Deep neural networks have shown promising performance for automatic segmentation of Gross Tumor Volume (GTV) of Nasopharyngeal Carcinoma (NPC). However, models trained with images from a source domain often underperform when deployed to new clinical centers, especially those with different imaging modalities with large domain shifts. In this paper, we focus on source-free cross-modality domain adaptation to address this challenge by adapting a source model to unlabeled target domains of different modalities without access to the source-domain data. We propose Reliable Pseudo Labels for Source-Free cross-modality Domain Adaptation (RPL-SFDA) with three key components: 1) A target-domain pseudo label generation module using test-time intensity augmentation based on Biezier Curves; 2) A reliable pseudo label selection module that rejects low-quality pseudo labels based on uncertainty; and 3) An uncertainty-guided training procedure in the unlabeled target domain where the model is supervised by reliable pseudo labels and regularized by entropy minimization. Experimental results on a cross-modality NPC MRI dataset demonstrate that, compared to the baseline, our method significantly improved the average Dice scores in two different target modalities by 5.95 and 6.48 percentage points, respectively. Furthermore, it outperformed existing source-free cross-modality adaptation methods.