diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt
index 1376c90fbcefe..8453da19ce3a6 100644
--- a/cmake/CMakeLists.txt
+++ b/cmake/CMakeLists.txt
@@ -1412,6 +1412,10 @@ endif()
 if (onnxruntime_USE_CUDA)
   set(CMAKE_CUDA_RUNTIME_LIBRARY Shared)
   set(CMAKE_CUDA_STANDARD 17)
+  if(onnxruntime_CUDA_HOME)
+    file(TO_CMAKE_PATH CUDAToolkit_ROOT ${onnxruntime_CUDA_HOME})
+  endif()
+  find_package(CUDAToolkit REQUIRED)
   if(onnxruntime_CUDNN_HOME)
     file(TO_CMAKE_PATH ${onnxruntime_CUDNN_HOME} onnxruntime_CUDNN_HOME)
   endif()
diff --git a/cmake/adjust_global_compile_flags.cmake b/cmake/adjust_global_compile_flags.cmake
index 8161ea574b8cc..d3f9256105127 100644
--- a/cmake/adjust_global_compile_flags.cmake
+++ b/cmake/adjust_global_compile_flags.cmake
@@ -205,7 +205,7 @@ endif()
 
 
 macro(check_nvcc_compiler_flag _FLAG _RESULT)
-    execute_process(COMMAND ${onnxruntime_CUDA_HOME}/bin/nvcc "${_FLAG}" RESULT_VARIABLE NVCC_OUT ERROR_VARIABLE NVCC_ERROR)
+    execute_process(COMMAND ${CUDAToolkit_BIN_DIR}/nvcc "${_FLAG}" RESULT_VARIABLE NVCC_OUT ERROR_VARIABLE NVCC_ERROR)
     message("NVCC_ERROR = ${NVCC_ERROR}")
     message("NVCC_OUT = ${NVCC_OUT}")
     if ("${NVCC_OUT}" MATCHES "0")
diff --git a/cmake/external/onnxruntime_external_deps.cmake b/cmake/external/onnxruntime_external_deps.cmake
index 22d12b128dc1f..09d57164b4ee1 100644
--- a/cmake/external/onnxruntime_external_deps.cmake
+++ b/cmake/external/onnxruntime_external_deps.cmake
@@ -556,16 +556,15 @@ message("Finished fetching external dependencies")
 set(onnxruntime_LINK_DIRS )
 if (onnxruntime_USE_CUDA)
       #TODO: combine onnxruntime_CUDNN_HOME and onnxruntime_CUDA_HOME, assume they are the same
+      find_package(CUDAToolkit REQUIRED)
       if (WIN32)
         if(onnxruntime_CUDNN_HOME)
           list(APPEND onnxruntime_LINK_DIRS ${onnxruntime_CUDNN_HOME}/lib ${onnxruntime_CUDNN_HOME}/lib/x64)
         endif()
-        list(APPEND onnxruntime_LINK_DIRS ${onnxruntime_CUDA_HOME}/x64/lib64)
       else()
         if(onnxruntime_CUDNN_HOME)
           list(APPEND onnxruntime_LINK_DIRS  ${onnxruntime_CUDNN_HOME}/lib ${onnxruntime_CUDNN_HOME}/lib64)
         endif()
-        list(APPEND onnxruntime_LINK_DIRS ${onnxruntime_CUDA_HOME}/lib64)
       endif()
 endif()
 
