Xiaonan Nie, Xupeng Miao, Zilong Wang, Zichao Yang, Jilong Xue, Lingxiao Ma, Gang Cao, Bin Cui
With the increasing data volume, there is a trend of using large-scale pre-trained models to store the knowledge into an enormous number of model parameters. The training of these models is composed of lots of dense algebras, requiring a huge amount of hardware resources. Recently, sparsely-gated Mixture-of-Experts (MoEs) are becoming more popular and have demonstrated impressive pretraining scalability in various downstream tasks. However, such a sparse conditional computation may not be effective as expected in practical systems due to the routing imbalance and fluctuation problems. Generally, MoEs are becoming a new data analytics paradigm in the data life cycle and suffering from unique challenges at scales, complexities, and granularities never before possible. In this paper, we propose a novel DNN training framework, FlexMoE, which systematically and transparently address the inefficiency caused by dynamic dataflow. We first present an empirical analysis on the problems and opportunities of training MoE models, which motivates us to overcome the routing imbalance and fluctuation problems by a dynamic expert management and device placement mechanism. Then we introduce a novel scheduling module over the existing DNN runtime to monitor the data flow, make the scheduling plans, and dynamically adjust the model-to-hardware mapping guided by the real-time data traffic. A simple but efficient heuristic algorithm is exploited to dynamically optimize the device placement during training. We have conducted experiments on both NLP models (e.g., BERT and GPT) and vision models (e.g., Swin). And results show FlexMoE can achieve superior performance compared with existing systems on real-world workloads -- FlexMoE outperforms DeepSpeed by 1.70x on average and up to 2.10x, and outperforms FasterMoE by 1.30x on average and up to 1.45x.
Xiaonan Nie, Xupeng Miao, Shijie Cao, Lingxiao Ma, Qibin Liu, Jilong Xue, Youshan Miao, Yi Liu, Zhi Yang, Bin Cui
Mixture-of-experts (MoE) is becoming popular due to its success in improving the model quality, especially in Transformers. By routing tokens with a sparse gate to a few experts (i.e., a small pieces of the full model), MoE can easily increase the model parameters to a very large scale while keeping the computation cost in a constant level. Most existing works just initialize some random experts, set a fixed gating strategy (e.g., Top-k), and train the model from scratch in an ad-hoc way. We identify that these MoE models are suffering from the immature experts and unstable sparse gate, which are harmful to the convergence performance. In this paper, we propose an efficient end-to-end MoE training framework called EvoMoE. EvoMoE starts from training one single expert and gradually evolves into a large and sparse MoE structure. EvoMoE mainly contains two phases: the expert-diversify phase to train the base expert for a while and spawn multiple diverse experts from it, and the gate-sparsify phase to learn an adaptive sparse gate and activate a dynamic number of experts. EvoMoE naturally decouples the joint learning of both the experts and the sparse gate and focuses on learning the basic knowledge with a single expert at the early training stage. Then it diversifies the experts and continues to train the MoE with a novel Dense-to-Sparse gate (DTS-Gate). Specifically, instead of using a permanent sparse gate, DTS-Gate begins as a dense gate that routes tokens to all experts, then gradually and adaptively becomes sparser while routes to fewer experts. Evaluations are conducted on three popular models and tasks, including RoBERTa for masked language modeling task, GPT for language modeling task and Transformer for machine translation task. The results show that EvoMoE outperforms existing baselines, including Switch, BASE Layer, Hash Layer and StableMoE.
Xiaonan Nie, Pinxue Zhao, Xupeng Miao, Tong Zhao, Bin Cui
As giant dense models advance quality but require large amounts of GPU budgets for training, the sparsely gated Mixture-of-Experts (MoE), a kind of conditional computation architecture, is proposed to scale models while keeping their computation constant. Specifically, the input tokens are routed by the gate network and only activates part of the expert network. Existing MoE training systems only support part of mainstream MoE models (e.g. Top k) training under expensive high-bandwidth GPU clusters. In this paper, we present HetuMoE, a high-performance large-scale sparse MoE training system built on Hetu. HetuMoE provides multiple gating strategies and efficient GPU kernel implementations. To further improve the training efficiency on commodity GPU clusters (e.g, with only 1 NiC), we introduce the hierarchical AllToAll communication that combines hierarchical networks and aggregating messages. Compared with existing state-of-the-art MoE systems, HetuMoE obtains at least 15% speedup. Specifically, HetuMoE outperforms DeepSpeed-MoE up to 8.1x under the switch gate with a batch size of 32. Our code is available at: https://github.com/PKU-DAIR/Hetu.
