fix fetch op handle

Author: chengduoZH

paddle/fluid/framework/details/fetch_op_handle.cc

@@ -48,17 +48,18 @@ void FetchOpHandle::RunImpl() {
   WaitInputVarGenerated(platform::CPUPlace());
 
   tensors_.resize(inputs_.size());
-  auto *var_handle = static_cast<VarHandle *>(inputs_[0]);
-  auto &var_name = var_handle->name_;
   platform::CPUPlace cpu;
   auto &scopes = *local_scopes_;
 
-  for (size_t i = 0; i < scopes.size(); ++i) {
-    auto &scope = scopes[i];
-    auto *var =
-        scope->FindVar(kLocalExecScopeName)->Get<Scope *>()->FindVar(var_name);
+  for (size_t i = 0; i < inputs_.size(); ++i) {
+    auto *var_handle = static_cast<VarHandle *>(inputs_[i]);
+    auto &scope = scopes.at(var_handle->scope_idx_);
+    auto *var = scope->FindVar(kLocalExecScopeName)
+                    ->Get<Scope *>()
+                    ->FindVar(var_handle->name_);
     PADDLE_ENFORCE_NOT_NULL(var, "Cannot find variable %s in execution scope",
-                            var_name);
+                            var_handle->name_);
+
     auto &t = var->Get<framework::LoDTensor>();
     if (platform::is_gpu_place(t.place())) {
 #ifdef PADDLE_WITH_CUDA

python/paddle/fluid/tests/unittests/test_parallel_executor.py

@@ -764,7 +764,7 @@ def test_update_dense_parameter(self):
 
     def test_update_sparse_parameter_with_new_strategy(self):
         self.check_network_convergence(
-            is_sparse=False, balance_parameter_opt_between_cards=True)
+            is_sparse=True, balance_parameter_opt_between_cards=True)
 
     def test_update_dense_parameter_with_new_strategy(self):
         self.check_network_convergence(
@@ -836,7 +836,7 @@ def parallel_exe(self, train_inputs, seed):
                     assert not math.isnan(np.sum(ret[i])) and \
                            not math.isinf(np.sum(ret[i]))
 
-    def test_update_sparse_parameter(self):
+    def test_fetch_op(self):
         tst_reader = paddle.batch(flowers.test(use_xmap=False), batch_size=16)
         tst_reader_iter = tst_reader()