v0.1.0a3
Highlights
- Added clearer trial progress and failure diagnostics.
- Added request-quality eligibility and transparent ranking evidence.
- Expanded the complete GuideLLM configuration examples.
Release artifacts
| Platform | Install |
|---|---|
| PyPI | pip install "vtune==0.1.0a3" |
| Linux/WSL runtime | pip install "vtune[runtime]==0.1.0a3" |
The universal wheel and source distribution are attached to the GitHub release.
Runtime and diagnostics
- Added one-based trial IDs, structured lifecycle progress, best-so-far output, default failure details, and richer timeout diagnostics.
- Accepted duration strings for startup timeouts and underscore-style vLLM arguments; excessive trial requests now warn instead of stopping the run.
Ranking and reports
- Added GuideLLM request-success and error accounting.
- Excluded workloads when more than half their requests failed or remained incomplete, then ranked eligible trials by quality and objective value.
- Added eligibility explanations to reports.
Validation and limitations
This release retained the alpha hardware and backend-comparability limitations documented by the preceding releases.
Full changelog: v0.1.0a2...v0.1.0a3.