fix asyn update error on pserver

paddle/fluid/operators/listen_and_serv_op.cc

@@ -163,7 +163,8 @@ void ListenAndServOp::RunSyncLoop(
 }
 
 void ListenAndServOp::RunAsyncLoop(framework::Executor *executor,
-                                   framework::ProgramDesc *program) const {
+                                   framework::ProgramDesc *program,
+                                   framework::Scope *recv_scope) const {
   VLOG(3) << "RunAsyncLoop in";
   // grad name to block id
   std::unordered_map<std::string, int32_t> grad_to_block_id;
@@ -191,6 +192,10 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor,
     block_list.push_back(blkid);
   }
   auto optimize_prepared = executor->Prepare(*program, block_list);
+  // execute global block if needed
+  if (block_list[0] == 1 && id_to_grad.count(1) == 0) {
+    executor->RunPreparedContext(optimize_prepared[0].get(), recv_scope);
+  }
   std::unordered_map<std::string,
                      std::shared_ptr<framework::ExecutorPrepareContext>>
       grad_to_prepared_ctx;
@@ -319,7 +324,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
   if (sync_mode) {
     RunSyncLoop(&executor, program, &recv_scope, prefetch_block_id_list);
   } else {
-    RunAsyncLoop(&executor, program);
+    RunAsyncLoop(&executor, program, &recv_scope);
   }
 }
 

paddle/fluid/operators/listen_and_serv_op.h

@@ -50,7 +50,8 @@ class ListenAndServOp : public framework::OperatorBase {
                    const std::vector<int>& prefetch_block_id_list) const;
 
   void RunAsyncLoop(framework::Executor* executor,
-                    framework::ProgramDesc* program) const;
+                    framework::ProgramDesc* program,
+                    framework::Scope* recv_scope) const;
 
   void SavePort() const;
 

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -1299,16 +1299,6 @@ class DistributeTranspiler(object):
                     ufind.union(op1, op2)
         return ufind
 
-    def _is_opt_role_op(self, op):
-        # NOTE: depend on oprole to find out whether this op is for
-        # optimize
-        op_maker = core.op_proto_and_checker_maker
-        optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize
-        if op_maker.kOpRoleAttrName() in op.attrs and \
-            int(op.attrs[op_maker.kOpRoleAttrName()]) == int(optimize_role):
-            return True
-        return False
-
     def _is_optimizer_op(self, op):
         if "Param" in op.input_names and \
             "LearningRate" in op.input_names:
@@ -1399,7 +1389,10 @@ class DistributeTranspiler(object):
         params_grads = []
         origin_var_dict = self.origin_program.global_block().vars
         for op in block.ops:
-            if self._is_opt_role_op(op):
+            # NOTE(Yancey1989): we can not use op role to distinguish an optimizer op
+            # or not, because all ops in optimizer sub-graph would
+            # sign the optimizer op role
+            if self._is_optimizer_op(op):
                 opt_ops.append(op)
                 # HACK(wuyi): if we find grad vars from input of optimize
                 # ops, we may get the output of clip op. Use syntax "@GRAD"