Enable Llama3 Multi-turn conversation (#4721)

Summary:
Pull Request resolved: #4721

To provide more conversational type output, we now include previous prompt/responses as part of our input into the `generate()` by the [Llama3 prompt formatting ](https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-3/).
1. Currently we look back at the past 2 conversations (i.e. `CONVERSATION_HISTORY_MESSAGE_LOOKBACK`)
2. This is for text only prompt and responses.
3. Supports if user closes + re-opens app again.
3. As part of this needed to separate out how `prompt` is placed in the `generate()` function since system prompt is always first, followed by conversation history (if present) and then current prompt.

Multi-turn format (with example from [Llama 3 Model Card](https://llama.meta.com/docs/model-cards-and-prompt-formats/meta-llama-3)):
```
<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a helpful AI assistant for travel tips and recommendations<|eot_id|><|start_header_id|>user<|end_header_id|>

What is France's capital?<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Bonjour! The capital of France is Paris!<|eot_id|><|start_header_id|>user<|end_header_id|>

What can I do there?<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Paris, the City of Light, offers a romantic getaway with must-see attractions like the Eiffel Tower and Louvre Museum, romantic experiences like river cruises and charming neighborhoods, and delicious food and drink options, with helpful tips for making the most of your trip.<|eot_id|><|start_header_id|>user<|end_header_id|>

Give me a detailed list of the attractions I should visit, and time it takes in each one, to plan my trip accordingly.<|eot_id|><|start_header_id|>assistant<|end_header_id|>
```

Reviewed By: Riandy

Differential Revision: D61134262

Author: cmodi-meta

examples/demo-apps/android/LlamaDemo/app/build.gradle.kts

@@ -57,7 +57,7 @@ dependencies {
   implementation("androidx.constraintlayout:constraintlayout:2.2.0-alpha12")
   implementation("com.facebook.fbjni:fbjni:0.5.1")
   implementation("com.google.code.gson:gson:2.8.6")
-  implementation(files("libs/executorch-llama.aar"))
+  implementation(files("libs/modified-executorch-llama.aar"))
   implementation("com.google.android.material:material:1.12.0")
   implementation("androidx.activity:activity:1.9.0")
   testImplementation("junit:junit:4.13.2")

examples/demo-apps/android/LlamaDemo/app/src/main/java/com/example/executorchllamademo/MainActivity.java

@@ -70,6 +70,9 @@ public class MainActivity extends AppCompatActivity implements Runnable, LlamaCa
   private SettingsFields mCurrentSettingsFields;
   private Handler mMemoryUpdateHandler;
   private Runnable memoryUpdater;
+  private int promptID = 0;
+
+  private static final int CONVERSATION_HISTORY_MESSAGE_LOOKBACK = 2;
 
   @Override
   public void onResult(String result) {
@@ -183,6 +186,11 @@ private void populateExistingMessages(String existingMsgJSON) {
     mMessageAdapter.notifyDataSetChanged();
   }
 
+  private int setPromptID() {
+
+    return mMessageAdapter.getMaxPromptID() + 1;
+  }
+
   @Override
   protected void onCreate(Bundle savedInstanceState) {
     super.onCreate(savedInstanceState);
@@ -204,6 +212,7 @@ protected void onCreate(Bundle savedInstanceState) {
     String existingMsgJSON = mDemoSharedPreferences.getSavedMessages();
     if (!existingMsgJSON.isEmpty()) {
       populateExistingMessages(existingMsgJSON);
+      promptID = setPromptID();
     }
     mSettingsButton = requireViewById(R.id.settings);
     mSettingsButton.setOnClickListener(
@@ -540,6 +549,48 @@ private void addSelectedImagesToChatThread(List<Uri> selectedImageUri) {
     mMessageAdapter.notifyDataSetChanged();
   }
 
