Merge pull request #766 from intel/niharika/mild_weight_sharing

Mild weight as input implemented to keep quantization parameters as initializers for QDQ nodes

Author: n1harika

onnxruntime/core/providers/openvino/qdq_transformations/qdq_stripping.cc

@@ -677,6 +677,27 @@ static void AddInitializerAsInput(onnxruntime::Graph& dst_graph,
   }
 }
 
+// To check if the input parameters of a DQ or Q node are quantization parameters
+// Scale and Zero point parameters are quantization parameters
+static bool IsQuantizationParameter(const std::string& initializer_name,
+                                   const onnxruntime::GraphViewer& src_graph) {
+    // Check if this initializer is used as scale or zero_point in any DQ/Q node
+    for (auto& node_idx : src_graph.GetNodesInTopologicalOrder()) {
+        const auto* node = src_graph.GetNode(node_idx);
+        if (node->OpType() == "DequantizeLinear" || node->OpType() == "QuantizeLinear") {
+            const auto& input_defs = node->InputDefs();
+            // Check if this initializer is used as scale (input 1) or zero_point (input 2)
+            if (input_defs.size() >= 2 && input_defs[1]->Name() == initializer_name) {
+                return true; // This is a scale parameter
+            }
+            if (input_defs.size() >= 3 && input_defs[2]->Name() == initializer_name) {
+                return true; // This is a zero_point parameter
+            }
+        }
+    }
+    return false;
+}
+
 // Creates a new model without the DQ/Q operators in the src graph.
 Status CreateModelWithStrippedQDQNodes(const GraphViewer& src_graph,
                                        const logging::Logger& logger,
@@ -845,19 +866,31 @@ Status CreateModelWithStrippedQDQNodes(const GraphViewer& src_graph,
     if (!init_with_data &&
         utils::HasExternalData(initializer_tensor) &&
         enable_ovep_weight_sharing) {
-      insert_metadata(initializer_tensor);
 
-      // Add initializer with external data as input
-      AddInitializerAsInput(dst_graph, accumulated_inputs, src_graph, name);
-    } else {
-      // Add as an initialized tensor if it does not have external data
-      if (initializers_to_keep.count(name) > 0) {
-        if (init_with_data) {
-          dst_graph.AddInitializedTensor(*init_with_data);
+        // Only convert to input if it's not a quantization parameter
+        bool is_quant_param = IsQuantizationParameter(name, src_graph);
+
+        if (!is_quant_param) {
+            // This is actual weight data - so to convert to input for weight sharing
+            insert_metadata(initializer_tensor);
+            AddInitializerAsInput(dst_graph, accumulated_inputs, src_graph, name);
         } else {
-          dst_graph.AddInitializedTensor(initializer_tensor);
+            // This is a quantization parameter - keep as initializer even if external
+
+            if (initializers_to_keep.count(name) > 0) {
+
+                dst_graph.AddInitializedTensor(initializer_tensor);
+            }
+        }
+    } else {
+        // Add as an initialized tensor if it does not have external data
+        if (initializers_to_keep.count(name) > 0) {
+            if (init_with_data) {
+                dst_graph.AddInitializedTensor(*init_with_data);
+            } else {
+                dst_graph.AddInitializedTensor(initializer_tensor);
+            }
         }
-      }
     }
 
     current_scope_initializer_set.insert(name);