InferPilot v0.1.0 — evidence before configuration changes

InferPilot v0.1.0 — evidence before configuration changes

InferPilot is an evidence-first vLLM screening and controlled capacity-analysis tool. It helps an
operator choose a configuration experiment, measure it compatibly, and abstain when the evidence is
insufficient.

Try it

No GPU or repository checkout:

uvx inferpilot demo

Against a running vLLM server:

uvx inferpilot doctor --url http://localhost:8000

doctor is read-only. It uses two Prometheus snapshots to screen KV pressure, preemption, and
queueing. A screening result nominates an experiment; it is not a proven bottleneck diagnosis or a
performance prediction.

Included

• Live vLLM metric screening with explicit insufficient-evidence states
• Immutable benchmark bundles with environment and effective-configuration provenance
• Closed-loop, Poisson, and batched-Poisson workloads
• SLO, Pareto, blocked-study, integrity, and compatibility gates
• SLO-constrained capacity frontiers and cost-per-token reporting
• Bound quality-evidence gates for precision-changing candidates
• GPU-free synthetic demo using the real decision contracts
• Structured field-report form for anonymized operator evidence

One honest measured result

In the latest preregistered Qwen2.5-7B/A10G study, FP8 KV improved throughput by 33.6% on the
KV-pressured workload and 0.2% on the unpressured control. The registered recomputation-reduction
mechanism did not hold, so the verdict remained INCONCLUSIVE.

Read the complete result.

Boundaries

InferPilot is not an autonomous production controller. It does not change a deployment, invent an
SLO, claim generality from one model/GPU, or call FP8 quality-safe without task and long-context
evidence. Real GPU execution uses the separately pinned vLLM 0.29.0 benchmark environment.

The repository includes the complete audit trail, including invalid runs and negative results.