Kavier

Simulating the performance, sustainability, and efficiency of LLM ecosystems under inference and training.

Kavier is a physics-driven simulator. It predicts inference latency, training throughput, GPU utilization and MFU, energy, carbon emissions, and cost per token or sample.

Install

pip install kavier    # from PyPI
uv add kavier         # into a uv project

Python 3.11 or newer.

Quick start

From a clone of this repository:

uv sync
uv run kavier inference --trace src/kavier/sdk/inference/data/input/input_example.csv
uv run kavier --help

The subcommands are inference, training, cluster, energy, carbon, and calibrate (needs the [calibration] extra). Each documents its flags with --help.

Documentation

https://atlarge-research.github.io/Kavier/. Build it locally with uv run --group docs mkdocs serve.

Development

uv sync installs the dev tools. CI runs these gates on every pull request (ci.yml):

uv run pre-commit run --all-files --show-diff-on-failure   # ruff check, ruff format, whitespace
uv run mypy --strict --show-error-codes \
  -p kavier.cli -p kavier.sdk.co2 -p kavier.sdk.cluster
uv run mypy --strict --follow-imports=skip --show-error-codes \
  src/kavier/__init__.py src/kavier/__main__.py \
  src/kavier/sdk/training/calibration/__init__.py \
  src/kavier/sdk/training/core/engine.py
uv run pytest --cov
uv run --group docs mkdocs build --strict

The calibration and plot tests are skipped unless you sync with --extra calibration --extra plot. Run uv run pre-commit install once per clone to get the ruff and whitespace hooks on commit.

Citation

See CITATION.cff.

License

MIT. See LICENSE.txt.