Skip to main content
Name a Hugging Face model and say what you want from it. RunInfra plans a run, measures candidates on real GPUs, and returns a ranked list with the evidence behind every number. You pick the winner.
An optimization runyou act at the ends01 PlanYour prompt becomes a planOne priority, plus every constraint you named in the request.you decide02 FilterConstraints drop candidatesA candidate that violates a hard constraint never reaches the ranking.automatic03 MeasureEvery candidate is measuredThe baseline and the surviving candidates, measured on real GPUs.automatic04 RankRanked on your priorityOther dimensions are reported, they do not drive the ranking.automatic05 PickYou pick the winnerOr roll back to an earlier version later, without re-benchmarking.you decideRedeployThe artifact that won the run is what gets deployed, never a generic rebuild.A run that satisfies no constraint returns zero winners and costs nothing.
An optimization runyou act at the ends01 PlanYour prompt becomes a planOne priority, plus every constraint you named in the request.you decide02 FilterConstraints drop candidatesA candidate that violates a hard constraint never reaches the ranking.automatic03 MeasureEvery candidate is measuredThe baseline and the surviving candidates, measured on real GPUs.automatic04 RankRanked on your priorityOther dimensions are reported, they do not drive the ranking.automatic05 PickYou pick the winnerOr roll back to an earlier version later, without re-benchmarking.you decideRedeployThe artifact that won the run is what gets deployed, never a generic rebuild.A run that satisfies no constraint returns zero winners and costs nothing.

​
Start one

The agent turns that into a plan and prices it. Nothing runs and nothing is charged until you accept the plan.

​
Say what to rank on

One priority decides the winner. Every other dimension is still measured and reported, it just does not drive the ranking. Constraints are hard filters rather than preferences. Name as many as you like in one sentence, the agent reads the commas as “and” and applies every one of them.

​
What gets searched

Serving configuration is frequently the largest measured change, and often larger than the quantization itself. A serving value copied from the selected configuration is labeled from recorded configuration. Measured on your model is reserved for a value produced by a trusted measurement, and a calculation from measured inputs is labeled derived from measured inputs. If no trusted measurement exists, the result says No measured value instead of hiding the row or presenting the configured value as evidence.
Quantized weights are not searched by default. Quantization trades output quality for speed and memory, so it runs only when you ask for it, when the model does not fit the resolved GPU, or when the target is local or on-device hardware. The plan states the decision either way. See Lossless first.

​
Compare instead of optimize

Comparisons are measured, not advisory. Ask in chat and RunInfra runs real benchmarks before it says anything about a winner.
Comparisons stay honest by construction:
  • A model never borrows another model’s baseline. Absent data renders as a gap, never as a borrowed number.
  • A candidate that failed to benchmark shows as failed, never as a zero.
  • A candidate with no recorded measurement shows as not measured, never as an estimate.
  • A winner is declared only when the measured candidates ran on the same GPU. When the comparison would be unfair, RunInfra vetoes the winner and says why.
  • Cancel mid-comparison and the candidates that already finished keep their measured rows, and the run reads canceled rather than completed.
Experiment mode is a measure-only bake-off. It compares candidates side by side, never declares or applies a winner, and keeps deploy and export locked until you switch the pipeline back to Optimize.

​
Versions and rollback

Every run creates a pipeline version you can compare against or return to. Rollback is a metadata change, so nothing is re-benchmarked.
Rolling back does not touch a running deployment until you redeploy the version you rolled back to. Deploying a winner carries the exact artifact that won: the pipeline files, model revision, serving backend, quantization, and version identity. If that snapshot no longer matches the deployable pipeline, the deploy stops before any GPU is provisioned instead of quietly serving a stock build.

​
What you pay for

Every run is quoted before it starts and held against your balance. The estimate you accepted is a ceiling, the settled charge is never higher, and a failed or cancelled run releases the whole hold. There are no free optimization runs at any level of spend. See Pricing and credits. Settlement is measured or downgraded:
  • Every phase declares the evidence it must produce before it runs. A benchmark phase declares its latency and memory numbers, a quantization phase declares its quality score.
  • A phase that finishes without the evidence it declared is marked downgraded, and the part of the charge tied to it is removed. You pay for what was measured, not for what was planned.
  • A run that produced no measured work at all ends failed and costs nothing. It never settles as a green result with a minimum charge.
  • A run carrying downgrades never promotes a version, and the result lists each downgraded phase with its missing outputs and the reason.
A run is a durable server-side job. Closing the tab does not stop it, and Stop aborts the GPU work and finalizes the bill. See Execution lifecycle.

​
Limits

Optimization is best-effort. If no candidate satisfies your constraints, the run returns zero winners with a recommendation to relax one, usually cost or latency.
  • Run duration depends on model size, the number of candidates, and hardware availability. The session shows live phase progress and live cost instead of a promised finish time.
  • Quality evidence depends on the modality and the available eval path. Where no measured gate exists, the run reports quality as not measured instead of treating an estimate as a pass.

​
Common questions

A candidate is excluded when a pre-quantized checkpoint does not exist on Hugging Face for that model and method, when it violates one of your hard constraints, or when it needs a GPU your workspace does not unlock.
Yes. The run happens on isolated benchmark hardware while your deployment keeps serving the variant it already has. When you redeploy, RunInfra rolls replicas one at a time with connection draining, so requests in flight finish on the old replica.
Never. Your deployment serves the previously chosen variant for the whole run. You decide when, and whether, to switch.

​
Next steps

Reading your result

What a published speedup means, and the five reasons one is refused.

Lossless first

The four conditions that permit quantization, and nothing else.

Benchmark reports

Named public benchmarks against a settled winner and its baseline.

Deploy the winner

Turn it into an endpoint you can call.