Comment by Joky

TensorFlow is also a runtime, yet we model its dataflow graph (the input to the runtime) as a dialect, same for ONNX. TensorRT isn't that different actually.