PyTorch model with Nvidia Inference Server

I want to run my PyTorch model over Nvidia Inference Server with high performance. Instead of convert PyTorch model to Torch Script, is it possible to convert pytorch -> onnx -> tensorrt -> nvidia inference server? Which one is better for performance?

If I’m not mistaken, ONNX uses TorchScript under the hood, so you won’t be able to directly export to ONNX.
Please correct me, if I’m wrong.