Merge remote-tracking branch 'ups/develop' into refine/jit/gru

test=develop

Author: tensor-tang

cmake/generic.cmake

@@ -261,6 +261,13 @@ function(cc_library TARGET_NAME)
         add_dependencies(${TARGET_NAME} mklml)
         target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
       endif()
+      # remove link to python, see notes at:
+      # https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually
+      if("${cc_library_DEPS};" MATCHES "python;")
+        list(REMOVE_ITEM cc_library_DEPS python)
+        add_dependencies(${TARGET_NAME} python)
+        target_link_libraries(${TARGET_NAME} "-Wl,-undefined,dynamic_lookup")
+      endif()
       target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
       add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
     endif()

paddle/fluid/framework/details/var_handle.h

@@ -49,6 +49,8 @@ struct VarHandleBase {
 
   void AddOutput(OpHandleBase* out, ir::Node* node) {
     if (pending_ops_.find(out) == pending_ops_.end()) {
+      PADDLE_ENFORCE(out != nullptr, "The output of %s should not be nullptr",
+                     this->Node()->Name());
       pending_ops_.insert(out);
       node_->outputs.push_back(node);
     }

paddle/fluid/framework/parallel_executor.cc

@@ -299,6 +299,12 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
 }
 
 ParallelExecutor::~ParallelExecutor() {
+  const auto dev_ctxs =
+      platform::DeviceContextPool::Instance().GetAllDeviceContexts();
+  for (auto &dev_ctx : dev_ctxs) {
+    dev_ctx->Wait();
+  }
+
   if (member_->own_local_scope_) {
     for (size_t i = 1; i < member_->local_scopes_.size(); ++i) {
       Scope *local_scope = member_->local_scopes_[i];