I usually do onnx or tensorRT for deployment and use whatever I want for deployment. Most libraries can serialize down to an ONNX model or a tensorRT model.