Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Dong, Zhenyuan, Zhang, Sai Qian
Format: Preprint
Veröffentlicht: 2024
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2409.07756
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866910711910760448
author Dong, Zhenyuan
Zhang, Sai Qian
author_facet Dong, Zhenyuan
Zhang, Sai Qian
contents Diffusion Transformers (DiTs) have recently attracted significant interest from both industry and academia due to their enhanced capabilities in visual generation, surpassing the performance of traditional diffusion models that employ U-Net. However, the improved performance of DiTs comes at the expense of higher parameter counts and implementation costs, which significantly limits their deployment on resource-constrained devices like mobile phones. We propose DiTAS, a data-free post-training quantization (PTQ) method for efficient DiT inference. DiTAS relies on the proposed temporal-aggregated smoothing techniques to mitigate the impact of the channel-wise outliers within the input activations, leading to much lower quantization error under extremely low bitwidth. To further enhance the performance of the quantized DiT, we adopt the layer-wise grid search strategy to optimize the smoothing factor. Moreover, we integrate a training-free LoRA module for weight quantization, leveraging alternating optimization to minimize quantization errors without additional fine-tuning. Experimental results demonstrate that our approach enables 4-bit weight, 8-bit activation (W4A8) quantization for DiTs while maintaining comparable performance as the full-precision model.
format Preprint
id arxiv_https___arxiv_org_abs_2409_07756
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle DiTAS: Quantizing Diffusion Transformers via Enhanced Activation Smoothing
Dong, Zhenyuan
Zhang, Sai Qian
Computer Vision and Pattern Recognition
Diffusion Transformers (DiTs) have recently attracted significant interest from both industry and academia due to their enhanced capabilities in visual generation, surpassing the performance of traditional diffusion models that employ U-Net. However, the improved performance of DiTs comes at the expense of higher parameter counts and implementation costs, which significantly limits their deployment on resource-constrained devices like mobile phones. We propose DiTAS, a data-free post-training quantization (PTQ) method for efficient DiT inference. DiTAS relies on the proposed temporal-aggregated smoothing techniques to mitigate the impact of the channel-wise outliers within the input activations, leading to much lower quantization error under extremely low bitwidth. To further enhance the performance of the quantized DiT, we adopt the layer-wise grid search strategy to optimize the smoothing factor. Moreover, we integrate a training-free LoRA module for weight quantization, leveraging alternating optimization to minimize quantization errors without additional fine-tuning. Experimental results demonstrate that our approach enables 4-bit weight, 8-bit activation (W4A8) quantization for DiTs while maintaining comparable performance as the full-precision model.
title DiTAS: Quantizing Diffusion Transformers via Enhanced Activation Smoothing
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2409.07756