soarsmu / Avatar

Replication Package for "Greening Large Language Models of Code", ICSE 2024 Software Engineering in Society

Avatar

This replication package contains the source code for applying Avatar to optimize and train models. The obtained model checkpoints can be found at this link.

Environment configuration

To reproduce our experiments, machines with GPUs and NVIDIA CUDA toolkit are required.

We provide a Dockerfile to help build the experimental environment. Please run the following scripts to compile a docker image:

docker build -t YOUR_CUSTOM_TAG .

NOTE⚠️: Be careful with the torch version that you need to use. In our case, we use torch==1.8.1+cu101. Pls modify the Dockerfile according to your cuda version if needed.

Then, please run the docker container:

dokcer run -it -v YOUR_LOCAL_REPO_PATH:/root/Avatar --gpus all YOUR_CUSTOM_TAG

How to run

Please use docker attach to go inside the docker container.

Then, for each experiment in our paper, the scripts and instructions are in the README.md files under each subfolder called Avatar.

Misc

Due to the random nature of neural networks and our GA algorithm, users may obtain slightly different results. Please note that such results usually can be tolerated, i.e., they mostly do not conflict with the conclusions of the paper.

If you meet any problems when using our code, please contact Jieke SHI by [e-mail]. Many thanks!

This paper has been accepted by the Software Engineering in Society Track of the 46th IEEE/ACM International Conference on Software Engineering (ICSE '24). The arXiv version can be viewed by this link. If you use any code from our repo in your paper, pls cite:

@misc{shi2023smaller,
      title={Greening Large Language Models of Code},
      author={Jieke Shi and Zhou Yang and Hong Jin Kang and Bowen Xu and Junda He and David Lo},
      year={2023},
      eprint={2309.04076},
      archivePrefix={arXiv},
      primaryClass={cs.SE}
}