Failed to find nvToolsExt

C++
13

Good news is that I did find a user workaround that does not require tweaking any of the CMake modules that come with a PyTorch distribution. Assuming that PYTORCH_VERSION and PYTORCH_CUDA_VERSION are available, with these two CMake variables defined as values from running the following Python script and reading the output:

from torch import version


if __name__ == "__main__":
    print(version.__version__.split("+")[0])  # for PYTORCH_VERSION
    print(version.cuda)                       # for PYTORCH_CUDA_VERSION

Right before your find_package(Torch CONFIG) call you can do the following:

# assumes find_package(CUDAToolkit) was already done
if(
    PYTORCH_VERSION VERSION_GREATER_EQUAL 2.5.0 AND
    PYTORCH_CUDA_VERSION VERSION_GREATER_EQUAL 12
)
    message(STATUS "PyTorch NVTX headers workaround: Yes")
    # only do this if nvToolsExt is not defined and CUDA::nvtx3 exists
    if(NOT TARGET CUDA::nvToolsExt AND TARGET CUDA::nvtx3)
        add_library(CUDA::nvToolsExt INTERFACE IMPORTED)
        # ensure that PyTorch is told to use NVTX3 headers
        target_compile_definitions(
            CUDA::nvToolsExt INTERFACE
            TORCH_CUDA_USE_NVTX3
        )
        target_link_libraries(CUDA::nvToolsExt INTERFACE CUDA::nvtx3)
    endif()
else()
    message(STATUS "PyTorch NVTX headers workaround: No")
endif()
# find Torch C++ as usual
find_package(Torch CONFIG)

This has pacified CMake nicely for me so hopefully it will be of help to others too.

3 Likes