Merge branch 'master' into tokenizer-fix-decode

openvinotoolkit · Oct 12, 2023 · 4b57fcc · 4b57fcc
2 parents 2ee3707 + b21bc7c
commit 4b57fcc
Show file tree

Hide file tree

Showing 19 changed files with 512 additions and 436 deletions.
diff --git a/modules/nvidia_plugin/CMakeLists.txt b/modules/nvidia_plugin/CMakeLists.txt
@@ -3,7 +3,7 @@
 #
 
 cmake_minimum_required(VERSION 3.13)
-project(InferenceEngineNVIDIAGpuPlugin CXX CUDA)
+project(OpenVINONVIDIAGpuPlugin CXX CUDA)
 
 # Initialize CMAKE_CUDA_ARCHITECTURES when CMAKE_CUDA_COMPILER_ID is NVIDIA.
 # Raise an error if CUDA_ARCHITECTURES is empty.
@@ -26,16 +26,14 @@ if (ENABLE_CUDNN_BACKEND_API)
     add_definitions(-DENABLE_CUDNN_BACKEND_API)
 endif()
 
-set(IE_MAIN_CUDA_PLUGIN_SOURCE_DIR ${InferenceEngineNVIDIAGpuPlugin_SOURCE_DIR})
-
 find_package(OpenVINODeveloperPackage REQUIRED
              PATHS "${InferenceEngineDeveloperPackage_DIR}")
 
 include(cmake/features.cmake)
 
 set(HAS_REQUIRED REQUIRED)
 
-if(${CMAKE_VERSION} VERSION_LESS "3.17.0")
+if(CMAKE_VERSION VERSION_LESS 3.17.0)
     set(HAS_REQUIRED)
     find_package(CUDA REQUIRED)
     set(CUDAToolkit_VERSION_MAJOR ${CUDA_VERSION_MAJOR})
@@ -92,14 +90,23 @@ endif()
 find_library(CUDNN_PATH
              NAMES cudnn
              ${HAS_REQUIRED}
-             HINTS "$ENV{CUDNN_PATH}"  "${CUDA_TOOLKIT_ROOT_DIR}" "${CUDA_PATH}" "$ENV{CUDA_PATH}"
+             HINTS "$ENV{CUDNN_PATH}" "${CUDA_TOOLKIT_ROOT_DIR}" "${CUDA_PATH}" "$ENV{CUDA_PATH}"
              PATH_SUFFIXES nvidia/current lib64 lib/x64 lib)
 
+if(CUDAToolkit_VERSION_MAJOR LESS_EQUAL 10)
+    set(CUTENSOR_LIB_DIRS lib/10.2 libcutensor/10.2)
+elseif(CUDAToolkit_VERSION_MAJOR GREATER_EQUAL 11 AND CUDAToolkit_VERSION_MINOR LESS_EQUAL 0)
+    set(CUTENSOR_LIB_DIRS lib/11.0 libcutensor/11.0)
+elseif(CUDAToolkit_VERSION_MAJOR GREATER_EQUAL 11 AND CUDAToolkit_VERSION_MINOR GREATER_EQUAL 1)
+    set(CUTENSOR_LIB_DIRS lib/11 libcutensor/11)
+else()
+    set(CUTENSOR_LIB_DIRS lib/${CUDAToolkit_VERSION_MAJOR} libcutensor/${CUDAToolkit_VERSION_MAJOR})
+endif()
 find_library(CUTENSOR_PATH
              NAMES cutensor
              ${HAS_REQUIRED}
              HINTS "$ENV{CUTENSOR_PATH}" "${CUDA_TOOLKIT_ROOT_DIR}" "${CUDA_PATH}" "$ENV{CUDA_PATH}" 
-             PATH_SUFFIXES nvidia/current lib64 lib/x64 lib lib/11)
+             PATH_SUFFIXES nvidia/current lib64 lib/x64 lib ${CUTENSOR_LIB_DIRS})
 get_filename_component(CUTENSOR_INCLUDE_DIR "${CUTENSOR_PATH}" DIRECTORY)
 get_filename_component(CUTENSOR_INCLUDE_DIR "${CUTENSOR_INCLUDE_DIR}/../../include" REALPATH)
 
@@ -121,11 +128,11 @@ set_target_properties(CUDA::cutensor PROPERTIES IMPORTED_LOCATION "${CUTENSOR_PA
 set_target_properties(CUDA::cutensor PROPERTIES IMPORTED_IMPLIB "${CUTENSOR_PATH}")
 
