!15799 Fixing multi graph suspend for debugger in GPU

From: @parastooashtari
Reviewed-by: @john_tzanakakis,@tom__chen
Signed-off-by: @john_tzanakakis

mindspore/ccsrc/backend/session/gpu_session.cc

@@ -443,7 +443,7 @@ void GPUSession::PostExecuteGraph(const std::shared_ptr<KernelGraph> &kernel_gra
     DumpJsonParser::GetInstance().UpdateDumpIter();
   }
   if (debugger_) {
-    debugger_->PostExecute();
+    debugger_->PostExecute(kernel_graph);
   }
 }
 

mindspore/ccsrc/debug/debugger/debugger.cc

@@ -328,9 +328,6 @@ void Debugger::PreExecute(const KernelGraphPtr &graph_ptr, uint32_t graph_sum) {
           graph_ptr_ = dbg_graph_ptr;
           SendMultiGraphsAndSuspend(graph_proto_list_, graph_sum);
           graph_proto_list_.clear();
-        } else if (graph_id == rungraph_id_list_.front() && device_target_ == kGPUDevice) {
-          // stop only when receive the first sub run graph for each step
-          CommandLoop();
         }
       }
     }
@@ -343,7 +340,7 @@ void Debugger::PreExecute(const KernelGraphPtr &graph_ptr, uint32_t graph_sum) {
   }
 }
 
-void Debugger::PostExecute() {
+void Debugger::PostExecute(const KernelGraphPtr &graph_ptr) {
   // access lock for public method
   std::lock_guard<std::mutex> a_lock(access_lock_);
   if (pipeline::ExecutorPy::GetDebugTerminate()) {
@@ -355,9 +352,15 @@ void Debugger::PostExecute() {
       if (device_target_ != kGPUDevice) {
         num_step_++;
       }
-
       SendWatchpoints(CheckWatchpoints());
-      CommandLoop();
+      if (graph_ptr != nullptr && device_target_ == kGPUDevice) {
+        auto graph_id = graph_ptr->graph_id();
+        if (graph_id == rungraph_id_list_.front()) {
+          CommandLoop();
+        }
+      } else {
+        CommandLoop();
+      }
     }
     // Only keep parameters in the current map
     debug_services_->ResetLoadedTensors();

mindspore/ccsrc/debug/debugger/debugger.h

@@ -80,7 +80,7 @@ class Debugger : public std::enable_shared_from_this<Debugger> {
 
   // analyze tensors and wait for command
   // don't need a graph_ptr because it is saved during pre_execute
-  void PostExecute();
+  void PostExecute(const KernelGraphPtr &graph_ptr = nullptr);
 
   bool ReadNodeDataRequired(const CNodePtr &kernel);