This paper presents a method for embedding classical data in Hilbert space using vectorized quantum block encoding, thereby improving the efficiency of quantum models such as the Quantum Transformer (QT), which replaces classical self-attention mechanisms with quantum circuit simulations. Conventional QTs rely on deeply parameterized quantum circuits (PQCs), which are susceptible to QPU noise and suffer from performance degradation. In this paper, we propose a vectorized quantum transformer (VQT), which enables efficient training via a vectorized nonlinear quantum encoder and supports the computation of an ideal masked attention matrix via quantum approximate simulation. This achieves shot-efficient, gradient-free quantum circuit simulation (QCS) with reduced classical sampling overhead. Comparing the accuracy of quantum circuit simulations on IBM and IonQ, and benchmarking natural language processing tasks on IBM's state-of-the-art high-fidelity Kingston QPU, we demonstrate competitive results. This noise-robust, medium-scale, quantum-friendly VQT approach presents a novel architecture for end-to-end machine learning in quantum computing.