+  private String getConversationHistory() {
+    String conversationHistory = "";
+
+    ArrayList<Message> conversations =
+        mMessageAdapter.getRecentSavedTextMessages(CONVERSATION_HISTORY_MESSAGE_LOOKBACK);
+    if (conversations.isEmpty()) {
+      return conversationHistory;
+    }
+
+    int prevPromptID = conversations.get(0).getPromptID();
+    String conversationFormat =
+        PromptFormat.getConversationFormat(mCurrentSettingsFields.getModelType());
+    String format = conversationFormat;
+    for (int i = 0; i < conversations.size(); i++) {
+      Message conversation = conversations.get(i);
+      int currentPromptID = conversation.getPromptID();
+      if (currentPromptID != prevPromptID) {
+        conversationHistory = conversationHistory + format;
+        format = conversationFormat;
+        prevPromptID = currentPromptID;
+      }
+      if (conversation.getIsSent()) {
+        format = format.replace(PromptFormat.USER_PLACEHOLDER, conversation.getText());
+      } else {
+        format = format.replace(PromptFormat.ASSISTANT_PLACEHOLDER, conversation.getText());
+      }
+    }
+    conversationHistory = conversationHistory + format;
+
+    return conversationHistory;
+  }
+
+  private String getTotalFormattedPrompt(String conversationHistory, String rawPrompt) {
+    if (conversationHistory.isEmpty()) {
+      return mCurrentSettingsFields.getFormattedSystemAndUserPrompt(rawPrompt);
+    }
+
+    return mCurrentSettingsFields.getFormattedSystemPrompt()
+        + conversationHistory
+        + mCurrentSettingsFields.getFormattedUserPrompt(rawPrompt);
+  }
+
   private void onModelRunStarted() {
     mSendButton.setClickable(false);
     mSendButton.setImageResource(R.drawable.baseline_stop_24);
@@ -575,20 +626,21 @@ private void onModelRunStopped() {
                             + " bytes size = "
                             + image.getBytes().length);
               });
+          String conversationHistory = getConversationHistory();
           String rawPrompt = mEditTextMessage.getText().toString();
-          String prompt = mCurrentSettingsFields.getFormattedSystemAndUserPrompt(rawPrompt);
           // We store raw prompt into message adapter, because we don't want to show the extra
           // tokens from system prompt
-          mMessageAdapter.add(new Message(rawPrompt, true, MessageType.TEXT, 0));
+          mMessageAdapter.add(new Message(rawPrompt, true, MessageType.TEXT, promptID));
           mMessageAdapter.notifyDataSetChanged();
           mEditTextMessage.setText("");
-          mResultMessage = new Message("", false, MessageType.TEXT, 0);
+          mResultMessage = new Message("", false, MessageType.TEXT, promptID);
           mMessageAdapter.add(mResultMessage);
           // Scroll to bottom of the list
           mMessagesView.smoothScrollToPosition(mMessageAdapter.getCount() - 1);
           // After images are added to prompt and chat thread, we clear the imageURI list
           // Note: This has to be done after imageURIs are no longer needed by LlamaModule
           mSelectedImageUri = null;
+          promptID++;
           Runnable runnable =
               new Runnable() {
                 @Override
@@ -600,9 +652,10 @@ public void run() {
                           onModelRunStarted();
                         }
                       });
-                  ETLogging.getInstance().log("Running inference.. prompt=" + prompt);
+                  String finalPrompt = getTotalFormattedPrompt(conversationHistory, rawPrompt);
+                  ETLogging.getInstance().log("Running inference.. prompt=" + finalPrompt);
                   long generateStartTime = System.currentTimeMillis();
-                  mModule.generate(prompt, false, MainActivity.this);
+                  mModule.generate(finalPrompt, 700,false, MainActivity.this);
                   long generateDuration = System.currentTimeMillis() - generateStartTime;
                   mResultMessage.setTotalGenerationTime(generateDuration);
                   runOnUiThread(

examples/demo-apps/android/LlamaDemo/app/src/main/java/com/example/executorchllamademo/MessageAdapter.java

