Haibin Lin

Guangming Sheng, Chi Zhang, Zilingfeng Ye et al.

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 dataflow. 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 1.53$\times$~20.57$\times$ throughput improvement when running various RLHF algorithms using HybridFlow, as compared with state-of-the-art baselines. HybridFlow source code will be available at https://github.com/volcengine/verl.

Weiqi Feng, Yangrui Chen, Shaoyu Wang et al.

Multimodal large language models (MLLMs) have extended the success of large language models (LLMs) to multiple data types, such as image, text and audio, achieving significant performance in various domains, including multimodal translation, visual question answering and content generation. Nonetheless, existing systems are inefficient to train MLLMs due to substantial GPU bubbles caused by the heterogeneous modality models and complex data dependencies in 3D parallelism. This paper proposes Optimus, a distributed MLLM training system that reduces end-to-end MLLM training time. Optimus is based on our principled analysis that scheduling the encoder computation within the LLM bubbles can reduce bubbles in MLLM training. To make scheduling encoder computation possible for all GPUs, Optimus searches the separate parallel plans for encoder and LLM, and adopts a bubble scheduling algorithm to enable exploiting LLM bubbles without breaking the original data dependencies in the MLLM model architecture. We further decompose encoder layer computation into a series of kernels, and analyze the common bubble pattern of 3D parallelism to carefully optimize the sub-millisecond bubble scheduling, minimizing the overall training time. Our experiments in a production cluster show that Optimus accelerates MLLM training by 20.5%-21.3% with ViT-22B and GPT-175B model over 3072 GPUs compared to baselines.

Juntao Zhao, Borui Wan, Yanghua Peng et al.

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.

Ziheng Jiang, Haibin Lin, Yinmin Zhong et al.

We present the design, implementation and engineering experience in building and deploying MegaScale, a production system for training large language models (LLMs) at the scale of more than 10,000 GPUs. Training LLMs at this scale brings unprecedented challenges to training efficiency and stability. We take a full-stack approach that co-designs the algorithmic and system components across model block and optimizer design, computation and communication overlapping, operator optimization, data pipeline, and network performance tuning. Maintaining high efficiency throughout the training process (i.e., stability) is an important consideration in production given the long extent of LLM training jobs. Many hard stability issues only emerge at large scale, and in-depth observability is the key to address them. We develop a set of diagnosis tools to monitor system components and events deep in the stack, identify root causes, and derive effective techniques to achieve fault tolerance and mitigate stragglers. MegaScale achieves 55.2% Model FLOPs Utilization (MFU) when training a 175B LLM model on 12,288 GPUs, improving the MFU by 1.34x compared to Megatron-LM. We share our operational experience in identifying and fixing failures and stragglers. We hope by articulating the problems and sharing our experience from a systems perspective, this work can inspire future LLM systems research.

Juntao Zhao, Borui Wan, Yanghua Peng et al.

Recent breakthroughs in Large-scale language models (LLMs) have demonstrated impressive performance on various tasks. The immense sizes of LLMs have led to very 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. There is a lack of designs to support efficient LLM serving using a heterogeneous cluster, while the current solutions focus on model partition and uniform compression among homogeneous devices. 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. We carefully decide on mixed-precision model quantization together with phase-aware model partition and micro-batch sizing in distributed LLM serving with an efficient algorithm, to greatly enhance inference throughput while fulfilling user-specified model quality targets. Extensive experiments on production inference workloads in 11 different clusters demonstrate that LLM-PQ achieves up to 2.88x (2.26x on average) throughput improvement in inference, showing great advantages over state-of-the-art works.

Ruidong Zhu, Ziheng Jiang, Chao Jin et al.

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. We present MegaScale-Infer, an efficient and cost-effective system for serving large-scale MoE models. MegaScale-Infer disaggregates attention and FFN modules within each model layer, enabling independent scaling, tailored parallelism strategies, and heterogeneous deployment for both modules. To fully exploit disaggregation in the presence of MoE's sparsity, MegaScale-Infer introduces ping-pong pipeline parallelism, which partitions a request batch into micro-batches and shuttles them between attention and FFNs for inference. Combined with distinct model parallelism for each module, MegaScale-Infer effectively hides communication overhead and maximizes GPU utilization. To adapt to disaggregated attention and FFN modules and minimize data transmission overhead (e.g., token dispatch), MegaScale-Infer provides a high-performance M2N communication library that eliminates unnecessary GPU-to-CPU data copies, group initialization overhead, and GPU synchronization. Experimental results indicate that MegaScale-Infer achieves up to 1.90x higher per-GPU throughput than state-of-the-art solutions.