 if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-    if(${CUDAToolkit_VERSION_MAJOR} LESS_EQUAL 10)
+    if(CUDAToolkit_VERSION_MAJOR LESS_EQUAL 10)
         set(CMAKE_CUDA_ARCHITECTURES "30;35;50;60;72")
-    elseif(${CUDAToolkit_VERSION_MAJOR} EQUAL 11 AND ${CUDAToolkit_VERSION_MINOR} LESS_EQUAL 7)
+    elseif(CUDAToolkit_VERSION_MAJOR EQUAL 11 AND CUDAToolkit_VERSION_MINOR LESS_EQUAL 7)
         set(CMAKE_CUDA_ARCHITECTURES "35;50;60-virtual;61;70-virtual;75;86")
-    elseif(${CUDAToolkit_VERSION_MAJOR} EQUAL 11 AND ${CUDAToolkit_VERSION_MINOR} EQUAL 8)
+    elseif(CUDAToolkit_VERSION_MAJOR EQUAL 11 AND CUDAToolkit_VERSION_MINOR EQUAL 8)
         set(CMAKE_CUDA_ARCHITECTURES "35;50;60-virtual;61;70-virtual;75;86;89-virtual;90-virtual")
     else()
         set(CMAKE_CUDA_ARCHITECTURES "50;60-virtual;61;70-virtual;75;86;89-virtual;90-virtual")
@@ -140,7 +147,7 @@ include(FetchContent)
 FetchContent_Declare(GSL
                      GIT_REPOSITORY "https://github.com/microsoft/GSL"
                      GIT_TAG "v3.1.0")
-if(${CMAKE_VERSION} VERSION_LESS "3.14.0")
+if(CMAKE_VERSION VERSION_LESS 3.14.0)
     FetchContent_GetProperties(GSL)
     if(NOT GSL_POPULATED)
         FetchContent_Populate(GSL)
@@ -156,7 +163,7 @@ set_property(TARGET GSL PROPERTY INTERFACE_COMPILE_FEATURES)
 FetchContent_Declare(fmt-header-only
                      GIT_REPOSITORY "https://github.com/fmtlib/fmt"
                      GIT_TAG "7.1.3")
-if(${CMAKE_VERSION} VERSION_LESS "3.14.0")
+if(CMAKE_VERSION VERSION_LESS 3.14.0)
     FetchContent_GetProperties(fmt-header-only)
     if(NOT fmt-header-only_POPULATED)
         FetchContent_Populate(fmt-header-only)
@@ -185,4 +192,4 @@ endif()
 # install
 
 # ATTENTION: uncomment to install component
-# ie_cpack(template)
+# ov_cpack(template)
diff --git a/modules/nvidia_plugin/src/CMakeLists.txt b/modules/nvidia_plugin/src/CMakeLists.txt
@@ -43,7 +43,7 @@ ov_mark_target_as_cc(${TARGET_NAME})
 
 set_property(TARGET ${OBJ_NAME} PROPERTY CUDA_ARCHITECTURES ${CMAKE_CUDA_ARCHITECTURES})
 
-ieTargetLinkWholeArchive(${TARGET_NAME} ${OBJ_NAME})
+ov_target_link_whole_archive(${TARGET_NAME} ${OBJ_NAME})
 
 target_include_directories(${OBJ_NAME}
                            SYSTEM PUBLIC
@@ -53,7 +53,7 @@ target_include_directories(${OBJ_NAME}
                            PUBLIC
                            "${CUTENSOR_INCLUDE_DIR}"
                            "${CMAKE_CURRENT_SOURCE_DIR}"
-                           "${IE_MAIN_CUDA_PLUGIN_SOURCE_DIR}/include"
+                           "${OpenVINONVIDIAGpuPlugin_SOURCE_DIR}/include"
                            PRIVATE $<TARGET_PROPERTY:openvino::conditional_compilation,INTERFACE_INCLUDE_DIRECTORIES>
 )
 

diff --git a/modules/nvidia_plugin/src/cuda/float16.hpp b/modules/nvidia_plugin/src/cuda/float16.hpp
@@ -20,7 +20,7 @@
 #define CUDA_HAS_BF16_MATH
 #endif  // defined (CUDA_HAS_BF16_TYPE) && (__CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__))
 
-#ifdef __CUDACC__
+#if defined(__CUDACC__) && CUDA_VERSION < 12000
 #if !defined(CUDA_HAS_HALF_MATH)
 /* Some basic arithmetic operations expected of a builtin */
 __device__ __forceinline__ __half operator+(const __half &lh, const __half &rh) {
@@ -192,4 +192,4 @@ __device__ __forceinline__ bool operator<=(const __nv_bfloat16 &lh, const __nv_b
     return static_cast<float>(lh) <= static_cast<float>(rh);
 }
 #endif  /* defined(CUDA_HAS_BF16_TYPE) && !defined(CUDA_HAS_BF16_MATH) */
-#endif  // __CUDACC__
+#endif  // defined(__CUDACC__) && CUDA_VERSION < 12000
diff --git a/modules/nvidia_plugin/tests/functional/CMakeLists.txt b/modules/nvidia_plugin/tests/functional/CMakeLists.txt
@@ -3,14 +3,12 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
-# [cmake:functional_tests]
 set(TARGET_NAME ov_nvidia_func_tests)
 
-if (ENABLE_INTEL_CPU)
+# TODO: remove dependency on CPU plugin after migration to tests for API 2.0
+if(CMAKE_SOURCE_DIR STREQUAL OpenVINO_SOURCE_DIR AND ENABLE_INTEL_CPU)
     set(OPENVINO_ADDITIONAL_DEPENDENCIES openvino_intel_cpu_plugin)
-else(ENABLE_INTEL_CPU)
-    set(OPENVINO_ADDITIONAL_DEPENDENCIES)
-endif (ENABLE_INTEL_CPU)
+endif()
 
 ov_add_test_target(
         NAME
@@ -34,4 +32,3 @@ ov_add_test_target(
 if(ENABLE_PROXY)
     target_compile_definitions(${TARGET_NAME} PUBLIC PROXY_PLUGIN_ENABLED)
 endif()
-# [cmake:functional_tests]