diff --git a/cmake/onnxruntime_providers_cuda.cmake b/cmake/onnxruntime_providers_cuda.cmake
index 9887d615c92d7..0f6d48bdb6ec8 100644
--- a/cmake/onnxruntime_providers_cuda.cmake
+++ b/cmake/onnxruntime_providers_cuda.cmake
@@ -178,15 +178,16 @@
     add_dependencies(${target} onnxruntime_providers_shared ${onnxruntime_EXTERNAL_DEPENDENCIES})
     if(onnxruntime_CUDA_MINIMAL)
       target_compile_definitions(${target} PRIVATE USE_CUDA_MINIMAL)
-      target_link_libraries(${target} PRIVATE ${ABSEIL_LIBS} ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 safeint_interface)
+      target_link_libraries(${target} PRIVATE ${ABSEIL_LIBS} ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 safeint_interface CUDA::cudart)
     else()
-      target_link_libraries(${target} PRIVATE cublasLt cublas cudnn curand cufft ${ABSEIL_LIBS} ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 safeint_interface)
+      target_link_libraries(${target} PRIVATE CUDA::cublasLt CUDA::cublas cudnn CUDA::curand CUDA::cufft CUDA::cudart
+              ${ABSEIL_LIBS} ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 safeint_interface)
       if(onnxruntime_CUDNN_HOME)
           target_include_directories(${target} PRIVATE ${onnxruntime_CUDNN_HOME}/include)
           target_link_directories(${target} PRIVATE ${onnxruntime_CUDNN_HOME}/lib)
       endif()
     endif()
-
+    
     if (onnxruntime_USE_TRITON_KERNEL)
       # compile triton kernel, generate .a and .h files
       include(onnxruntime_compile_triton_kernel.cmake)
@@ -196,25 +197,24 @@
       target_include_directories(${target} PRIVATE ${triton_kernel_header_dir})
       target_link_libraries(${target} PUBLIC -Wl,--whole-archive ${triton_kernel_obj_file} -Wl,--no-whole-archive)
       # lib cuda needed by cuLaunchKernel
-      target_link_libraries(${target} PRIVATE cuda)
+      target_link_libraries(${target} PRIVATE CUDA::cuda_driver)
     endif()
 
     include(cutlass)
     target_include_directories(${target} PRIVATE ${cutlass_SOURCE_DIR}/include ${cutlass_SOURCE_DIR}/examples)
 
-    target_include_directories(${target} PRIVATE ${ONNXRUNTIME_ROOT} ${CMAKE_CURRENT_BINARY_DIR}  ${eigen_INCLUDE_DIRS} ${TVM_INCLUDES} PUBLIC ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+    target_include_directories(${target} PRIVATE ${ONNXRUNTIME_ROOT} ${CMAKE_CURRENT_BINARY_DIR}  ${eigen_INCLUDE_DIRS} ${TVM_INCLUDES}
+     PUBLIC ${CUDAToolkit_INCLUDE_DIRS})
     # ${CMAKE_CURRENT_BINARY_DIR} is so that #include "onnxruntime_config.h" inside tensor_shape.h is found
     set_target_properties(${target} PROPERTIES LINKER_LANGUAGE CUDA)
     set_target_properties(${target} PROPERTIES FOLDER "ONNXRuntime")
 
     if (onnxruntime_ENABLE_CUDA_PROFILING) # configure cupti for cuda profiling
-      target_include_directories(${target} PRIVATE ${onnxruntime_CUDA_HOME}/extras/CUPTI/include)
-      target_link_directories(${target} PRIVATE ${onnxruntime_CUDA_HOME}/extras/CUPTI/lib64)
-      target_link_libraries(${target} PRIVATE cupti)
+      target_link_libraries(${target} PRIVATE CUDA::cupti)
     endif()
 
-    if (onnxruntime_ENABLE_NVTX_PROFILE AND NOT WIN32)
-      target_link_libraries(${target} PRIVATE nvToolsExt)
+    if (onnxruntime_ENABLE_NVTX_PROFILE)
+      target_link_libraries(${target} PRIVATE CUDA::nvtx3)
     endif()
 
