Distributed Training

Distributed training methods are crucial for large language models (LLMs). However, existing distributed training methods often suffer from communication bottlenecks, stragglers, and limited elasticity, particularly in heterogeneous or large-scale environments. Local SGD methods have been proposed to address these issues, but their effectiveness remains limited to small-scale training due to additional memory overhead and lack of concerns on efficiency and stability. To tackle these issues, we propose EDiT, an innovative Efficient Distributed Training method that combines a tailored Local SGD approach with model sharding techniques to enhance large-scale training efficiency. EDiT performs layer-wise parameter synchronization during forward pass, reducing communication and memory overhead and enabling overlap. Besides, EDiT employs a pseudo gradient penalty strategy to suppress loss spikes, which ensures training stability and improves performance. Additionally, we introduce A-EDiT, a fully asynchronous variant of EDiT that accommodates heterogeneous clusters. Building on EDiT/A-EDiT, we conduct a series of experiments to validate large-scale asynchronous training for LLMs, accompanied by comprehensive analyses. Experimental results demonstrate the superior performance of EDiT/A-EDiT, establishing them as robust solutions for distributed LLM training in diverse computational ecosystems. The code is available at Atorch codebase: https://github.com/intelligent-machine-learning/atorch/tree/main/atorch/local_sgd.

This paper presents a novel approach to address the challenge of online sequence learning for decision making under uncertainty in non-stationary, partially observable environments. The proposed algorithm, Distributed Hebbian Temporal Memory (DHTM), is based on the factor graph formalism and a multi-component neuron model. DHTM aims to capture sequential data relationships and make cumulative predictions about future observations, forming Successor Features (SFs). Inspired by neurophysiological models of the neocortex, the algorithm uses distributed representations, sparse transition matrices, and local Hebbian-like learning rules to overcome the instability and slow learning of traditional temporal memory algorithms such as RNN and HMM. Experimental results show that DHTM outperforms LSTM, RWKV and a biologically inspired HMM-like algorithm, CSCG, on non-stationary data sets. Our results suggest that DHTM is a promising approach to address the challenges of online sequence learning and planning in dynamic environments.

Prompts to large language models (LLMs) have evolved beyond simple user questions. For LLMs to solve complex problems, today’s practices are to include domain-specific instructions, illustration of tool usages, and/or long context such as textbook chapters in prompts. As such, many parts of prompts are repetitive across requests. Recent works propose to cache and reuse KV state of prompts. However, they are all confined to a single- GPU optimization, while production LLM serving systems are distributed by nature. This paper proposes Preble, the first distributed LLM serving platform that targets and op- timizes for prompt sharing. We designed a distributed scheduling system that co-optimizes KV state reuse and computation load-balancing with a new scheduling algorithm and a hierarchical scheduling mechanism. Our evaluation of Preble with real workloads and re- quest arrival patterns on two open-source LLMs shows that Preble outperforms the SOTA serving systems by 1.5× to 14.5× on average latency and 2× to 10× on p99 latency.

We analyze two variants of Local Gradient Descent applied to distributed logistic regression with heterogeneous, separable data and show convergence at the rate $O(1/KR)$ for $K$ local steps and sufficiently large $R$ communication rounds. In contrast, all existing convergence guarantees for Local GD applied to any problem are at least $\Omega(1/R)$, meaning they fail to show the benefit of local updates. The key to our improved guarantee is showing progress on the logistic regression objective when using a large stepsize $\eta \gg 1/K$, whereas prior analysis depends on $\eta \leq 1/K$.

Decentralized training of deep neural networks has attracted significant attention for its theoretically superior scalability compared to synchronous data-parallel methods like All-Reduce. However, realizing this potential in multi-node training is challenging due to the complex design space that involves communication topologies, computation patterns, and optimization algorithms. This paper identifies three key factors that can lead to speedups over All-Reduce training and constructs a runtime model to determine when and how decentralization can shorten the per-iteration runtimes. To support the decentralized training of transformer-based models, we introduce a decentralized Adam algorithm that overlaps communications with computations, prove its convergence, and propose an accumulation technique to mitigate the high variance caused by small local batch sizes. We deploy our solution in clusters with up to 64 GPUs, demonstrating its practical advantages in both runtime and generalization performance under a fixed iteration budget. The experiment code is open-source at [https://github.com/WangZesen/Decentralized-Training-Exp](https://github.com/WangZesen/Decentralized-Training-Exp), and the extension code is open-source at [https://github.com/WangZesen/Decent-DP](https://github.com/WangZesen/Decent-DP).

