Do we really need tensorrt dependency?

[RobertLou]

We're currently using the PyTorch backend as our main inference path.
On my local machine, downloading and compiling TensorRT is quite slow, so I’d prefer to avoid it if possible.

However, I noticed that several modules (e.g., the logger) still depend on TensorRT.
Is there any plan—or a feasible way—to make TensorRT optional or remove this dependency entirely for users who only need PyTorch?