     if (onnxruntime_ENABLE_TRAINING_OPS)
diff --git a/cmake/onnxruntime_providers_tensorrt.cmake b/cmake/onnxruntime_providers_tensorrt.cmake
index 686a993de3a4a..15ffc29e79ff4 100644
--- a/cmake/onnxruntime_providers_tensorrt.cmake
+++ b/cmake/onnxruntime_providers_tensorrt.cmake
@@ -8,7 +8,7 @@
   set(BUILD_LIBRARY_ONLY 1)
   add_definitions("-DONNX_ML=1")
   add_definitions("-DONNX_NAMESPACE=onnx")
-  set(CUDA_INCLUDE_DIRS ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+  set(CUDA_INCLUDE_DIRS ${CUDAToolkit_INCLUDE_DIRS})
   set(TENSORRT_ROOT ${onnxruntime_TENSORRT_HOME})
   set(OLD_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS})
   set(PROTOBUF_LIBRARY ${PROTOBUF_LIB})
@@ -58,7 +58,7 @@
       URL_HASH SHA1=${DEP_SHA1_onnx_tensorrt}
     )
     if (NOT CUDA_INCLUDE_DIR)
-      set(CUDA_INCLUDE_DIR ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}) # onnx-tensorrt repo needs this variable to build
+      set(CUDA_INCLUDE_DIR ${CUDAToolkit_INCLUDE_DIRS}) # onnx-tensorrt repo needs this variable to build
     endif()
     # The onnx_tensorrt repo contains a test program, getSupportedAPITest, which doesn't support Windows. It uses
     # unistd.h. So we must exclude it from our build. onnxruntime_fetchcontent_makeavailable is for the purpose.
@@ -102,11 +102,12 @@
   onnxruntime_add_include_to_target(onnxruntime_providers_tensorrt onnxruntime_common onnx flatbuffers::flatbuffers Boost::mp11 safeint_interface)
   add_dependencies(onnxruntime_providers_tensorrt onnxruntime_providers_shared ${onnxruntime_EXTERNAL_DEPENDENCIES})
   if (onnxruntime_USE_TENSORRT_BUILTIN_PARSER)
-    target_link_libraries(onnxruntime_providers_tensorrt PRIVATE ${trt_link_libs} cudart ${ONNXRUNTIME_PROVIDERS_SHARED} ${PROTOBUF_LIB} flatbuffers::flatbuffers Boost::mp11 safeint_interface ${ABSEIL_LIBS})
+    target_link_libraries(onnxruntime_providers_tensorrt PRIVATE ${trt_link_libs} ${ONNXRUNTIME_PROVIDERS_SHARED} ${PROTOBUF_LIB} flatbuffers::flatbuffers Boost::mp11 safeint_interface ${ABSEIL_LIBS} PUBLIC CUDA::cudart)
   else()
-    target_link_libraries(onnxruntime_providers_tensorrt PRIVATE ${onnxparser_link_libs} ${trt_link_libs} cudart ${ONNXRUNTIME_PROVIDERS_SHARED} ${PROTOBUF_LIB} flatbuffers::flatbuffers ${ABSEIL_LIBS})
+    target_link_libraries(onnxruntime_providers_tensorrt PRIVATE ${onnxparser_link_libs} ${trt_link_libs} ${ONNXRUNTIME_PROVIDERS_SHARED} ${PROTOBUF_LIB} flatbuffers::flatbuffers ${ABSEIL_LIBS} PUBLIC CUDA::cudart)
   endif()
-  target_include_directories(onnxruntime_providers_tensorrt PRIVATE ${ONNXRUNTIME_ROOT} ${CMAKE_CURRENT_BINARY_DIR} ${eigen_INCLUDE_DIRS} PUBLIC ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+  target_include_directories(onnxruntime_providers_tensorrt PRIVATE ${ONNXRUNTIME_ROOT} ${CMAKE_CURRENT_BINARY_DIR} ${eigen_INCLUDE_DIRS}
+    PUBLIC ${CUDAToolkit_INCLUDE_DIRS})