We address the challenge of minimizing "true risk" in multi-node distributed learning.\footnote{We use the term node to refer to a client, FPGA, APU, CPU, GPU, or worker.} These systems are frequently exposed to both inter-node and intra-node "label shifts", which present a critical obstacle to effectively optimizing model performance while ensuring that data remains confined to each node. To tackle this, we propose the Versatile Robust Label Shift (VRLS) method, which enhances the maximum likelihood estimation of the test-to-train label importance ratio. VRLS incorporates Shannon entropy-based regularization and adjusts the importance ratio during training to better handle label shifts at the test time. In multi-node learning environments, VRLS further extends its capabilities by learning and adapting importance ratios across nodes, effectively mitigating label shifts and improving overall model performance. Experiments conducted on MNIST, Fashion MNIST, and CIFAR-10 demonstrate the effectiveness of VRLS, outperforming baselines by up to 20\% in imbalanced settings. These results highlight the significant improvements VRLS offers in addressing label shifts. Our theoretical analysis further supports this by establishing high-probability bounds on estimation errors.

As large language models (LLMs) grow more powerful, they also become more difficult to trust. They could be either aligned with human intentions, or exhibit "subversive misalignment" -- introducing subtle errors that bypass safety checks. Although individual errors may not immediately cause harm, each increases the risk of an eventual safety failure. With this uncertainty, model deployment often grapples with the tradeoff between ensuring safety and harnessing the capabilities of untrusted models. In this work, we introduce the ``Diffuse Risk Management'' problem, aiming to balance the average-case safety and usefulness in the deployment of untrusted models over a large sequence of tasks. We approach this problem by developing a two-level framework: the single-task level (micro-protocol) and the whole-scenario level (macro-protocol). At the single-task level, we develop various \textit{micro}-protocols that use a less capable, but extensively tested (trusted) model to harness and monitor the untrusted model. At the whole-scenario level, we find an optimal \textit{macro}-protocol that uses an adaptive estimate of the untrusted model's risk to choose between micro-protocols. To evaluate the robustness of our method, we follow \textit{control evaluations} in a code generation testbed, which involves a red team attempting to generate subtly backdoored code with an LLM whose deployment is safeguarded by a blue team. Experiment results show that our approach retains 99.6\% usefulness of the untrusted model while ensuring near-perfect safety, significantly outperforming existing deployment methods. Our approach also demonstrates robustness when the trusted and untrusted models have a large capability gap. Our findings demonstrate the promise of managing diffuse risks in the deployment of increasingly capable but untrusted LLMs.

Quadratic programming (QP) forms a crucial foundation in optimization, appearing in a broad spectrum of domains and serving as the basis for more advanced algorithms. Consequently, as the scale and complexity of modern applications continue to grow, the development of efficient and reliable QP algorithms becomes increasingly vital. In this context, this paper introduces a novel deep learning-aided distributed optimization architecture designed for tackling large-scale QP problems. First, we combine the state-of-the-art Operator Splitting QP (OSQP) method with a consensus approach to derive **DistributedQP**, a new method tailored for network-structured problems, with convergence guarantees to optimality. Subsequently, we unfold this optimizer into a deep learning framework, leading to **DeepDistributedQP**, which leverages learned policies to accelerate reaching to desired accuracy within a restricted amount of iterations. Our approach is also theoretically grounded through Probably Approximately Correct (PAC)-Bayes theory, providing generalization bounds on the expected optimality gap for unseen problems. The proposed framework, as well as its centralized version **DeepQP**, significantly outperform their standard optimization counterparts on a variety of tasks such as randomly generated problems, optimal control, linear regression, transportation networks and others. Notably, DeepDistributedQP demonstrates strong generalization by training on small problems and scaling to solve much larger ones (up to 50K variables and 150K constraints) using the same policy. Moreover, it achieves orders-of-magnitude improvements in wall-clock time compared to OSQP. The certifiable performance guarantees of our approach are also demonstrated, ensuring higher-quality solutions over traditional optimizers.

This paper introduces *distributed speculative inference (DSI)*, a novel inference algorithm that is provably faster than speculative inference (SI) [leviathan2023, chen2023, miao2024, sun2025, timor2025] and standard autoregressive inference (non-SI). Like other SI algorithms, DSI operates on frozen language models (LMs), requiring no training or architectural modifications, and it preserves the target distribution. Prior studies on SI have demonstrated empirical speedups over non-SI—but rely on sufficiently fast and accurate drafters, which are often unavailable in practice. We identify a gap where SI can be slower than non-SI if drafters are too slow or inaccurate. We close this gap by proving that DSI is faster than both SI and non-SI—given any drafters. DSI is therefore not only faster than SI, but also unlocks the acceleration of LMs for which SI fails. DSI leverages *speculation parallelism (SP)*, a novel type of task parallelism, to orchestrate target and drafter instances that overlap in time, establishing a new foundational tradeoff between computational resources and latency. Our simulations show that DSI is 1.29-1.92x faster than SI in single-node setups for various off-the-shelf LMs and tasks. We open-source all our code.

