[Build] linking against header-only library cudnn_frontend

[ConnorBaker]

Describe the issue

I'm packaging onnxruntime for the Nixpkgs CUDA ecosystem (https://github.com/connorbaker/cuda-packages).

When building with nix, I ran into errors while linking because the linker was unable to find -lcudnn_frontend.

My understanding was that cudnn-frontend was a header-only library (https://github.com/NVIDIA/cudnn-frontend) -- why does onnxruntime try to link against a library?

onnxruntime/cmake/onnxruntime_unittests.cmake

Line 70 in ac9c135

target_link_libraries(${_UT_TARGET} PRIVATE CUDA::cudart cudnn_frontend)

onnxruntime/cmake/onnxruntime_providers_cuda.cmake

Line 208 in ac9c135

target_link_libraries(${target} PRIVATE CUDA::cublasLt CUDA::cublas CUDNN::cudnn_all cudnn_frontend CUDA::curand CUDA::cufft CUDA::cudart

Urgency

No response

Target platform

x86_64-linux with TensorRT

Build script

nix build -L --builders '' --no-link github:connorbaker/cuda-packages/fabe15378f8cf5a0870d6189115170a042a802f5#pkgsCuda.sm_89.cudaPackages_12.onnxruntime

Error / output

cuda12.6-onnxruntime> /nix/store/va7kw1i822h95im4jacci19v0cqajfyf-binutils-2.43.1/bin/ld: cannot find -lcudnn_frontend: No such file or directory
cuda12.6-onnxruntime> collect2: error: ld returned 1 exit status
cuda12.6-onnxruntime> ninja: build stopped: subcommand failed.

Visual Studio Version

No response

GCC / Compiler Version

13.3.0

[tianleiwu]

cudnn_frontend is an interface library. Need use target_link_libraries() in CMake. Interface libraries in CMake are special in that they do not generate any compiled output themselves but propagate build properties, such as include directories or other linked libraries, to targets that depend on them.

Example to use it:

onnxruntime/cmake/onnxruntime_providers_cuda.cmake

Lines 200 to 210 in bbe7c87

	include(cudnn_frontend) # also defines CUDNN::*
	if (onnxruntime_USE_CUDA_NHWC_OPS)
	if(CUDNN_MAJOR_VERSION GREATER 8)
	add_compile_definitions(ENABLE_CUDA_NHWC_OPS)
	else()
	message( WARNING "To compile with NHWC ops enabled please compile against cuDNN 9 or newer." )
	endif()
	endif()
	target_link_libraries(${target} PRIVATE CUDA::cublasLt CUDA::cublas CUDNN::cudnn_all cudnn_frontend CUDA::curand CUDA::cufft CUDA::cudart
	${ABSEIL_LIBS} ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 safeint_interface)
	endif()

Here is related doc about build: https://onnxruntime.ai/docs/build/eps.html#cuda

[snnn]

cudnn_frontend is a cmake target. It's ok to do so, because the target_link_libraries command will add extra compile flags for the target. In this case, for example, CUDNN_FRONTEND_SKIP_JSON_LIB. But, we may also change it to onnxruntime_add_include_to_target.

onnxruntime_add_include_to_target(${target} cudnn_frontend)

Let me know if it would solve the problem.

[ConnorBaker]

This is largely the result of the way I was bringing in cudnn-frontend.

With Nixpkgs, I was including the result of cudnn-frontend's installation, which is essentially an include directory. cudnn-frontend doesn't provide a CMake configuration, so while my patched build was able to find the header files, it was failing during the linking phase because CMake didn't know it was an interface library.

The dependency fetching setup you've got in-repo will fetch the cudnn-frontend repo and unpack it and include it directly. I believe that means the cudnn_frontend target is created as an interface library (since it uses the repo's CMakeLists.txt).

In terms of my packaging, I've vendored a replacement for cudnn-frontend's CMakeLists.txt which creates a CMake configuration file so other projects can find it: ConnorBaker/cuda-packages@57e8841.

I then patch onnxruntime by replacing cmake/external/cudnn_frontend.cmake with (ConnorBaker/cuda-packages@bb6bf09):

find_package(cudnn_frontend REQUIRED)

and all works again.

I would like it if NVIDIA projects which are built and installed by CMake provide CMake configurations, but not all of them do. As such, I'm closing this issue since it's not a problem caused by onnxruntime.