ML-EcoLyzer

A framework for measuring the environmental impact of ML inference. Tracks CO2 emissions, energy consumption, and water usage across different hardware setups.

Why?

Training gets all the attention, but inference runs 24/7 in production. We built this to answer: "How much does running this model actually cost the environment?"

Install

Available on PyPI:

pip install ml-ecolyzer

With framework-specific dependencies:

pip install ml-ecolyzer[huggingface]  # transformers, diffusers
pip install ml-ecolyzer[pytorch]       # torchvision, torchaudio
pip install ml-ecolyzer[all]           # everything

Quick Start

from mlecolyzer import EcoLyzer

config = {
    "project": "my_analysis",
    "models": [{"name": "gpt2", "task": "text"}],
    "datasets": [{"name": "wikitext", "task": "text", "limit": 100}]
}

eco = EcoLyzer(config)
results = eco.run()

print(f"CO2: {results['final_report']['analysis_summary']['total_co2_emissions_kg']:.6f} kg")
print(f"Energy: {results['final_report']['analysis_summary']['total_energy_kwh']:.6f} kWh")

What it measures

• CO2 emissions - Based on power draw and regional carbon intensity
• Energy usage - Via NVIDIA-SMI, psutil, or RAPL
• Water footprint - Cooling overhead varies by hardware tier
• ESS (Environmental Sustainability Score) - Parameters per gram of CO2, useful for comparing models

ESS = Effective Parameters (M) / CO2 (g)

Higher ESS = more efficient. INT8 models typically score ~74% higher than FP32.

Supported setups

• GPUs: A100, T4, RTX series, GTX series
• CPU-only works too
• Frameworks: HuggingFace, PyTorch, scikit-learn

Config file

project: "benchmark_run"

models:
  - name: "facebook/opt-350m"
    task: "text"
    quantization:
      enabled: true
      target_dtype: "int8"

datasets:
  - name: "wikitext"
    task: "text"
    limit: 500

hardware:
  device_profile: "auto"

output:
  output_dir: "./results"
  export_formats: ["json", "csv"]

CLI

# Single run
mlecolyzer analyze --model gpt2 --dataset wikitext --task text

# System info
mlecolyzer info

Benchmarks

Ran 1,500+ inference configs across:

• Hardware: GTX 1650, RTX 4090, Tesla T4, A100
• Models: GPT-2, OPT, Qwen, LLaMA, Phi, Whisper, ViT
• Precisions: FP32, FP16, INT8

Key findings:

• A100 has poor ESS when underutilized (overkill for small batches)
• Consumer GPUs (RTX/T4) often more efficient for single-batch inference
• Quantization helps a lot, especially INT8

Contributing

See CONTRIBUTING.md. PRs welcome.

# Dev setup
pip install -e ".[dev]"
pytest

Citation

@misc{mlecolyzer2026,
  title={ML-EcoLyzer: A Framework for Quantifying the Environmental Impact of Machine Learning Inference},
  author={Minoza, Jose Marie Antonio and Laylo, Rex Gregor and Villarin, Christian and Ibanez, Sebastian},
  year={2026},
  note={AAAI Workshop on AI for Environmental Science},
  eprint={2511.06694},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  doi={10.48550/arXiv.2511.06694}
}

License

MIT