arXiv subjects

Peter Richtárik

Publications and source records attributed to Peter Richtárik.

At least 55 records · Page 3Linked to original sources

arXiv2025

Drop-Muon: Update Less, Converge Faster

Conventional wisdom in deep learning optimization dictates updating all layers at every step-a principle followed by all recent state-of-the-art optimizers such as Muon. In this work, we challenge this assumption, showing that full-network updates can be fundamentally suboptimal, both in theory and in practice. We introduce a non-Euclidean Randomized Progressive Training method-Drop-Muon-a simple yet powerful framework that updates only a subset of layers per step according to a randomized schedule, combining the efficiency of progressive training with layer-specific non-Euclidean updates for top-tier performance. We provide rigorous convergence guarantees under both layer-wise smoothness and layer-wise $(L^0, L^1)$-smoothness, covering deterministic and stochastic gradient settings, marking the first such results for progressive training in the stochastic and non-smooth regime. Our cost analysis further reveals that full-network updates are not optimal unless a very specific relationship between layer smoothness constants holds. Through controlled CNN experiments, we empirically demonstrate that Drop-Muon consistently outperforms full-network Muon, achieving the same accuracy up to $1.4\times$ faster in wall-clock time. Together, our results suggest a shift in how large-scale models can be efficiently trained, challenging the status quo and offering a highly efficient, theoretically grounded alternative to full-network updates.

cs.LG↗

arXiv2025

Error Feedback for Muon and Friends

Recent optimizers like Muon, Scion, and Gluon have pushed the frontier of large-scale deep learning by exploiting layer-wise linear minimization oracles (LMOs) over non-Euclidean norm balls, capturing neural network structure in ways traditional algorithms cannot. Yet, no principled distributed framework exists for these methods, and communication bottlenecks remain unaddressed. The very few distributed variants are heuristic, with no convergence guarantees in sight. We introduce EF21-Muon, the first communication-efficient, non-Euclidean LMO-based optimizer with rigorous convergence guarantees. EF21-Muon supports stochastic gradients, momentum, and bidirectional compression with error feedback-marking the first extension of error feedback beyond the Euclidean setting. It recovers Muon/Scion/Gluon when compression is off and specific norms are chosen, providing the first efficient distributed implementation of this powerful family. Our theory covers non-Euclidean smooth and the more general $(L^0, L^1)$-smooth setting, matching best-known Euclidean rates and enabling faster convergence under suitable norm choices. We further extend the analysis to layer-wise (generalized) smoothness regimes, capturing the anisotropic structure of deep networks. Experiments on NanoGPT benchmarking EF21-Muon against uncompressed Muon/Scion/Gluon demonstrate up to $7\times$ communication savings with no accuracy degradation.

cs.LG↗

arXiv2025

Non-Euclidean Broximal Point Method: A Blueprint for Geometry-Aware Optimization

The recently proposed Broximal Point Method (BPM) [Gruntkowska et al., 2025] offers an idealized optimization framework based on iteratively minimizing the objective function over norm balls centered at the current iterate. It enjoys striking global convergence guarantees, converging linearly and in a finite number of steps for proper, closed and convex functions. However, its theoretical analysis has so far been confined to the Euclidean geometry. At the same time, emerging trends in deep learning optimization, exemplified by algorithms such as Muon [Jordan et al., 2024] and Scion [Pethick et al., 2025], demonstrate the practical advantages of minimizing over balls defined via non-Euclidean norms which better align with the underlying geometry of the associated loss landscapes. In this note, we ask whether the convergence theory of BPM can be extended to this more general, non-Euclidean setting. We give a positive answer, showing that most of the elegant guarantees of the original method carry over to arbitrary norm geometries. Along the way, we clarify which properties are preserved and which necessarily break down when leaving the Euclidean realm. Our analysis positions Non-Euclidean BPM as a conceptual blueprint for understanding a broad class of geometry-aware optimization algorithms, shedding light on the principles behind their practical effectiveness.

