refine pe when exception raises, test=develop (#20894)

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -77,9 +77,9 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
     // run the recorded operators directly. This strategy could make the
     // execution faster.
     VLOG(3) << "Run the traced ops.";
-    RunTracedOps(traced_ops_);
+    bool is_exception_free =
-    RunTracedOps(fetch_ops);
+        RunTracedOps(traced_ops_) && RunTracedOps(fetch_ops);
-    if (exception_.IsCaught()) {
+    if (!is_exception_free) {
       ExecutionFinal(&fetch_ops);
     }
   } else {
@@ -259,25 +259,25 @@ void FastThreadedSSAGraphExecutor::ExecutionFinal(
   exception_.ReThrow();
 }
 
-void FastThreadedSSAGraphExecutor::RunTracedOps(
+bool FastThreadedSSAGraphExecutor::RunTracedOps(
     const std::vector<OpHandleBase *> &traced_ops) {
   for (auto &op : traced_ops) {
-    if (exception_.IsCaught()) {
+    if (!RunOpSync(op)) return false;
-      return;
-    }
-    RunOpSync(op);
   }
+  return true;
 }
 
-void FastThreadedSSAGraphExecutor::RunOpSync(OpHandleBase *op) {
+bool FastThreadedSSAGraphExecutor::RunOpSync(OpHandleBase *op) {
   try {
     VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
     if (LIKELY(!strategy_.dry_run_)) {
       op->Run(strategy_.use_cuda_);
     }
     VLOG(10) << op << " " << op->Name() << " Done ";
+    return true;
   } catch (...) {
     exception_.Catch(std::current_exception());
+    return false;
   }
 }
 

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h

@@ -78,9 +78,9 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
 
   inline void ExecutionFinal(std::vector<OpHandleBase *> *fetch_ops);
 
-  inline void RunOpSync(OpHandleBase *op);
+  inline bool RunOpSync(OpHandleBase *op);
 
-  void RunTracedOps(const std::vector<OpHandleBase *> &traced_ops);
+  bool RunTracedOps(const std::vector<OpHandleBase *> &traced_ops);
 
   void InsertFetchOps(
       const std::vector<std::string> &fetch_tensors, FeedFetchList *fetches,

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -81,9 +81,9 @@ inline FeedFetchList ThreadedSSAGraphExecutor::RunImpl(
     // run the recorded operators directly. This strategy could make the
     // execution faster.
     VLOG(3) << "Run the traced ops.";
-    RunTracedOps(traced_ops_);
+    bool is_exception_free =
-    RunTracedOps(fetch_ops);
+        RunTracedOps(traced_ops_) && RunTracedOps(fetch_ops);
-    if (exception_holder_.IsCaught()) {
+    if (!is_exception_free) {
       ExecutionFinal(&fetch_ops);
     }
   } else {
@@ -308,25 +308,25 @@ void ThreadedSSAGraphExecutor::RunOp(
   RecordOps(op);
 }
 
-void ThreadedSSAGraphExecutor::RunTracedOps(
+bool ThreadedSSAGraphExecutor::RunTracedOps(
     const std::vector<OpHandleBase *> &traced_ops) {
   for (auto &op : traced_ops) {
-    if (exception_holder_.IsCaught()) {
+    if (!RunOpSync(op)) return false;
-      return;
-    }
-    RunOpSync(op);
   }
+  return true;
 }
 
-void ThreadedSSAGraphExecutor::RunOpSync(OpHandleBase *op) {
+bool ThreadedSSAGraphExecutor::RunOpSync(OpHandleBase *op) {
   try {
     VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
     if (LIKELY(!strategy_.dry_run_)) {
       op->Run(strategy_.use_cuda_);
     }
     VLOG(10) << op << " " << op->Name() << " Done ";
+    return true;
   } catch (...) {
     exception_holder_.Catch(std::current_exception());
+    return false;
   }
 }
 

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -109,9 +109,9 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
 
   inline void ExecutionFinal(std::vector<OpHandleBase *> *fetch_ops);
 
-  inline void RunOpSync(OpHandleBase *op);
+  inline bool RunOpSync(OpHandleBase *op);
 
-  void RunTracedOps(const std::vector<OpHandleBase *> &traced_ops);
+  bool RunTracedOps(const std::vector<OpHandleBase *> &traced_ops);
 };
 
 }  // namespace details