It appears that the FindTENSORRT.cmake script does not look in the same folders as done by tensorrt.cmake

Consider changing

find_path(
    TENSORRT_INCLUDE_DIR NvInfer.h
    HINTS ${TENSORRT_DIR}
    PATH_SUFFIXES include)

To the following as done by tensorrt.cmake, so we are consistent.

find_path(
        TENSORRT_INCLUDE_DIR NvInfer.h
        HINTS ${TENSORRT_DIR} ${CUDA_TOOLKIT_ROOT_DIR}
        PATH_SUFFIXES include)

Otherwise, TENSORRT_DIR must always be set during build (or set as an env. variable).

It appears this issue was introduced in MMdeploy 0.4.0 as the issue occurred after upgrading.