The ML.ENERGY Leaderboard

Source code for The ML.ENERGY Leaderboard, which is a web leaderboard that displays results of The ML.ENERGY benchmark.

Running the Leaderboard Web App

Building the Data

After running the benchmark and collecting results, you can build the data for the leaderboard using the following command (with path adjustments as necessary):

BASE_DIR="/path/to/results"
uv run --with numpy --with PyYAML scripts/build_data.py \
  --results-dir "$BASE_DIR/llm/h100/run" \
  --results-dir "$BASE_DIR/llm/b200/run" \
  --results-dir "$BASE_DIR/diffusion/h100/run" \
  --results-dir "$BASE_DIR/diffusion/b200/run" \
  --output-dir public/data \
  --llm-config-dir ../benchmark/configs/vllm \
  --diffusion-config-dir ../benchmark/configs/xdit

Web App Preview

npm run dev

Citation

@inproceedings{mlenergy-neuripsdb25,
    title={The {ML.ENERGY Benchmark}: Toward Automated Inference Energy Measurement and Optimization}, 
    author={Jae-Won Chung and Jeff J. Ma and Ruofan Wu and Jiachen Liu and Oh Jun Kweon and Yuxuan Xia and Zhiyu Wu and Mosharaf Chowdhury},
    year={2025},
    booktitle={NeurIPS Datasets and Benchmarks},
}