diff --git a/examples/models/llama/main.cpp b/examples/models/llama/main.cpp
index 5179bf28fc7..d75b152be1f 100644
--- a/examples/models/llama/main.cpp
+++ b/examples/models/llama/main.cpp
@@ -20,6 +20,8 @@ DEFINE_string(
     "llama2.pte",
     "Model serialized in flatbuffer format.");
 
+DEFINE_string(data_path, "", "Data file for the model.");
+
 DEFINE_string(tokenizer_path, "tokenizer.bin", "Tokenizer stuff.");
 
 DEFINE_string(prompt, "The answer to the ultimate question is", "Prompt.");
@@ -49,6 +51,11 @@ int32_t main(int32_t argc, char** argv) {
   // and users can create their own DataLoaders to load from arbitrary sources.
   const char* model_path = FLAGS_model_path.c_str();
 
+  std::optional<std::string> data_path = std::nullopt;
+  if (!FLAGS_data_path.empty()) {
+    data_path = FLAGS_data_path.c_str();
+  }
+
   const char* tokenizer_path = FLAGS_tokenizer_path.c_str();
 
   const char* prompt = FLAGS_prompt.c_str();
@@ -74,7 +81,7 @@ int32_t main(int32_t argc, char** argv) {
 #endif
   // create llama runner
   // @lint-ignore CLANGTIDY facebook-hte-Deprecated
-  example::Runner runner(model_path, tokenizer_path);
+  example::Runner runner(model_path, tokenizer_path, data_path);
 
   if (warmup) {
     // @lint-ignore CLANGTIDY facebook-hte-Deprecated