Knowledge graph question generation (KGQG) refers to the task of generating natural language questions from knowledge graphs (KGs). Although this problem has been deeply studied in the past few years, at present, the training of small models heavily relies on a large amount of labeled data, and large language models (LLMs) also require a large number of parameters and high training costs. In reality, a significant amount of time and financial resources are required for manual annotation, which results in a scarcity of annotated data in practice. Therefore, the methods of these models are not very realistic. To address the above problems, we propose the Few-Shot Knowledge Graph Question Generation via LLMs Abstraction-to-Instantiation Conversion and Small Models Collaboration model (AIC-KGQG), which automatically generates labeled data through few-shot examples and collaborates between large and small model. We compare this framework with four mainstream knowledge graph question generation methods. The results show that AIC-KGQG achieves state-of-the-art performance under few-shot conditions, balancing data dependency and deployment efficiency, and providing a practical solution for KGQG in resource-constrained fields.