, inefficiencies, and programming bugs.

Chen Zhang, Jianghui Wang, Bingyang Cheng et al.

Attention learners, neural networks built on the attention mechanism, e.g., transformers, excel at learning the implicit relationships that relate sequences to their corresponding properties, e.g., mapping a given sequence of tokens to the probability of the next token. However, the learning process tends to be costly. To address this, we present a novel paradigm named Attention Neural Teaching (AtteNT) that reinterprets the learning process through a nonparametric teaching perspective. Specifically, the latter provides a theoretical framework for teaching mappings that are implicitly defined (i.e., nonparametric) via example selection. Such an implicit mapping is embodied through a dense set of sequence-property pairs, with the AtteNT teacher selecting a subset to accelerate convergence in attention learner training. By analytically investigating the role of attention on parameter-based gradient descent during training, and recasting the evolution of attention learners, shaped by parameter updates, through functional gradient descent in nonparametric teaching, we show for the first time that teaching attention learners is consistent with teaching importance-adaptive nonparametric learners. These new findings readily commit AtteNT to enhancing learning efficiency of attention learners. Specifically, we observe training time reductions of 13.01% for LLMs and 20.58% for ViTs, spanning both fine-tuning and training-from-scratch regimes. Crucially, these gains are achieved without compromising accuracy; in fact, performance is consistently preserved and often enhanced across a diverse set of downstream tasks.

Salma Kharrat, Fares Fourati, Marco Canini

The effectiveness of Large Language Models (LLMs) in solving tasks depends significantly on the quality of their instructions, which often require substantial human effort to craft. This underscores the need for automated instruction optimization. However, optimizing instructions is particularly challenging when working with black-box LLMs, where model parameters and gradients are inaccessible. We introduce ACING, an actor-critic reinforcement learning framework that formulates instruction optimization as a stateless, continuous-action problem, enabling exploration of infinite instruction spaces using only black-box feedback. ACING automatically discovers prompts that outperform human-written prompts in 76% of instruction-induction tasks, with gains of up to 33 points and a 10-point median improvement over the best automatic baseline in 33 tasks spanning instruction-induction, summarization, and chain-of-thought reasoning. Extensive ablations highlight its robustness and efficiency. An implementation of ACING is available at https://github.com/salmakh1/ACING.

Keyu Wang, Tian Lyu, Guinan Su et al.

Layer pruning has emerged as a widely adopted technique for improving the efficiency of large language models (LLMs). Although existing methods demonstrate strong performance retention on general knowledge tasks, their effect on long-chain reasoning, a more brittle yet crucial capability, remains largely unexplored. In this work, we study the impact of layer pruning on long-chain reasoning through the lens of test-time scaling, a key mechanism in modern LLMs that enables strong reasoning capacity by allocating more computation at inference time. With extensive experiments, we demonstrate that pruning even one or two layers can severely impair test-time scaling, with performance collapsing drastically on long reasoning benchmarks even when performance on knowledge-intensive and shallow reasoning tasks remains stable. Furthermore, we find that standard supervised fine-tuning remedies fail to recover test-time scaling once it has deteriorated. Through in-depth analyses, we identify the mechanisms underlying this fragility of test-time scaling and highlight the fundamental risks of applying layer pruning to reasoning-intensive LLMs. These findings call for a rethinking of layer pruning strategies and provide insights for developing methods that preserve the robustness of reasoning. We open-source the codebase in \href{https://github.com/keyu-wang-2002/Layer-Pruning-Harms-Inference-Scaling}{https://github.com/keyu-wang-2002/Layer-Pruning-Harms-Inference-Scaling}.

Jihao Xin, Tian Lyu, Qilong Pan et al.

Post-training compression reduces LLM parameter counts but often produces irregular tensor dimensions that degrade GPU performance -- a phenomenon we call \emph{dimensional misalignment}. We present a full-stack analysis tracing root causes at three levels: framework, library, and hardware. The key insight is that model inference becomes slower because the resulting dimensions are unfriendly with the GPU execution stack. For example, compressing Llama-3-8B with activation-aware singular value decomposition (ASVD) has 15\% fewer parameters yet runs no faster than the uncompressed baseline, because 95\% of its dimensions are misaligned. We propose \textbf{GAC} (GPU-Aligned Compression), a new compression paradigm that wraps any dimension-reducing compressor and re-selects hardware-aligned dimensions via multi-choice knapsack optimization under the same parameter budget. We evaluate GAC on Llama-3-8B with ASVD and LLM-Pruner, achieving 100\% alignment and recovering up to 1.5$\times$ speedup while preserving model quality.

