fix some bugs

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -862,7 +862,7 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(
       if (node->Op()->Type() == "fetch_barrier") {
         outvar_dev_id =
             GetVarDeviceID(*result, output->Name(), *sharded_var_device);
-        PADDLE_ENFORCE_NE(outvar_dev_id, -1);
+        PADDLE_ENFORCE_NE(outvar_dev_id, -1, "output name %s", output->Name());
       }
       p = places_[outvar_dev_id];
       ir::Node *new_node = nullptr;

paddle/fluid/operators/distributed/parameter_prefetch.cc

@@ -100,7 +100,7 @@ inline void SplitIdsIntoMultipleVarsBySection(
   }
 }
 
-inline void MergeMultipleVarsIntoOnBySection(
+inline void MergeMultipleVarsIntoOneBySection(
     const std::string& id_name, const std::string& out_name,
     const std::vector<std::string>& out_var_names,
     const std::vector<int64_t>& height_section,
@@ -125,25 +125,30 @@ inline void MergeMultipleVarsIntoOnBySection(
   for (size_t section_idx = 0; section_idx < out_var_names.size();
        ++section_idx) {
     auto& ids_in_this_section = splited_ids[section_idx];
-    auto& prefetch_out_var =
+    if (!ids_in_this_section.empty()) {
-        scope->Var(out_var_names[section_idx])->Get<framework::LoDTensor>();
+      auto& prefetch_out_var =
-    const auto* out_var_data = prefetch_out_var.data<float>();
+          scope->Var(out_var_names[section_idx])->Get<framework::LoDTensor>();
-    auto& dims = prefetch_out_var.dims();
+      const auto* out_var_data = prefetch_out_var.data<float>();
-
+      auto& dims = prefetch_out_var.dims();
-    PADDLE_ENFORCE_EQ(dims.size(), 2, "");
+
-    PADDLE_ENFORCE_EQ(ids_in_this_section.size(), dims[0]);
+      PADDLE_ENFORCE_EQ(dims.size(), 2, "");
-
+      PADDLE_ENFORCE_EQ(ids_in_this_section.size(), dims[0]);
-    auto row_numel = dims[1];
+
-
+      auto row_numel = dims[1];
-    for (size_t i = 0; i < dims[0]; ++i) {
+
-      auto id = ids_in_this_section[i];
+      for (size_t i = 0; i < dims[0]; ++i) {
-      auto origin_id = id + abs_sections[section_idx];
+        auto id = ids_in_this_section[i];
-      auto& offsets = id_to_offset[origin_id];
+        auto origin_id = id + abs_sections[section_idx];
-      for (auto& offset : offsets) {
+        auto& offsets = id_to_offset[origin_id];
-        // should support GPU tensor
+        for (auto& offset : offsets) {
-        memory::Copy(cpu_place, out_tensor_data + offset * row_numel, cpu_place,
+          // should support GPU tensor
-                     out_var_data + i * row_numel, sizeof(float) * row_numel);
+          memory::Copy(cpu_place, out_tensor_data + offset * row_numel,
+                       cpu_place, out_var_data + i * row_numel,
+                       sizeof(float) * row_numel);
+        }
       }
+    } else {
+      VLOG(30) << "ids in this section is empty";
     }
   }
 }
@@ -190,13 +195,14 @@ void prefetch(const std::string& id_name, const std::string& out_name,
       VLOG(30) << "don't send no-initialied variable: " << out_var_names[i];
     }
   }
+
   for (size_t i = 0; i < rets.size(); i++) {
     PADDLE_ENFORCE(rets[i]->Wait(), "internal error in RPCClient");
   }
 
-  MergeMultipleVarsIntoOnBySection(id_name, out_name, out_var_names,
+  MergeMultipleVarsIntoOneBySection(id_name, out_name, out_var_names,
-                                   height_sections, splited_ids, context,
+                                    height_sections, splited_ids, context,
-                                   &local_scope);
+                                    &local_scope);
 
   context.scope().DeleteScope(&local_scope);
 }

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -444,7 +444,7 @@ class DistributeTranspiler(object):
                 # connect deps to send op in async mode
                 recv_dep_in = self.grad_name_to_send_dummy_out[
                     self.param_name_to_grad_name[param_varname]]
-            all_recv_outputs.extend(splited_var)
+
             # get recv op_role_var, if not splited, the grad should have .trainer suffix
             # if splited, grad should be the original grad var name. ParallelExecutor
             # will use op_role_var to get expected device place to run this op.
@@ -460,6 +460,7 @@ class DistributeTranspiler(object):
                 self._update_remote_sparse_update_op(param_varname,
                                                      height_sections, eps)
             else:
+                all_recv_outputs.extend(splited_var)
                 program.global_block().append_op(
                     type="recv",
                     inputs={"X": [recv_dep_in]},