Fix CUDA/cuDNN DLL preload paths for CUDA 13 consolidated wheel layout

[tianleiwu]

Description

Fix #29198.

NVIDIA restructured the CUDA Python wheels starting with CUDA 13: the per-component CUDA Toolkit packages (cublas, cufft, cuda_runtime, cuda_nvrtc, curand, ...) were consolidated into a single nvidia/cu{major} package and the -cuNN suffix was dropped from those package names. This PR updates the DLL/shared-library preload logic and the wheel dependency metadata so onnxruntime-gpu (and onnxruntime-trt-rtx) keep working on both the legacy CUDA 12 layout and the new CUDA 13 consolidated layout.

Summary of Changes

Preload logic (onnxruntime/__init__.py)

File	Change
onnxruntime/__init__.py	_get_nvidia_dll_paths now detects the CUDA 13+ consolidated layout and resolves CUDA libraries under nvidia/cu{major} — Windows uses an architecture sub-folder (bin/<arch>, e.g. bin/x86_64), Linux uses a flat lib. The legacy CUDA 12 per-component paths are preserved.
onnxruntime/__init__.py	Added build_cuda_version and arch parameters (for testability/arch override); cuDNN paths factored out since cuDNN keeps its own nvidia/cudnn package layout in both schemes.
onnxruntime/__init__.py	print_debug_info drops the -cuNN suffix from CUDA Toolkit package names for CUDA 13+ (cuDNN keeps its suffixed name).

Wheel dependency metadata (setup.py)

File	Change
setup.py	onnxruntime-gpu cuda extras drop the -cuNN suffix for CUDA 13+ (nvidia-cuda-nvrtc, nvidia-cuda-runtime, nvidia-cufft, nvidia-curand); cuDNN dependency keeps the suffixed name.
setup.py	onnxruntime-trt-rtx CUDA Runtime dependency drops the -cuNN suffix for CUDA 13+.

Tests (onnxruntime/test/python/onnxruntime_test_python_preload_dlls.py)

• New unit tests pin the expected relative paths for the CUDA 12 (legacy) and CUDA 13 (consolidated) layouts on both Windows and Linux, the Windows arch override, the Linux flat-lib layout, the unchanged cuDNN layout, and the cuda/cudnn toggles.

Testing

• Run the new tests: python -m pytest onnxruntime/test/python/onnxruntime_test_python_preload_dlls.py (or python -m unittest onnxruntime.test.python.onnxruntime_test_python_preload_dlls).
• Backward compatibility: CUDA 12 paths and the cuDNN layout are unchanged; only CUDA 13+ takes the new consolidated paths and unsuffixed package names.
• Build in Linux and Windows, and pip install onnxruntime-gpu*.whl[cuda,cudnn], then import onnxruntime; onnxruntime.preload_dlls() can run successfully in python.

Checklist

• Tests added/updated
• No breaking changes (CUDA 12 behavior preserved)

[Copilot]

Pull request overview

Updates ONNX Runtime’s Python CUDA/cuDNN preload behavior and wheel dependency metadata to support NVIDIA’s CUDA 13+ consolidated Python wheel layout (nvidia/cu{major}) while preserving the legacy CUDA 12 per-component layout.

Changes:

• Updated onnxruntime.__init__._get_nvidia_dll_paths() to resolve CUDA libraries from consolidated nvidia/cu{major} paths for CUDA 13+ (with Windows bin/<arch> handling) while keeping CUDA 12 behavior intact.
• Adjusted setup.py extras/dependencies to use unsuffixed NVIDIA CUDA toolkit package names for CUDA 13+ (cuDNN remains suffixed).
• Added Python unit tests that pin expected relative preload paths for both CUDA 12 and CUDA 13 layouts across Windows/Linux.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
onnxruntime/__init__.py	Detect CUDA 13+ consolidated wheel layout for preload paths; adjust debug package name reporting accordingly.
setup.py	Update CUDA-related dependency names/extras for CUDA 13 consolidated wheels; keep cuDNN naming scheme.
onnxruntime/test/python/onnxruntime_test_python_preload_dlls.py	Add unit tests covering legacy vs consolidated CUDA layouts and cudnn/cuda toggles.
cmake/onnxruntime_providers_cuda.cmake	Narrow/rename CCCL header patch workaround to CUDA 13.3-specific logic on UNIX.

[github-actions]

You can commit the suggested changes from lintrunner.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.