Enable QDQ cleanup pass for OVEP

- Add QDQ cleanup as a Level 1 optimization only for OpenVINO
- Fix QDQ cleanup pass so it removes all Q-DQ pairs (by removing all
output edges of DQ nodes and not assume only one output)
- ORT removes remaining DQ layers and adds quantization weights as
intializers of Conv

onnxruntime/core/optimizer/graph_transformer_utils.cc

@@ -250,15 +250,20 @@ InlinedVector<std::unique_ptr<GraphTransformer>> GenerateTransformers(
       // run TransposeOptimizer last as it works in a slightly different way by moving Transpose nodes around.
       // shouldn't affect the end result - just easier to debug any issue if it's last.
       transformers.emplace_back(std::make_unique<TransposeOptimizer>(std::move(cpu_allocator)));
+#if defined(USE_OPENVINO)
+      const bool enable_quant_qdq_cleanup =
+          session_options.config_options.GetConfigOrDefault(kOrtSessionOptionsEnableQuantQDQCleanup, "0") == "1";
+      // The QDQFinalCleanupTransformer must run AFTER other transformers that fuse Q/DQ nodes. Otherwise, their
+      // fusions might be prevented if this one removes a Q/DQ node too early.
+      transformers.emplace_back(std::make_unique<QDQFinalCleanupTransformer>(enable_quant_qdq_cleanup));
+#endif
     } break;
 
     case TransformerLevel::Level2: {
       // we run TransposeOptimizer again in Level2 for some CPU EP specific optimizations that can only be
       // applied once nodes are assigned to the CPU EP (which happens between level 1 and level 2).
       transformers.emplace_back(std::make_unique<TransposeOptimizer>(std::move(cpu_allocator), kCpuExecutionProvider));
 
-      const bool enable_quant_qdq_cleanup =
-          session_options.config_options.GetConfigOrDefault(kOrtSessionOptionsEnableQuantQDQCleanup, "0") == "1";
 #if !defined(DISABLE_CONTRIB_OPS)
       const bool qdq_is_int8_allowed =
           session_options.config_options.GetConfigOrDefault(kOrtSessionOptionsQDQIsInt8Allowed,
@@ -354,9 +359,13 @@ InlinedVector<std::unique_ptr<GraphTransformer>> GenerateTransformers(
 #endif
 
 #endif  // !defined(DISABLE_CONTRIB_OPS)
+#if !defined(USE_OPENVINO)
+      const bool enable_quant_qdq_cleanup =
+          session_options.config_options.GetConfigOrDefault(kOrtSessionOptionsEnableQuantQDQCleanup, "0") == "1";
       // The QDQFinalCleanupTransformer must run AFTER other transformers that fuse Q/DQ nodes. Otherwise, their
       // fusions might be prevented if this one removes a Q/DQ node too early.
       transformers.emplace_back(std::make_unique<QDQFinalCleanupTransformer>(enable_quant_qdq_cleanup));
+#endif
 
     } break;
 

onnxruntime/core/optimizer/qdq_transformer/qdq_final_cleanup.cc

@@ -56,12 +56,12 @@ bool CleanUpNodeSequence(NodeSequence node_sequence_type, Graph& graph, NodeInde
 
   // we have a node sequence to clean up
 
-  // we support a second_node that produces a graph output if it has no output edges, or a second_node with one output edge.
+  // we support a second_node that produces a graph output if it has no output edges, or 
+  // a second_node with one or more output edges
   const bool produces_graph_output = graph.NodeProducesGraphOutput(second_node);
   const auto output_edges_count = second_node.GetOutputEdgesCount();
 
-  if ((produces_graph_output && output_edges_count != 0) ||
-      (!produces_graph_output && output_edges_count != 1)) {
+  if (produces_graph_output && output_edges_count != 0) {
     return false;
   }
 
@@ -92,20 +92,20 @@ bool CleanUpNodeSequence(NodeSequence node_sequence_type, Graph& graph, NodeInde
 
   if (!produces_graph_output) {
     // remove edge to downstream node
-    const Node::EdgeEnd& output_edge = *second_node.OutputEdgesBegin();
-    downstream_node_idx = output_edge.GetNode().Index();
-    downstream_arg_idx = output_edge.GetDstArgIndex();
-
-    // source arg idx is 0 as Q/DQ only has one output
-    graph.RemoveEdge(second_node.Index(), downstream_node_idx, 0, downstream_arg_idx);
-
-    // replace input on downstream node
-    Node& downstream_node = *graph.GetNode(downstream_node_idx);
-    downstream_node.MutableInputDefs()[downstream_arg_idx] = first_node.MutableInputDefs()[0];
-
-    // create edge between src_node (if available) and downstream node
-    if (input_edge) {
-      graph.AddEdge(src_node_idx, downstream_node_idx, src_arg_idx, downstream_arg_idx);
+    for (unsigned int idx = 0; idx < output_edges_count; idx++) {
+      const Node::EdgeEnd& output_edge = *second_node.OutputEdgesBegin();
+      downstream_node_idx = output_edge.GetNode().Index();
+      downstream_arg_idx = output_edge.GetDstArgIndex();
+      graph.RemoveEdge(second_node.Index(), downstream_node_idx, 0, downstream_arg_idx);
+
+      // replace input on downstream node
+      Node& downstream_node = *graph.GetNode(downstream_node_idx);
+      downstream_node.MutableInputDefs()[downstream_arg_idx] = first_node.MutableInputDefs()[0];
+
+      // create edge between src_node (if available) and downstream node
+      if (input_edge) {
+        graph.AddEdge(src_node_idx, downstream_node_idx, src_arg_idx, downstream_arg_idx);
+      }
     }
   } else {
     NodeArg* graph_output_nodearg = second_node.MutableOutputDefs()[0];