Yangtao Deng, Xiang Shi, Zhuo Jiang et al.

Large-scale distributed model training requires simultaneous training on up to thousands of machines. Faulty machine detection is critical when an unexpected fault occurs in a machine. From our experience, a training task can encounter two faults per day on average, possibly leading to a halt for hours. To address the drawbacks of the time-consuming and labor-intensive manual scrutiny, we propose Minder, an automatic faulty machine detector for distributed training tasks. The key idea of Minder is to automatically and efficiently detect faulty distinctive monitoring metric patterns, which could last for a period before the entire training task comes to a halt. Minder has been deployed in our production environment for over one year, monitoring daily distributed training tasks where each involves up to thousands of machines. In our real-world fault detection scenarios, Minder can accurately and efficiently react to faults within 3.6 seconds on average, with a precision of 0.904 and F1-score of 0.893.

Jinda Jia, Cong Xie, Hanlin Lu et al.

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. In addition 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$\times$ speedup in end-to-end throughput on a scale of 128 GPUs.

Jinkun Lin, Ziheng Jiang, Zuquan Song et al.

Large language model (LLM) training is one of the most demanding distributed computations today, often requiring thousands of GPUs with frequent synchronization across machines. Such a workload pattern makes it susceptible to stragglers, where the training can be stalled by few slow workers. At ByteDance we find stragglers are not trivially always caused by hardware failures, but can arise from multiple complex factors. This work aims to present a comprehensive study on the straggler issues in LLM training, using a five-month trace collected from our ByteDance LLM training cluster. The core methodology is what-if analysis that simulates the scenario without any stragglers and contrasts with the actual case. We use this method to study the following questions: (1) how often do stragglers affect training jobs, and what effect do they have on job performance; (2) do stragglers exhibit temporal or spatial patterns; and (3) what are the potential root causes for stragglers?

Hao Ge, Junda Feng, Qi Huang et al.

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. In this work, we introduce ByteScale, an efficient, flexible, and scalable LLM training framework for large-scale mixed training of long and short sequences. The core of ByteScale is a novel parallelism strategy, namely Hybrid Data Parallelism (HDP), which unifies the inter- and intra-data partitioning with a dynamic mesh design. In particular, we build a communication optimizer, which eliminates the redundant communication for short sequences by data-aware sharding and dynamic communication, and further compresses the communication cost for long sequences by selective offloading. Besides, we also develop a balance scheduler to mitigate the imbalanced computation by parallelism-aware data assignment. We evaluate ByteScale with the model sizes ranging from 7B to 141B, context lengths from 256K to 2048K, on a production cluster with more than 12,000 GPUs. Experiment results show that ByteScale outperforms the state-of-the-art training system by up to 7.89x.

Ziyi Zhang, Ziheng Jiang, Chengquan Jiang et al.

Low-latency decoding for large language models (LLMs) is crucial for applications like chatbots and code assistants, yet generating long outputs remains slow in single-query settings. Prior work on speculative decoding (which combines a small draft model with a larger target model) and tensor parallelism has each accelerated decoding. However, conventional approaches fail to apply both simultaneously due to imbalanced compute requirements (between draft and target models), KV-cache inconsistencies, and communication overheads under small-batch tensor-parallelism. This paper introduces SwiftSpec, a system that targets ultra-low latency for LLM decoding. SwiftSpec redesigns the speculative decoding pipeline in an asynchronous and disaggregated manner, so that each component can be scaled flexibly and remove draft overhead from the critical path. To realize this design, SwiftSpec proposes parallel tree generation, tree-aware KV cache management, and fused, latency-optimized kernels to overcome the challenges listed above. Across 5 model families and 6 datasets, SwiftSpec achieves an average of 1.75x speedup over state-of-the-art speculative decoding systems and, as a highlight, serves Llama3-70B at 348 tokens/s on 8 Nvidia Hopper GPUs, making it the fastest known system for low-latency LLM serving at this scale.

