Instead of fine-tuning models per repository or retrieving long context, generate lightweight adapters dynamically—this scales better and adapts to code changes automatically.
Code2LoRA uses a hypernetwork to generate repository-specific LoRA adapters for code models, injecting project knowledge without slowing down inference. It handles both stable codebases and evolving projects by tracking code changes, outperforming simpler fine-tuning approaches on a new 604-repository benchmark.