diff --git a/hipamd/src/hip_graph.cpp b/hipamd/src/hip_graph.cpp
index fcd7cfd02e..fa0b56087e 100644
--- a/hipamd/src/hip_graph.cpp
+++ b/hipamd/src/hip_graph.cpp
@@ -42,7 +42,9 @@ inline hipError_t ihipGraphUpload(hipGraphExec_t graphExec, hipStream_t stream)
   if (!hip::isValid(stream)) {
     return hipErrorContextIsDestroyed;
   }
-
+  if (!hip::GraphExec::isGraphExecValid(reinterpret_cast<hip::GraphExec*>(graphExec))) {
+    return hipErrorInvalidValue;
+  }
   return hipSuccess;
 }
 
diff --git a/hipamd/src/hip_graph_internal.cpp b/hipamd/src/hip_graph_internal.cpp
index eac63876f2..eb627b90f3 100644
--- a/hipamd/src/hip_graph_internal.cpp
+++ b/hipamd/src/hip_graph_internal.cpp
@@ -356,6 +356,7 @@ hipError_t GraphExec::Init() {
     // For graph nodes capture AQL packets to dispatch them directly during graph launch.
     status = CaptureAQLPackets();
   }
+  instantiateDeviceId_ = hip::getCurrentDevice()->deviceId();
   return status;
 }
 
@@ -589,7 +590,8 @@ hipError_t GraphExec::Run(hipStream_t stream) {
     repeatLaunch_ = true;
   }
 
-  if (parallelLists_.size() == 1) {
+  if (parallelLists_.size() == 1 &&
+      instantiateDeviceId_ == hip_stream->DeviceId()) {
     amd::AccumulateCommand* accumulate = nullptr;
     bool isLastPacketKernel = false;
     if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
@@ -631,6 +633,13 @@ hipError_t GraphExec::Run(hipStream_t stream) {
       accumulate->enqueue();
       accumulate->release();
     }
+  } else if (parallelLists_.size() == 1 &&
+             instantiateDeviceId_ != hip_stream->DeviceId()) {
+    for (int i = 0; i < topoOrder_.size(); i++) {
+      topoOrder_[i]->SetStream(hip_stream, this);
+      status = topoOrder_[i]->CreateCommand(topoOrder_[i]->GetQueue());
+      topoOrder_[i]->EnqueueCommands(stream);
+    }
   } else {
     UpdateStream(parallelLists_, hip_stream, this);
     amd::Command* rootCommand = nullptr;
diff --git a/hipamd/src/hip_graph_internal.hpp b/hipamd/src/hip_graph_internal.hpp
index e14ce6b6b2..ac9f775069 100644
--- a/hipamd/src/hip_graph_internal.hpp
+++ b/hipamd/src/hip_graph_internal.hpp
@@ -569,6 +569,7 @@ struct GraphExec {
   std::vector<address> kernarg_graph_;
   uint32_t kernarg_graph_cur_offset_ = 0;
   uint32_t kernarg_graph_size_ = 128 * Ki;
+  int instantiateDeviceId_ = -1;
 
  public:
   GraphExec(std::vector<Node>& topoOrder, std::vector<std::vector<Node>>& lists,