Keyword spotting (KWS) is an essential in speech recognition, particularly for smart device interfaces. However, it is challenging to design KWS models that are both efficient and accurate, especially for mobile devices with limited resources. In this paper, we propose RepTor-family models, where a structural re-parameterization technique was applied to temporal convolutions in order to improve the efficiency of KWS systems in various environments. The re-parameterized RepTor models have a single-branch plain feed-forward architecture without any residual connections and, therefore, can be efficiently computed on mobile devices. We further propose a latency-aware differentiable kernel search method to find the optimal kernel size for each convolution layer. The proposed RepTor-A~E models achieve 97.9% top-1 accuracy on the Google Speech Command dataset v2 with a latency of 183ms on the Galaxy S10. They outperform state-of-the-art KWS models regarding the accuracy-latency trade-off.