Skip to content

v0.1.0a18 — Reliable results and reproducible runs

Highlights

  • Results now compare exact workload datasets and configurations, retain failed repeats, reject non-finite scores, and preserve baseline identity.
  • Runs snapshot external vLLM YAML before execution, and retries use the accepted finalist-validation manifest and repeat budget.
  • CLI aliases follow deterministic precedence, and list-valued arguments keep every configured value.
  • Server launch and readiness use one resolved endpoint, listener ownership is checked by process and address, and Random/TPE search spaces reject GPU requirements that no configured worker supports.

Release artifacts

Platform Install
PyPI pip install "vllm-optimizer==0.1.0a18"
Linux/WSL runtime pip install "vllm-optimizer[runtime]==0.1.0a18"

The universal wheel, source distribution, SBOM, and build provenance are produced by the protected release workflow.

Result integrity

  • Workload identity includes dataset and backend configuration, preventing mismatched workloads from being compared or reclassified together.
  • Failed repeats remain in persisted and regenerated results instead of disappearing from eligibility decisions.
  • NaN and infinite scores cannot enter rankings.
  • Reclassified baselines retain their baseline role.

Reproducibility and configuration

  • External vLLM YAML is captured before each trial and copied launch commands use the captured settings.
  • Retry plans follow the accepted finalist-validation artifacts and repeat budgets, including multi-trial runs with separate YAML snapshot paths.
  • Underscore, hyphen, and optional -- aliases resolve consistently, with tuned and runtime values overriding fixed settings as intended.
  • List arguments emit one flag followed by the complete value list.

Execution and validation

  • Launch, endpoint guarding, and readiness share the resolved host and port.
  • Readiness verifies that the addressed IPv4 or IPv6 listener belongs to the managed vLLM process session.
  • Random and TPE preflight validate every possible tensor-parallel requirement against explicit GPU workers, independent of the requested trial budget.

Limitation

External YAML secrets are redacted in persisted manifests and cannot be safely reconstructed. A retry containing such a value is rejected with an error that identifies the unavailable setting; start a new run with the original secret.

Validation

The package and runtime version are 0.1.0a18. The release workflow validates Python 3.11 and 3.12, coverage, formatting, typing, documentation, dependencies, wheel installation, package metadata, SBOM generation, and provenance.

Full changelog: v0.1.0a17...v0.1.0a18.