Why use ONNX with Triton Inference Server? Why use ONNX in general?

Since Triton can support TensorFlow and PyTorch via torchscript. I was wondering why you would want to convert your model to ONNX? Is it simply to use TensorRT?

Also just wanted to know why use ONNX in general? What are the main advantages?

advantages computervision general inference onnx pytorch server support tensorflow tensorrt triton triton inference server