math.OC↗

arXiv2025

Local SGD and Federated Averaging Through the Lens of Time Complexity

We revisit the classical Local SGD and Federated Averaging (FedAvg) methods for distributed optimization and federated learning. While prior work has primarily focused on iteration complexity, we analyze these methods through the lens of time complexity, taking into account both computation and communication costs. Our analysis reveals that, despite its favorable iteration complexity, the time complexity of canonical Local SGD is provably worse than that of Minibatch SGD and Hero SGD (locally executed SGD). We introduce a corrected variant, Dual Local SGD, and further improve it by increasing the local step sizes, leading to a new method called Decaying Local SGD. Our analysis shows that these modifications, together with Hero SGD, are optimal in the nonconvex setting (up to logarithmic factors), closing the time complexity gap. Finally, we use these insights to improve the theory of a number of other asynchronous and local methods.

math.OC↗

arXiv2025

FedComLoc: Communication-Efficient Distributed Training of Sparse and Quantized Models

Federated Learning (FL) has garnered increasing attention due to its unique characteristic of allowing heterogeneous clients to process their private data locally and interact with a central server, while being respectful of privacy. A critical bottleneck in FL is the communication cost. A pivotal strategy to mitigate this burden is Local Training, which involves running multiple local stochastic gradient descent iterations between communication phases. Our work is inspired by the innovative Scaffnew algorithm, which has considerably advanced the reduction of communication complexity in FL. We introduce FedComLoc (Federated Compressed and Local Training), integrating practical and effective compression into Scaffnew to further enhance communication efficiency. Extensive experiments, using the popular TopK compressor and quantization, demonstrate its prowess in substantially reducing communication overheads in heterogeneous settings.

cs.LG↗

arXiv2025

Bernoulli-LoRA: A Theoretical Framework for Randomized Low-Rank Adaptation

Parameter-efficient fine-tuning (PEFT) has emerged as a crucial approach for adapting large foundational models to specific tasks, particularly as model sizes continue to grow exponentially. Among PEFT methods, Low-Rank Adaptation (LoRA) (arXiv:2106.09685) stands out for its effectiveness and simplicity, expressing adaptations as a product of two low-rank matrices. While extensive empirical studies demonstrate LoRA's practical utility, theoretical understanding of such methods remains limited. Recent work on RAC-LoRA (arXiv:2410.08305) took initial steps toward rigorous analysis. In this work, we introduce Bernoulli-LoRA, a novel theoretical framework that unifies and extends existing LoRA approaches. Our method introduces a probabilistic Bernoulli mechanism for selecting which matrix to update. This approach encompasses and generalizes various existing update strategies while maintaining theoretical tractability. Under standard assumptions from non-convex optimization literature, we analyze several variants of our framework: Bernoulli-LoRA-GD, Bernoulli-LoRA-SGD, Bernoulli-LoRA-PAGE, Bernoulli-LoRA-MVR, Bernoulli-LoRA-QGD, Bernoulli-LoRA-MARINA, and Bernoulli-LoRA-EF21, establishing convergence guarantees for each variant. Additionally, we extend our analysis to convex non-smooth functions, providing convergence rates for both constant and adaptive (Polyak-type) stepsizes. Through extensive experiments on various tasks, we validate our theoretical findings and demonstrate the practical efficacy of our approach. This work is a step toward developing theoretically grounded yet practically effective PEFT methods.

cs.LG↗

arXiv2025

The Ball-Proximal (="Broximal") Point Method: a New Algorithm, Convergence Theory, and Applications

Non-smooth and non-convex global optimization poses significant challenges across various applications, where standard gradient-based methods often struggle. We propose the Ball-Proximal Point Method, Broximal Point Method, or Ball Point Method (BPM) for short - a novel algorithmic framework inspired by the classical Proximal Point Method (PPM) (Rockafellar, 1976), which, as we show, sheds new light on several foundational optimization paradigms and phenomena, including non-convex and non-smooth optimization, acceleration, smoothing, adaptive stepsize selection, and trust-region methods. At the core of BPM lies the ball-proximal ("broximal") operator, which arises from the classical proximal operator by replacing the quadratic distance penalty by a ball constraint. Surprisingly, and in sharp contrast with the sublinear rate of PPM in the nonsmooth convex regime, we prove that BPM converges linearly and in a finite number of steps in the same regime. Furthermore, by introducing the concept of ball-convexity, we prove that BPM retains the same global convergence guarantees under weaker assumptions, making it a powerful tool for a broader class of potentially non-convex optimization problems. Just like PPM plays the role of a conceptual method inspiring the development of practically efficient algorithms and algorithmic elements, e.g., gradient descent, adaptive step sizes, acceleration (Ahn & Sra, 2020), and "W" in AdamW (Zhuang et al., 2022), we believe that BPM should be understood in the same manner: as a blueprint and inspiration for further development.