Xiaonan Nie, Qibin Liu, Fangcheng Fu, Shenhan Zhu, Xupeng Miao, Xiaoyang Li, Yang Zhang, Shouda Liu, Bin Cui
Larger transformer models always perform better on various tasks but require more costs to scale up the model size. To efficiently enlarge models, the mixture-of-experts (MoE) architecture is widely adopted, which consists of a gate network and a series of experts and keep the training cost constant by routing the input data to a fixed number of experts instead of all. In existing large-scale MoE training systems, experts would be distributed among different GPUs for parallelization, and thus input data requires additional all-to-all communications to access the target experts and conduct corresponding computations. However, upon evaluating the training process of three mainstream MoE models on commonly used GPU clusters, we found that the all-to-all communication ratio averaged around 45%, which significantly hinders the efficiency and scalability of training MoE models. In this paper, we propose LSH-MoE, a communication-efficient MoE training framework using locality-sensitive hashing (LSH). We first present the problems of scaling MoE training in existing systems and highlight the potential of exploiting token similarity to facilitate data compression. Then, we introduce an efficient LSH-based compression technique, which utilizes the cross-polytope hashing for rapid clustering and implements a residual-based error compensation scheme to alleviate the adverse impact of compression. To verify the effectiveness of our methods, we conduct experiments on both language models (e.g., RoBERTa, GPT, and T5) and vision models (e.g., Swin) for pre-training and fine-tuning tasks. The results demonstrate that our method substantially outperforms its counterparts across different tasks by 1.28x - 2.2x of speedup.
Xiaonan Nie, Yi Liu, Fangcheng Fu, Jinbao Xue, Dian Jiao, Xupeng Miao, Yangyu Tao, Bin Cui
Recent years have witnessed the unprecedented achievements of large-scale pre-trained models, especially the Transformer models. Many products and services in Tencent Inc., such as WeChat, QQ, and Tencent Advertisement, have been opted in to gain the power of pre-trained models. In this work, we present Angel-PTM, a productive deep learning system designed for pre-training and fine-tuning Transformer models. Angel-PTM can train extremely large-scale models with hierarchical memory efficiently. The key designs of Angel-PTM are the fine-grained memory management via the Page abstraction and a unified scheduling method that coordinate the computations, data movements, and communications. Furthermore, Angel-PTM supports extreme model scaling with SSD storage and implements the lock-free updating mechanism to address the SSD I/O bandwidth bottlenecks. Experimental results demonstrate that Angel-PTM outperforms existing systems by up to 114.8% in terms of maximum model scale as well as up to 88.9% in terms of training throughput. Additionally, experiments on GPT3-175B and T5-MoE-1.2T models utilizing hundreds of GPUs verify the strong scalability of Angel-PTM.
Xupeng Miao, Yujie Wang, Youhe Jiang, Chunan Shi, Xiaonan Nie, Hailin Zhang, Bin Cui
Transformer models have achieved state-of-the-art performance on various domains of applications and gradually becomes the foundations of the advanced large deep learning (DL) models. However, how to train these models over multiple GPUs efficiently is still challenging due to a large number of parallelism choices. Existing DL systems either rely on manual efforts to make distributed training plans or apply parallelism combinations within a very limited search space. In this approach, we propose Galvatron, a new system framework that incorporates multiple popular parallelism dimensions and automatically finds the most efficient hybrid parallelism strategy. To better explore such a rarely huge search space, we 1) involve a decision tree to make decomposition and pruning based on some reasonable intuitions, and then 2) design a dynamic programming search algorithm to generate the optimal plan. Evaluations on four representative Transformer workloads show that Galvatron could perform automatically distributed training with different GPU memory budgets. Among all evluated scenarios, Galvatron always achieves superior system throughput compared to previous work with limited parallelism.
