v0.1.0a12
Highlights
- Added live benchmark output and workload-aware progress.
- Added configurable request-failure eligibility with complete request counts.
- Added immutable offline reclassification under a different failure policy.
Release artifacts
| Platform | Install |
|---|---|
| PyPI | pip install "vllm-optimizer==0.1.0a12" |
| Linux/WSL runtime | pip install "vllm-optimizer[runtime]==0.1.0a12" |
The universal wheel, source distribution, and SBOM are attached to the release.
Live execution
- Flushed subprocess output into
benchmark.logwhile workloads run. - Displayed request progress or elapsed duration and appended fast workloads' exact final count.
Request quality and reports
- Added
benchmark.max_failure_percentageandbenchmark.accept_any_request_failures. - Added successful, failed, errored, incomplete, and failure-percentage fields to JSON, CSV, and HTML.
- Stored backend-provided failure details in
failed_requests.json.
Offline reclassification
- Added
vllm-opt reclassifyto create new reports from stored evidence without launching processes or modifying the source run.
Validation and limitations
Two real repeats of GuideLLM 0.7.3 and vLLM Bench Serve completed 10/10 requests on an RTX 3080 WSL2 host. vLLM used the installed 0.19.0 V0 engine because 0.28.0 could not start when WSL reported unavailable unified virtual addressing. Native Linux vLLM 0.28.0 remained outstanding.
Full changelog: v0.1.0a11...v0.1.0a12.