Juntao Zhao, Qi Lu, Wei Jia et al.

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. Under multisource preprocessing, two fundamental challenges exist. 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. We present Omniload, an industrial-grade distributed data loading architecture for LFMs, with four innovations: (1) Disaggregated data preprocessing via role-specific actors (Source Loaders/Data Constructors) to eliminate source and parallelism redundant data access and ensure multisource scalability. (2) Centralized and declarative data plane for elastic multisource orchestration, such as long-short context, multimodality, and curriculum learning. (3) Multi-level auto-partitioning and scaling mechanism for source loaders under heterogeneous preprocessing costs. (4) Shadow loaders with differential checkpointing for fault recovery without workflow interruption. Deployed on production clusters scaling to multi-thousand GPUs, Omniload achieves: (1) 4.5x end-to-end training throughput improvement, (2) 13.5x reduction in CPU memory usage.

Guangming Sheng, Yuxuan Tong, Borui Wan et al.

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.

Youjie Li, Cheng Wan, Zhiqi Lin et al.

Large Language Models (LLMs) have scaled rapidly in size and complexity, requiring increasingly intricate parallelism for distributed training, such as 3D parallelism. This sophistication motivates a shift toward simpler, more debuggable programming paradigm like Single Program Multiple Data (SPMD). However, SPMD in eager execution introduces two key challenges: ensuring consistency with single-device execution and achieving high performance at scale. In this paper, we introduce veScale, an eager-mode training system that fully embraces SPMD paradigm to democratize distributed tensor programming. veScale addresses the prevalent issue of inconsistent results in systems like PyTorch by introducing a novel algorithm of distributed Random Number Generation (RNG) compatible with arbitrary sharded operators. veScale also significantly boosts training performance by reducing PyTorch primitive's overhead and improving communication efficiency. Evaluations show that veScale delivers up to 2.2x speedup over the state-of-the-art training systems, like TorchTitan, and cuts code complexity by 78.4%, while preserving single-device-equivalent results.

Li-Wen Chang, Wenlei Bao, Qi Hou et al.

Large deep learning models have demonstrated strong ability to solve many tasks across a wide range of applications. Those large models typically require training and inference to be distributed. Tensor parallelism is a common technique partitioning computation of an operation or layer across devices to overcome the memory capacity limitation of a single processor, and/or to accelerate computation to meet a certain latency requirement. However, this kind of parallelism introduces additional communication that might contribute a significant portion of overall runtime. Thus limits scalability of this technique within a group of devices with high speed interconnects, such as GPUs with NVLinks in a node. This paper proposes a novel method, Flux, to significantly hide communication latencies with dependent computations for GPUs. Flux over-decomposes communication and computation operations into much finer-grained operations and further fuses them into a larger kernel to effectively hide communication without compromising kernel efficiency. Flux can potentially overlap up to 96% of communication given a fused kernel. Overall, it can achieve up to 1.24x speedups for training over Megatron-LM on a cluster of 128 GPUs with various GPU generations and interconnects, and up to 1.66x and 1.30x speedups for prefill and decoding inference over vLLM on a cluster with 8 GPUs with various GPU generations and interconnects.

Yuchen Zhong, Cong Xie, Shuai Zheng et al.

Communication overhead severely hinders the scalability of distributed machine learning systems. Recently, there has been a growing interest in using gradient compression to reduce the communication overhead of the distributed training. However, there is little understanding of applying gradient compression to adaptive gradient methods. Moreover, its performance benefits are often limited by the non-negligible compression overhead. In this paper, we first introduce a novel adaptive gradient method with gradient compression. We show that the proposed method has a convergence rate of $\mathcal{O}(1/\sqrt{T})$ for non-convex problems. In addition, we develop a scalable system called BytePS-Compress for two-way compression, where the gradients are compressed in both directions between workers and parameter servers. BytePS-Compress pipelines the compression and decompression on CPUs and achieves a high degree of parallelism. Empirical evaluations show that we improve the training time of ResNet50, VGG16, and BERT-base by 5.0%, 58.1%, 23.3%, respectively, without any accuracy loss with 25 Gb/s networking. Furthermore, for training the BERT models, we achieve a compression rate of 333x compared to the mixed-precision training.

