PerfGuard of ICLR2026

This is the official code for PerfGuard in ICLR 2026

Environment Setup

1. Install Conda Environment

conda create -n PerfGuard python=3.10
conda activate PerfGuard

2. Install MetaGPT

3. Please go to the following GitHub repository to install the necessary tool dependencies.

https://github.com/HaozheZhao/UltraEdit
https://github.com/HuiZhang0812/CreatiLayout
https://github.com/stepfun-ai/Step1X-Edit
https://github.com/weichow23/AnySD
https://github.com/bytedance/DreamO
https://github.com/Xiaojiu-z/EasyControl
https://github.com/tencent-ailab/IP-Adapter

4. Modify the relevant code to avoid environment conflicts.

# Insert the following code into './Step1X_Edit/inference.py'
import sys
sys.path.append("Step1X_Edit")
# Insert the following code into  './AnySD/anysd/src/model.py'
import sys
sys.path.append("AnySD")

5. Install the remaining dependencies

pip install -r requirement.txt

6. Download the planner checkpoint zip file and unzip it.

link: https://pan.baidu.com/s/1DyK9rZeTebdwAKfl_pl6DQ?pwd=ziy2 passward: ziy2 

7. Deploy LLM locally using vLLM

export VLLM_USE_MODELSCOPE=True 
vllm serve Qwen/Qwen3-14B   --port 8001
vllm serve your_planner_checkpoint  --port 8002

8. Modify the LLM config file under ./config

Run the program

python run.py