[StandaloneExe] improving sequentialRun mode of standaloneExecutor (#52111)

* Verify SequentialRun Model of StandaloneExecutor

* fix

* fix

* fix

* remove redundant code

* fix CI

* fix CI

* recover multi-step dependency

paddle/fluid/framework/new_executor/interpreter/dependency_builder.cc

@@ -297,12 +297,10 @@ void DependencyBuilder::AddDependencyForReadOp() {
 void DependencyBuilder::AddDependencyForSequentialRun() {
   size_t dependence_op_idx = ULLONG_MAX;
   for (size_t op_idx = 0; op_idx < op_num_; ++op_idx) {
-    if (!IsCpuOp(instructions_->at(op_idx))) {
-      if (dependence_op_idx != ULLONG_MAX) {
-        AddDownstreamOp(dependence_op_idx, op_idx);
-      }
-      dependence_op_idx = op_idx;
+    if (dependence_op_idx != ULLONG_MAX) {
+      AddDownstreamOp(dependence_op_idx, op_idx);
     }
+    dependence_op_idx = op_idx;
   }
 }
 

paddle/fluid/framework/new_executor/interpreter/dependency_builder.h

@@ -19,6 +19,8 @@
 
 #include "paddle/fluid/framework/new_executor/new_executor_defs.h"
 
+DECLARE_bool(new_executor_sequential_run);
+
 namespace paddle {
 namespace framework {
 namespace interpreter {

paddle/fluid/framework/new_executor/interpretercore.cc

@@ -360,8 +360,8 @@ void InterpreterCore::reset_scope(Scope* new_scope) {
     const auto& var_name = var_scope_.GetNameById(i);
     var_list[i] = new_scope->FindVar(var_name);
   }
-  // The index should assured valid, cause the InterpreterCore may not be fully
-  // built, but was still cached and used. For example, see unit test
+  // The index should be assured valid, cause the InterpreterCore may not be
+  // fully built, but was still cached and used. For example, see unit test
   // `test_assert.py`, it may exit before `InterpreterCore::Convert`, but still
   // was cached and used by later tests.
   for (size_t i = 0; i < std::min(refs_.size(), var_list.size()); i++) {