Distributed machine learning training and inference is common today because today's large models require more memory and compute than can be provided by a single GPU. Distributed models are generally produced by programmers who take a sequential model specification and apply several distribution strategies to distribute state and computation across GPUs. Unfortunately, bugs can be introduced in the process, and a distributed model implementation's outputs might differ from the sequential model's outputs. In this paper, we describe an approach to statically identify such bugs by checking model refinement, that is, can the sequential model's outputs be reconstructed from the distributed model's outputs? Our approach, implemented in Entangle, uses iterative rewriting to prove model refinement. Our approach can scale to today's large models and deployments: we evaluate it using GPT and Llama-3. Further, it provides actionable outputs that aids in bug localization.

Low-latency, single-request decoding of large language models is critical for interactive systems with tight SLA demands. Prior work reduces latency through speculative decoding (combining a small draft model with a larger target model), but the draft model remains on the critical path, and communication overhead limits scaling across GPUs due to the small batch size associated with single-request decoding. To address these limitations, this paper introduces SwiftSpec: a system architecture that disaggregates draft and target models across homogeneous GPUs within a single node and utilizes NCCL-low-latency primitives directly to improve the performance of core GEMM and attention kernels. Our implementation includes 3k lines of custom CUDA for fused kernels and an evolving tree cache for KV-cache consistency and maximized reuse between draft and target models. On a single 8×H800 GPU node, SwiftSpec achieves 347 tokens/s for Llama-3-70B---1.3× faster than NVIDIA's own benchmarks on a higher-performance 8×H200 setup---and averages 1.75× faster decoding than state-of-the-art speculative decoding across five model families and six datasets. Specifically, we find that for Llama-3-70B SwiftSpec is significantly faster across all 480 tested queries, showing 1.7× speedup over the best open-source baseline for 95th percentile requests. Code for SwiftSpec will be available at https://github.com/ByteDance-Seed/SwiftSpec

We present MegaScale-MoE, a production system tailored for the efficient training of large-scale mixture-of-experts (MoE) models. MoE emerges as a promising architecture to scale large language models (LLMs) to unprecedented sizes, thereby enhancing model performance. However, existing MoE training systems experience a degradation in training efficiency, exacerbated by the escalating scale of MoE models and the continuous evolution of hardware.

Reinforcement learning (RL) post-training for Large Language Models (LLMs) is now scaling to large clusters and running for extended durations to enhance model reasoning performance. However, the scalability of existing RL frameworks is limited, as extreme long-tail skewness in RL trajectory generation causes severe GPU underutilization. Current asynchronous RL systems attempt to mitigate this, but they rely on global weight synchronization between the actor and all rollouts, which creates a rigid model update schedule. This global synchronization is ill-suited for the highly skewed and evolving distribution of trajectory generation latency in RL training, crippling training efficiency. Our key insight is that efficient scaling requires breaking this lockstep through trajectory-level asynchrony, which generates and consumes each trajectory independently. We propose Laminar, a scalable and robust RL post-training system built on a fully decoupled architecture. First, we replace global updates with a tier of relay workers acting as a distributed parameter service. This enables asynchronous and fine-grained weight synchronization, allowing rollouts to pull the latest weight anytime without stalling the actor's training loop. Second, a dynamic repack mechanism consolidates long-tail trajectories onto a few dedicated rollouts, maximizing generation throughput. The fully decoupled design also isolates failures, ensuring robustness for long-running jobs. Our evaluation on a 1024-GPU cluster shows that Laminar achieves up to 5.48$\times$ training throughput speedup over state-of-the-art systems, while reducing model convergence time.

As the foundational component of versatile AI applications, training an multimodal large language model (MLLM) relies on multimodal datasets with dynamic modality mixture proportions and sample length distributions. However, existing MLLM systems remain inefficient under dynamic workloads, due to statically coupled decisions of resource allocation and model parallelization between encoders and the LLM backbone. This paper presents MegaScale-Omni, an industrial-grade MLLM training system tailored for dynamic workload adaption and hyper-scale deployment. MegaScale-Omni is built upon the training scheme of encoder-LLM multiplexing with three key innovations: (1) Decoupled parallelism strategies with long-short sequence parallelism for encoders to process variable-length samples, and full-fledged 5D parallelism for the LLM backbone, both organized under a communication-efficient parallelization layout. (2) Unified encoder-LLM representations for flexible, extensible colocation, and a new paradigm of encoder-LLM joint pipeline with workload resilience. (3) Workload balancing techniques via decentralized grouped reordering in data loaders and adaptive resharding from encoder to LLM ranks. MegaScale-Omni is deployed as the foundation of our in-house large-scale MLLM training tasks with thousands of GPUs. Our experimental results demonstrate 1.27×–7.57× throughput improvement under production-grade dynamic workloads, as compared to four state-of-the-art systems.

