How could I export self-defined operator to onnx?

Hi,

I learned from docs that we can export our model to onnx with torch.onnx, however, this only support pytorch inner operator. If I created a new operator and implemented it with cuda, how could I export this to onnx please ?