PipeDiT: Accelerating Diffusion Transformers in Video Generation with Task Pipelining and Model Decoupling

Video generation has been advancing rapidly, and diffusion transformer (DiT) based models have demonstrated remarkable capabilities. However, their practical deployment is often hindered by slow inference speeds and high memory consumption. In this paper, we propose a novel pipelining framework named PipeDiT to accelerate video generation, which is equipped with three main innovations. First, we design a pipelining algorithm (PipeSP) for sequence parallelism (SP) to enable the computation of latent generation and communication among multiple GPUs to be pipelined, thus reducing inference latency. Second, we propose DeDiVAE to decouple the diffusion module and the variational autoencoder (VAE) module into two GPU groups, the executions of which can also be pipelined to reduce memory consumption and inference latency. Third, to better utilize the GPU resources in the VAE group, we propose an attention co-processing (Aco) method to further reduce the overall video generation latency. We integrate our PipeDiT into both OpenSoraPlan and Hun-yuanVideo, two state-of-the-art open-source video generation frameworks, and conduct extensive experiments on two 8-GPU systems. Experimental results show that, under many common resolution and timestep configurations, our PipeDiT achieves 1.06× to 4.02× speedups over OpenSoraPlan and HunyuanVideo.