Scalable generation of multipartite entanglement in noisy quantum devices is a key challenge for quantum technologies. Drawing inspiration from the established experimental ability in entanglement generation of photonic memristors, we introduce an unsupervised optimization framework that realizes a quantum transformer architectures with non-linear parametrization of the input for preparing highly entangled multipartite states. The architecture building blocks are simulated noisy two-level photonic quantum memristor dynamics with wave input; this architecture has already found application for realizing a quantum vision transformer. Unlike the previous vision transformer, we add a nonlinearity inside the manifold space by adding a sinusoidal-type function. By reshaping the manifold space, we show that the architecture improves optimization and enhances entanglement generation compared to its linear counterpart. We then address realistic noisy settings, where the optimization is driven by the experimentally accessible Meyer–Wallach global entanglement metric, used as a scalable cost function. The generated states are independently certified through bipartite negativity and tested for genuine multipartite entanglement via PPT-mixture semidefinite-programming criterion. Our results indicate that nonlinearly parametrized quantum transformer architectures provide a robust and scalable route to noisy multipartite entanglement generation.