ORT-TRT backend uses too much CPU memory

[ShuaiShao93]

Description
When using ORT-TRT backend on GPU, the CPU memory usage is as high as the usage when we use CPU inference.

Triton Information
What version of Triton are you using?
2.45.0

Are you using the Triton container or did you build it yourself?
container

To Reproduce

• Use any ONNX model like deberta
• Use ONNX backend plus TensorRT EP
• Start the server on T4 machine with docker run
• Verify the model uses GPU
• Check CPU memory usage, it's 14GB
• Force model to use CPU
• Check CPU memory usage, it's not higher than 14GB

Expected behavior
The CPU memory usage should be very low when model uses ORT-TRT backend on GPU

[ShuaiShao93]

A similar issue was reported before: #5392