SOLUTIONS
Explore two connected jobs: tuning each model and testing whether models can safely share a device.
TWO QUESTIONS. ONE CONNECTED LOOP.
The best configuration for an isolated model may not be the right configuration for a shared GPU. Sera keeps both questions in view.
Set the workload, quality floor, and progress threshold. The swarm proposes, reviews, and measures candidates until a valid stop condition. The demo has no fixed total trial cap.
You get A tested configuration, a usable runner, a saved report, and replay downloads.
Revisit the ledger for viable single-device configurations. Check memory fit, run the models together, and evaluate each model against its own limits.
You get A sharing verdict, with evidence behind the decision.
WHEN THE GOAL CHANGES
Consolidation asks a different question from solo speed. The demo can optimize latency, then throughput. Each stage starts from the previous winner and remeasures it. Set k=3 to allow at most 3% regression in the earlier objective.
Trial one specialist’s lever group at a time, then consider combinations after the individual effects have been measured.
Meet the specialists ↗Memory headroom cannot establish latency under contention. A joint trial measures what happens when the workloads compete for the same device.
Explore the demo pairing ↗SEE THE LOOP IN ACTION
useseraai.com/notebook — a live marimo notebook running the real loop.