v0.1.0a10
Highlights
- Prevented incomplete benchmark requests from entering rankings.
- Added vLLM server-drain evidence and minimum-repeat enforcement.
- Added repeat uncertainty, finalist drift detection, and stronger release gates.
Release artifacts
| Platform | Install |
|---|---|
| PyPI | pip install "vtune==0.1.0a10" |
| Linux/WSL runtime | pip install "vtune[runtime]==0.1.0a10" |
Benchmark correctness
- Required exact, clean request totals for GuideLLM request-count and vLLM Bench Serve workloads.
- Added a conservative one-hour implicit timeout for request-limited GuideLLM runs and a server-drain gate after warmups and measured repeats.
- Enforced
benchmark.min_repeatsfor every configured benchmark.
Reports and release engineering
- Added uncertainty and drift evidence with sequential finalist reruns.
- Added built-wheel testing, separate SBOM output, and provenance using an immutable action revision.
Validation and limitations
Public/private behavior, strict docs, source/wheel, Twine, clean-install, and publish simulation were release gates. Backend normalization did not make workload semantics identical, and output-token fulfillment still depended on backend evidence. Hardware validation remained outside the release check.
Full changelog: v0.1.0a9...v0.1.0a10.