Pinxue Zhao, Hailin Zhang, Fangcheng Fu, Xiaonan Nie, Qibin Liu, Fang Yang, Yuanbo Peng, Dian Jiao, Shuaipeng Li, Jinbao Xue, Yangyu Tao, Bin Cui
Nowadays, Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications. However, long context training poses great challenges considering the constraint of GPU memory. It not only leads to substantial activation memory consumption during training, but also incurs considerable memory fragmentation. To facilitate long context training, existing frameworks have adopted strategies such as recomputation and various forms of parallelisms. Nevertheless, these techniques rely on redundant computation or extensive communication, resulting in low Model FLOPS Utilization (MFU). In this paper, we propose MEMO, a novel LLM training framework designed for fine-grained activation memory management. Given the quadratic scaling of computation and linear scaling of memory with sequence lengths when using FlashAttention, we offload memory-consuming activations to CPU memory after each layer's forward pass and fetch them during the backward pass. To maximize the swapping of activations without hindering computation, and to avoid exhausting limited CPU memory, we implement a token-wise activation recomputation and swapping mechanism. Furthermore, we tackle the memory fragmentation issue by employing a bi-level Mixed Integer Programming (MIP) approach, optimizing memory reuse across transformer layers. Empirical results demonstrate that MEMO achieves an average of 1.97x and 1.80x MFU compared to Megatron-LM and DeepSpeed, respectively. This improvement is attributed to MEMO's ability to minimize memory fragmentation, reduce recomputation and intensive communication, and circumvent the delays associated with the memory reorganization process due to fragmentation. By leveraging fine-grained activation memory management, MEMO facilitates efficient training of 7B LLM with 1 million sequence length on just 8 A800 GPUs, achieving an MFU of 52.30%.
Hailin Zhang, Xiaodong Ji, Yilin Chen, Fangcheng Fu, Xupeng Miao, Xiaonan Nie, Weipeng Chen, Bin Cui
As the field of Large Language Models (LLMs) continues to evolve, the context length in inference is steadily growing. Key-Value Cache (KVCache), the intermediate representations of tokens within LLM inference, has now become the primary memory bottleneck due to limited GPU memory. Current methods selectively determine suitable keys and values for self-attention computation in LLMs to address the issue. However, they either fall short in maintaining model quality or result in high serving latency. Drawing inspiration from advanced embedding retrieval techniques prevalent in the data management community, we consider the storage and retrieval of KVCache as a typical embedding retrieval problem. We propose PQCache, which employs Product Quantization (PQ) to manage KVCache, maintaining model quality while ensuring low serving latency. During the prefilling phase, we apply PQ to tokens' keys for each LLM layer and head. During the autoregressive decoding phase, we use PQ codes and centroids to approximately identify important preceding tokens, then fetch the corresponding key-value pairs for self-attention computation. Through meticulous design of overlapping and caching, we minimize any additional computation and communication overhead during both phases. Extensive experiments demonstrate that PQCache achieves both effectiveness and efficiency, with 4.60% score improvement over existing methods on InfiniteBench and low system latency in both prefilling and decoding.
Ceyuan Yang, Zhijie Lin, Yang Zhao, Fei Xiao, Hao He, Qi Zhao, Chaorui Deng, Kunchang Li, Zihan Ding, Yuwei Guo, Fuyun Wang, Fangqi Zhu, Xiaonan Nie, Shenhan Zhu, Shanchuan Lin, Hongsheng Li, Weilin Huang, Guang Shi, Haoqi Fan
We present Omni, a unified multimodal model natively trained on diverse modalities, including text, images, videos, 3D geometry, and hidden representations. We find that such training enables Context Unrolling, where the model explicitly reasons across multiple modal representations before producing predictions. This process enables the model to aggregate complementary information across heterogeneous modalities, facilitating a more faithful approximation of the shared multimodal knowledge manifold and improving downstream reasoning fidelity. As a result, Omni achieves strong performance on both multimodal generation and understanding benchmarks, while demonstrating advanced multimodal reasoning capabilities, including in-context generation of text, image, video, and 3D geometry.
