Skip to content

Consumer Runtime

Choose a dual-model setup or share one resident backbone between PEFT roles to make room for training on a single GPU. Padded trajectory updates improve update throughput without changing the effective optimizer batch or the strict-OPD freshness contract.

Measured throughput and peak-memory tradeoffs for shared-backbone and dual-model training across trajectory batch sizes.

The figure measures throughput and memory for one RTX 4080 workload. Read the full data-bound Consumer Runtime v1 report for the matrix, profiler evidence, equivalence gate and hardware limits.

Next: configure a matching PyTorch build and recipe with the single-GPU guide, or inspect failure recovery in RecoveryBench.