Skip to content

v0.1.0a13

Highlights

  • Aligned search, final ranking, baseline comparison, and reports on one metric.
  • Made repeat defaults and small-sample confidence claims statistically honest.
  • Added recursive secret redaction and coordinator-level crash consistency.
  • Added synthetic lifecycle coverage and fail-closed release automation.

Release artifacts

Platform Install
PyPI pip install "vllm-optimizer==0.1.0a13"
Linux/WSL runtime pip install "vllm-optimizer[runtime]==0.1.0a13"

The universal wheel, source distribution, and SBOM are attached to the release.

Search and measurement

  • Made the configured metric primary among eligible trials, with failure rate and count as deterministic tie-breakers.
  • Defaulted to one warmup and three measured repeats; smaller explicit smoke runs are labelled exploratory.
  • Added Student's t confidence intervals and honest unavailable-drift messages.

Security and lifecycle

  • Recursively redacted nested configuration, arguments, headers, environment, normalized results, terminal output, and reports without changing execution.
  • Atomically finalized coordinator failures, preserved completed trials, and distinguished internal failure from task or keyboard interruption.

Offline reporting and release engineering

  • Persisted the effective repeat, warmup, drift, and request-failure policy for regeneration and reclassification, with legacy-manifest fallback.
  • Added deterministic Ruff, mypy, pytest, coverage, docs, audit, SBOM, package, Python 3.11/3.12 wheel, version, and private-regression gates.
  • Split production modules by responsibility; each remains at most 150 lines.

Validation

  • The public and available private suites passed: 152 tests and 12 subtests.
  • Python 3.11/3.12 package workflows, clean wheels, 84% repository coverage, all 34 risk-focused thresholds, lint, types, strict docs, audit, Twine, and archive inspection passed.

Hardware and performance

Native-Linux vLLM 0.28, L40/H100, tensor-parallel, multi-GPU, real long-run and interruption cleanup, cross-backend raw comparison, and GPU smoke execution are hardware validation pending. No claim is made that those checks passed.

Full changelog: v0.1.0a12...v0.1.0a13.