Guosheng Dong, Da Pan, Yiding Sun, Shusen Zhang, Zheng Liang, Xin Wu, Yanjun Shen, Fan Yang, Haoze Sun, Tianpeng Li, Mingan Lin, Jianhua Xu, Yufan Zhang, Xiaonan Nie, Lei Su, Bingning Wang, Wentao Zhang, Jiaxin Mao, Zenan Zhou, Weipeng Chen
The general capabilities of Large Language Models (LLM) highly rely on the composition and selection on extensive pretraining datasets, treated as commercial secrets by several institutions. To mitigate this issue, we open-source the details of a universally applicable data processing pipeline and validate its effectiveness and potential by introducing a competitive LLM baseline. Specifically, the data processing pipeline consists of broad collection to scale up and reweighting to improve quality. We then pretrain a 7B model BaichuanSEED with 3T tokens processed by our pipeline without any deliberate downstream task-related optimization, followed by an easy but effective supervised fine-tuning stage. BaichuanSEED demonstrates consistency and predictability throughout training and achieves comparable performance on comprehensive benchmarks with several commercial advanced large language models, such as Qwen1.5 and Llama3. We also conduct several heuristic experiments to discuss the potential for further optimization of downstream tasks, such as mathematics and coding.
Haoyang Li, Fangcheng Fu, Hao Ge, Sheng Lin, Xuanyu Wang, Jiawen Niu, Yujie Wang, Hailin Zhang, Xiaonan Nie, Bin Cui
As the scale of models and training data continues to grow, there is an expanding reliance on more GPUs to train large-scale models, which inevitably increases the likelihood of encountering dynamic stragglers that some devices lag behind in performance occasionally. However, hybrid parallel training, one of the de facto paradigms to train large models, is typically sensitive to the stragglers. This paper presents Malleus, a straggler-resilient hybrid parallel training framework for large-scale models. Malleus quantifies the stragglers at the nuanced, per-GPU granularity during training, and develops a novel planning algorithm to deduce the optimal parallelization of GPU devices, pipeline stages, model layers, and training data, maximizing training efficiency when stragglers exist. In addition, once a shift in the straggler situation is detected, Malleus adaptively adjusts the parallelization via a re-planning process, and seamlessly and efficiently migrates the model states on the fly, without sacrificing the stability of the training tasks. Empirical results on large language models with up to 110B parameters show that Malleus consistently outperforms existing parallel training frameworks under various straggler situations, delivering on average 2.63-5.28 times of efficiency improvement.
Yu Gao, Haoyuan Guo, Tuyen Hoang, Weilin Huang, Lu Jiang, Fangyuan Kong, Huixia Li, Jiashi Li, Liang Li, Xiaojie Li, Xunsong Li, Yifu Li, Shanchuan Lin, Zhijie Lin, Jiawei Liu, Shu Liu, Xiaonan Nie, Zhiwu Qing, Yuxi Ren, Li Sun, Zhi Tian, Rui Wang, Sen Wang, Guoqiang Wei, Guohong Wu, Jie Wu, Ruiqi Xia, Fei Xiao, Xuefeng Xiao, Jiangqiao Yan, Ceyuan Yang, Jianchao Yang, Runkai Yang, Tao Yang, Yihang Yang, Zilyu Ye, Xuejiao Zeng, Yan Zeng, Heng Zhang, Yang Zhao, Xiaozheng Zheng, Peihao Zhu, Jiaxin Zou, Feilong Zuo
Notable breakthroughs in diffusion modeling have propelled rapid improvements in video generation, yet current foundational model still face critical challenges in simultaneously balancing prompt following, motion plausibility, and visual quality. In this report, we introduce Seedance 1.0, a high-performance and inference-efficient video foundation generation model that integrates several core technical improvements: (i) multi-source data curation augmented with precision and meaningful video captioning, enabling comprehensive learning across diverse scenarios; (ii) an efficient architecture design with proposed training paradigm, which allows for natively supporting multi-shot generation and jointly learning of both text-to-video and image-to-video tasks. (iii) carefully-optimized post-training approaches leveraging fine-grained supervised fine-tuning, and video-specific RLHF with multi-dimensional reward mechanisms for comprehensive performance improvements; (iv) excellent model acceleration achieving ~10x inference speedup through multi-stage distillation strategies and system-level optimizations. Seedance 1.0 can generate a 5-second video at 1080p resolution only with 41.4 seconds (NVIDIA-L20). Compared to state-of-the-art video generation models, Seedance 1.0 stands out with high-quality and fast video generation having superior spatiotemporal fluidity with structural stability, precise instruction adherence in complex multi-subject contexts, native multi-shot narrative coherence with consistent subject representation.
