kavier 0.4.0

Kavier: Simulating the Performance, Sustainability, and Efficiency of LLM Ecosystems under Inference and Training

Kavier

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

---

This repository is the home of Kavier, the first scientific instrument for predicting performance, sustainability, and efficiency of LLM ecosystems under inference and training.

Kavier helps operators, researchers, and engineers predict:

• Performance — inference latencies, training throughput, GPU utilization
• Sustainability — energy consumption, carbon emissions (gCO2/Mtoken)
• Efficiency — financial and energy cost per token/sample given GPU-hour prices

Quick start

git clone https://github.com/atlarge-research/kavier.git
cd kavier

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
python -m pip install -U pip
pip install -e ".[dev]"

Run your first simulation against the tiny bundled synthetic example trace:

kavier-perf --trace src/kavier_inference/data/input/input_example.csv

Congrats! You have just run your first simulation with Kavier! 🎉

Or skip the flags — launch the interactive UI and pick a simulator, model and GPU from guided menus, then chain into energy/carbon or export OpenDC:

kavier

If you installed Kavier from PyPI (pip install kavier) you have no src/ directory; the same synthetic example trace ships inside the package, so resolve its path via importlib.resources:

TRACE=$(python -c "from importlib.resources import files; print(files('kavier_inference')/'data/input/input_example.csv')")
kavier-perf --trace "$TRACE"

Structure

Kavier is organized into the following first-party packages:

src/
├── kavier/              # Umbrella facade (re-exports the sub-packages)
├── kavier_inference/    # Inference simulation (kavier-perf)
├── kavier_training/     # Training simulation (kavier-train)
├── kavier_energy/       # Energy calculator (kavier-energy)
├── kavier_co2/          # Carbon emissions (kavier-co2)
├── kavier_library/      # Shared GPU & LLM specifications
├── kavier_io/           # Shared I/O utilities
│   └── opendc/          # OpenDC workload export (tasks/fragments)
├── kavier_ui/           # Interactive REPL (the `kavier` command)
└── tests/               # Test suites

Documentation

See docs/index.html for the main documentation: getting started, the Kavier CLIs (kavier interactive UI, kavier-perf, kavier-train, kavier-energy, kavier-co2), the YAML --config input, structure, and the contributing guide.

Contributing

Questions, suggestions and contributions are welcome and appreciated! Please refer to the contributing guide for more details.

License

Kavier is distributed under the MIT license. See LICENSE.txt.