Weiyan Wang

Zhimin Li, Jianwei Zhang, Qin Lin et al.

We present Hunyuan-DiT, a text-to-image diffusion transformer with fine-grained understanding of both English and Chinese. To construct Hunyuan-DiT, we carefully design the transformer structure, text encoder, and positional encoding. We also build from scratch a whole data pipeline to update and evaluate data for iterative model optimization. For fine-grained language understanding, we train a Multimodal Large Language Model to refine the captions of the images. Finally, Hunyuan-DiT can perform multi-turn multimodal dialogue with users, generating and refining images according to the context. Through our holistic human evaluation protocol with more than 50 professional human evaluators, Hunyuan-DiT sets a new state-of-the-art in Chinese-to-image generation compared with other open-source models. Code and pretrained models are publicly available at github.com/Tencent/HunyuanDiT

Weiyan Wang, Yilun Jin, Yiming Zhang et al.

Due to high accuracy, BERT-like models have been widely adopted by text mining and web searching. However, large BERT-like models suffer from inefficient online inference, facing the following two problems on GPUs: (1) their high accuracy relies on the large model depth, which linearly increases the sequential computation on GPUs; (2) stochastic and dynamic online workloads cause extra costs from batching and paddings. Therefore, we present \sys for the real-world setting of GPU inference on online workloads. At its core, \sys adopts stacking distillation and boosting ensemble, distilling the original deep model into a group of shallow but virtually stacked student models running in parallel. This enables \sys to achieve a lower model depth (e.g., two layers) than the others and the lowest inference latency while maintaining accuracy. In addition, adaptive student pruning realizes dynamic student numbers according to changing online workloads. Especially for occasional workload bursts, it can temporarily decrease the student number with minimal accuracy loss to improve system throughput. We conduct comprehensive experiments to verify the effectiveness, whose results show that \sys outperforms the baselines by $4.1\times\sim 1.6\times$ in latency while maintaining accuracy and achieves up to $22.27\times$ higher throughput for workload bursts.

Weiyan Wang, Xingjian Shi, Ruiqi Shu et al.

In practice, physical spatiotemporal forecasting can suffer from data scarcity, because collecting large-scale data is non-trivial, especially for extreme events. Hence, we propose \method{}, a novel probabilistic framework to realize iterative self-training with new self-ensemble strategies, achieving better physical consistency and generalization on extreme events. Following any base forecasting model, we can encode its deterministic outputs into a latent space and retrieve multiple codebook entries to generate probabilistic outputs. Then BeamVQ extends the beam search from discrete spaces to the continuous state spaces in this field. We can further employ domain-specific metrics (e.g., Critical Success Index for extreme events) to filter out the top-k candidates and develop the new self-ensemble strategy by combining the high-quality candidates. The self-ensemble can not only improve the inference quality and robustness but also iteratively augment the training datasets during continuous self-training. Consequently, BeamVQ realizes the exploration of rare but critical phenomena beyond the original dataset. Comprehensive experiments on different benchmarks and backbones show that BeamVQ consistently reduces forecasting MSE (up to 39%), enhancing extreme events detection and proving its effectiveness in handling data scarcity.

Hao Wu, Yuan Gao, Xingjian Shi et al.

To address the dual challenges of inherent stochasticity and non-differentiable metrics in physical spatiotemporal forecasting, we propose Spatiotemporal Forecasting as Planning (SFP), a new paradigm grounded in Model-Based Reinforcement Learning. SFP constructs a novel Generative World Model to simulate diverse, high-fidelity future states, enabling an "imagination-based" environmental simulation. Within this framework, a base forecasting model acts as an agent, guided by a beam search-based planning algorithm that leverages non-differentiable domain metrics as reward signals to explore high-return future sequences. These identified high-reward candidates then serve as pseudo-labels to continuously optimize the agent's policy through iterative self-training, significantly reducing prediction error and demonstrating exceptional performance on critical domain metrics like capturing extreme events.

Hao Wang, Jingrong Chen, Xinchen Wan et al.