Modern frameworks for training large foundation models (LFMs) employ dataloaders in a data-parallel manner, with each loader processing a disjoint subset of training data. When preparing data for LFM training that originates from multiple, distinct sources, two fundamental challenges arise. First, due to the quadratic computational complexity of the attention operator, the non-uniform sample distribution over data-parallel ranks leads to significant workload imbalance among dataloaders, degrading the training efficiency. Second, supporting diverse data sources requires per-dataset file access states that are redundantly replicated across parallel loaders, consuming excessive memory. This also hinders dynamic data mixing (e.g., curriculum learning) and causes redundant access/memory overhead in hybrid parallelism.

Modern large language models (LLMs) serving systems address distributed deployment challenges through two key techniques: distributed model partitioning for parallel computation across accelerators and quantization for reducing parameter size. While existing systems assume homogeneous GPU environments, we reveal significant untapped potential in heterogeneous systems with mixed-capacity accelerators where two critical limitations persist: (1) uniform partitioning and quantization strategies fail to adapt to hardware heterogeneity, exacerbating resource imbalance, and (2) decoupled optimization of partitioning and quantization overlooks critical performance synergies between these techniques. We present SplitQuant, a phase-aware distributed serving system that co-optimizes mixedprecision quantization, phase-aware model partitioning, and micro-batch sizing for heterogeneous environments. Our approach combines analytical modeling of quality-runtime tradeoffs with a lightweight planning algorithm to maximize throughput while preserving user-specified model quality targets. Evaluations across 10 production clusters show SplitQuant achieves up to$2.34 \times(1.61 \times$mean) higher throughput than state-of-theart approaches without violating accuracy targets. Our results underscore the value of co-designing quantization and model partitioning strategies for heterogeneous environments.

Reinforcement Learning from Human Feedback (RLHF) is widely used in Large Language Model (LLM) alignment. Traditional RL can be modeled as a dataflow, where each node represents computation of a neural network (NN) and each edge denotes data dependencies between the NNs. RLHF complicates the dataflow by expanding each node into a distributed LLM training or generation program, and each edge into a many-to-many multicast. Traditional RL frameworks execute the dataflow using a single controller to instruct both intra-node computation and inter-node communication, which can be inefficient in RLHF due to large control dispatch overhead for distributed intra-node computation. Existing RLHF systems adopt a multi-controller paradigm, which can be inflexible due to nesting distributed computation and data communication. We propose HybridFlow, which combines single-controller and multi-controller paradigms in a hybrid manner to enable flexible representation and efficient execution of the RLHF data flow. We carefully design a set of hierarchical APIs that decouple and encapsulate computation and data dependencies in the complex RLHF dataflow, allowing efficient operation orchestration to implement RLHF algorithms and flexible mapping of the computation onto various devices. We further design a 3D-HybridEngine for efficient actor model resharding between training and generation phases, with zero memory redundancy and significantly reduced communication overhead. Our experimental results demonstrate 1.53x~20.57× throughput improvement when running various RLHF algorithms using HybridFlow, as compared with state-of-the-art baselines. HybridFlow source code is available at https://github.com/volcengine/verl

Distributed training often suffers from high communication overhead due to large-scale gradient synchronization. Although gradient compression—particularly at 4-bit or even lower precision—significantly reduces transfer volume, it typically results in sacrifice in precision and degradation of the final model accuracy. In this work, we introduce DUO, a distributed training framework designed to mitigate accuracy degradation incurred by gradient compression without involving additional overhead. DUO achieves this by inserting an additional high-precision gradient synchronization step into a previously computation-only phase, so that its communication is fully hidden by computation. We provide a comprehensive theoretical proof of convergence for DUO and validate its effectiveness through extensive pre-training experiments on GPT models. Our results indicate that DUO effectively restores accuracy when using 4-bit gradient compression, achieving performance comparable to uncompressed training. Remarkably, DUO maintains minimal accuracy degradation even under extreme compression scenarios, including 1-bit gradients or complete omission of the low-precision gradient communication step (0-bit transmission).

