Skip to content

v0.1.0a7

Highlights

  • Persisted each trial's resolved sequential or parallel execution assignment.
  • Stopped labelling backend averages as median or P99 measurements.
  • Strengthened offline validation and artifact-integrity warnings.

Release artifacts

Platform Install
PyPI pip install "vtune==0.1.0a7"
Linux/WSL runtime pip install "vtune[runtime]==0.1.0a7"

Results and offline reporting

  • Stored execution mode and, for parallel trials, worker, device IDs, and port.
  • Recomputed a6 derived summaries from normalized evidence without modifying source runs; lost a6 execution assignments cannot be reconstructed.
  • Validated required report data and warned about changed optional artifacts.

Security

  • Clarified that append-only application behavior is not cryptographic immutability and name-based redaction cannot identify arbitrary secret values.

Known issues

  • The published a7 package does not import on Python 3.11 because the original TrialReport.execution default is incompatible with that interpreter.
  • Parallel mode remained experimental and real multi-GPU execution was not tested.

Validation

  • All 49 available private tests and 12 subtests passed.
  • Python 3.12 package and documentation checks passed.
  • vLLM 0.28.0 and GuideLLM 0.7.3 were inspected on one RTX 3080 WSL host, but the native server run did not complete in that CUDA environment.

Full changelog: v0.1.0a6...v0.1.0a7.