v0.1.0a7

August 31, 2026. Install the experiment runtime with pip install "vtune[runtime]==0.1.0a7" on Linux or WSL.

This correctness release records each trial's resolved execution assignment in the typed execution object. Sequential trials record their mode; parallel trials additionally record the worker, integer device IDs, and port.

It also stops treating an average as a median or P99. Older a6 reports may therefore show inaccurate derived median/P99 fields. Offline regeneration with a7 recomputes those derived summaries from normalized trial data without modifying the source run. Lost a6 worker/device/port assignments cannot be recovered.

Required structured report data is validated during regeneration. Missing or changed optional logs and raw diagnostics are reported as integrity warnings. Runs are append-only by application behavior; this is not cryptographic immutability. Optional LLM summaries require HTTPS except for loopback HTTP; name-based redaction reduces accidental disclosure but is not a guarantee that arbitrary user values contain no secrets.

Validation and limitations

All 49 available private tests and 12 subtests passed, including execution-assignment serialization and offline-regeneration coverage. Documentation and Python 3.12 package checks passed. The public Python 3.11 workflow built and installed the wheel successfully but failed while importing vtune because Python 3.11 rejects the original TrialReport.execution default. The source fix was verified by the public Python 3.11 workflow after the a7 tag; published a7 artifacts are immutable, so users needing Python 3.11 must wait for the next package release. vLLM 0.28.0 and GuideLLM 0.7.3 were installed and inspected on one RTX 3080 WSL host, but a native server run could not complete because of the available CUDA environment. Simultaneous multi-GPU execution was not tested because only one physical GPU was available. Parallel mode remains experimental; rerun finalists sequentially before production use.