math.OC↗

arXiv2025

Quantize Once, Train Fast: Allreduce-Compatible Compression with Provable Guarantees

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.

cs.LG↗

arXiv2025

EF21 with Bells & Whistles: Six Algorithmic Extensions of Modern Error Feedback

First proposed by Seide (2014) as a heuristic, error feedback (EF) is a very popular mechanism for enforcing convergence of distributed gradient-based optimization methods enhanced with communication compression strategies based on the application of contractive compression operators. However, existing theory of EF relies on very strong assumptions (e.g., bounded gradients), and provides pessimistic convergence rates (e.g., while the best known rate for EF in the smooth nonconvex regime, and when full gradients are compressed, is $O(1/T^{2/3})$, the rate of gradient descent in the same regime is $O(1/T)$). Recently, Richtárik et al. (2021) proposed a new error feedback mechanism, EF21, based on the construction of a Markov compressor induced by a contractive compressor. EF21 removes the aforementioned theoretical deficiencies of EF and at the same time works better in practice. In this work we propose six practical extensions of EF21, all supported by strong convergence theory: partial participation, stochastic approximation, variance reduction, proximal setting, momentum, and bidirectional compression. To the best of our knowledge, several of these techniques have not been previously analyzed in combination with EF, and in cases where prior analysis exists -- such as for bidirectional compression -- our theoretical convergence guarantees significantly improve upon existing results.

cs.LG↗

arXiv2025

Collaborative Value Function Estimation Under Model Mismatch: A Federated Temporal Difference Analysis

Federated reinforcement learning (FedRL) enables collaborative learning while preserving data privacy by preventing direct data exchange between agents. However, many existing FedRL algorithms assume that all agents operate in identical environments, which is often unrealistic. In real-world applications, such as multi-robot teams, crowdsourced systems, and large-scale sensor networks, each agent may experience slightly different transition dynamics, leading to inherent model mismatches. In this paper, we first establish linear convergence guarantees for single-agent temporal difference learning (TD(0)) in policy evaluation and demonstrate that under a perturbed environment, the agent suffers a systematic bias that prevents accurate estimation of the true value function. This result holds under both i.i.d. and Markovian sampling regimes. We then extend our analysis to the federated TD(0) (FedTD(0)) setting, where multiple agents, each interacting with its own perturbed environment, periodically share value estimates to collaboratively approximate the true value function of a common underlying model. Our theoretical results indicate the impact of model mismatch, network connectivity, and mixing behavior on the convergence of FedTD(0). Empirical experiments corroborate our theoretical gains, highlighting that even moderate levels of information sharing significantly mitigate environment-specific errors.

cs.LG↗

arXiv2025

MindFlayer SGD: Efficient Parallel SGD in the Presence of Heterogeneous and Random Worker Compute Times

We investigate the problem of minimizing the expectation of smooth nonconvex functions in a distributed setting with multiple parallel workers that are able to compute stochastic gradients. A significant challenge in this context is the presence of arbitrarily heterogeneous and stochastic compute times among workers, which can severely degrade the performance of existing parallel stochastic gradient descent (SGD) methods. While some parallel SGD algorithms achieve optimal performance under deterministic but heterogeneous delays, their effectiveness diminishes when compute times are random - a scenario not explicitly addressed in their design. To bridge this gap, we introduce MindFlayer SGD, a novel parallel SGD method specifically designed to handle stochastic and heterogeneous compute times. Through theoretical analysis and empirical evaluation, we demonstrate that MindFlayer SGD consistently outperforms existing baselines, particularly in environments with heavy-tailed noise. Our results highlight its robustness and scalability, making it a compelling choice for large-scale distributed learning tasks.

