add CUDABlas related API tests

[youge325]

暂时无法对齐，先跳过

[Copilot]

Pull request overview

Adds CUDA BLAS–related API coverage to the compatibility test suite, with Torch/Paddle behavioral differences captured via output placeholders to keep result files line-aligned for comparison.

Changes:

• Introduce test/ops/CUDABlasTest.cpp covering getCurrentCUDABlasHandle and at::cuda::blas::gemm<T> (with Torch-side stubs when symbols aren’t linkable).
• Extend CMake to locate CUDA Toolkit headers and add CUDA include paths to Torch/Paddle test targets.
• Add Paddle-specific compile/link adjustments for CUDA compat headers (define PADDLE_WITH_CUDA, link cudart when available).

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
test/ops/CUDABlasTest.cpp	New CUDA BLAS API tests plus Torch stub outputs for non-exported symbols
cmake/build.cmake	Adds Paddle CUDA compile definitions and links cudart when available
CMakeLists.txt	Detects CUDA Toolkit and threads CUDA include dirs into Torch/Paddle include paths

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

This unconditionally defines PADDLE_WITH_CUDA for all Paddle builds (USE_PADDLE_API=1), even if the detected Paddle installation is CPU-only or ROCm/HIP-based. That can force inclusion of CUDA-only types/headers and break compilation. Please gate PADDLE_WITH_CUDA (and the cudart linking) on actually having CUDA headers/libraries available (e.g., CUDAToolkit_FOUND / TARGET CUDA::cudart), or make the CUDA-dependent tests opt-in/out explicitly.

Suggested change

	# Paddle's CUDA compat headers (CUDAContextLight.h, CUDAFunctions.h)
	# require PADDLE_WITH_CUDA to be defined so that GPU type aliases
	# (gpuStream_t, cudaDeviceProp) are resolved via cuda_runtime.h.
	target_compile_definitions(${_test_name} PRIVATE PADDLE_WITH_CUDA)
	# Link libcudart for CUDA runtime symbols used by the Paddle CUDA compat
	# layer.
	if(TARGET CUDA::cudart)
	target_link_libraries(${_test_name} CUDA::cudart)
	elseif(EXISTS "/usr/local/cuda/lib64/libcudart.so")
	target_link_libraries(${_test_name}
	"/usr/local/cuda/lib64/libcudart.so")
	endif()
	# Only enable Paddle's CUDA pathways when CUDA is actually available.
	# This prevents defining PADDLE_WITH_CUDA for CPU-only or ROCm builds,
	# which would otherwise pull in CUDA-only headers/types and break
	# compilation.
	set(_paddle_cuda_available OFF)
	if(CUDAToolkit_FOUND)
	set(_paddle_cuda_available ON)
	elseif(TARGET CUDA::cudart)
	set(_paddle_cuda_available ON)
	elseif(EXISTS "/usr/local/cuda/include/cuda.h"
	AND EXISTS "/usr/local/cuda/lib64/libcudart.so")
	set(_paddle_cuda_available ON)
	endif()
	if(_paddle_cuda_available)
	# Paddle's CUDA compat headers (CUDAContextLight.h, CUDAFunctions.h)
	# require PADDLE_WITH_CUDA to be defined so that GPU type aliases
	# (gpuStream_t, cudaDeviceProp) are resolved via cuda_runtime.h.
	target_compile_definitions(${_test_name} PRIVATE PADDLE_WITH_CUDA)
	# Link libcudart for CUDA runtime symbols used by the Paddle CUDA
	# compat layer.
	if(TARGET CUDA::cudart)
	target_link_libraries(${_test_name} CUDA::cudart)
	elseif(EXISTS "/usr/local/cuda/lib64/libcudart.so")
	target_link_libraries(${_test_name}
	"/usr/local/cuda/lib64/libcudart.so")
	endif()
	else()
	message(WARNING
	"USE_PADDLE_API is enabled, but CUDA toolkit was not detected; "
	"skipping PADDLE_WITH_CUDA definition and cudart linkage for "
	"test target ${_test_name}.")
	endif()

[Copilot]

CUDA_INCLUDE_DIRS is set to an empty string when CUDA headers aren’t found, but ${CUDA_INCLUDE_DIRS}/cccl will then expand to /cccl and be added as an include directory. Also, if CUDA_INCLUDE_DIRS is a CMake list (multiple include dirs), appending /cccl this way will only apply to the last element. Please guard this append (only add the cccl path when CUDA include dir is non-empty and exists) and derive the cccl subdir from a single concrete CUDA include path (e.g., the first element).

Suggested change

	"${CUDA_INCLUDE_DIRS}"
	"${CUDA_INCLUDE_DIRS}/cccl")
	"${CUDA_INCLUDE_DIRS}")
	# Append CUDA cccl include directory if available.
	if(CUDA_INCLUDE_DIRS)
	# CUDA_INCLUDE_DIRS may be a list; use the first element as the base path.
	list(GET CUDA_INCLUDE_DIRS 0 CUDA_INCLUDE_DIRS_FIRST)
	if(CUDA_INCLUDE_DIRS_FIRST AND EXISTS "${CUDA_INCLUDE_DIRS_FIRST}")
	list(APPEND PADDLE_INCLUDE_DIR "${CUDA_INCLUDE_DIRS_FIRST}/cccl")
	endif()
	endif()

[Copilot]

find_package(CUDAToolkit QUIET) falls back to a warning and continues even when CUDA headers aren’t present, but test/ops/CUDABlasTest.cpp includes CUDA-specific headers and is always part of TEST_SRC_FILES, so the build can still hard-fail at compile time. Consider either making CUDA headers a hard requirement for building this test (use REQUIRED or a dedicated option) or conditionally excluding CUDABlasTest.cpp from TEST_SRC_FILES when CUDA isn’t found.

Suggested change

	message(WARNING "CUDA headers not found; CUDA tests may not compile.")
	set(CUDA_INCLUDE_DIRS "")
	message(STATUS
	"CUDA headers not found; CUDA-specific tests (e.g. CUDABlasTest) will be skipped.")
	set(CUDA_INCLUDE_DIRS "")
	# Avoid compiling CUDA-dependent tests when CUDA is not available.
	list(REMOVE_ITEM TEST_SRC_FILES
	"${PROJECT_SOURCE_DIR}/test/ops/CUDABlasTest.cpp")