[BUG] [CMake] Package config references a non-existent header: cmake/NvInferVersion.h

[patrikhuber]

Describe the bug

cmake/TensorRT-RTXConfig.cmake adds a header file set to TRTRTX::tensorrt_headers with a file that the package
does not contain:

  target_sources(TRTRTX::tensorrt_headers
    INTERFACE
      FILE_SET "HEADERS"
      TYPE "HEADERS"
      BASE_DIRS "${_IMPORT_PREFIX}/cmake"
      FILES "${_IMPORT_PREFIX}/cmake/NvInferVersion.h"
  )

NvInferVersion.h is only in include/. CMake 4.3 reports the missing file at generate time when a file API client
requests the code model. Visual Studio does this on every configure.

Steps to reproduce

1. Extract TensorRT-RTX-1.6.1.120-Windows-amd64-cuda-13.4-Release-external. Create an empty
   bin/tensorrt_player.exe in it, or find_package fails earlier on a different missing file (see [CMake] Windows package config references missing bin/tensorrt_player.exe (tensorrt_capture_player) TensorRT#4851).

2. Create a folder with this CMakeLists.txt, with <package> set to the extracted directory:

    cmake_minimum_required(VERSION 3.27)
    project(trt_probe LANGUAGES CXX)
    set(TensorRT-RTX_DIR "<package>/cmake")
    find_package(CUDAToolkit REQUIRED)
    find_package(TensorRT-RTX CONFIG REQUIRED)
    file(WRITE "${CMAKE_BINARY_DIR}/main.cpp" "int main() {}\n")
    add_executable(probe "${CMAKE_BINARY_DIR}/main.cpp")
    target_link_libraries(probe PRIVATE TRTRTX::tensorrt)
   

3. Open the folder in Visual Studio 2026 Insiders (18.10, bundled CMake 4.3.1) with File > Open > Folder. The CMake
   configure runs automatically. The Output window (CMake) shows:

    CMake Error in CMakeLists.txt:
      Cannot find source file:
        <package>/cmake/NvInferVersion.h
   

To reproduce on the command line instead: Visual Studio requests the CMake code model through the file API, and the
error only appears with that request. Write {"requests":[{"kind":"codemodel","version":2}]} to
build/.cmake/api/v1/query/client-x/query.json, then run cmake -S . -B build -G Ninja with CMake 4.3.1. CMake
4.2.3, and CMake 4.3.1 without the query, do not report the error.

Expected behavior

The file set uses BASE_DIRS "${_IMPORT_PREFIX}/include" and FILES "${_IMPORT_PREFIX}/include/NvInferVersion.h".
TensorRT-Enterprise 11.3.0.99 already does this: its header file set lists include/ headers, and the same repro
reports no error.

Environment

• TensorRT-RTX version: 1.6.1.120
• GPU: NVIDIA RTX 3500 Ada Generation Laptop GPU (driver 596.86)
• Operating system: Windows 11 Pro 10.0.26200
• CUDA version: 13.3 (toolkit 13.3.73)
• CPU architecture: x86_64

Screenshots

N/A

Additional context

• The same issue was present in TensorRT-Enterprise 11.2.1.2 and it was fixed in 11.3.0.99.