math.OC↗

arXiv2025

GradSkip: Communication-Accelerated Local Gradient Methods with Better Computational Complexity

We study a class of distributed optimization algorithms that aim to alleviate high communication costs by allowing clients to perform multiple local gradient-type training steps before communication. In a recent breakthrough, Mishchenko et al. (2022) proved that local training, when properly executed, leads to provable communication acceleration, and this holds in the strongly convex regime without relying on any data similarity assumptions. However, their ProxSkip method requires all clients to take the same number of local training steps in each communication round. We propose a redesign of the ProxSkip method, allowing clients with ``less important'' data to get away with fewer local training steps without impacting the overall communication complexity of the method. In particular, we prove that our modified method, GradSkip, converges linearly under the same assumptions and has the same accelerated communication complexity, while the number of local gradient steps can be reduced relative to a local condition number. We further generalize our method by extending the randomness of probabilistic alternations to arbitrary unbiased compression operators and by considering a generic proximable regularizer. This generalization, which we call GradSkip+, recovers several related methods in the literature as special cases. Finally, we present an empirical study on carefully designed toy problems that confirm our theoretical claims.

cs.LG↗

arXiv2025

Ringmaster ASGD: The First Asynchronous SGD with Optimal Time Complexity

Asynchronous Stochastic Gradient Descent (Asynchronous SGD) is a cornerstone method for parallelizing learning in distributed machine learning. However, its performance suffers under arbitrarily heterogeneous computation times across workers, leading to suboptimal time complexity and inefficiency as the number of workers scales. While several Asynchronous SGD variants have been proposed, recent findings by Tyurin & Richtárik (NeurIPS 2023) reveal that none achieve optimal time complexity, leaving a significant gap in the literature. In this paper, we propose Ringmaster ASGD, a novel Asynchronous SGD method designed to address these limitations and tame the inherent challenges of Asynchronous SGD. We establish, through rigorous theoretical analysis, that Ringmaster ASGD achieves optimal time complexity under arbitrarily heterogeneous and dynamically fluctuating worker computation times. This makes it the first Asynchronous SGD method to meet the theoretical lower bounds for time complexity in such scenarios.

cs.LG↗

arXiv2025

ATA: Adaptive Task Allocation for Efficient Resource Management in Distributed Machine Learning

Asynchronous methods are fundamental for parallelizing computations in distributed machine learning. They aim to accelerate training by fully utilizing all available resources. However, their greedy approach can lead to inefficiencies using more computation than required, especially when computation times vary across devices. If the computation times were known in advance, training could be fast and resource-efficient by assigning more tasks to faster workers. The challenge lies in achieving this optimal allocation without prior knowledge of the computation time distributions. In this paper, we propose ATA (Adaptive Task Allocation), a method that adapts to heterogeneous and random distributions of worker computation times. Through rigorous theoretical analysis, we show that ATA identifies the optimal task allocation and performs comparably to methods with prior knowledge of computation times. Experimental results further demonstrate that ATA is resource-efficient, significantly reducing costs compared to the greedy approach, which can be arbitrarily expensive depending on the number of workers.

cs.LG↗

arXiv2025

Gluon: Making Muon & Scion Great Again! (Bridging Theory and Practice of LMO-based Optimizers for LLMs)

Recent developments in deep learning optimization have brought about radically new algorithms based on the Linear Minimization Oracle (LMO) framework, such as $\sf Muon$ and $\sf Scion$. After over a decade of $\sf Adam$'s dominance, these LMO-based methods are emerging as viable replacements, offering several practical advantages such as improved memory efficiency, better hyperparameter transferability, and most importantly, superior empirical performance on large-scale tasks, including LLM training. However, a significant gap remains between their practical use and our current theoretical understanding: prior analyses (1) overlook the layer-wise LMO application of these optimizers in practice, and (2) rely on an unrealistic smoothness assumption, leading to impractically small stepsizes. To address both, we propose a new LMO-based method called $\sf Gluon$, capturing prior theoretically analyzed methods as special cases, and introduce a new refined generalized smoothness model that captures the layer-wise geometry of neural networks, matches the layer-wise practical implementation of $\sf Muon$ and $\sf Scion$, and leads to convergence guarantees with strong practical predictive power. Unlike prior results, our theoretical stepsizes closely match the fine-tuned values reported by Pethick et al. (2025). Our experiments with NanoGPT and CNN confirm that our assumption holds along the optimization trajectory, ultimately closing the gap between theory and practice.

