diff --git a/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_t5.h b/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_t5.h
index 5360bfd8f4..b2eed7922e 100644
--- a/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_t5.h
+++ b/onnxruntime/contrib_ops/cpu/transformers/beam_search_impl_t5.h
@@ -143,7 +143,7 @@ Status BeamSearchT5<T>::Execute(const FeedsFetchesManager& encoder_feeds_fetches
 
 #ifdef DEBUG_BEAM_SEARCH
   const IConsoleDumper* dumper = this->GetConsoleDumper();
-  for (size_t i = 0; i < encoder_feeds.size(); i++) {
+  for (int i = 0; i < this->encoder_subgraph_.num_subgraph_inputs; i++) {
     dumper->Print("encoder_feeds", static_cast<int>(i), true);
     dumper->Print("", encoder_feeds[i]);
   }
diff --git a/onnxruntime/python/tools/transformers/convert_beam_search.py b/onnxruntime/python/tools/transformers/convert_beam_search.py
index 1a724ec194..0a35286f08 100644
--- a/onnxruntime/python/tools/transformers/convert_beam_search.py
+++ b/onnxruntime/python/tools/transformers/convert_beam_search.py
@@ -677,7 +677,7 @@ def remove_shared_initializers(
         if value_info.name in mapping_initializers_2:
             value_info.name = mapping_initializers_2[value_info.name]
 
-    # Rename nodes inputs in graph 1:
+    # Rename nodes inputs in graph 2:
     for node in graph2.node:
         for j in range(len(node.input)):
             if node.input[j] in mapping_initializers_2:
@@ -854,8 +854,6 @@ def convert_model(args: argparse.Namespace):
     else:
         node.attribute.append(onnx.helper.make_attribute("decoder", decoder_model.graph))
 
-    from onnx import TensorProto
-
     # graph inputs
     input_ids = onnx.helper.make_tensor_value_info("input_ids", TensorProto.INT32, ["batch_size", "sequence_length"])
     max_length = onnx.helper.make_tensor_value_info("max_length", TensorProto.INT32, [1])