Can we directly use TensorRT to accelerate TorchScript without ONNX?
You could try to use tensor2trt to directly use TensorRT with the Python API. However, as far as I know, the supported methods might be limited, so you would have to check, if your current model uses only supported layers.
A bit unrelated to your question, but might also be interesting: you could have a look ar this blogpost to see how to deploy Tacotron2 and Waveglow to TensorRT7.
I will check it, thank you very much:)