cs.LG↗

arXiv2025

The Stochastic Multi-Proximal Method for Nonsmooth Optimization

Stochastic gradient descent type methods are ubiquitous in machine learning, but they are only applicable to the optimization of differentiable functions. Proximal algorithms are more general and applicable to nonsmooth functions. We propose a new stochastic and variance-reduced algorithm, the Stochastic Multi-Proximal Method (SMPM), in which the proximity operators of a (possibly empty) random subset of functions are called at every iteration, according to an arbitrary sampling distribution. Several existing algorithms, including Point-SAGA (2016), Proxskip (2022) and RandProx-Minibatch (2023) are recovered as particular cases. We derive linear convergence results in presence of strong convexity and smoothness or similarity of the functions. We prove convergence in the general convex case and accelerated O(1/t2) convergence with varying stepsizes in presence of strong convexity solely. Our results are new even for the above special cases. Moreover, we show an application to distributed optimization with compressed communication, outperforming existing methods.

math.OC↗

arXiv2025

Methods with Local Steps and Random Reshuffling for Generally Smooth Non-Convex Federated Optimization

Non-convex Machine Learning problems typically do not adhere to the standard smoothness assumption. Based on empirical findings, Zhang et al. (2020b) proposed a more realistic generalized $(L_0, L_1)$-smoothness assumption, though it remains largely unexplored. Many existing algorithms designed for standard smooth problems need to be revised. However, in the context of Federated Learning, only a few works address this problem but rely on additional limiting assumptions. In this paper, we address this gap in the literature: we propose and analyze new methods with local steps, partial participation of clients, and Random Reshuffling without extra restrictive assumptions beyond generalized smoothness. The proposed methods are based on the proper interplay between clients' and server's stepsizes and gradient clipping. Furthermore, we perform the first analysis of these methods under the Polyak-Ł ojasiewicz condition. Our theory is consistent with the known results for standard smooth problems, and our experimental results support the theoretical insights.

math.OC↗

arXiv2025

BurTorch: Revisiting Training from First Principles by Coupling Autodiff, Math Optimization, and Systems

In this work, we introduce BurTorch, a compact high-performance framework designed to optimize Deep Learning (DL) training on single-node workstations through an exceptionally efficient CPU-based backpropagation (Rumelhart et al., 1986; Linnainmaa, 1970) implementation. Although modern DL frameworks rely on compilerlike optimizations internally, BurTorch takes a different path. It adopts a minimalist design and demonstrates that, in these circumstances, classical compiled programming languages can play a significant role in DL research. By eliminating the overhead of large frameworks and making efficient implementation choices, BurTorch achieves orders-of-magnitude improvements in performance and memory efficiency when computing $\nabla f(x)$ on a CPU. BurTorch features a compact codebase designed to achieve two key goals simultaneously. First, it provides a user experience similar to script-based programming environments. Second, it dramatically minimizes runtime overheads. In large DL frameworks, the primary source of memory overhead for relatively small computation graphs $f(x)$ is due to feature-heavy implementations. We benchmarked BurTorch against widely used DL frameworks in their execution modes: JAX (Bradbury et al., 2018), PyTorch (Paszke et al., 2019), TensorFlow (Abadi et al., 2016); and several standalone libraries: Autograd (Maclaurin et al., 2015), Micrograd (Karpathy, 2020), Apple MLX (Hannun et al., 2023). For small compute graphs, BurTorch outperforms best-practice solutions by up to $\times 2000$ in runtime and reduces memory consumption by up to $\times 3500$. For a miniaturized GPT-3 model (Brown et al., 2020), BurTorch achieves up to a $\times 20$ speedup and reduces memory up to $\times 80$ compared to PyTorch.

cs.LG↗