The widespread use of large-scale knowledge graphs (KG) in recommendation systems (RS) has made efficiently processing heterogeneous data and capturing higher-order semantic relationships significant challenges. Traditional methods often fail to leverage the complex relationships and dynamic changes in heterogeneous data, especially in terms of graph sampling and propagation depth, resulting in reduced accuracy and flexibility. To address these challenges, we propose a novel model, DyHGSampler and AdaSGC-based knowledge graph diffusion for recommendation (DAGR). Specifically, DAGR employs DyHGSampler to dynamically perform relationship-aware node sampling by leveraging a Gumbel-Softmax-based strategy, thus flexibly incorporating relational semantics from heterogeneous graphs. This resolves the limitations of traditional sampling methods in precisely capturing these semantics. The AdaSGC module adaptively adjusts the propagation depth according to node characteristics, effectively avoiding high-order information loss or redundancy. Additionally, DAGR reduces noise interference through the diffusion model and uses contrastive learning to optimize node embeddings, further enhancing robustness and performance. We have carried out extensive experiments on two public datasets. The results demonstrate that DAGR surpasses state-of-the-art methods in various metrics.