Motivated by recent advances in serverless cloud computing, in particular the ``function as a service'' (FaaS) model, we consider the problem of minimizing a convex function in a massively parallel fashion, where communication between workers is limited. Focusing on the case of a twice-differentiable objective subject to an L2 penalty, we propose a scheme where the central node (server) effectively runs a Newton method, offloading its high per-iteration cost---stemming from the need to invert the Hessian---to the workers. In our solution, workers produce independently coarse but low-bias estimates of the inverse Hessian, using an adaptive sketching scheme. The server then averages the descent directions produced by the workers, yielding a good approximation for the exact Newton step. The main component of our adaptive sketching scheme is a low-complexity procedure for selecting the sketching dimension, an issue that was left largely unaddressed in the existing literature on Hessian sketching for distributed optimization. Our solution is based on ideas from asymptotic random matrix theory, specifically the Marchenko-Pastur law. For Gaussian sketching matrices, we derive non asymptotic guarantees for our algorithm which do not depend on the condition number of the Hessian nor a priori require the sketching dimension to be proportional to the dimension, as is often the case in asymptotic random matrix theory. Lastly, when the objective is self-concordant, we provide convergence guarantees for the approximate Newton's method with noisy Hessians, which may be of independent interest beyond the setting considered in this paper.

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-Łojasiewicz condition. Our theory is consistent with the known results for standard smooth problems, and our experimental results support the theoretical insights.

The development of large language models (LLMs) has been instrumental in advancing state-of-the-art natural language processing applications. Training LLMs with billions of parameters and trillions of tokens requires sophisticated distributed systems that enable composing and comparing several state-of-the-art techniques in order to efficiently scale across thousands of accelerators. However, existing solutions are complex, scattered across multiple libraries/repositories, lack interoperability, and are cumbersome to maintain. Thus, curating and empirically comparing training recipes requires non-trivial engineering effort. This paper introduces **TORCHTITAN**$^1$, a PyTorch-native distributed training system that unifies and advances state-of-the-art techniques, streamlining integration and reducing engineering overhead. TORCHTITAN enables seamless application of 4D parallelism in a modular and composable manner, while featuring elastic scaling to adapt to changing computational requirements. The system provides comprehensive logging, efficient checkpointing, and debugging tools, ensuring production-ready training. Moreover, TORCHTITAN incorporates innovative hardware-software co-designed solutions, leveraging cutting-edge features like Float8 training and SymmetricMemory to maximize hardware utilization. As a flexible experimental test bed, TORCHTITAN facilitates the curation and comparison of custom recipes for diverse training contexts. By leveraging TORCHTITAN, we developed optimized training recipes for the Llama 3.1 family and provide actionable guidance on selecting and combining distributed training techniques to maximize training efficiency, based on our hands-on experiences. We thoroughly assess TORCHTITAN on the Llama 3.1 family of LLMs, spanning 8 billion to 405 billion parameters, and showcase its exceptional performance, modular composability, and elastic scalability. By stacking training optimizations, we demonstrate accelerations ranging from 65.08% on Llama 3.1 8B at 128 GPU scale (1D), 12.59% on Llama 3.1 70B at 256 GPU scale (2D), to 30% on Llama 3.1 405B at 512 GPU scale (3D) on NVIDIA H100 GPUs over optimized baselines. We also demonstrate the effectiveness of 4D parallelism in enabling long context training. $^1$ GitHub: [https://github.com/pytorch/torchtitan](https://github.com/pytorch/torchtitan)

In $D$istributed optimization and $L$earning, and even more in the modern framework of federated learning, communication, which is slow and costly, is critical. We introduce LoCoDL, a communication-efficient algorithm that leverages the two popular and effective techniques of $Lo$cal training, which reduces the communication frequency, and $Co$mpression, in which short bitstreams are sent instead of full-dimensional vectors of floats. LoCoDL works with a large class of unbiased compressors that includes widely-used sparsification and quantization methods. LoCoDL provably benefits from local training and compression and enjoys a doubly-accelerated communication complexity, with respect to the condition number of the functions and the model dimension, in the general heterogeneous regime with strongly convex functions. This is confirmed in practice, with LoCoDL outperforming existing algorithms.

