On the Convergence of DP-SGD with Adaptive Clipping

Stochastic Gradient Descent (SGD) with gradient clipping is a powerful technique for enabling differentially private optimization. Although prior works extensively investigated clipping with a constant threshold, private training remains highly sensitive to threshold selection, which can be expensive or even infeasible to tune. This sensitivity motivates the development of adaptive approaches, such as quantile clipping, which have demonstrated empirical success but lack a solid theoretical understanding. This paper provides the first comprehensive convergence analysis of SGD with quantile clipping (QC-SGD). We demonstrate that QC-SGD suffers from a bias problem similar to constant-threshold clipped SGD but show how this can be mitigated through a carefully designed quantile and step size schedule. Our analysis reveals crucial relationships between quantile selection, step size, and convergence behavior, providing practical guidelines for parameter selection. We extend these results to differentially private optimization, establishing the first theoretical guarantees for DP-QC-SGD. Our findings provide theoretical foundations for widely used adaptive clipping heuristic and highlight open avenues for future research.

Differentially Private Random Block Coordinate Descent

Coordinate Descent (CD) methods have gained significant attention in machine learning due to their effectiveness in solving high-dimensional problems and their ability to decompose complex optimization tasks. However, classical CD methods were neither designed nor analyzed with data privacy in mind, a critical concern when handling sensitive information. This has led to the development of differentially private CD methods, such as DP-CD (Differentially Private Coordinate Descent) proposed by Mangold et al. (ICML 2022), yet a disparity remains between non-private CD and DP-CD methods. In our work, we propose a differentially private random block coordinate descent method that selects multiple coordinates with varying probabilities in each iteration using sketch matrices. Our algorithm generalizes both DP-CD and the classical DP-SGD (Differentially Private Stochastic Gradient Descent), while preserving the same utility guarantees. Furthermore, we demonstrate that better utility can be achieved through importance sampling, as our method takes advantage of the heterogeneity in coordinate-wise smoothness constants, leading to improved convergence rates.

MARINA-P: Superior Performance in Non-smooth Federated Optimization with Adaptive Stepsizes

Non-smooth communication-efficient federated optimization is crucial for many machine learning applications, yet remains largely unexplored theoretically. Recent advancements have primarily focused on smooth convex and non-convex regimes, leaving a significant gap in understanding the non-smooth convex setting. Additionally, existing literature often overlooks efficient server-to-worker communication (downlink), focusing primarily on worker-to-server communication (uplink). We consider a setup where uplink costs are negligible and focus on optimizing downlink communication by improving state-of-the-art schemes like EF21-P (arXiv:2209.15218) and MARINA-P (arXiv:2402.06412) in the non-smooth convex setting. We extend the non-smooth convex theory of EF21-P [Anonymous, 2024], originally developed for single-node scenarios, to the distributed setting, and extend MARINA-P to the non-smooth convex setting. For both algorithms, we prove an optimal $O(1/\sqrt{T})$ convergence rate and establish communication complexity bounds matching classical subgradient methods. We provide theoretical guarantees under constant, decreasing, and adaptive (Polyak-type) stepsizes. Our experiments demonstrate that MARINA-P with correlated compressors outperforms other methods in both smooth non-convex and non-smooth convex settings. This work presents the first theoretical results for distributed non-smooth optimization with server-to-worker compression, along with comprehensive analysis for various stepsize schemes.

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-{\L} ojasiewicz condition. Our theory is consistent with the known results for standard smooth problems, and our experimental results support the theoretical insights.

Pushing the Limits of Large Language Model Quantization via the Linearity Theorem

Quantizing large language models has become a standard way to reduce their memory and computational costs. Typically, existing methods focus on breaking down the problem into individual layer-wise sub-problems, and minimizing per-layer error, measured via various metrics. Yet, this approach currently lacks theoretical justification and the metrics employed may be sub-optimal. In this paper, we present a "linearity theorem" establishing a direct relationship between the layer-wise $\ell_2$ reconstruction error and the model perplexity increase due to quantization. This insight enables two novel applications: (1) a simple data-free LLM quantization method using Hadamard rotations and MSE-optimal grids, dubbed HIGGS, which outperforms all prior data-free approaches such as the extremely popular NF4 quantized format, and (2) an optimal solution to the problem of finding non-uniform per-layer quantization levels which match a given compression constraint in the medium-bitwidth regime, obtained by reduction to dynamic programming. On the practical side, we demonstrate improved accuracy-compression trade-offs on Llama-3.1 and 3.2-family models, as well as on Qwen-family models. Further, we show that our method can be efficiently supported in terms of GPU kernels at various batch sizes, advancing both data-free and non-uniform quantization for LLMs.

