MTMD inference & UI: Fix garbage output in MTMD inference and resolve complex scenarios on the AIResearch page

android/kantvplayer-lib/src/main/java/kantvai/ai/KANTVAIUtils.java

@@ -316,6 +316,13 @@ public static boolean isMTMD_ImageModel(String name) {
 
      public static boolean isAudioFile(String filename) {
          //naive method
+         //IMPORTANT: must match what mtmd-helper actually supports.
+         //mtmd-helper.cpp uses miniaudio to decode audio and only accepts
+         //wav/mp3/flac. aac and ac3 (e.g. from m4a containers) are NOT
+         //decoded by miniaudio and will be rejected inside mtmd, so we
+         //should not even route them to the audio MTMD path. If a future
+         //build adds aac support, add it here and update the list of
+         //supported audio types accordingly.
          String suffix = filename.substring(filename.lastIndexOf(".") + 1);
          if (suffix.contains("wav")) {
              return true;
@@ -325,11 +332,7 @@ public static boolean isAudioFile(String filename) {
              return true;
          }
 
-         if (suffix.contains("aac")) {
-             return true;
-         }
-
-         if (suffix.contains("ac3")) {
+         if (suffix.contains("flac")) {
              return true;
          }
 

android/kantvplayer-lib/src/main/java/kantvai/ai/ggmljava.java

@@ -112,6 +112,24 @@ public final class ggmljava {
 
     public static native void  llm_finalize();
 
+    /**
+     * Release the llama backend (DSP + quant tables) installed at JNI init.
+     * Pair of {@link #backendInit()}. Safe to call multiple times. The
+     * intended call site is AIResearchFragment.onDestroy() so the Hexagon
+     * DSP gets released on activity teardown instead of holding resources
+     * until the OS kills the process.
+     */
+    public static native void  backendCleanup();
+
+    /**
+     * Release the currently loaded model (and mtmd/vision context if any).
+     * Subsequent inference calls will trigger a fresh 4GB read from disk
+     * - use this to free RAM when the user switches models in
+     * LLMSettingFragment, or to drop the model on activity teardown
+     * alongside {@link #backendCleanup()}. Idempotent.
+     */
+    public static native void  unloadModel();
+
     // ============================================================================================
     // LLM/MTMD benchmark singleton state queries (UI helpers).
     //

android/kantvplayer/src/main/java/com/kantvai/kantvplayer/ui/fragment/settings/LLMSettingFragment.java

@@ -51,6 +51,13 @@ public class LLMSettingFragment extends BaseSettingsFragment {
 
     private LLMSettingHeaderPreference mHeaderPreference;
 
+    // Tracks the model path that was last seen by the pref.llmmodel
+    // change listener. The listener fires both for the user's dropdown
+    // choice AND for the "heal on access" setValue() during fragment
+    // creation (see line 133), so we need to compare against the last
+    // seen value to avoid spurious unloadModel() calls.
+    private String mLastSelectedModelPath = null;
+
 
      @Override
      public String getTitle() {
@@ -221,6 +228,34 @@ public void onSharedPreferenceChanged(SharedPreferences sharedPreferences, Strin
                     KANTVLog.g(TAG, "LLM model name: " + KANTVAIModelMgr.getInstance().getModelName(mSettings.getLLMModel()));
                     String modelPath = KANTVUtils.getSDCardDataPath() + KANTVAIModelMgr.getInstance().getModelName(mSettings.getLLMModel());
                     KANTVLog.g(TAG, "modelPath:" + modelPath);
+
+                    // If the user actually switched to a *different* model
+                    // (not just the heal-on-access setValue that fires
+                    // during fragment creation) and no inference is
+                    // currently running, unload the model on the native
+                    // side right now.
+                    //
+                    // Why here and not in ensure_model_loaded: the C++
+                    // cache-miss path also has an unload + 500ms sleep
+                    // (Phase 2.5 / A), so the safety net is in place.
+                    // The point of doing it here is *latency* - by the
+                    // time the user navigates back to AI Research and
+                    // fires an inference, the 4GB has already been
+                    // released and the page-reclaimer has had seconds
+                    // (not milliseconds) to give those pages back to
+                    // the OS. The inference itself then takes the
+                    // cache-miss path's fast new-load branch instead
+                    // of unloading+loading in one go (which used to
+                    // OOM on a 12GB OnePlus, see project memory).
+                    if (mLastSelectedModelPath != null
+                            && !mLastSelectedModelPath.equals(modelPath)
+                            && !ggmljava.llm_is_running_state()) {
+                        KANTVLog.j(TAG, "Model changed from '"
+                                + mLastSelectedModelPath + "' to '" + modelPath
+                                + "', unloading old model proactively");
+                        ggmljava.unloadModel();
+                    }
+                    mLastSelectedModelPath = modelPath;
                 } catch (Exception ex) {
                     KANTVLog.g(TAG, "error: " + ex.toString());
                     KANTVUtils.showMsgBox(mActivity, "error: " + ex.toString());