Skip to content

v0.1.0a19 — Audited runtime and reporting correctness

Highlights

  • Offline reclassification now applies the same request-completion safeguards as live GuideLLM and vLLM runs.
  • Configuration merging, resolved YAML exports, IPv6 endpoints, retry redaction, search validation, and parallel cleanup are hardened against reproduced edge cases.
  • Historical median scoring remains compatible with saved runs while rejecting mismatched workloads.
  • Terminal, JSON, HTML, and CSV outputs now agree about winners, comparisons, and baseline evidence.
  • The GPU smoke workflow now performs an actual small-model serving and benchmarking cycle.

Release artifacts

Platform Install
PyPI pip install "vllm-optimizer==0.1.0a19"
Linux/WSL runtime pip install "vllm-optimizer[runtime]==0.1.0a19"

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

Runtime and configuration fixes

  • Each external and experiment configuration layer is normalized before merge, preventing duplicate aliases such as tensor_parallel_size and tensor-parallel-size from conflicting across layers.
  • Explicit null experiment values use the same omission semantics in exported resolved YAML as they do on the command line, so inherited external values remain visible and active.
  • IPv6 literals are bracketed when constructing HTTP endpoints.
  • Retry validation detects redaction placeholders embedded in supported header strings, including Authorization: <redacted>.
  • Parallel interruption closes and awaits the scheduler before run finalization.
  • Grid bounds and steps must be finite before ranges are expanded.
  • Cleanup completes even if interruption arrives during shutdown, so owned processes still reach forced termination when their grace period expires.
  • Multi-trial retry validates the selected saved configurations directly instead of expanding their reconstructed Cartesian search space.
  • Retry rejects unrecoverable redaction placeholders in inline fixed and selected server arguments as well as external YAML.

Result integrity and reporting

  • Reclassification checks configured GuideLLM request counts and normalized vLLM request totals against successful, errored, and incomplete requests.
  • Legacy median scoring keeps the original mean-per-repeat, then median order without allowing repeats from different workload identities to be combined.
  • Overall and best-tuned winners are distinguished consistently. JSON exposes both overall-winner and best-tuned baseline comparisons, while terminal output labels the tuned comparison explicitly.
  • CSV baseline rows include their score and settings.
  • Resolved YAML normalizes aliases per layer and applies selected-null omission after merging, matching the command that was actually launched.

GPU smoke validation

The manually dispatched native-Linux workflow serves OPT-125M with vLLM, sends a real one-request benchmark through vLLM Bench Serve, verifies JSON, CSV, and HTML outputs, and confirms clean process shutdown. The release checklist keeps this hardware run separate from portable software validation.

Validation

The package and runtime version are 0.1.0a19. 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: v0.1.0a18...v0.1.0a19.