DoRA decomposes pre-trained weights into magnitude and direction components, using LoRA for directional updates to reduce trainable parameters while improving learning capacity and training stability. The method is implemented in PyTorch and consistently outperforms standard LoRA on downstream tasks including commonsense reasoning, visual instruction tuning, and image/video-text understanding when applied to models like LLaMA, LLaVA, and VL-BART. No additional inference overhead is introduced compared to the base models.