Inference scaling empowers LLMs with unprecedented reasoning ability, with reinforcement learning as the core technique to elicit complex reasoning. However, key technical details of state-of-the-art reasoning LLMs are concealed (such as in OpenAI o1 blog and DeepSeek R1 technical report), thus the community still struggles to reproduce their RL training results. We propose the **D**ecoupled Clip and **D**ynamic s**A**mpling **P**olicy **O**ptimization (**DAPO**) algorithm, and fully open-source a state-of-the-art large-scale RL system that achieves 50 points on AIME 2024 using Qwen2.5-32B base model. Unlike previous works that withhold training details, we introduce four key techniques of our algorithm that make large-scale LLM RL a success. In addition, we open-source our training code, which is built on the verl framework, along with a carefully curated and processed dataset. These components of our open-source system enhance reproducibility and support future research in large-scale LLM RL.

Scaling long-context ability is essential for Large Language Models (LLMs). To amortize the memory consumption across multiple devices in long-context training, inter-data partitioning (a.k.a. Data Parallelism) and intra-data partitioning (a.k.a. Context Parallelism) are commonly used. Current training frameworks predominantly treat the two techniques as orthogonal, and establish static communication groups to organize the devices as a static mesh (e.g., a 2D mesh). However, the sequences for LLM training typically vary in lengths, no matter for texts, multi-modalities or reinforcement learning. The mismatch between data heterogeneity and static mesh causes redundant communication and imbalanced computation, degrading the training efficiency.

The development of large language models (LLMs) poses new challenges in data center network topology design. To assist in exploring topology design, we propose ATOP, an Automated Topology Optimization Pipeline, which models network topology as a set of hyperparameters, enabling the discovery of potential topologies. With various optimization algorithms and customizable optimization objectives, ATOP achieves automated topology optimization on a scale of tens of thousands of GPUs. We apply ATOP on network topologies for 256, 1024, 4096, and 16384 GPUs, optimizing performance under LLMs training traffic patterns, collective communication performance, fault tolerance, and network cost. We also evaluate ATOP in different scenarios: building, optimizing, and expanding a data center. From ATOP's results, we discover a new topology — ZCube, which reaches the highest cost-effectiveness across various GPU scales. Simulation results show that ZCube, compared to the previous state-of-the-art topologies, including Rail-optimized Fat-tree (ROFT), Rail-only, and HPN, improves end-to-end LLM training speed by 3% to 7% and reduces network hardware costs by 26% to 46%. We also construct ZCube on a real-world testbed. Results show that ZCube reduces hardware costs by 25% compared to Rail-Optimized Topology while maintaining the same all-reduce and all-to-all performance.

Mixture-of-Experts (MoE) showcases tremendous potential to scale large language models (LLMs) with enhanced performance and reduced computational complexity. However, its sparsely activated architecture shifts feed-forward networks (FFNs) from being compute-intensive to memory-intensive during inference, leading to substantially lower GPU utilization and increased operational costs.

The training scale of large language models (LLMs) has reached tens of thousands of GPUs and is still continuously expanding, enabling faster learning of larger models. Accompanying the expansion of the resource scale is the prevalence of failures (CUDA error, NaN values, job hang, etc.), which poses significant challenges to training stability. Any large-scale LLM training infrastructure should strive for minimal training interruption, efficient fault diagnosis, and effective failure tolerance to enable highly efficient continuous training. This paper presents ByteRobust, a large-scale GPU infrastructure management system tailored for robust and stable training of LLMs. It exploits the uniqueness of LLM training process and gives top priorities to detecting and recovering failures in a routine manner. Leveraging parallelisms and characteristics of LLM training, ByteRobust enables high-capacity fault tolerance, prompt fault demarcation, and localization with an effective data-driven approach, comprehensively ensuring continuous and efficient training of LLM tasks. ByteRobust is deployed on a production GPU platform with over 200,000 GPUs and advances the state of the art in training robustness by achieving 97% ETTR for a three-month training job on 9,600 GPUs.

Deep Neural Networks (DNNs) have shown excellent performance in a wide range of machine learning applications. Knowing the latency of running a DNN model or tensor program on a specific device is useful in various tasks, such as DNN graph- or tensor-level optimization and device selection. Considering the large space of DNN models and devices that impedes direct profiling of all combinations, recent efforts focus on building a predictor to model the performance of DNN models on different devices. However, none of the existing attempts have achieved a cost model that can accurately predict the performance of various tensor programs while supporting both training and inference accelerators. We propose CDMPP, an efficient tensor program latency prediction framework for both cross-model and cross-device prediction. We design an informative but efficient representation of tensor programs, called compact ASTs, and a pre-order-based positional encoding method, to capture the internal structure of tensor programs. We develop a domain-adaption-inspired method to learn domain-invariant representations and devise a KMeans-based sampling algorithm, for the predictor to learn from different domains (i.e., different DNN operators and devices). Our extensive experiments on a diverse range of DNN models and devices demonstrate that CDMPP significantly outperforms state-of-the-art baselines with 14.03% and 10.85% prediction error for cross-model and cross-device prediction, respectively, and one order of magnitude higher training efficiency. The implementation and the expanded dataset are available at https://github.com/joapolarbear/cdmpp.

