Merge pull request #774 from intel/sahar/psu_lora

DeQuantizeLinear is dangling which needs to be handled in capability.cc

Author: n1harika

onnxruntime/core/providers/openvino/ov_versions/capability.cc

@@ -166,17 +166,33 @@ std::vector<std::unique_ptr<ComputeCapability>> GetCapability::Execute() {
     auto connected_clusters = GetConnectedClusters(graph_viewer_, ng_clusters);
 
     int no_of_clusters = 0;
+    std::vector<NodeIndex> prev_cluster;
+    bool try_next_cluster = false;
 
     for (auto this_cluster : connected_clusters) {
+      bool omit_subgraph = false;
+      if (try_next_cluster) {
+        // no need to check previous cluster
+        for (auto idx : prev_cluster) {
+          if ((std::find(this_cluster.begin(), this_cluster.end(), idx)) == this_cluster.end()) {
+            this_cluster.emplace_back(idx);
+          }
+        }
+        try_next_cluster = false;
+      }
+
       // If subgraph has less then three, graph is considered trivial unless its an epctx cluster
-      if (this_cluster.size() < 3) {
+      if (!try_next_cluster && this_cluster.size() < 3) {
         bool is_epctx_node = false;
         for (auto node_idx : this_cluster) {
           if (graph_viewer_.GetNode(node_idx)->OpType() == "EPContext")
             is_epctx_node = true;
         }
-        if (!is_epctx_node)
-          continue;
+        if (!is_epctx_node) {
+          omit_subgraph = true;
+          prev_cluster = this_cluster;
+          try_next_cluster = true;
+        }
       }
 
       std::vector<std::string> cluster_graph_inputs, cluster_inputs, cluster_outputs;
@@ -188,7 +204,7 @@ std::vector<std::unique_ptr<ComputeCapability>> GetCapability::Execute() {
                                 cluster_inputs,
                                 cluster_outputs);
 
-      bool omit_subgraph = false;
+
       // Omitting zero dim subgraphs
       for (auto index : this_cluster) {
         const Node* node = graph_viewer_.GetNode(index);