We present TFGM (Training Free Graph Matching), a framework to boost the performance of Graph Neural Networks (GNNs) based graph matching without training. TFGM sidesteps two crucial problems when training GNNs: 1) the limited supervision due to expensive annotation, and 2) training's computational cost. A basic framework, BasicTFGM, is first proposed by adopting the inference stage of graph matching methods. Our analysis shows that the BasicTFGM is a linear relaxation to the quadratic assignment formulation of graph matching. This guarantees the preservation of structure compatibility and an efficient polynomial complexity. Empirically, we further improve the BasicTFGM by handcrafting two types of matching priors into the architecture of GNNs: comparing node neighborhoods of different localities and utilizing annotation data if available. For evaluation, we conduct extensive experiments on a broad set of settings, including supervised keypoint matching between images, semi-supervised entity alignment between knowledge graphs, and unsupervised alignment between protein interaction networks. Applying TFGM on various GNNs shows promising improvements over baselines. Further ablation studies demonstrate the effective and efficient training-free property of TFGM. Our code is available at https://github.com/acharkq/Training-Free-Graph-Matching.