How to contribute

EcoCompute is an open research project. There are five ways to contribute, and you do not need a GPU to get started.

Measure Run a measurement and submit a replication

The most valuable contribution: run the container on your GPU and submit your energy.json as a public replication. This is how the dataset grows.

1. Run the container. See the Container guide for the three-command quick start. The default run measures NF4 + FP16 baseline on TinyLlama-1.1B.
2. Validate your report. Run python3 tools/validate.py ecocompute-out/energy.json. Aim for protocol-conformant or dataset-eligible.
3. Submit at quantenergy.tech/replications. Your energy.json is encoded in the URL — nothing is uploaded to a server.

What makes a high-value replication

• A different GPU architecture — Turing, Ampere, Ada, Hopper, Blackwell are covered; anything else is valuable.
• A different model size — especially in the crossover region (2B–8B).
• A different backend — currently almost all data is bitsandbytes. llama.cpp, vLLM, TensorRT-LLM are all welcome.
• Batch sizes > 1 or context lengths ≠ 2048 — the entire variable matrix is empty for these. This is the biggest data gap.

Code Contribute code

Code contributions are welcome. See CONTRIBUTING.md for the full development guide.

Quick start

git clone https://github.com/hongping-zh/ecocompute-mlcube.git
cd ecocompute-mlcube
pip install -r requirements-dev.txt
python -m pytest tests/ -q

Good first issues

These issues are labeled good first issue and are suitable for new contributors:

• Add batch_size > 1 support to entrypoint good first issue
• Add context_length parameter to measurement good first issue

See all good first issues and help wanted on GitHub.

Report Report a bug or issue

Found something wrong? Open an issue at github.com/hongping-zh/ecocompute-mlcube/issues.

A good bug report includes: the exact command you ran, your GPU model and driver version (nvidia-smi output), the energy.json output, and whether you used Docker or native mode.

Protocol Improve the protocol or dataset

The measurement protocol lives at quantenergy.tech/method/. If you have suggestions for the protocol itself (measurement window, thermal requirements, randomization rules), open an issue or discuss in the SSRN preprint comments.

Share Spread the word

• Cite the project in your papers — see quantenergy.tech/cite/ for BibTeX.
• Share your replication results — a public replication is the strongest form of endorsement.
• Mention EcoCompute when discussing LLM quantization energy in blogs, talks, or social media.

Decision process

Questions?