A UAV-swarm-enabled collaborative content caching scheme based on non-orthogonal multiple access (NOMA) was proposed in order to address the communication requirement of computation-intensive and latency-sensitive service in highly dynamic internet of vehicle (IoV) scenario. NOMA technology was integrated to achieve spectrum sharing among multiple vehicle by deploying UAV swarm as edge node. A K-Means++ based dynamic clustering mechanism was designed to periodically partition vehicle cluster and optimize the deployment location of UAV cluster head. A graph convolutional network (GCN) was constructed for the UAV swarm network in order to aggregate the topological relationship, caching status and content popularity feature through a dynamic adjacency matrix. Then the capability of cross-node information sharing was enhanced. The cooperative caching problem was formulated as a decentralized partially observable Markov decision process (Dec-POMDP). An attention mechanism was introduced into the Qmix deep reinforcement learning algorithm. The attention mechanism was utilized to perform weighted fusion of neighboring UAV state in order to maximize the long-term cache hit rate. The simulation results showed that NOMA achieved nearly 60% improvement in latency and throughput performance compared with traditional orthogonal multiple access (OMA). The proposed scheme outperformed other caching schemes across various vehicle density scenario, showing enhancement in key performance metrics such as cache hit rate, average content retrieval latency and energy consumption. The efficiency and robustness of the proposed scheme in dynamic IoV environment were validated.