Android JNI llama cache temperature in class

Author: kirklandsign

extension/android/jni/jni_layer_llama.cpp

@@ -114,6 +114,7 @@ class ExecuTorchLlmCallbackJni
 class ExecuTorchLlmJni : public facebook::jni::HybridClass<ExecuTorchLlmJni> {
  private:
   friend HybridBase;
+  float temperature_;
   int model_type_category_;
   std::unique_ptr<llm::IRunner> runner_;
   std::unique_ptr<llm::MultimodalRunner> multi_modal_runner_;
@@ -159,6 +160,7 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass<ExecuTorchLlmJni> {
 #endif
 
     model_type_category_ = model_type_category;
+    temperature_ = temperature;
     if (model_type_category == MODEL_TYPE_CATEGORY_MULTIMODAL) {
       multi_modal_runner_ = std::make_unique<example::LlavaRunner>(
           model_path->toStdString().c_str(),
@@ -181,8 +183,7 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass<ExecuTorchLlmJni> {
     } else if (model_type_category == MODEL_TYPE_MEDIATEK_LLAMA) {
       runner_ = std::make_unique<MTKLlamaRunner>(
           model_path->toStdString().c_str(),
-          tokenizer_path->toStdString().c_str(),
-          temperature);
+          tokenizer_path->toStdString().c_str());
       // Interpret the model type as LLM
       model_type_category_ = MODEL_TYPE_CATEGORY_LLM;
 #endif
@@ -222,6 +223,7 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass<ExecuTorchLlmJni> {
       executorch::extension::llm::GenerationConfig config{
           .echo = static_cast<bool>(echo),
           .seq_len = seq_len,
+          .temperature = temperature_,
       };
       runner_->generate(
           prompt->toStdString(),