Sign language generation (SLG) plays a pivotal role in assisting the communication between deaf or Hard-of-Hearing (DHH) and hearing communities, while addressing the challenges DHH individuals face in reading text. Existing methods primarily rely on Multi-Head Attention (MHA) decoder that often struggle with prediction drift and computational inefficiency for real time SL gesture generation. To address these limitations, we propose a novel Multimodal Gated Attention Units (MultiGAU) decoder. The proposed decoder replaces MHA’s quadratic complexity with a linear formulation, drastically reducing computational overhead while reducing the prediction drift. Experimental results show that the proposed method achieves double the token throughput and improves the temporal alignment. These results establish a new standard for efficient and temporally accurate SLG systems.