Driver drowsiness is a major cause of road accidents, making its detection a key objective in intelligent transportation systems. Vision-based methods, especially those using facial landmark detection, offer a non-invasive and efficient means of identifying fatigue through signs like eye closure and yawning. While traditional two-stage facial landmark detection systems are accurate, their computational complexity limits real-time use in embedded environments. To overcome this, single-stage models such as YOLOFaceMark integrate face and landmark detection in a unified pipeline. However, its RepStem module suffers from mismatched feature dimensions that hinder learning. This work introduces a redesigned stem module that aligns outputs for better early-stage feature fusion, enhancing performance while retaining real-time capabilities. The improved model outperforms the baseline and effectively detects drowsiness cues, making it suitable for practical driver monitoring applications.