InfLoRA is a parameter-efficient fine-tuning (PEFT) method designed for continual learning, aiming to maintain model performance on old tasks while adapting to new tasks. The method introduces a small number of parameters to reparameterize pre-trained weights, allowing fine-tuning of these parameters to be equivalent to fine-tuning the pre-trained weights within a subspace. InfLoRA designs this subspace to eliminate interference from new tasks on old tasks, achieving a good balance between stability and plasticity. Experimental results show that InfLoRA outperforms existing state-of-the-art continual learning methods on multiple datasets. The method is compatible with the class-incremental scenario, where task identities are unknown during inference. InfLoRA uses a dimensionality reduction matrix designed before learning a new task to ensure that the subspace does not interfere with the performance on old tasks. The method leverages gradient projection memory (DualGPM) to approximate gradient spaces for new and old tasks, enabling the design of the subspace. InfLoRA's architecture allows for efficient parameter expansion and integration of old branches into pre-trained weights, reducing the number of parameters needed for subsequent tasks. The method is evaluated on ImageNet-R, CIFAR100, and DomainNet, demonstrating superior performance compared to other PEFT-based continual learning methods. Ablation studies show that InfLoRA effectively eliminates interference from new tasks on old tasks, achieving a better trade-off between stability and plasticity. The method is also combined with classifier alignment to further improve performance.InfLoRA is a parameter-efficient fine-tuning (PEFT) method designed for continual learning, aiming to maintain model performance on old tasks while adapting to new tasks. The method introduces a small number of parameters to reparameterize pre-trained weights, allowing fine-tuning of these parameters to be equivalent to fine-tuning the pre-trained weights within a subspace. InfLoRA designs this subspace to eliminate interference from new tasks on old tasks, achieving a good balance between stability and plasticity. Experimental results show that InfLoRA outperforms existing state-of-the-art continual learning methods on multiple datasets. The method is compatible with the class-incremental scenario, where task identities are unknown during inference. InfLoRA uses a dimensionality reduction matrix designed before learning a new task to ensure that the subspace does not interfere with the performance on old tasks. The method leverages gradient projection memory (DualGPM) to approximate gradient spaces for new and old tasks, enabling the design of the subspace. InfLoRA's architecture allows for efficient parameter expansion and integration of old branches into pre-trained weights, reducing the number of parameters needed for subsequent tasks. The method is evaluated on ImageNet-R, CIFAR100, and DomainNet, demonstrating superior performance compared to other PEFT-based continual learning methods. Ablation studies show that InfLoRA effectively eliminates interference from new tasks on old tasks, achieving a better trade-off between stability and plasticity. The method is also combined with classifier alignment to further improve performance.