Federated learning (FL) has been considered as a promising paradigm for enabling distributed training/learning in many machine-learning services without revealing users’ local data. Driven by the growing interests in exploiting FL in wireless networks, this paper studies the Non-orthogonal Multiple Access (NOMA) assisted FL in which a group of end-devices (EDs) form a NOMA cluster to send their locally trained models to the cellular base station (BS) for model aggregation. In particular, we consider that the BS adopts wireless power transfer (WPT) to power the EDs (for their data transmission and local training) in each round of FL iteration, and formulate a joint optimization of the BS’s WPT for different EDs, the EDs’ NOMA-transmission for sending the local models to the BS, the BS’s broadcasting of the aggregated model to all EDs, the processing-rates of the BS and EDs, as well as the training-accuracy of the FL, with the objective of minimizing the system-wise cost accounting for the total energy consumption as well as the FL convergence latency. In spite of the strict non-convexity of the joint optimization problem, we analytically characterize the BS’s and all EDs’ optimal processing-rates, based on which we propose a layered algorithm for finding the optimal solutions for the joint optimization problem via exploiting monotonic optimization. Numerical results validate that our algorithm can achieve the optimal solution as LINGO’s global-solver (i.e., a commercial optimization package) while significantly reducing the computation-time. Moreover, the results also demonstrate that our NOMA assisted FL can reduce the system cost compared to the benchmark FL scheme with the fixed local training-accuracy by more than 70% and the conventional frequency division multiple access (FDMA) based FL by 78%.