Memory-Efficient Large Language Models for Program Repair with Semantic-Guided Patch Generation

Fixing software bugs is crucial yet demands significant resources from developers. Automated Program Repair (APR) is a promising solution to address this challenging task. The emergence of Large Language Models (LLMs) has opened a new era of LLM-based APR, substantially advancing the APR field further. LLM-based APR methods face significant challenges regarding memory inefficiency, hindering their scalability and effectiveness. This is largely due to the beam search utilized in the patch generation phase of LLM-based APR, which requires large beam sizes to search for more potentially good repair candidates.