@@ -16,6 +16,7 @@
 import android.widget.ImageView;
 import android.widget.TextView;
 import java.util.ArrayList;
+import java.util.Collections;
 
 public class MessageAdapter extends ArrayAdapter<Message> {
 
@@ -90,4 +91,41 @@ public void clear() {
   public ArrayList<Message> getSavedMessages() {
     return savedMessages;
   }
+
+  public ArrayList<Message> getRecentSavedTextMessages(int numOfLatestPromptMessages) {
+    ArrayList<Message> recentMessages = new ArrayList<Message>();
+    int lastIndex = savedMessages.size() - 1;
+    Message messageToAdd = savedMessages.get(lastIndex);
+    int oldPromptID = messageToAdd.getPromptID();
+
+    for (int i = 0; i < savedMessages.size(); i++) {
+      messageToAdd = savedMessages.get(lastIndex - i);
+      if (messageToAdd.getMessageType() != MessageType.SYSTEM) {
+        if (messageToAdd.getPromptID() != oldPromptID) {
+          numOfLatestPromptMessages--;
+          oldPromptID = messageToAdd.getPromptID();
+        }
+        if (numOfLatestPromptMessages > 0) {
+          if (messageToAdd.getMessageType() == MessageType.TEXT) {
+            recentMessages.add(messageToAdd);
+          }
+        } else {
+          break;
+        }
+      }
+    }
+
+    // To place the order in [input1, output1, input2, output2...]
+    Collections.reverse(recentMessages);
+    return recentMessages;
+  }
+
+  public int getMaxPromptID() {
+    int maxPromptID = -1;
+    for (Message msg : savedMessages) {
+
+      maxPromptID = Math.max(msg.getPromptID(), maxPromptID);
+    }
+    return maxPromptID;
+  }
 }

examples/demo-apps/android/LlamaDemo/app/src/main/java/com/example/executorchllamademo/PromptFormat.java

@@ -12,6 +12,7 @@ public class PromptFormat {
 
   public static final String SYSTEM_PLACEHOLDER = "{{ system_prompt }}";
   public static final String USER_PLACEHOLDER = "{{ user_prompt }}";
+  public static final String ASSISTANT_PLACEHOLDER = "{{ assistant_response }}";
 
   public static String getSystemPromptTemplate(ModelType modelType) {
     switch (modelType) {
@@ -32,8 +33,20 @@ public static String getUserPromptTemplate(ModelType modelType) {
       case LLAMA_3_1:
         return "<|start_header_id|>user<|end_header_id|>\n"
             + USER_PLACEHOLDER
-            + "<|eot_id|>\n"
+            + "<|eot_id|>"
             + "<|start_header_id|>assistant<|end_header_id|>";
+
+      case LLAVA_1_5:
+      default:
+        return USER_PLACEHOLDER;
+    }
+  }
+
+  public static String getConversationFormat(ModelType modelType) {
+    switch (modelType) {
+      case LLAMA_3:
+      case LLAMA_3_1:
+        return getUserPromptTemplate(modelType) + "\n" + ASSISTANT_PLACEHOLDER + "<|eot_id|>";
       case LLAVA_1_5:
       default:
         return USER_PLACEHOLDER;

examples/demo-apps/android/LlamaDemo/app/src/main/java/com/example/executorchllamademo/SettingsFields.java

@@ -38,12 +38,12 @@ public String getFormattedSystemAndUserPrompt(String prompt) {
     return getFormattedSystemPrompt() + getFormattedUserPrompt(prompt);
   }
 
-  private String getFormattedSystemPrompt() {
+  public String getFormattedSystemPrompt() {
     return PromptFormat.getSystemPromptTemplate(modelType)
         .replace(PromptFormat.SYSTEM_PLACEHOLDER, systemPrompt);
   }
 
-  private String getFormattedUserPrompt(String prompt) {
+  public String getFormattedUserPrompt(String prompt) {
     return userPrompt.replace(PromptFormat.USER_PLACEHOLDER, prompt);
   }