v0.1.0a1
Highlights
The first public alpha turns YAML into a local vLLM optimization run with owned server lifecycle, GuideLLM benchmarking, deterministic search, immutable artifacts, and decision-focused JSON, CSV, and HTML reports.
Release artifacts
| Platform | Install |
|---|---|
| PyPI | pip install "vtune==0.1.0a1" |
| Linux/WSL runtime | pip install "vtune[runtime]==0.1.0a1" |
The universal wheel and source distribution are attached to the GitHub release.
Experiment engine
- Added typed YAML configuration, arbitrary vLLM flags and environment values, readiness polling, process ownership, trial management, baselines, retries, intelligent timeouts, and failure classification.
- Added GuideLLM execution and normalized benchmark metrics.
Search, reports, and reproduction
- Added Grid, Random, and Optuna TPE search with deterministic seeds, persistent Optuna storage, and duplicate prevention.
- Stored commands, environment metadata, versions, timing, checksums, logs, and exit state for reproduction.
- Added parameter effects, baseline comparison, score history, and throughput/latency views.
Security
- Added loopback-only defaults, safe experiment names, and name-based secret redaction for persisted commands and environment values.
Validation and limitations
- All 106 private tests passed.
- One vLLM 0.28.0 and GuideLLM 0.7.3 compatibility experiment completed; this was not a general performance claim.
- The wheel passed Ubuntu 24.04 and Windows Python 3.12 checks.
- Experiment execution remains Linux-only.
Full changelog: first public release through v0.1.0a1.