Skip to content

Tune vLLM with evidence

vTune runs controlled vLLM experiments from one YAML file. You choose the settings and workload; vTune handles process lifecycle, unique search trials, failures, persistence, and a decision-focused report.

pip install "vtune[runtime]"
vtune --config experiment.yaml

Use plain pip install vtune for configuration and report inspection without local vLLM execution. Understand the two installation choices.

What vTune gives you

  • A baseline and Grid, Random, or TPE exploration of your chosen settings.
  • One or more GuideLLM or vLLM Bench Serve workloads per server configuration.
  • Immutable trial artifacts, concise terminal progress, and detailed logs.
  • A self-contained HTML report showing the best configuration and observed parameter effects.
  • Exact, secret-redacted commands for reproducing completed trials.

Start here

Run your first experiment Read the configuration guide

Alpha software

Experiment execution targets Linux with NVIDIA GPUs. The universal Python package can be installed elsewhere for configuration and result inspection.