Youhe Jiang, Fangcheng Fu, Xupeng Miao, Xiaonan Nie, Bin Cui
Large-scale deep learning models contribute to significant performance improvements on varieties of downstream tasks. Current data and model parallelism approaches utilize model replication and partition techniques to support the distributed training of ultra-large models. However, directly deploying these systems often leads to sub-optimal training efficiency due to the complex model architectures and the strict device memory constraints. In this paper, we propose Optimal Sharded Data Parallel (OSDP), an automated parallel training system that combines the advantages from both data and model parallelism. Given the model description and the device information, OSDP makes trade-offs between the memory consumption and the hardware utilization, thus automatically generates the distributed computation graph and maximizes the overall system throughput. In addition, OSDP introduces operator splitting to further alleviate peak memory footprints during training with negligible overheads, which enables the trainability of larger models as well as the higher throughput. Extensive experimental results of OSDP on multiple different kinds of large-scale models demonstrate that the proposed strategy outperforms the state-of-the-art in multiple regards. Our code is available at https://github.com/Youhe-Jiang/OptimalShardedDataParallel.
Keer Lu, Xiaonan Nie, Zheng Liang, Da Pan, Shusen Zhang, Keshi Zhao, Weipeng Chen, Zenan Zhou, Guosheng Dong, Bin Cui, Wentao Zhang
In recent years, Large Language Models (LLMs) have demonstrated significant improvements across a variety of tasks, one of which is the long-context capability. The key to improving long-context performance lies in effective data organization and management strategies that integrate data from multiple domains and optimize the context window during training. Through extensive experimental analysis, we identified three key challenges in designing effective data management strategies that enable the model to achieve long-context capability without sacrificing performance in other tasks: (1) a shortage of long documents across multiple domains, (2) effective construction of context windows, and (3) efficient organization of large-scale datasets. To address these challenges, we introduce DataSculpt, a novel data management framework designed for long-context training. We first formulate the organization of training data as a multi-objective combinatorial optimization problem, focusing on attributes including relevance, homogeneity, integrity, and efficiency. Specifically, our approach utilizes a coarse-to-fine methodology to optimize training data organization both efficiently and effectively. We begin by clustering the data based on semantic similarity (coarse), followed by a multi-objective greedy search within each cluster to score and concatenate documents into various context windows (fine). Our comprehensive evaluations demonstrate that DataSculpt significantly enhances long-context training performance, resulting in improvements of 18.09% in retrieval augmentation, 21.23% in summarization, 21.27% in reading comprehension, and a 3.81% increase in code completion, while also maintaining overall model proficiency with a 4.88% improvement.
Ran Yan, Youhe Jiang, Xiaonan Nie, Fangcheng Fu, Bin Cui, Binhang Yuan
Training large language model (LLM) is a computationally intensive task, which is typically conducted in data centers with homogeneous high-performance GPUs. We explore an alternative approach by deploying training computations across heterogeneous GPUs to enable better flexibility and efficiency for heterogeneous resource utilization. To achieve this goal, we propose a novel system, HexiScale, that can flexibly support asymmetric partition of training computations in the scope of data-, pipeline-, and tensor model parallelism. We further formalize the allocation of asymmetric partitioned training computations over a set of heterogeneous GPUs as a constrained optimization problem and propose an efficient hierarchical graph partitioning algorithm. Our approach effectively allocates training computations across GPUs, fully leveraging the available computational power. We conduct empirical studies to evaluate the performance of HexiScale with state-of-the-art homogeneous and heterogeneous training systems. When training LLMs at different scales (from 7B to 30B), HexiScale achieves comparable MFU when running over heterogeneous GPUs compared to state-of-the-art training systems running over homogeneous high-performance GPUs with the same total peak FLOPS. The percentage gaps in MFU between HexiScale and comparable homogeneous settings are as low as $0.3\%$, with an average of $3.5\%$.