Scaling of deep neural networks, especially Transformers, is pivotal for their surging performance and has further led to the emergence of sophisticated reasoning capabilities in foundation models. Such scaling generally requires training large models from scratch with random initialization, failing to leverage the knowledge acquired by their smaller counterparts, which are already resource-intensive to obtain. To tackle this inefficiency, we present $\textbf{L}$ossl$\textbf{E}$ss $\textbf{MO}$del Expansio$\textbf{N}$ (LEMON), a recipe to initialize scaled models using the weights of their smaller but pre-trained counterparts. This is followed by model training with an optimized learning rate scheduler tailored explicitly for the scaled models, substantially reducing the training time compared to training from scratch. Notably, LEMON is versatile, ensuring compatibility with various network structures, including models like Vision Transformers and BERT. Our empirical results demonstrate that LEMON reduces computational costs by 56.7\% for Vision Transformers and 33.2\% for BERT when compared to training from scratch.

A number of production deep learning clusters have attempted to explore inference hardware for DNN training, at the off-peak serving hours with many inference GPUs idling. Conducting DNN training with a combination of heterogeneous training and inference GPUs, known as hybrid device training, presents considerable challenges due to disparities in compute capability and significant differences in memory capacity. We propose QSync, a training system that enables efficient synchronous data-parallel DNN training over hybrid devices by strategically exploiting quantized operators. According to each device’s available resource capacity, QSync selects a quantization-minimized setting for operators in the distributed DNN training graph, minimizing model accuracy degradation but keeping the training efficiency brought by quantization. We carefully design a predictor with a bi-directional mixed-precision indicator to reflect the sensitivity of DNN layers on fixed-point and floating-point low-precision operators, a replayer with a neighborhood-aware cost mapper to accurately estimate the latency of distributed hybrid mixed-precision training, and then an allocator that efficiently synchronizes workers with minimized model accuracy degradation. QSync bridges the computational graph on PyTorch to an optimized backend for quantization kernel performance and flexible support for various GPU architectures. Extensive experiments show that QSync’s predictor can accurately simulate distributed mixed-precision training with < 5% error, with a consistent 0.27 − 1.03% accuracy improvement over the from-scratch training tasks compared to uniform precision.

Recent years have witnessed a clear trend towards language models with an ever-increasing number of parameters, as well as the growing training overhead and memory usage. Distributed training, particularly through Sharded Data Parallelism (ShardedDP) which partitions optimizer states among workers, has emerged as a crucial technique to mitigate training time and memory usage. Yet, a major challenge in the scalability of ShardedDP is the intensive communication of weights and gradients. While compression techniques can alleviate this issue, they often result in worse accuracy. Driven by this limitation, we propose SDP4Bit (Toward 4Bit Communication Quantization in Sharded Data Parallelism for LLM Training), which effectively reduces the communication of weights and gradients to nearly 4 bits via two novel techniques: quantization on weight differences, and two-level gradient smooth quantization. Furthermore, SDP4Bit presents an algorithm-system co-design with runtime optimization to minimize the computation overhead of compression. Additional to the theoretical guarantees of convergence, we empirically evaluate the accuracy of SDP4Bit on the pre-training of GPT models with up to 6.7 billion parameters, and the results demonstrate a negligible impact on training loss. Furthermore, speed experiments show that SDP4Bit achieves up to 4.08× speedup in end-to-end throughput on a scale of 128 GPUs.

The immense sizes of Large-scale language models (LLMs) have led to high resource demand and cost for running the models. Though the models are largely served using uniform high-caliber GPUs nowadays, utilizing a heterogeneous cluster with a mix of available high- and low-capacity GPUs can potentially substantially reduce the serving cost. This paper proposes LLM-PQ, a system that advocates adaptive model quantization and phase-aware partition to improve LLM serving efficiency on heterogeneous GPU clusters. Extensive experiments on production inference workloads demonstrate throughput improvement in inference, showing great advantages over state-of-the-art works.

