fix nccl wrapper on windows

test=develop

paddle/fluid/framework/fleet/nccl_wrapper.cc

@@ -24,7 +24,7 @@ std::shared_ptr<NCCLWrapper> NCCLWrapper::s_instance_ = NULL;
 bool NCCLWrapper::is_initialized_ = false;
 
 void NCCLWrapper::InitNCCL() {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   PADDLE_ENFORCE(platform::dynload::ncclCommInitRank(
       &(nccl_info_.comm_), nccl_info_.global_ranks_, nccl_info_.nccl_id_,
       nccl_info_.my_global_rank_));
@@ -33,14 +33,14 @@ void NCCLWrapper::InitNCCL() {
 }
 
 void NCCLWrapper::SetNCCLId(const NCCLInfo& nccl_info) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   nccl_info_.nccl_id_ = nccl_info.nccl_id_;
 #endif
   return;
 }
 
 NCCLInfo NCCLWrapper::GetNCCLId() {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   PADDLE_ENFORCE(platform::dynload::ncclGetUniqueId(&(nccl_info_.nccl_id_)));
 #endif
   return nccl_info_;
@@ -48,7 +48,7 @@ NCCLInfo NCCLWrapper::GetNCCLId() {
 
 void NCCLWrapper::SetRankInfo(const int local_rank, const int global_rank,
                               const int ranks) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   nccl_info_.local_rank_ = local_rank;
   nccl_info_.my_global_rank_ = global_rank;
   nccl_info_.global_ranks_ = ranks;
@@ -60,7 +60,7 @@ void NCCLWrapper::SetRankInfo(const int local_rank, const int global_rank,
 
 void NCCLWrapper::SyncVar(const int root_rank, const Scope& scope,
                           const std::vector<std::string>& var_names) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   for (auto& name : var_names) {
     auto var = scope.FindVar(name);
     LoDTensor* tensor = var->GetMutable<LoDTensor>();

paddle/fluid/framework/fleet/nccl_wrapper.h

@@ -24,7 +24,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/variable_helper.h"
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/dynload/nccl.h"
 #endif
 #include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
@@ -41,7 +41,7 @@ class NCCLInfo {
   int local_rank_;
   int global_ranks_;
   int my_global_rank_;
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   ncclUniqueId nccl_id_;
   ncclComm_t comm_;
   cudaStream_t stream_;

paddle/fluid/pybind/pybind.cc

@@ -58,7 +58,9 @@ limitations under the License. */
 #include "paddle/fluid/pybind/imperative.h"
 #include "paddle/fluid/pybind/inference_api.h"
 #include "paddle/fluid/pybind/ir.h"
+#ifndef _WIN32
 #include "paddle/fluid/pybind/nccl_wrapper_py.h"
+#endif
 #include "paddle/fluid/pybind/protobuf.h"
 #include "paddle/fluid/pybind/pybind.h"  // NOLINT
 #include "paddle/fluid/pybind/reader_py.h"
@@ -1406,7 +1408,9 @@ All parameter, weight, gradient are variables in Paddle.
   BindRecordIOWriter(&m);
   BindAsyncExecutor(&m);
   BindFleetWrapper(&m);
+#ifndef _WIN32
   BindNCCLWrapper(&m);
+#endif
   BindGraph(&m);
   BindNode(&m);
   BindInferenceApi(&m);