Siming Huang, Tianhao Cheng, Jason Klein Liu, Weidi Xu, Jiaran Hao, Liuyihan Song, Yang Xu, Jian Yang, Jiaheng Liu, Chenchen Zhang, Linzheng Chai, Ruifeng Yuan, Xianzhen Luo, Qiufeng Wang, YuanTao Fan, Qingfu Zhu, Zhaoxiang Zhang, Yang Gao, Jie Fu, Qian Liu, Houyi Li, Ge Zhang, Yuan Qi, Xu Yinghui, Wei Chu, Zili Wang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Gradient compression has been widely adopted in data-parallel distributed training of deep neural networks to reduce communication overhead. Some literatures have demonstrated that large gradients are more important than small ones because they contain more information, such as Top-k compressor. Other mainstream methods, like random-k compressor and gradient quantization, usually treat all gradients equally. Different from all of them, we regard large and small gradients selection as the exploitation and exploration of gradient information, respectively. And we find taking both of them into consideration is the key to boost the final accuracy. So, we propose a novel gradient compressor: Gradient Sampling with Bayes Prior in this paper. Specifically, we sample important/large gradients based on the global gradient distribution, which is periodically updated across multiple workers. Then we introduce Bayes Prior into distribution model to further explore the gradients. We prove the convergence of our method for smooth non-convex problems in the distributed system. Compared with methods that running after high compression ratio at the expense of accuracy, we pursue no loss of accuracy and the actual acceleration benefit in practice. Experimental comparisons on a variety of computer vision tasks (e.g. image classification and object detection) and backbones (ResNet, MobileNetV2, InceptionV3 and AlexNet) show that our approach outperforms the state-of-the-art techniques in terms of both speed and accuracy, with the limitation of 100× compression ratio.

Thanks to the popularity of GPU and the growth of its computational power, more and more deep learning tasks, such as face recognition, image retrieval and word embedding, can take advantage of extreme classification to improve accuracy. However, it remains a big challenge to train a deep model with millions of classes efficiently due to the huge memory and computation consumption in the last layer. By sampling a small set of classes to avoid the total classes calculation, sampling-based approaches have been proved to be an effective solution. But most of them suffer from the following two issues: i) the important classes are ignored or only partly sampled, such as the methods using random sampling scheme or retrieval techniques of low recall (e.g., locality-sensitive hashing), resulting in the degradation of accuracy; ii) inefficient implementation owing to incompatibility with GPU, like selective softmax. It uses hashing forest to help select classes, but the search process is implemented in CPU. To address the above issues, we propose a new sampling-based softmax called ANN Softmax in this paper. Specifically, we employ binary quantization with inverted file system to improve the recall of important classes. With the help of dedicated kernel design, it can be totally parallelized in mainstream training framework. Then, we find the size of important classes that are recalled by each training sample has a great impact on the final accuracy, so we introduce sample grouping optimization to well approximate the full classes training. Experimental evaluations on two tasks (Embedding Learning and Classification) and ten datasets (e.g., MegaFace, ImageNet, SKU datasets) demonstrate our proposed method maintains the same precision as Full Softmax for different loss objectives, including cross entropy loss, ArcFace, CosFace and D-Softmax loss, with only 1/10 sampled classes, which outperforms the state-of-the-art techniques. Moreover, we implement ANN Soft-max in a complete GPU pipeline that can accelerate the training more than 4.3X. Equipped our method with a 256 GPUs cluster, the time of training a classifier of 300 million classes on our SKU-300M dataset can be reduced to ten days.

Deep neural networks (DNNs) have gained tremendous attractions as compelling solutions for applications such as image classification, object detection, speech recognition, and so forth. Its great success comes with excessive trainings to make sure the model accuracy is good enough for those applications. Nowadays, it becomes challenging to train a DNN model because of 1) the model size and data size keep increasing, which usually needs more iterations to train; 2) DNN algorithms evolve rapidly, which requires the training phase to be short for a quick deployment. To address those challenges, distributed training platforms have been proposed to leverage massive server nodes for training, with the hope of significant training time reduction. Therefore, scalability is a critical performance metric to evaluate a distributed training platform. Nevertheless, our analysis reveals that traditional server clusters have poor scalability for training due to the traffic congestions within the server and beyond. The intra-server traffic on the I/O fabric can result in severe congestions and skewed quality of service as high performance devices are competing with each other. Moreover, the traffic congestions on the Ethernet for inter-server communication could also incur significant performance degradation. In this work, we devise a novel distributed training platform, EFLOPS, that adopts an algorithm and system co-design methodology to achieve good scalability. A new server architecture is proposed to alleviate the intra-server congestions. Moreover, a new network topology, BiGraph, is proposed to divide the network into two separate parts, so that there is always a direct connection between any nodes from different parts. Finally, accompany with BiGraph, a topology-aware allreduce algorithm is proposed to eliminate the traffic congestion on the direct connection. The experimental results show that eliminating the congestions on network interface can gain up to 11.3xcommunication speedup. The proposed algorithm and topology can provide further improvement up to 6.08x. The overall performance of ResNet-50 training achieves near-linear scalability, and is competitive to the top-rankings of MLPerf results.

In the last decades, extreme classification has become an essential topic for deep learning. It has achieved great success in many areas, especially in computer vision and natural language processing (NLP). However, it is very challenging to train a deep model with millions of classes due to the memory and computation explosion in the last output layer. In this paper, we propose a large-scale training system to address these challenges. First, we build a hybrid parallel training framework to make the training process feasible. Second, we propose a novel softmax variation named KNN softmax, which reduces both the GPU memory consumption and computation costs and improves the throughput of training. Then, to eliminate the communication overhead, we propose a new overlapping pipeline and a gradient sparsification method. Furthermore, we design a fast continuous convergence strategy to reduce total training iterations by adaptively adjusting learning rate and updating model parameters. With the help of all the proposed methods, we gain 3.9× throughput of our training system and reduce almost 60% of training iterations. The experimental results show that using an in-house 256 GPUs cluster, we could train a classifier of 100 million classes on Alibaba Retail Product Dataset in about five days while achieving a comparable accuracy with the naive softmax training process.