Gradient compression (GC) is a promising approach to addressing the communication bottleneck in distributed deep learning (DDL). It saves the communication time, but also incurs additional computation overheads. The training throughput of compression-enabled DDL is determined by the compression strategy, including whether to compress each tensor, the type of compute resources (e.g., CPUs or GPUs) for compression, the communication schemes for compressed tensor, and so on. However, it is challenging to find the optimal compression strategy for applying GC to DDL because of the intricate interactions among tensors. To fully unleash the benefits of GC, two questions must be addressed: 1) How to express any compression strategies and the corresponding interactions among tensors of any DDL training job? 2) How to quickly select a near-optimal compression strategy?

Data parallelism across multiple machines is widely adopted for accelerating distributed deep learning, but it is hard to achieve linear speedup due to the heavy communication. In this paper, we propose SAPipe, a performant system that pushes the training speed of data parallelism to its fullest extent. By introducing partial staleness, the communication overlaps the computation with minimal staleness in SAPipe. To mitigate additional problems incurred by staleness, SAPipe adopts staleness compensation techniques including weight prediction and delay compensation with provably lower error bounds. Additionally, SAPipe presents an algorithm-system co-design with runtime optimization to minimize system overhead for the staleness training pipeline and staleness compensation. We have implemented SAPipe in the BytePS framework, compatible to both TensorFlow and PyTorch. Our experiments show that SAPipe achieves up to 157% speedups over BytePS (non-stale), and outperforms PipeSGD in accuracy by up to 13.7%.

Personalized real-time recommendation has had a profound impact on retail, media, entertainment and other industries. However, developing recommender systems for every use case is costly, time consuming and resource-intensive. To fill this gap, we present a black-box recommender system that can adapt to a diverse set of scenarios without the need for manual tuning. We build on techniques that go beyond simple matrix factorization to incorporate important new sources of information: the temporal order of events [Hidasi et al., 2015], contextual information to bootstrap cold-start users, metadata information about items [Rendle 2012] and the additional information surrounding each event. Additionally, we address two fundamental challenges when putting recommender systems in the real-world: how to efficiently train them with even millions of unique items and how to cope with changing item popularity trends [Wu et al., 2017]. We introduce a compact model, which we call hierarchical recurrent network with meta data (HRNN-meta) to address the real-time and diverse metadata needs; we further provide efficient training techniques via importance sampling that can scale to millions of items with little loss in performance. We report significant improvements on a wide range of real-world datasets and provide intuition into model capabilities with synthetic experiments. Parts of HRNN-meta have been deployed in production at scale for customers to use at Amazon Web Services and serves as the underlying recommender engine for thousands of websites.

The scalability of Distributed Stochastic Gradient Descent (SGD) is today limited by communication bottlenecks. We propose a novel SGD variant: \underline{C}ommunication-efficient \underline{S}GD with \underline{E}rror \underline{R}eset, or \underline{CSER}. The key idea in CSER is first a new technique called ``error reset'' that adapts arbitrary compressors for SGD, producing bifurcated local models with periodic reset of resulting local residual errors. Second we introduce partial synchronization for both the gradients and the models, leveraging advantages from them. We prove the convergence of CSER for smooth non-convex problems. Empirical results show that when combined with highly aggressive compressors, the CSER algorithms accelerate the distributed training by nearly $10\times$ for CIFAR-100, and by $4.5\times$ for ImageNet.

We present GluonCV and GluonNLP, the deep learning toolkits for computer vision and natural language processing based on Apache MXNet (incubating). These toolkits provide state-of-the-art pre-trained models, training scripts, and training logs, to facilitate rapid prototyping and promote reproducible research. We also provide modular APIs with flexible building blocks to enable efficient customization. Leveraging the MXNet ecosystem, the deep learning models in GluonCV and GluonNLP can be deployed onto a variety of platforms with different programming languages. The Apache 2.0 license has been adopted by GluonCV and GluonNLP to allow for software distribution, modification, and usage.

In the literature of industrial designing and digital entertainment, the resizing of 3D complex models is a popular and useful operation. One of the key issues in model resizing is that some visually salient features of the underlying model should be preserved as much as possible. Based on the surface deformation scheme, a novel visual saliency guided content-aware model resizing approach is presented in this paper. Firstly, the edge sensitivity measure is determined by edge saliency measure and slippage value. Owing to the edge sensitivity, a quadratic energy function can thus be constructed to guide salient feature-preserving model resizing. Finally, the original model can be non-homogeneously scaled according to the per-edge scales solved by iteratively minimizing the quadratic energy function. Our visual saliency guided model resizing scheme can be benefit not only for the global resizing procedure but also for the local resizing operation. The experimental results demonstrate that our proposed model resizing scheme can effectively preserve the visually salient features of the underlying model in a content-aware manner.