Communication overhead poses an important obstacle to distributed DNN training and draws increasing attention in recent years. Despite continuous efforts, prior solutions such as gradient compression/reduction, compute/communication overlapping and layer-wise flow scheduling, etc., are still coarse-grained and insufficient for an efficient distributed training especially when the network is under pressure. We present DLCP, a novel solution exploiting the domain-specific properties of deep learning to optimize communication overhead of DNN training in a fine-grained manner. At its heart, DLCP comprises of several key innovations beyond prior work: e.g., it exploits {\em bounded loss tolerance} of SGD-based training to improve tail communication latency which cannot be avoided purely through gradient compression. It then performs fine-grained packet-level prioritization and dropping, as opposed to flow-level scheduling, based on layers and magnitudes of gradients to further speedup model convergence without affecting accuracy. In addition, it leverages inter-packet order-independency to perform per-packet load balancing without causing classical re-ordering issues. DLCP works with both Parameter Server and collective communication routines. We have implemented DLCP with commodity switches, integrated it with various training frameworks including TensorFlow, MXNet and PyTorch, and deployed it in our small-scale testbed with 10 Nvidia V100 GPUs. Our testbed experiments and large-scale simulations show that DLCP delivers up to $84.3\%$ additional training acceleration over the best existing solutions.

Weiyan Wang, Cengguang Zhang, Liu Yang et al.

Bulk synchronous parallel (BSP) is the de-facto paradigm for distributed DNN training in today's production clusters. However, due to the global synchronization nature, its performance can be significantly influenced by network bottlenecks caused by either static topology heterogeneity or dynamic bandwidth contentions. Existing solutions, either system-level optimizations strengthening BSP (e.g., Ring or Hierarchical All-reduce) or algorithmic optimizations replacing BSP (e.g., ASP or SSP, which relax the global barriers), do not completely solve the problem, as they may still suffer from communication inefficiency or risk convergence inaccuracy. In this paper, we present a novel divide-and-shuffle synchronization (DS-Sync) to realize communication efficiency without sacrificing convergence accuracy for distributed DNN training. At its heart, by taking into account the network bottlenecks, DS-Sync improves communication efficiency by dividing workers into non-overlap groups to synchronize independently in a bottleneck-free manner. Meanwhile, it maintains convergence accuracy by iteratively shuffling workers among different groups to ensure a global consensus. We theoretically prove that DS-Sync converges properly in non-convex and smooth conditions like DNN. We further implement DS-Sync and integrate it with PyTorch, and our testbed experiments show that DS-Sync can achieve up to $94\%$ improvements on the end-to-end training time with existing solutions while maintaining the same accuracy.

Qinghe Jing, Weiyan Wang, Junxue Zhang et al.

The scarcity of data and isolated data islands encourage different organizations to share data with each other to train machine learning models. However, there are increasing concerns on the problems of data privacy and security, which urges people to seek a solution like Federated Transfer Learning (FTL) to share training data without violating data privacy. FTL leverages transfer learning techniques to utilize data from different sources for training, while achieving data privacy protection without significant accuracy loss. However, the benefits come with a cost of extra computation and communication consumption, resulting in efficiency problems. In order to efficiently deploy and scale up FTL solutions in practice, we need a deep understanding on how the infrastructure affects the efficiency of FTL. Our paper tries to answer this question by quantitatively measuring a real-world FTL implementation FATE on Google Cloud. According to the results of carefully designed experiments, we verified that the following bottlenecks can be further optimized: 1) Inter-process communication is the major bottleneck; 2) Data encryption adds considerable computation overhead; 3) The Internet networking condition affects the performance a lot when the model is large.

Weiyan Wang, Yuxiang WU, Yu Zhang et al.

Task-oriented dialogue systems can efficiently serve a large number of customers and relieve people from tedious works. However, existing task-oriented dialogue systems depend on handcrafted actions and states or extra semantic labels, which sometimes degrades user experience despite the intensive human intervention. Moreover, current user simulators have limited expressive ability so that deep reinforcement Seq2Seq models have to rely on selfplay and only work in some special cases. To address those problems, we propose a uSer and Agent Model IntegrAtion (SAMIA) framework inspired by an observation that the roles of the user and agent models are asymmetric. Firstly, this SAMIA framework model the user model as a Seq2Seq learning problem instead of ranking or designing rules. Then the built user model is used as a leverage to train the agent model by deep reinforcement learning. In the test phase, the output of the agent model is filtered by the user model to enhance the stability and robustness. Experiments on a real-world coffee ordering dataset verify the effectiveness of the proposed SAMIA framework.