Mohammed Aljahdali, Ahmed M. Abdelmoniem, Marco Canini et al.

In Federated Learning (FL), forgetting, or the loss of knowledge across rounds, hampers algorithm convergence, particularly in the presence of severe data heterogeneity among clients. This study explores the nuances of this issue, emphasizing the critical role of forgetting in FL's inefficient learning within heterogeneous data contexts. Knowledge loss occurs in both client-local updates and server-side aggregation steps; addressing one without the other fails to mitigate forgetting. We introduce a metric to measure forgetting granularly, ensuring distinct recognition amid new knowledge acquisition. Leveraging these insights, we propose Flashback, an FL algorithm with a dynamic distillation approach that is used to regularize the local models, and effectively aggregate their knowledge. Across different benchmarks, Flashback outperforms other methods, mitigates forgetting, and achieves faster round-to-target-accuracy, by converging in 6 to 16 rounds.

Jihao Xin, Ivan Ilin, Shunkang Zhang et al.

In distributed training, communication often emerges as a bottleneck. In response, we introduce Kimad, a solution that offers adaptive gradient compression. By consistently monitoring bandwidth, Kimad refines compression ratios to match specific neural network layer requirements. Our exhaustive tests and proofs confirm Kimad's outstanding performance, establishing it as a benchmark in adaptive compression for distributed deep learning.

Norah Alballa, Ahmed M. Abdelmoniem, Marco Canini

This research investigates the enhancement of knowledge distillation (KD) processes in pre-trained models, an emerging field in knowledge transfer with significant implications for distributed training and federated learning environments. These environments benefit from reduced communication demands and accommodate various model architectures. Despite the adoption of numerous KD approaches for transferring knowledge among pre-trained models, a comprehensive understanding of KD's application in these scenarios is lacking. Our study conducts an extensive comparison of multiple KD techniques, including standard KD, tuned KD (via optimized temperature and weight parameters), deep mutual learning, and data partitioning KD. We assess these methods across various data distribution strategies to identify the most effective contexts for each. Through detailed examination of hyperparameter tuning, informed by extensive grid search evaluations, we pinpoint when adjustments are crucial to enhance model performance. This paper sheds light on optimal hyperparameter settings for distinct data partitioning scenarios and investigates KD's role in improving federated learning by minimizing communication rounds and expediting the training process. By filling a notable void in current research, our findings serve as a practical framework for leveraging KD in pre-trained models within collaborative and federated learning frameworks.

Banruo Liu, Mubarak Adetunji Ojewale, Yuhan Ding et al.

We propose NeuronaBox, a flexible, user-friendly, and high-fidelity approach to emulate DNN training workloads. We argue that to accurately observe performance, it is possible to execute the training workload on a subset of real nodes and emulate the networked execution environment along with the collective communication operations. Initial results from a proof-of-concept implementation show that NeuronaBox replicates the behavior of actual systems with high accuracy, with an error margin of less than 1% between the emulated measurements and the real system.

Muhammad Bilal, Zafar Qazi, Marco Canini

The rise of Generative AI Search is fundamentally transforming how users and intelligent systems interact with the Internet. LLMs increasingly act as intermediaries between humans and web information. Yet the web remains optimized for human browsing rather than AI-driven semantic retrieval, resulting in wasted network bandwidth, lower information quality, and unnecessary complexity for developers. We introduce the concept of an AI-Native Internet, a web architecture in which servers expose semantically relevant information chunks rather than full documents, supported by a Web-native semantic resolver that allows AI applications to discover relevant information sources before retrieving fine-grained chunks. Through motivational experiments, we quantify the inefficiencies of current HTML-based retrieval, and outline architectural directions and open challenges for evolving today's document-centric web into an AI-oriented substrate that better supports semantic access to web content.

Norah Alballa, Wenxuan Zhang, Ziquan Liu et al.