Error Feedback under $(L_0,L_1)$-Smoothness: Normalization and Momentum

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

Tighter Performance Theory of FedExProx

We revisit FedExProx - a recently proposed distributed optimization method designed to enhance convergence properties of parallel proximal algorithms via extrapolation. In the process, we uncover a surprising flaw: its known theoretical guarantees on quadratic optimization tasks are no better than those offered by the vanilla Gradient Descent (GD) method. Motivated by this observation, we develop a novel analysis framework, establishing a tighter linear convergence rate for non-strongly convex quadratic problems. By incorporating both computation and communication costs, we demonstrate that FedExProx can indeed provably outperform GD, in stark contrast to the original analysis. Furthermore, we consider partial participation scenarios and analyze two adaptive extrapolation strategies - based on gradient diversity and Polyak stepsizes - again significantly outperforming previous results. Moving beyond quadratics, we extend the applicability of our analysis to general functions satisfying the Polyak-Lojasiewicz condition, outperforming the previous strongly convex analysis while operating under weaker assumptions. Backed by empirical results, our findings point to a new and stronger potential of FedExProx, paving the way for further exploration of the benefits of extrapolation in federated learning.

Unlocking FedNL: Self-Contained Compute-Optimized Implementation

Federated Learning (FL) is an emerging paradigm that enables intelligent agents to collaboratively train Machine Learning (ML) models in a distributed manner, eliminating the need for sharing their local data. The recent work (arXiv:2106.02969) introduces a family of Federated Newton Learn (FedNL) algorithms, marking a significant step towards applying second-order methods to FL and large-scale optimization. However, the reference FedNL prototype exhibits three serious practical drawbacks: (i) It requires 4.8 hours to launch a single experiment in a sever-grade workstation; (ii) The prototype only simulates multi-node setting; (iii) Prototype integration into resource-constrained applications is challenging. To bridge the gap between theory and practice, we present a self-contained implementation of FedNL, FedNL-LS, FedNL-PP for single-node and multi-node settings. Our work resolves the aforementioned issues and reduces the wall clock time by x1000. With this FedNL outperforms alternatives for training logistic regression in a single-node -- CVXPY (arXiv:1603.00943), and in a multi-node -- Apache Spark (arXiv:1505.06807), Ray/Scikit-Learn (arXiv:1712.05889). Finally, we propose two practical-orientated compressors for FedNL - adaptive TopLEK and cache-aware RandSeqK, which fulfill the theory of FedNL.

Randomized Asymmetric Chain of LoRA: The First Meaningful Theoretical Framework for Low-Rank Adaptation

Fine-tuning has become a popular approach to adapting large foundational models to specific tasks. As the size of models and datasets grows, parameter-efficient fine-tuning techniques are increasingly important. One of the most widely used methods is Low-Rank Adaptation (LoRA), with adaptation update expressed as the product of two low-rank matrices. While LoRA was shown to possess strong performance in fine-tuning, it often under-performs when compared to full-parameter fine-tuning (FPFT). Although many variants of LoRA have been extensively studied empirically, their theoretical optimization analysis is heavily under-explored. The starting point of our work is a demonstration that LoRA and its two extensions, Asymmetric LoRA and Chain of LoRA, indeed encounter convergence issues. To address these issues, we propose Randomized Asymmetric Chain of LoRA (RAC-LoRA) -- a general optimization framework that rigorously analyzes the convergence rates of LoRA-based methods. Our approach inherits the empirical benefits of LoRA-style heuristics, but introduces several small but important algorithmic modifications which turn it into a provably convergent method. Our framework serves as a bridge between FPFT and low-rank adaptation. We provide provable guarantees of convergence to the same solution as FPFT, along with the rate of convergence. Additionally, we present a convergence analysis for smooth, non-convex loss functions, covering gradient descent, stochastic gradient descent, and federated learning settings. Our theoretical findings are supported by experimental results.

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

We study the problem of minimizing the expectation of smooth nonconvex functions with the help of several parallel workers whose role is to compute stochastic gradients. In particular, we focus on the challenging situation where the workers' compute times are arbitrarily heterogeneous and random. In the simpler regime characterized by arbitrarily heterogeneous but deterministic compute times, Tyurin and Richt\'arik (NeurIPS 2023) recently designed the first theoretically optimal asynchronous SGD method, called Rennala SGD, in terms of a novel complexity notion called time complexity. The starting point of our work is the observation that Rennala SGD can have arbitrarily bad performance in the presence of random compute times -- a setting it was not designed to handle. To advance our understanding of stochastic optimization in this challenging regime, we propose a new asynchronous SGD method, for which we coin the name MindFlayer SGD. Our theory and empirical results demonstrate the superiority of MindFlayer SGD over existing baselines, including Rennala SGD, in cases when the noise is heavy tailed.