GPU clusters underpin modern deep learning, yet studies across industry and academia consistently report widespread GPU underutilization. Prior work and our own analysis indicate that inefficiency often stems from recurring patterns in code, job scripts, and runtime behaviour that users rarely detect. We argue that addressing this issue is an emerging SE and MSR challenge: it requires mining inefficiency patterns, combining static and dynamic signals for actionable feedback, validating job-submission artefacts, and developing privacy-aware datasets linking code, configuration, and runtime metrics. These directions point toward user-centered tools that can prevent underutilization before jobs reach the queue.