diff --git a/tensorflow/lite/delegates/gpu/common/model_builder.cc b/tensorflow/lite/delegates/gpu/common/model_builder.cc
index b33c006edf21124d35b095e442f9919da76fcb3a..b501808811ee42d5407ef0c85ea674a35262cb47 100644
--- a/tensorflow/lite/delegates/gpu/common/model_builder.cc
+++ b/tensorflow/lite/delegates/gpu/common/model_builder.cc
@@ -3124,6 +3124,33 @@ TfLiteIntArray* GetOpsToReplace(TfLiteContext* context, bool allow_quant_ops) {
   return ConvertVectorToTfLiteIntArray(ops_to_replace);
 }
 
+// Creates inputs and outputs passed by io_tensors parameters in the resulting
+// graph. We force it to make sure that delegated subgraph has same order of
+// inputs and outputs with the original one. When delegated model is built from
+// the tflite model representation tensors are created lazily, so there is no
+// guarantee that the order will match the source model tensors order.
+absl::Status PrecreateIOTensors(
+    TfLiteContext* context, GraphFloat32* graph, TfLiteIntArray* io_tensors,
+    std::unordered_map<int, Value<TensorRef<BHWC>>*>* tensor_to_value) {
+  for (int i = 0; i < io_tensors->size; ++i) {
+    const int tensor_index = io_tensors->data[i];
+    if (tensor_to_value->find(tensor_index) != tensor_to_value->end()) {
+      return absl::AlreadyExistsError(absl::StrCat(
+          "Tensor with tflite index ", tensor_index, " was already created."));
+    }
+    const TfLiteTensor& tflite_tensor = context->tensors[tensor_index];
+    if (tflite_tensor.allocation_type == TfLiteAllocationType::kTfLiteMmapRo) {
+      continue;
+    }
+    Value<TensorRef<BHWC>>* value = graph->NewValue();
+    RETURN_IF_ERROR(
+        ConvertTfLiteTensorToTensorRef(tflite_tensor, &value->tensor));
+    value->tensor.ref = tensor_index;
+    (*tensor_to_value)[tensor_index] = value;
+  }
+  return absl::OkStatus();
+}
+
 absl::Status BuildModel(TfLiteContext* context,
                         const TfLiteDelegateParams* delegate_params,
                         GraphFloat32* graph,
@@ -3154,6 +3181,10 @@ absl::Status BuildModel(TfLiteContext* context,
     tflite_nodes.push_back(i);
   }
   std::unordered_map<int, Value<TensorRef<BHWC>>*> tensor_to_value;
+  RETURN_IF_ERROR(PrecreateIOTensors(
+      context, graph, delegate_params->input_tensors, &tensor_to_value));
+  RETURN_IF_ERROR(PrecreateIOTensors(
+      context, graph, delegate_params->output_tensors, &tensor_to_value));
   for (int i = 0; i < operations.size(); ++i) {
     TfLiteNode* tflite_node;
     TfLiteRegistration* registration;