v0.1.0a6
Highlights
- Added timestamped live progress, repeat scores, and session duration.
- Prevented zero-completion GuideLLM results from becoming valid evidence.
- Expanded report methodology and optional OpenAI-compatible summaries.
Release artifacts
| Platform | Install |
|---|---|
| PyPI | pip install "vtune==0.1.0a6" |
| Linux/WSL runtime | pip install "vtune[runtime]==0.1.0a6" |
The universal wheel and source distribution are attached to the GitHub release.
GuideLLM completion safety
- Restored GuideLLM's normal in-flight draining lifecycle.
- Classified a zero-exit run with no completed requests as
benchmark_no_completed_requestswith a direct log reference.
Progress and reports
- Added TTY spinner, active stage, trial bar, repeat aggregates, and warnings for ineligible benchmark evidence.
- Added ordered benchmark rankings, labelled axes, standard latency/throughput summaries, scoring methodology, and exploratory importance explanations.
- Made
schema_versionoptional while retaining schema 1 compatibility.
Security
- Optional summaries read their API key only from a named environment variable.
- Loopback HTTP is allowed; non-loopback summary endpoints require HTTPS.
Validation and limitations
- All 156 private tests and 28 subtests passed.
- Compilation, whitespace, CLI, configuration, reporting, and GuideLLM completion checks passed.
- Native vLLM execution remained Linux-only.
Full changelog: v0.1.0a5...v0.1.0a6.