Conventional wisdom attributes the success of Generative Flow Networks (GFlowNets) to their ability to exploit the compositional structure of the sample space for learning generalizable flow functions (Bengio et al., 2021). Despite the abundance of empirical evidence, formalizing this belief with verifiable non-vacuous statistical guarantees has remained elusive. We address this issue with the first data-dependent generalization bounds for GFlowNets. We also elucidate the negative impact of the state space size on the generalization performance of these models via Azuma-Hoeffding-type oracle PAC-Bayesian inequalities. We leverage our theoretical insights to design a novel distributed learning algorithm for GFlowNets, which we call *Subgraph Asynchronous Learning* (SAL). In a nutshell, SAL utilizes a divide-and-conquer strategy: multiple GFlowNets are trained in parallel on smaller subnetworks of the flow network, and then aggregated with an additional GFlowNet that allocates appropriate flow to each subnetwork. Our experiments with synthetic and real-world problems demonstrate the benefits of SAL over centralized training in terms of mode coverage and distribution matching.

We propose a novel asynchronous bundle method to solve distributed learning problems. Compared to existing asynchronous methods, our algorithm computes the next iterate based on a more accurate approximation of the objective function and does not require any prior information about the maximal information delay in the system. This makes the proposed method fast and easy to tune. We prove that the algorithm converges in both deterministic and stochastic (mini-batch) settings, and quantify how the convergence times depend on the level of asynchrony. The practical advantages of our method are illustrated through numerical experiments on classification problems of varying complexities and scales.

Mixture of Experts (MoE) is a widely used technique to expand model sizes for better model quality while maintaining the computation cost constant. In a nutshell, an MoE model consists of multiple experts in each model layer and routes the training tokens to only a fixed number of experts rather than all. In distributed training, as experts are distributed among different GPUs, All-to-All communication is necessary to exchange the training tokens among the GPUs after each time of expert routing. Due to the frequent and voluminous data exchanges, All-to-All communication has become a notable challenge to training efficiency. In this paper, we manage to accelerate All-to-All communication in MoE models from the training sample perspective, which is unexplored so far. In particular, we put forward the observation that tokens in the same training sample have certain levels of locality in expert routing. Motivated by this, we develop NetMoE, which takes such locality into account and dynamically rearranges the placement of training samples to minimize All-to-All communication costs. Specifically, we model the All-to-All communication given the sample placement and formulate an integer programming problem to deduce the optimal placement in polynomial time. Experiments with 32 GPUs show that NetMoE achieves a maximum efficiency improvement of $1.67 \times$ compared with current MoE training frameworks.

Reversible architectures have been shown to be capable of performing on par with their non-reversible architectures, being applied in deep learning for memory savings and generative modeling. In this work, we show how reversible architectures can solve challenges in parallelizing deep model training. We introduce PETRA, a novel alternative to backpropagation for parallelizing gradient computations. PETRA facilitates effective model parallelism by enabling stages (i.e., a set of layers) to compute independently on different devices, while only needing to communicate activations and gradients between each other. By decoupling the forward and backward passes and keeping a single updated version of the parameters, the need for weight stashing is also removed. We develop a custom autograd-like training framework for PETRA, and we demonstrate its effectiveness on standard computer vision benchmarks, achieving competitive accuracies comparable to backpropagation using ResNet-18, ResNet-34, and ResNet-50 models.

Sign language pre-training has gained increasing attention for its ability to enhance performance across various sign language understanding (SLU) tasks. However, existing methods often suffer from a gap between pre-training and fine-tuning, leading to suboptimal results. To address this, we propose Uni-Sign, a unified pre-training framework that eliminates the gap between pre-training and downstream SLU tasks through a large-scale generative pre-training strategy and a novel fine-tuning paradigm. First, we introduce CSL-News, a large-scale Chinese Sign Language (CSL) dataset containing 1,985 hours of video paired with textual annotations, which enables effective large-scale pre-training. Second, Uni-Sign unifies SLU tasks by treating downstream tasks as a single sign language translation (SLT) task during fine-tuning, ensuring seamless knowledge transfer between pre-training and fine-tuning. Furthermore, we incorporate a prior-guided fusion (PGF) module and a score-aware sampling strategy to efficiently fuse pose and RGB information, addressing keypoint inaccuracies and improving computational efficiency. Extensive experiments across multiple SLU benchmarks demonstrate that Uni-Sign achieves state-of-the-art performance across multiple downstream SLU tasks. Dataset and code are available at github.com/ZechengLi19/Uni-Sign.