add WITH_NPU

Author: ShawnXuan

cmake/cuda.cmake

@@ -19,6 +19,8 @@ if(BUILD_CUDA)
     message(FATAL_ERROR "CUDA_NVCC_GENCODES is deprecated, use CMAKE_CUDA_ARCHITECTURES instead")
   endif()
   add_definitions(-DWITH_CUDA)
+  add_definitions(-DWITH_NPU)
+
   # NOTE: For some unknown reason, CUDAToolkit_VERSION may become empty when running cmake again
   set(CUDA_VERSION ${CUDAToolkit_VERSION} CACHE STRING "")
   if(NOT CUDA_VERSION)

oneflow/core/job/runtime.cpp

@@ -70,13 +70,15 @@ Runtime::Runtime(
     Singleton<RuntimeJobDescs>::Get()->AddPlan(plan);
     collective_boxing_scheduler_plan_token_ =
         Singleton<boxing::collective::Scheduler>::Get()->AddPlan(plan);
+#if defined(WITH_CUDA) || defined(WITH_NPU)
     const auto& vaild_ccl_comm_mgr_device_types =
         EagerCclCommMgrBuilder::Get().vaild_ccl_comm_mgr_device_types();
     if (!vaild_ccl_comm_mgr_device_types.empty() && !Singleton<EagerCclCommMgr>::Get()) {
       Singleton<EagerCclCommMgr>::SetAllocated(
           EagerCclCommMgrBuilder::Get().NewCclCommMgr(vaild_ccl_comm_mgr_device_types.front()));
     }
     Singleton<EagerCclCommMgr>::Get()->CreateCommFromPlan(plan);
+#endif // defined(WITH_CUDA) || defined(WITH_NPU)
   }
   std::vector<const TaskProto*> source_tasks;
   source_tasks.reserve(plan.task().size());