Hao Ge, Junda Feng, Qi Huang, Fangcheng Fu, Xiaonan Nie, Lei Zuo, Haibin Lin, Bin Cui, Xin Liu
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.
Team Seedance, Heyi Chen, Siyan Chen, Xin Chen, Yanfei Chen, Ying Chen, Zhuo Chen, Feng Cheng, Tianheng Cheng, Xinqi Cheng, Xuyan Chi, Jian Cong, Jing Cui, Qinpeng Cui, Qide Dong, Junliang Fan, Jing Fang, Zetao Fang, Chengjian Feng, Han Feng, Mingyuan Gao, Yu Gao, Dong Guo, Qiushan Guo, Boyang Hao, Qingkai Hao, Bibo He, Qian He, Tuyen Hoang, Ruoqing Hu, Xi Hu, Weilin Huang, Zhaoyang Huang, Zhongyi Huang, Donglei Ji, Siqi Jiang, Wei Jiang, Yunpu Jiang, Zhuo Jiang, Ashley Kim, Jianan Kong, Zhichao Lai, Shanshan Lao, Yichong Leng, Ai Li, Feiya Li, Gen Li, Huixia Li, JiaShi Li, Liang Li, Ming Li, Shanshan Li, Tao Li, Xian Li, Xiaojie Li, Xiaoyang Li, Xingxing Li, Yameng Li, Yifu Li, Yiying Li, Chao Liang, Han Liang, Jianzhong Liang, Ying Liang, Zhiqiang Liang, Wang Liao, Yalin Liao, Heng Lin, Kengyu Lin, Shanchuan Lin, Xi Lin, Zhijie Lin, Feng Ling, Fangfang Liu, Gaohong Liu, Jiawei Liu, Jie Liu, Jihao Liu, Shouda Liu, Shu Liu, Sichao Liu, Songwei Liu, Xin Liu, Xue Liu, Yibo Liu, Zikun Liu, Zuxi Liu, Junlin Lyu, Lecheng Lyu, Qian Lyu, Han Mu, Xiaonan Nie, Jingzhe Ning, Xitong Pan, Yanghua Peng, Lianke Qin, Xueqiong Qu, Yuxi Ren, Kai Shen, Guang Shi, Lei Shi, Yan Song, Yinglong Song, Fan Sun, Li Sun, Renfei Sun, Yan Sun, Zeyu Sun, Wenjing Tang, Yaxue Tang, Zirui Tao, Feng Wang, Furui Wang, Jinran Wang, Junkai Wang, Ke Wang, Kexin Wang, Qingyi Wang, Rui Wang, Sen Wang, Shuai Wang, Tingru Wang, Weichen Wang, Xin Wang, Yanhui Wang, Yue Wang, Yuping Wang, Yuxuan Wang, Ziyu Wang, Guoqiang Wei, Wanru Wei, Di Wu, Guohong Wu, Hanjie Wu, Jian Wu, Jie Wu, Ruolan Wu, Xinglong Wu, Yonghui Wu, Ruiqi Xia, Liang Xiang, Fei Xiao, XueFeng Xiao, Pan Xie, Shuangyi Xie, Shuang Xu, Jinlan Xue, Shen Yan, Bangbang Yang, Ceyuan Yang, Jiaqi Yang, Runkai Yang, Tao Yang, Yang Yang, Yihang Yang, ZhiXian Yang, Ziyan Yang, Songting Yao, Yifan Yao, Zilyu Ye, Bowen Yu, Jian Yu, Chujie Yuan, Linxiao Yuan, Sichun Zeng, Weihong Zeng, Xuejiao Zeng, Yan Zeng, Chuntao Zhang, Heng Zhang, Jingjie Zhang, Kuo Zhang, Liang Zhang, Liying Zhang, Manlin Zhang, Ting Zhang, Weida Zhang, Xiaohe Zhang, Xinyan Zhang, Yan Zhang, Yuan Zhang, Zixiang Zhang, Fengxuan Zhao, Huating Zhao, Yang Zhao, Hao Zheng, Jianbin Zheng, Xiaozheng Zheng, Yangyang Zheng, Yijie Zheng, Jiexin Zhou, Jiahui Zhu, Kuan Zhu, Shenhan Zhu, Wenjia Zhu, Benhui Zou, Feilong Zuo
Jie Liu, Zilyu Ye, Linxiao Yuan, Shenhan Zhu, Yu Gao, Jie Wu, Kunchang Li, Xionghui Wang, Xiaonan Nie, Weilin Huang, Wanli Ouyang
Unified models capable of interleaved generation have emerged as a promising paradigm, with the community increasingly converging on autoregressive modeling for text and flow matching for image generation. To advance this direction, we propose a unified reinforcement learning framework tailored for interleaved generation. We validate our approach on its fundamental unit: a single round of reasoning-driven image generation, where the model first expands the user prompt through reasoning, followed by image synthesis. Formulating this multimodal generation process as a Markov Decision Process with sparse terminal rewards, we introduce UniGRPO to jointly optimize text and image generation policies using GRPO. Adopting a minimalist methodology to avoid over-design, we leverage established training recipes for both modalities by seamlessly integrating standard GRPO for reasoning and FlowGRPO for visual synthesis. To ensure scalability to multi-round interleaved generation, we introduce two critical modifications to the original FlowGRPO: (1) eliminating classifier-free guidance to maintain linear, unbranched rollouts, which is essential for scaling to complex scenarios involving multi-turn interactions and multi-condition generation (e.g., editing); and (2) replacing the standard latent KL penalty with an MSE penalty directly on the velocity fields, providing a more robust and direct regularization signal to mitigate reward hacking effectively. Our experiments demonstrate that this unified training recipe significantly enhances image generation quality through reasoning, providing a robust and scalable baseline for the future post-training of fully interleaved models.
