vTune

vTune is a local-first experimentation and optimization tool for vLLM serving configurations. Users define the parameters and workloads they care about; vTune manages the server lifecycle, runs repeatable benchmarks, explores the search space, and reports which configurations performed best.

vTune is alpha software targeting Linux with NVIDIA GPUs and Python 3.11–3.12.

Documentation · Quick start · PyPI

The current code is verified with vLLM 0.28.0 and GuideLLM 0.7.3 on WSL2 with an RTX 3080. That host required VLLM_USE_V2_MODEL_RUNNER: "0" because UVA was unavailable and VLLM_USE_FLASHINFER_SAMPLER: "0" because the CUDA compiler toolkit was not installed. Native Linux systems may not require these settings.

Other combinations may work but are not yet verified.

The published py3-none-any wheel installs on Linux and Windows. Configuration validation and stored-result inspection work on Windows, but starting an experiment is supported only on Linux because vLLM has no native Windows runtime.

Quick start

Requirements: Linux, an NVIDIA GPU, a local model directory, and working vllm and guidellm commands.

Install the current code:

git clone https://github.com/brtydse100/vTune.git
cd vTune
pip install -e .

Create experiment.yaml:

schema_version: 1
experiment:
  name: first-run
server:
  model: /models/opt-125m
  gpu-memory-utilization: 0.8
tune:
  max-num-seqs:
    values: [8, 16]
benchmark:
  runs:
    - name: throughput
      profile:
        kind: throughput
        max_concurrency: 16
      constraints:
        - kind: max_requests
          count: 10
      data:
        - kind: synthetic_text
          prompt_tokens: 32
          output_tokens: 16
optimization:
  maximize: output_tokens_per_second
  sampler: tpe
  trials: 2

Run it:

vtune --config experiment.yaml

The short form is vtune -c experiment.yaml. The command validates the file, runs the experiment, persists results, and generates its exports and report. vTune binds vLLM to 127.0.0.1 by default. Set server.host explicitly only when the benchmark server must be reachable from another host.

Fixed vLLM flags go directly under server; tunable flags use top-level tune. Fixed and tunable environment variables use env and tune_env. See the configuration guide for categorical, boolean, integer-range, float-range, list, and environment examples.

Terminal output is concise by default. To stream vLLM and GuideLLM logs:

vtune --config experiment.yaml --verbose

The persistent equivalent uses GuideLLM's logging level names:

logging:
  level: DEBUG

Supported levels are DEBUG, INFO, WARNING, ERROR, and CRITICAL. Full per-trial log files are always saved. --verbose overrides the configured level with DEBUG for that invocation.

Retry one or more selected trials into a new immutable linked run:

vtune retry --run runs/EXPERIMENT/RUN_ID \
  --trial trial-0001 --trial trial-0004

The source run is never modified.

Display every stored vLLM and GuideLLM command for a trial without executing anything:

vtune reproduce --run runs/EXPERIMENT/RUN_ID --trial trial-0001

Each completed run also contains a self-contained report.html decision dashboard with the best observed configuration, baseline comparison, score history, throughput/latency tradeoff, and observed parameter effects.

Random and TPE runs never execute the same resolved configuration twice. optimization.trials cannot exceed the number of unique configurations in the declared search space.

Product documents

• First MVP specification
• Future implementation roadmap
• Architecture overview and early sketch
• Editable Draw.io architecture diagram
• Contributor guide
• Release notes

The MVP specification defines the first releasable version and its acceptance criteria. The roadmap describes capabilities that should be designed for now but implemented after the core experiment loop is reliable.

vTune is available under the MIT License.