Decentralized collaborative learning under data heterogeneity and privacy constraints has rapidly advanced. However, existing solutions like federated learning, ensembles, and transfer learning, often fail to adequately serve the unique needs of clients, especially when local data representation is limited. To address this issue, we propose a novel framework called Query-based Knowledge Transfer (QKT) that enables tailored knowledge acquisition to fulfill specific client needs without direct data exchange. QKT employs a data-free masking strategy to facilitate communication-efficient query-focused knowledge transfer while refining task-specific parameters to mitigate knowledge interference and forgetting. Our experiments, conducted on both standard and clinical benchmarks, show that QKT significantly outperforms existing collaborative learning methods by an average of 20.91\% points in single-class query settings and an average of 14.32\% points in multi-class query scenarios. Further analysis and ablation studies reveal that QKT effectively balances the learning of new and existing knowledge, showing strong potential for its application in decentralized learning.

Dong Chen, Alice Dethise, Istemi Ekin Akkus et al.

A collaboration between dataset owners and model owners is needed to facilitate effective machine learning (ML) training. During this collaboration, however, dataset owners and model owners want to protect the confidentiality of their respective assets (i.e., datasets, models and training code), with the dataset owners also caring about the privacy of individual users whose data is in their datasets. Existing solutions either provide limited confidentiality for models and training code, or suffer from privacy issues due to collusion. We present Citadel++, a collaborative ML training system designed to simultaneously protect the confidentiality of datasets, models and training code as well as the privacy of individual users. Citadel++ enhances differential privacy mechanisms to safeguard the privacy of individual user data while maintaining model utility. By employing Virtual Machine-level Trusted Execution Environments (TEEs) as well as the improved sandboxing and integrity mechanisms through OS-level techniques, Citadel++ effectively preserves the confidentiality of datasets, models and training code, and enforces our privacy mechanisms even when the models and training code have been maliciously designed. Our experiments show that Citadel++ provides model utility and performance while adhering to the confidentiality and privacy requirements of dataset owners and model owners, outperforming the state-of-the-art privacy-preserving training systems by up to 543x on CPU and 113x on GPU TEEs.

Salma Kharrat, Marco Canini, Samuel Horvath

This work tackles the challenges of data heterogeneity and communication limitations in decentralized federated learning. We focus on creating a collaboration graph that guides each client in selecting suitable collaborators for training personalized models that leverage their local data effectively. Our approach addresses these issues through a novel, communication-efficient strategy that enhances resource efficiency. Unlike traditional methods, our formulation identifies collaborators at a granular level by considering combinatorial relations of clients, enhancing personalization while minimizing communication overhead. We achieve this through a bi-level optimization framework that employs a constrained greedy algorithm, resulting in a resource-efficient collaboration graph for personalized learning. Extensive evaluation against various baselines across diverse datasets demonstrates the superiority of our method, named DPFL. DPFL consistently outperforms other approaches, showcasing its effectiveness in handling real-world data heterogeneity, minimizing communication overhead, enhancing resource efficiency, and building personalized models in decentralized federated learning scenarios.

Jihao Xin, Marco Canini, Peter Richtárik et al.

Distributed training enables large-scale deep learning, but suffers from high communication overhead, especially as models and datasets grow. Gradient compression, particularly quantization, is a promising approach to mitigate this bottleneck. However, existing quantization schemes are often incompatible with Allreduce, the dominant communication primitive in distributed deep learning, and many prior solutions rely on heuristics without theoretical guarantees. We introduce Global-QSGD, an Allreduce-compatible gradient quantization method that leverages global norm scaling to reduce communication overhead while preserving accuracy. Global-QSGD is backed by rigorous theoretical analysis, extending standard unbiased compressor frameworks to establish formal convergence guarantees. Additionally, we develop a performance model to evaluate its impact across different hardware configurations. Extensive experiments on NVLink, PCIe, and large-scale cloud environments show that Global-QSGD accelerates distributed training by up to 3.51% over baseline quantization methods, making it a practical and efficient solution for large-scale deep learning workloads.

Ahmed M. Abdelmoniem, Atal Narayan Sahu, Marco Canini et al.

Federated Learning (FL) enables distributed training by learners using local data, thereby enhancing privacy and reducing communication. However, it presents numerous challenges relating to the heterogeneity of the data distribution, device capabilities, and participant availability as deployments scale, which can impact both model convergence and bias. Existing FL schemes use random participant selection to improve fairness; however, this can result in inefficient use of resources and lower quality training. In this work, we systematically address the question of resource efficiency in FL, showing the benefits of intelligent participant selection, and incorporation of updates from straggling participants. We demonstrate how these factors enable resource efficiency while also improving trained model quality.

