Kavier

Kavier is a physics-driven simulator able to simulate LLM ecosystems under inference workloads¹, extended to support physics-driven simulation of LLM fine-tuning workloads².

It predicts the performance, sustainability, and efficiency of LLM ecosystems:

• Performance: inference latencies, training throughput, GPU utilization and Model FLOPs Utilization (MFU)
• Sustainability: energy consumption, carbon emissions
• Efficiency: financial and energy cost per token or sample

Kavier ports one of OpenDC's energy models and reproduces OpenDC's energy predictions, four orders of magnitude faster than OpenDC².

Install

pip install kavier
kavier --help

Pages

• Fine-tuning model: the analytical model and its calibration.
• Usage: CLI and Python API.
• Cluster simulator: FIFO/backfill queuing over a fixed cluster.

Cite

@software{kavier,
  author = {Nicolae, Radu and Lotito, Daniele and Trivedi, Animesh and Donkervliet, Jesse and Iosup, Alexandru},
  title  = {Kavier: Simulating the Performance, Sustainability, and Efficiency of LLM Ecosystems under Inference and Training},
  year   = {2026},
  doi    = {10.5281/zenodo.23110996},
  url    = {https://github.com/atlarge-research/Kavier}
}

The metadata is in CITATION.cff.

---

1. R. Nicolae, A. Iosup, A. Trivedi, J. Donkervliet. Kavier: Exploring Performance, Sustainability, and Efficiency of LLM Ecosystems under Inference through Cache-Aware Discrete-Event Simulation. BSc thesis, Vrije Universiteit Amsterdam, 2025. PDF ↩

2. R. Nicolae, D. Lotito, A. Iosup. Coastline: Exploring the impact of multi-objective, context-aware recommenders on performance and sustainability of datacenters under LLM fine-tuning workloads. MSc thesis, Vrije Universiteit Amsterdam, 2026. ↩↩