Shengnan Wang

Haoyue Tan, Shengnan Wang, Yulin Qiao et al.

Video diffusion transformers (DiTs) suffer from prohibitive inference latency due to quadratic attention complexity. Existing sparse attention methods either overlook semantic similarity or fail to adapt to heterogeneous token distributions across layers, leading to model performance degradation. We propose AdaCluster, a training-free adaptive clustering framework that accelerates the generation of DiTs while preserving accuracy. AdaCluster applies an angle-similarity-preserving clustering method to query vectors for higher compression, and designs a euclidean-similarity-preserving clustering method for keys, covering cluster number assignment, threshold-wise adaptive clustering, and efficient critical cluster selection. Experiments on CogVideoX-2B, HunyuanVideo, and Wan-2.1 on one A40 GPU demonstrate up to 1.67-4.31x speedup with negligible quality degradation.

Cheng Yang, Shengnan Wang, Chao Yang et al.

Pre-trained language models, such as BERT, have achieved significant accuracy gain in many natural language processing tasks. Despite its effectiveness, the huge number of parameters makes training a BERT model computationally very challenging. In this paper, we propose an efficient multi-stage layerwise training (MSLT) approach to reduce the training time of BERT. We decompose the whole training process into several stages. The training is started from a small model with only a few encoder layers and we gradually increase the depth of the model by adding new encoder layers. At each stage, we only train the top (near the output layer) few encoder layers which are newly added. The parameters of the other layers which have been trained in the previous stages will not be updated in the current stage. In BERT training, the backward computation is much more time-consuming than the forward computation, especially in the distributed training setting in which the backward computation time further includes the communication time for gradient synchronization. In the proposed training strategy, only top few layers participate in backward computation, while most layers only participate in forward computation. Hence both the computation and communication efficiencies are greatly improved. Experimental results show that the proposed method can achieve more than 110% training speedup without significant performance degradation.

Cheng Yang, Shengnan Wang, Yuechuan Li et al.

In recent years, BERT has made significant breakthroughs on many natural language processing tasks and attracted great attentions. Despite its accuracy gains, the BERT model generally involves a huge number of parameters and needs to be trained on massive datasets, so training such a model is computationally very challenging and time-consuming. Hence, training efficiency should be a critical issue. In this paper, we propose a novel coarse-refined training framework named CoRe to speed up the training of BERT. Specifically, we decompose the training process of BERT into two phases. In the first phase, by introducing fast attention mechanism and decomposing the large parameters in the feed-forward network sub-layer, we construct a relaxed BERT model which has much less parameters and much lower model complexity than the original BERT, so the relaxed model can be quickly trained. In the second phase, we transform the trained relaxed BERT model into the original BERT and further retrain the model. Thanks to the desired initialization provided by the relaxed model, the retraining phase requires much less training steps, compared with training an original BERT model from scratch with a random initialization. Experimental results show that the proposed CoRe framework can greatly reduce the training time without reducing the performance.