Heterogeneous device capabilities create severe synchronization bottlenecks in synchronous decentralized learning, making predefined model configurations inefficient across participating nodes. This letter proposes a resource-aware retention-ratio allocation framework that determines heterogeneous subnetworks before decentralized training by combining offline utility profiling with online resource-aware latency prediction. The resulting optimization minimizes the bottleneck latency under memory and aggregate learning-utility constraints. Although the learning utility is generally non-monotonic with respect to the retention ratio, we prove that the associated latency-feasibility problem remains monotonic, enabling an efficient binary-search solution. Experimental results verify that the proposed framework improves resource utilization, alleviates synchronization bottlenecks, and reduces overall training time while maintaining comparable learning performance.