IterIS: Iterative Inference-Solving Alignment for LoRA Merging (CVPR 2025)

• Code Contributor: Hongxu Chen, Bowei Zhu, Zhen Wang, Runshi Li, and Long Chen
• Paper: arXiv

Introduction

This is the official implementation of IterIS, accepted at CVPR 2025. The code requires GPU acceleration and can run on a single RTX 3090. Additionally, we recommend using the default hyperparameter settings for optimal performance.

To-do List

• Release code for merging in LLM merging experimetns
• Release code for merging in V&L experimetns
• Release code for merging in diffusion experimetns

Project Directory Structure

The complete file tree should be as follows:

├── config
│   ├── EMOTION-t5-lora-train-config
│   ├── GLUE-t5-lora-train-config
│   └── methods-config
├── data
│   ├── Emotion_v2.5
│   │   ├── emoint
│   │   ├── emotion-cause
│   │   ├── isear
│   │   └── tec
│   └── SENTICAP
│       ├── negative
│       ├── positive
│       └── val2014
├── images
├── loras
│   ├── EMOTION-lora-t5
│   │   ├── emoint
│   │   ├── emotion-cause
│   │   ├── isear
│   │   └── tec
│   ├── GLUE-lora-t5
│   │   ├── cola
│   │   ├── mnli
│   │   ├── mrpc
│   │   ├── qnli
│   │   ├── qqp
│   │   ├── rte
│   │   ├── sst2
│   ├── SENTICAP-lora-blip
│   │   ├── negative
│   │   ├── positive
│   ├── __pycache__
│   └── train_src
└── train

Setup

1. Install Dependencies

Run the following command to install all required packages:

python -m pip install -r requirements.txt -f https://download.pytorch.org/whl/cu121/torch_stable.html

2. Download the Dataset

Download and extract the images into ./data/SENTICAP/val2014:

wget http://images.cocodataset.org/zips/val2014.zip
unzip val2014.zip -d ./data/SENTICAP/val2014

---

Training, Merging, and Evaluation

3. Train Multiple LoRAs

(Taking the V&L experiments as an example)

python train/train_blip_negative.py  # Train NEG LoRA
python train/train_blip_positive.py  # Train POS LoRA

• You can modify the training settings in the source code or the YAML files located in ./config/.
• The trained LoRA checkpoints are available for download here.

After downloading, place them in the corresponding subfolder within ./LoRAs/.

---

4. Merge Different LoRAs

Use the following command to merge LoRAs based on the specific task:

python IterIS.py --task_type <TASK_NAME>

Supported task types:

• TASKS_blip_base → Merges POS and NEG LoRAs for V&L part
• GLUE_t5 → Merges GLUE LoRAs for LLM part (Multi-task)
• EMOTION_t5_large → Merges emotion classification LoRAs for LLM part (In-domain)

You can customize the merging settings in ./config/methods-config/iteris-config.yaml. Additional command-line parameters are available to control whether to save the merged model.

YAML Key Parameters and Usage:

• --max_iter: Maximum number of iterations for IterIS.
• --task_targets: A list of tasks to be merged.
• --with_pretrain_matrix: Whether to include the pre-trained weights in IterIS.
• --max_length: Maximum token length for padding or truncation.
• --lora_alpha: Alpha value for the target task's LoRA adaptation.
• --manual_coef: Manually adjusted weighting coefficient.
• --alpha_1: Regularization coefficient for one type of inner-product matrix in the closed-form solution.
• --alpha_2: Regularization coefficient for another type of inner-product matrix in the closed-form solution.
• --rank: Rank of the LoRA adaptation.
• --save: Whether to save the merged model.

5. Evaluate the Merged Model

After merging, the evaluation will run automatically using eval_model.py.

---

Additional Notes

• If you encounter any issues, please refer to the GitHub Issues section or contact the authors.
• For further customization, modify configurations in ./config/ and ./config/methods-config/.

---

This repository provides the official implementation of IterIS, and we welcome contributions and discussions! 🚀

Reference

Please cite our paper if you use our method in your works:

@article{chen2024iteris,
  title={IterIS: Iterative Inference-Solving Alignment for LoRA Merging},
  author={Chen, Hongxu and Li, Runshi and Zhu, Bowei and Wang, Zhen and Chen, Long},
  journal={arXiv preprint arXiv:2411.15231},
  year={2024}
}