Atal Narayan Sahu, Aritra Dutta, Ahmed M. Abdelmoniem et al.

Gradient compression is a widely-established remedy to tackle the communication bottleneck in distributed training of large deep neural networks (DNNs). Under the error-feedback framework, Top-$k$ sparsification, sometimes with $k$ as little as $0.1\%$ of the gradient size, enables training to the same model quality as the uncompressed case for a similar iteration count. From the optimization perspective, we find that Top-$k$ is the communication-optimal sparsifier given a per-iteration $k$ element budget. We argue that to further the benefits of gradient sparsification, especially for DNNs, a different perspective is necessary -- one that moves from per-iteration optimality to consider optimality for the entire training. We identify that the total error -- the sum of the compression errors for all iterations -- encapsulates sparsification throughout training. Then, we propose a communication complexity model that minimizes the total error under a communication budget for the entire training. We find that the hard-threshold sparsifier, a variant of the Top-$k$ sparsifier with $k$ determined by a constant hard-threshold, is the optimal sparsifier for this model. Motivated by this, we provide convex and non-convex convergence analyses for the hard-threshold sparsifier with error-feedback. Unlike with Top-$k$ sparsifier, we show that hard-threshold has the same asymptotic convergence and linear speedup property as SGD in the convex case and has no impact on the data-heterogeneity in the non-convex case. Our diverse experiments on various DNNs and a logistic regression model demonstrated that the hard-threshold sparsifier is more communication-efficient than Top-$k$.

Yuchen Jin, Tianyi Zhou, Liangyu Zhao et al.

The learning rate (LR) schedule is one of the most important hyper-parameters needing careful tuning in training DNNs. However, it is also one of the least automated parts of machine learning systems and usually costs significant manual effort and computing. Though there are pre-defined LR schedules and optimizers with adaptive LR, they introduce new hyperparameters that need to be tuned separately for different tasks/datasets. In this paper, we consider the question: Can we automatically tune the LR over the course of training without human involvement? We propose an efficient method, AutoLRS, which automatically optimizes the LR for each training stage by modeling training dynamics. AutoLRS aims to find an LR applied to every $τ$ steps that minimizes the resulted validation loss. We solve this black-box optimization on the fly by Bayesian optimization (BO). However, collecting training instances for BO requires a system to evaluate each LR queried by BO's acquisition function for $τ$ steps, which is prohibitively expensive in practice. Instead, we apply each candidate LR for only $τ'\llτ$ steps and train an exponential model to predict the validation loss after $τ$ steps. This mutual-training process between BO and the loss-prediction model allows us to limit the training steps invested in the BO search. We demonstrate the advantages and the generality of AutoLRS through extensive experiments of training DNNs for tasks from diverse domains using different optimizers. The LR schedules auto-generated by AutoLRS lead to a speedup of $1.22\times$, $1.43\times$, and $1.5\times$ when training ResNet-50, Transformer, and BERT, respectively, compared to the LR schedules in their original papers, and an average speedup of $1.31\times$ over state-of-the-art heavily-tuned LR schedules.

Ahmed M. Abdelmoniem, Chen-Yu Ho, Pantelis Papageorgiou et al.

Federated learning (FL) is becoming a popular paradigm for collaborative learning over distributed, private datasets owned by non-trusting entities. FL has seen successful deployment in production environments, and it has been adopted in services such as virtual keyboards, auto-completion, item recommendation, and several IoT applications. However, FL comes with the challenge of performing training over largely heterogeneous datasets, devices, and networks that are out of the control of the centralized FL server. Motivated by this inherent setting, we make a first step towards characterizing the impact of device and behavioral heterogeneity on the trained model. We conduct an extensive empirical study spanning close to 1.5K unique configurations on five popular FL benchmarks. Our analysis shows that these sources of heterogeneity have a major impact on both model performance and fairness, thus sheds light on the importance of considering heterogeneity in FL system design.

Ahmed M. Abdelmoniem, Ahmed Elzanaty, Mohamed-Slim Alouini et al.

