Skip to content

Trust plan

Current status: software checks and published single-GPU WSL2 evidence are available; native-Linux L40/H100 and multi-GPU validation remain pending.

Software gates completed

  • Request-count runs must account for their expected requests and meet the configured failure tolerance; missing or inconsistent totals cannot be ranked.
  • Explicit benchmark timeouts remain authoritative; duration-based limits are derived conservatively.
  • Server drain evidence is required and persisted before a trial completes.
  • Public synthetic tests cover configuration, execution, scheduling, retry, process ownership, crash finalization, redaction, scoring, reporting, and offline workflows.
  • The failure percentage is an eligibility gate. The configured maximize metric is the primary objective everywhere; request failure rate and count are deterministic tie-breakers.
  • Fixed-repeat defaults use no warmup and four measurements; adaptive search can stop weak candidates after its initial budget. Fresh finalist validation uses a fixed budget. Reports distinguish uncertainty, drift, and incomplete evidence.
  • Coordinator failures atomically finalize recoverable runs as failed; keyboard interruption finalizes them as interrupted.
  • Conventionally named credentials are recursively redacted from terminal output, manifests, run results, CSV, and HTML. Arbitrary values and raw backend files are not guaranteed secret-free.
  • CI owns Ruff, formatting, type, coverage, strict documentation, wheel installation, dependency audit, SBOM, package metadata, and tag/version gates on Python 3.11 and 3.12.

Hardware validation pending

The RTX 3080 WSL2 archive contains real vLLM 0.28.0 measurements and provenance. These are reporting demonstrations and do not establish native-Linux, L40/H100, multi-GPU, or production performance. The remaining evidence needed before broader hardware claims includes:

  • Native-Linux vLLM 0.28 end-to-end execution.
  • L40 and H100 runs, including tensor parallel and at least one multi-GPU run.
  • Real long-generation timeout and interruption cleanup.
  • Raw GuideLLM versus vllm bench serve comparison evidence using identical requests.
  • Execution of the repository GPU smoke workflow.

Record GPU, driver, CUDA, Python, vLLM, GuideLLM, model, configuration, raw artifacts, and dated pass/fail outcomes in the compatibility matrix. Until those checks succeed, the project remains explicitly hardware validation pending.