Cong Xie, Shuai Zheng, Oluwasanmi Koyejo et al.

The scalability of Distributed Stochastic Gradient Descent (SGD) is today limited by communication bottlenecks. We propose a novel SGD variant: Communication-efficient SGD with Error Reset, or 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 10x for CIFAR-100, and by 4.5x for ImageNet.

Shuai Zheng, Haibin Lin, Sheng Zha et al.

BERT has recently attracted a lot of attention in natural language understanding (NLU) and achieved state-of-the-art results in various NLU tasks. However, its success requires large deep neural networks and huge amount of data, which result in long training time and impede development progress. Using stochastic gradient methods with large mini-batch has been advocated as an efficient tool to reduce the training time. Along this line of research, LAMB is a prominent example that reduces the training time of BERT from 3 days to 76 minutes on a TPUv3 Pod. In this paper, we propose an accelerated gradient method called LANS to improve the efficiency of using large mini-batches for training. As the learning rate is theoretically upper bounded by the inverse of the Lipschitz constant of the function, one cannot always reduce the number of optimization iterations by selecting a larger learning rate. In order to use larger mini-batch size without accuracy loss, we develop a new learning rate scheduler that overcomes the difficulty of using large learning rate. Using the proposed LANS method and the learning rate scheme, we scaled up the mini-batch sizes to 96K and 33K in phases 1 and 2 of BERT pretraining, respectively. It takes 54 minutes on 192 AWS EC2 P3dn.24xlarge instances to achieve a target F1 score of 90.5 or higher on SQuAD v1.1, achieving the fastest BERT training time in the cloud.

Zhen Zhang, Chaokun Chang, Haibin Lin et al.

Recently there has been a surge of research on improving the communication efficiency of distributed training. However, little work has been done to systematically understand whether the network is the bottleneck and to what extent. In this paper, we take a first-principles approach to measure and analyze the network performance of distributed training. As expected, our measurement confirms that communication is the component that blocks distributed training from linear scale-out. However, contrary to the common belief, we find that the network is running at low utilization and that if the network can be fully utilized, distributed training can achieve a scaling factor of close to one. Moreover, while many recent proposals on gradient compression advocate over 100x compression ratio, we show that under full network utilization, there is no need for gradient compression in 100 Gbps network. On the other hand, a lower speed network like 10 Gbps requires only 2x--5x gradients compression ratio to achieve almost linear scale-out. Compared to application-level techniques like gradient compression, network-level optimizations do not require changes to applications and do not hurt the performance of trained models. As such, we advocate that the real challenge of distributed training is for the network community to develop high-performance network transport to fully utilize the network capacity and achieve linear scale-out.

Cong Xie, Oluwasanmi Koyejo, Indranil Gupta et al.

When scaling distributed training, the communication overhead is often the bottleneck. In this paper, we propose a novel SGD variant with reduced communication and adaptive learning rates. We prove the convergence of the proposed algorithm for smooth but non-convex problems. Empirical results show that the proposed algorithm significantly reduces the communication overhead, which, in turn, reduces the training time by up to 30% for the 1B word dataset.

Jian Guo, He He, Tong He et al.

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.

Haibin Lin, Hang Zhang, Yifei Ma et al.

With an increasing demand for training powers for deep learning algorithms and the rapid growth of computation resources in data centers, it is desirable to dynamically schedule different distributed deep learning tasks to maximize resource utilization and reduce cost. In this process, different tasks may receive varying numbers of machines at different time, a setting we call elastic distributed training. Despite the recent successes in large mini-batch distributed training, these methods are rarely tested in elastic distributed training environments and suffer degraded performance in our experiments, when we adjust the learning rate linearly immediately with respect to the batch size. One difficulty we observe is that the noise in the stochastic momentum estimation is accumulated over time and will have delayed effects when the batch size changes. We therefore propose to smoothly adjust the learning rate over time to alleviate the influence of the noisy momentum estimation. Our experiments on image classification, object detection and semantic segmentation have demonstrated that our proposed Dynamic SGD method achieves stabilized performance when varying the number of GPUs from 8 to 128. We also provide theoretical understanding on the optimality of linear learning rate scheduling and the effects of stochastic momentum.