Federated graph learning (FGL) enables collaborative model training without sharing local data, but suffers from severe client heterogeneity, such as shifts in label distributions and graph sizes. Pre-trained graph models (PGMs), learned from large-scale graph data, encode general structural and semantic priors, offering a promising solution to alleviate heterogeneity in FGL. To this end, we first explores the potential of PGMs for FGL. We empirically reveal that directly fine-tuning PGMs in federated settings often results in degraded performance due to misalignment between the global prior and heterogeneous local distributions. To address this misalignment, we introduce PeFGL, a PGM-enhanced FGL framework that effectively adapts pre-trained priors to heterogeneous local graphs. PeFGL is built on two key components. First, we employ conditional generative diffusion models to augment local graphs, enriching structural diversity and compensating for missing patterns that hinder PGM adaptation. We further design a pre-trained-knowledge-guided filtering mechanism to select generated samples that align with both the PGM prior and local distribution. Then, based on the augmented and filtered data, we extract invariant subgraphs that can be shared across clients to guide PGM fine-tuning, which suppresses noisy substructures while reducing cross-client heterogeneity. Extensive experiments on four real-world datasets demonstrate that our framework consistently achieves state-of-the-art performance.