The ZenDNN library and plug-ins have been developed to enable Deep Learning inference on AMD EPYC™ CPUs. The library offers optimized primitives, such as EmbeddingBag operators, Matrix multiplications and related fusions, Elementwise operations, Attention operators and Pool (Max and Average) that improve the performance of many transformer-based models, recommender system models, convolutional neural networks, and recurrent neural networks. For the primitives not supported by ZenDNN, execution will fall back to the native path of the framework.