diff --git a/onnxruntime/core/optimizer/insert_cast_transformer.cc b/onnxruntime/core/optimizer/insert_cast_transformer.cc
index 069bc0678b..d2781bcbea 100644
--- a/onnxruntime/core/optimizer/insert_cast_transformer.cc
+++ b/onnxruntime/core/optimizer/insert_cast_transformer.cc
@@ -254,7 +254,10 @@ Status InsertCastTransformer::ApplyImpl(onnxruntime::Graph& graph, bool& modifie
         ORT_ENFORCE(dtype_attribute->second.has_i(),
                     "InsertCastTransformer works on the assumption that `dtype` attribute holds an integer.");
 
-        dtype_attribute->second.set_i(TensorProto_DataType_FLOAT);
+        // Modify the dtype attribute (which defines the output type) to FLOAT if it is FLOAT16.
+        if (dtype_attribute->second.i() == TensorProto_DataType_FLOAT16) {
+          dtype_attribute->second.set_i(TensorProto_DataType_FLOAT);
+        }
       }
     }
 
diff --git a/onnxruntime/test/framework/insert_cast_transformer_test.cc b/onnxruntime/test/framework/insert_cast_transformer_test.cc
index 152a407e85..e493784fa5 100644
--- a/onnxruntime/test/framework/insert_cast_transformer_test.cc
+++ b/onnxruntime/test/framework/insert_cast_transformer_test.cc
@@ -134,8 +134,9 @@ TEST(TransformerTest, ThreeInARowRemoval) {
   ASSERT_TRUE(op_to_count["Cast"] == 2);
 }
 
-// test a case where the ONNX inferred type (float16) is different from the type bound
-// to the output NodeArg of the "RandomNormalLike" node because of the InsertCaseTransformer
+// test a case where the ONNX inferred output type (float16) is different from the type bound
+// to the output NodeArg of the "RandomNormalLike" node (input is float16) because of the InsertCaseTransformer
+// Here the ONNX inferred output type (float16) must be made float because that is what the kernel produces
 TEST(TransformerTest, RandomNormalLikeWithFloat16Inputs) {
   auto model_uri = MODEL_FOLDER ORT_TSTR("random_normal_like_float16.onnx");
   std::shared_ptr<Model> model;
@@ -153,5 +154,25 @@ TEST(TransformerTest, RandomNormalLikeWithFloat16Inputs) {
   EXPECT_TRUE(status.IsOK()) << status;
 }
 
+// A case where the ONNX inferred output type is int32 to a node that consumes float16 input
+// Here the InsertCastTransformer must not change the ONNX inferred output type and keep it
+// as is (int32)
+TEST(TransformerTest, MultinomialWithFloat16Input) {
+  auto model_uri = MODEL_FOLDER ORT_TSTR("multinomial_float16.onnx");
+  std::shared_ptr<Model> model;
+  auto status = Model::Load(model_uri, model, nullptr, DefaultLoggingManager().DefaultLogger());
+  ASSERT_TRUE(status.IsOK()) << status;
+
+  Graph& graph = model->MainGraph();
+  InsertCastTransformer transformer("Test");
+
+  bool modified = false;
+  status = transformer.Apply(graph, modified, DefaultLoggingManager().DefaultLogger());
+  EXPECT_TRUE(status.IsOK()) << status;
+  EXPECT_TRUE(modified) << "Transformer should have added some Cast nodes";
+  status = graph.Resolve();
+  EXPECT_TRUE(status.IsOK()) << status;
+}
+
 }  // namespace test
 }  // namespace onnxruntime
diff --git a/onnxruntime/test/testdata/transform/multinomial_float16.onnx b/onnxruntime/test/testdata/transform/multinomial_float16.onnx
new file mode 100644
index 0000000000..b72e2e21a2
Binary files /dev/null and b/onnxruntime/test/testdata/transform/multinomial_float16.onnx differ