Is HPI (High Performance Inference) in Paddle 3.1.0 worth sticking with CUDA 11.6, and why is TensorRT missing from the official Docker image?

[maakdan]

Hello,

In Paddle (or PaddleX) version 3.1.0, how much practical benefit does using the HPI (High Performance Inference) mode offer when working with PaddleOCR? Is the gain significant enough to justify sticking with CUDA 11.6 and giving up support for newer CUDA versions?

Also, I noticed that the official Docker image for the latest 3.1.0 release doesn't include TensorRT. Is there a specific reason for that?

[Bobholamovic]

Hi, please see my answer in this discussion: #15971 .