Intelligent reflecting surface (IRS)-assisted terahertz (THz) communication is becoming a key technology for next-generation wireless networks. Although IRS has great promise, it is severely hindered by the beam squint problem resulting from the frequency-independent nature of the passive reflecting elements, especially in ultra-wide THz bands. To address this problem, we introduce time delay modules to the IRS and focus on the weighted sum rate maximization problem through the joint optimization of transmit beamforming, IRS phase shifts, and time delays. To solve the formulated optimization problem, an alternating optimization algorithm that decomposes the original problem into three subproblems is proposed. Specifically, we apply semidefinite relaxation and successive convex approximation techniques to solve each subproblem. Simulation results demonstrate the superiority of the proposed scheme over recent literature works. Particularly, it achieves near-optimal performance with slightly increased hardware cost.