Add temperature to llama runner (#1969)

Summary:
Pull Request resolved: #1969

So we have deterministic output for CI testing.

Reviewed By: larryliu0820

Differential Revision: D53735445

fbshipit-source-id: 1a30f2edf366f2f3ca08a14cda20f75184b68801

Author: lucylq

examples/models/llama2/main.cpp

@@ -19,6 +19,11 @@ DEFINE_string(tokenizer_path, "tokenizer.bin", "Tokenizer stuff.");
 
 DEFINE_string(prompt, "The answer to the ultimate question is", "Prompt.");
 
+DEFINE_double(
+    temperature,
+    0.8f,
+    "Temperature; Default is 0.8f. 0 = greedy argmax sampling (deterministic). Lower temperature = more deterministic");
+
 int32_t main(int32_t argc, char** argv) {
   gflags::ParseCommandLineFlags(&argc, &argv, true);
 
@@ -31,8 +36,10 @@ int32_t main(int32_t argc, char** argv) {
 
   const char* prompt = FLAGS_prompt.c_str();
 
+  double temperature = FLAGS_temperature;
+
   // create llama runner
-  ::torch::executor::Runner runner(model_path, tokenizer_path);
+  ::torch::executor::Runner runner(model_path, tokenizer_path, temperature);
 
   // generate
   runner.generate(prompt);

examples/models/llama2/runner/runner.cpp

@@ -27,9 +27,11 @@
 namespace torch {
 namespace executor {
 
-Runner::Runner(const char* model_path, const char* tokenizer_path) {
+Runner::Runner(
+    const char* model_path,
+    const char* tokenizer_path,
+    float temperature) {
   // Constants definition
-  float temperature = 0.8f;
   float topp = 0.9f;
   unsigned long long rng_seed =
       (unsigned int)time(nullptr); // seed rng with time by default

examples/models/llama2/runner/runner.h

@@ -26,7 +26,10 @@ namespace executor {
 
 class Runner {
  public:
-  explicit Runner(const char* model_path, const char* tokenizer_path);
+  explicit Runner(
+      const char* model_path,
+      const char* tokenizer_path,
+      float temperature = 0.8f);
 
   Error generate(
       const std::string& prompt,