Clarification on TensorRT Binding Order and Guarantee of Input-Output Sequence (bindingIsInput() etc)

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

I am working on an inference application using NVIDIA TensorRT for a model that includes multiple inputs and outputs (e.g., 3 inputs and 5 outputs). During implementation, I observed that the input tensors are listed first in the bindings array, followed by the output tensors. However, I want to confirm if this order is guaranteed by TensorRT by default.

    1. Is it guaranteed by TensorRT that the bindings in the engine are always ordered with inputs listed first, followed by outputs?
    1. If this behavior is documented, could you please provide a reference to confirm this?
3

Hi @felhasznalo413 ,
Currently i am not sure if thats guaranteed, but Accessing tensors via their index has been deprecated and you should use the tensor name for accessing I/O.
Hope this helps.
Thanks