The recent many-fold increase in the size of deep neural networks makes efficient distributed training challenging. Many proposals exploit the compressibility of the gradients and propose lossy compression techniques to speed up the communication stage of distributed training. Nevertheless, compression comes at the cost of reduced model quality and extra computation overhead. In this work, we design an efficient compressor with minimal overhead. Noting the sparsity of the gradients, we propose to model the gradients as random variables distributed according to some sparsity-inducing distributions (SIDs). We empirically validate our assumption by studying the statistical characteristics of the evolution of gradient vectors over the training process. We then propose Sparsity-Inducing Distribution-based Compression (SIDCo), a threshold-based sparsification scheme that enjoys similar threshold estimation quality to deep gradient compression (DGC) while being faster by imposing lower compression overhead. Our extensive evaluation of popular machine learning benchmarks involving both recurrent neural network (RNN) and convolution neural network (CNN) models shows that SIDCo speeds up training by up to 41:7%, 7:6%, and 1:9% compared to the no-compression baseline, Topk, and DGC compressors, respectively.

Aritra Dutta, El Houcine Bergou, Ahmed M. Abdelmoniem et al.

Compressed communication, in the form of sparsification or quantization of stochastic gradients, is employed to reduce communication costs in distributed data-parallel training of deep neural networks. However, there exists a discrepancy between theory and practice: while theoretical analysis of most existing compression methods assumes compression is applied to the gradients of the entire model, many practical implementations operate individually on the gradients of each layer of the model. In this paper, we prove that layer-wise compression is, in theory, better, because the convergence rate is upper bounded by that of entire-model compression for a wide range of biased and unbiased compression methods. However, despite the theoretical bound, our experimental study of six well-known methods shows that convergence, in practice, may or may not be better, depending on the actual trained model and compression ratio. Our findings suggest that it would be advantageous for deep learning frameworks to include support for both layer-wise and entire-model compression.

Aritra Dutta, El Houcine Bergou, Yunming Xiao et al.

Optimization acceleration techniques such as momentum play a key role in state-of-the-art machine learning algorithms. Recently, generic vector sequence extrapolation techniques, such as regularized nonlinear acceleration (RNA) of Scieur et al., were proposed and shown to accelerate fixed point iterations. In contrast to RNA which computes extrapolation coefficients by (approximately) setting the gradient of the objective function to zero at the extrapolated point, we propose a more direct approach, which we call direct nonlinear acceleration (DNA). In DNA, we aim to minimize (an approximation of) the function value at the extrapolated point instead. We adopt a regularized approach with regularizers designed to prevent the model from entering a region in which the functional approximation is less precise. While the computational cost of DNA is comparable to that of RNA, our direct approach significantly outperforms RNA on both synthetic and real-world datasets. While the focus of this paper is on convex problems, we obtain very encouraging results in accelerating the training of neural networks.

Samuel Horvath, Chen-Yu Ho, Ludovit Horvath et al.

Modern deep learning models are often trained in parallel over a collection of distributed machines to reduce training time. In such settings, communication of model updates among machines becomes a significant performance bottleneck and various lossy update compression techniques have been proposed to alleviate this problem. In this work, we introduce a new, simple yet theoretically and practically effective compression technique: natural compression (NC). Our technique is applied individually to all entries of the to-be-compressed update vector and works by randomized rounding to the nearest (negative or positive) power of two, which can be computed in a "natural" way by ignoring the mantissa. We show that compared to no compression, NC increases the second moment of the compressed vector by not more than the tiny factor $\frac{9}{8}$, which means that the effect of NC on the convergence speed of popular training algorithms, such as distributed SGD, is negligible. However, the communications savings enabled by NC are substantial, leading to $3$-$4\times$ improvement in overall theoretical running time. For applications requiring more aggressive compression, we generalize NC to natural dithering, which we prove is exponentially better than the common random dithering technique. Our compression operators can be used on their own or in combination with existing operators for a more aggressive combined effect and offer new state-of-the-art both in theory and practice.

Amedeo Sapio, Marco Canini, Chen-Yu Ho et al.

Training machine learning models in parallel is an increasingly important workload. We accelerate distributed parallel training by designing a communication primitive that uses a programmable switch dataplane to execute a key step of the training process. Our approach, SwitchML, reduces the volume of exchanged data by aggregating the model updates from multiple workers in the network. We co-design the switch processing with the end-host protocols and ML frameworks to provide an efficient solution that speeds up training by up to 5.5$\times$ for a number of real-world benchmark models.