[new-exec] fit empty program for new executor (#41328)

paddle/fluid/framework/new_executor/interpretercore.cc

@@ -516,6 +516,12 @@ void InterpreterCore::RunInstruction(const Instruction& instr_node) {
 
 void InterpreterCore::ExecuteInstructionList(
     const std::vector<Instruction>& vec_instr) {
+  unfinished_op_numer_ = vec_instr.size();
+  if (unfinished_op_numer_ == 0) {
+    VLOG(4) << "No op to run, return";
+    return;
+  }
+
   // NOTE(zhiqiu): get the prepared deps from std::future, and async prepare
   // those for the next step
   auto atomic_deps = async_work_queue_->AtomicDeps();
@@ -524,8 +530,6 @@ void InterpreterCore::ExecuteInstructionList(
   async_work_queue_->PrepareAtomicDeps(dependecy_count_);
   async_work_queue_->PrepareAtomicVarRef(global_scope_->VecMetaInfo());
 
-  unfinished_op_numer_ = vec_instr.size();
-
   exception_holder_.Clear();
 
   for (size_t i = 0; i < dependecy_count_.size(); ++i) {

python/paddle/fluid/tests/unittests/interpreter/test_standalone_executor.py

@@ -277,6 +277,18 @@ class SwitchExecutorInterfaceWithFeed(unittest.TestCase):
         for x, y in zip(gt, res):
             self.assertTrue(np.array_equal(x, y))
 
+    def test_empty_program(self):
+        program = paddle.static.Program()
+        exe = paddle.static.Executor(self.place)
+        for i in range(10):
+            out = exe.run()  # old executor
+
+        for i in range(10):
+            print(i, flush=1)
+            os.environ['FLAGS_USE_STANDALONE_EXECUTOR'] = '1'
+            out = exe.run(program, feed=None)
+            del os.environ['FLAGS_USE_STANDALONE_EXECUTOR']
+
 
 class TestException(unittest.TestCase):
     def setUp(self):