Xupeng Miao, Hailin Zhang, Yining Shi, Xiaonan Nie, Zhi Yang, Yangyu Tao, Bin Cui
Embedding models have been an effective learning paradigm for high-dimensional data. However, one open issue of embedding models is that their representations (latent factors) often result in large parameter space. We observe that existing distributed training frameworks face a scalability issue of embedding models since updating and retrieving the shared embedding parameters from servers usually dominates the training cycle. In this paper, we propose HET, a new system framework that significantly improves the scalability of huge embedding model training. We embrace skewed popularity distributions of embeddings as a performance opportunity and leverage it to address the communication bottleneck with an embedding cache. To ensure consistency across the caches, we incorporate a new consistency model into HET design, which provides fine-grained consistency guarantees on a per-embedding basis. Compared to previous work that only allows staleness for read operations, HET also utilizes staleness for write operations. Evaluations on six representative tasks show that HET achieves up to 88% embedding communication reductions and up to 20.68x performance speedup over the state-of-the-art baselines.
Bin Xiao, Chunan Shi, Xiaonan Nie, Fan Yang, Xiangwei Deng, Lei Su, Weipeng Chen, Bin Cui
Large language models (LLMs) suffer from low efficiency as the mismatch between the requirement of auto-regressive decoding and the design of most contemporary GPUs. Specifically, billions to trillions of parameters must be loaded to the GPU cache through its limited memory bandwidth for computation, but only a small batch of tokens is actually computed. Consequently, the GPU spends most of its time on memory transfer instead of computation. Recently, parallel decoding, a type of speculative decoding algorithms, is becoming more popular and has demonstrated impressive efficiency improvement in generation. It introduces extra decoding heads to large models, enabling them to predict multiple subsequent tokens simultaneously and verify these candidate continuations in a single decoding step. However, this approach deviates from the training objective of next token prediction used during pre-training, resulting in a low hit rate for candidate tokens. In this paper, we propose a new speculative decoding algorithm, Clover, which integrates sequential knowledge into the parallel decoding process. This enhancement improves the hit rate of speculators and thus boosts the overall efficiency. Clover transmits the sequential knowledge from pre-speculated tokens via the Regressive Connection, then employs an Attention Decoder to integrate these speculated tokens. Additionally, Clover incorporates an Augmenting Block that modifies the hidden states to better align with the purpose of speculative generation rather than next token prediction. The experiment results demonstrate that Clover outperforms the baseline by up to 91% on Baichuan-Small and 146% on Baichuan-Large, respectively, and exceeds the performance of the previously top-performing method, Medusa, by up to 37% on Baichuan-Small and 57% on Baichuan-Large, respectively.