<p>In millimeter-wave communication, the frequency spectrum offers a large amount of bandwidth to alleviate the scarcity in the spectrum of existing cellular bands. The massive Multi-Input Multi-Output hybrid architecture systems are employed to estimate mmWave channels effectively. Errors in sparsity pattern estimation may limit the performance of channel estimation schemes. To recover the channel state information, off-grid compressive sensing methods are utilized, leveraging the sparsity nature in angular domains. Due to hardware imperfections in carrier frequency oscillators, random phase drifts occur in the hybrid architecture of mmWave off-grid systems. The radio frequency chains share the same phase offsets within a time frame, but they differ across different time frames. To achieve better reliability and reduced overhead in communication, the support vectors need to be effectively optimized. In this article, we propose a modified partially coherent compressive phase retrieval approach by incorporating singular value decomposition for effective support vector initialization for off-grid mmWave massive MIMO systems. This algorithm optimizes the initialization by identifying the best k-dimensional subspace from the sparsity elements by forming the singular vectors. The proposed Off-Grid SVD-PC-CPR algorithm effectively utilizes the partial coherence property along with the singular value decomposition technique to identify the dominant paths of mmWave channels from the phases of the measurements by selecting basis vectors. Based on the simulation results, the proposed algorithm achieves better normalized mean squared error, a multi-user sum-rate gain of 26 bits/s/Hz, and a 54% improvement in estimation accuracy over the conventional PC-CPR algorithm at 20 dB signal-to-noise ratio.</p>