v0.1.0a9
Highlights
- Renamed the public product to vLLM Config Tuner while retaining
vtune. - Normalized comparable GuideLLM and vLLM metrics without inventing data.
- Added command preflight, explicit ports, terminal color, and clearer reports.
Release artifacts
| Platform | Install |
|---|---|
| PyPI | pip install "vtune==0.1.0a9" |
| Linux/WSL runtime | pip install "vtune[runtime]==0.1.0a9" |
Benchmarking and reports
- Added canonical throughput, latency, and request-count fields while retaining backend-specific semantic differences.
- Added repeat, elapsed-time, median, P99, and metric-calculation notes.
- Validated complete YAML-derived commands before launching a run.
Known issues
- Long-generation GuideLLM evidence required manual inspection of request
totals, requested/observed tokens,
benchmark.log, andvllm.log. - Strict completion, output-length validation, and server draining were not yet implemented. No L40/H100 smoke host was available.
Validation
- All 59 private tests passed.
- Python 3.11/3.12, strict docs, wheel/sdist, metadata, clean-install, CLI, and configuration checks passed.
Full changelog: v0.1.0a8...v0.1.0a9.