   if(onnxruntime_CUDNN_HOME)
     target_include_directories(onnxruntime_providers_tensorrt PRIVATE ${onnxruntime_CUDNN_HOME}/include)
   endif()
diff --git a/cmake/onnxruntime_python.cmake b/cmake/onnxruntime_python.cmake
index 3f20787e87425..23c6e5e430875 100644
--- a/cmake/onnxruntime_python.cmake
+++ b/cmake/onnxruntime_python.cmake
@@ -282,10 +282,7 @@ if (WIN32)
     get_filename_component(CUDNN_DLL_NAME ${CUDNN_DLL_PATH} NAME_WE)
     string(REPLACE "cudnn64_" "" CUDNN_VERSION "${CUDNN_DLL_NAME}")
     if(NOT onnxruntime_CUDA_VERSION)
-      message("Reading json file ${onnxruntime_CUDA_HOME}/version.json")
-      set(CUDA_SDK_JSON_FILE_PATH "${onnxruntime_CUDA_HOME}/version.json")
-      file(READ ${CUDA_SDK_JSON_FILE_PATH} CUDA_SDK_JSON_CONTENT)
-      string(JSON onnxruntime_CUDA_VERSION GET ${CUDA_SDK_JSON_CONTENT} "cuda" "version")
+      set(onnxruntime_CUDA_VERSION ${CUDAToolkit_VERSION})
       message("onnxruntime_CUDA_VERSION=${onnxruntime_CUDA_VERSION}")
     endif()
     file(APPEND "${VERSION_INFO_FILE}"
diff --git a/cmake/onnxruntime_unittests.cmake b/cmake/onnxruntime_unittests.cmake
index 3ed695327c183..88f662075e177 100644
--- a/cmake/onnxruntime_unittests.cmake
+++ b/cmake/onnxruntime_unittests.cmake
@@ -67,7 +67,7 @@ function(AddTest)
     if(onnxruntime_USE_CUDA)
       #XXX: we should not need to do this. onnxruntime_test_all.exe should not have direct dependency on CUDA DLLs,
       # otherwise it will impact when CUDA DLLs can be unloaded.
-      target_link_libraries(${_UT_TARGET} PRIVATE cudart)
+      target_link_libraries(${_UT_TARGET} PRIVATE CUDA::cudart)
     endif()
     target_link_libraries(${_UT_TARGET} PRIVATE ${_UT_LIBS} GTest::gtest GTest::gmock ${onnxruntime_EXTERNAL_LIBRARIES})
   endif()
@@ -1268,7 +1268,7 @@ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
       list(APPEND onnxruntime_shared_lib_test_LIBS cpuinfo)
     endif()
     if (onnxruntime_USE_CUDA)
-      list(APPEND onnxruntime_shared_lib_test_LIBS cudart)
+      list(APPEND onnxruntime_shared_lib_test_LIBS CUDA::cudart)
     endif()
     if (onnxruntime_USE_ROCM)
       list(APPEND onnxruntime_shared_lib_test_LIBS hip::host)
diff --git a/onnxruntime/core/providers/cuda/nvtx_profile.cc b/onnxruntime/core/providers/cuda/nvtx_profile.cc
index 6c7c594066b86..867e7c1f24584 100644
--- a/onnxruntime/core/providers/cuda/nvtx_profile.cc
+++ b/onnxruntime/core/providers/cuda/nvtx_profile.cc
@@ -4,13 +4,8 @@
 #ifdef ENABLE_NVTX_PROFILE
 #include "nvtx_profile.h"
 #include "core/common/common.h"
-#if defined(_WIN32) || defined(WIN32) || defined(__CYGWIN__) || defined(__MINGW32__) || defined(__BORLANDC__)
 #include <nvtx3/nvToolsExt.h>
 #include <nvtx3/nvToolsExtCuda.h>
-#else
-#include <nvToolsExt.h>
-#include <nvToolsExtCuda.h>
-#endif
 
 namespace onnxruntime {
 namespace profile {