Merge remote-tracking branch 'origin/main' into add-preset-stories

Author: kirklandsign

.github/workflows/android-build.yml

@@ -31,6 +31,9 @@ jobs:
             path: llm/android/LlamaDemo
           - name: DeepLabV3Demo
             path: dl3/android/DeepLabV3Demo
+          - name: MV3Demo
+            path: mv3/android/MV3Demo
+
 
     name: Build ${{ matrix.name }}
     steps:
@@ -108,6 +111,7 @@ jobs:
             ## Apps included:
             - LlamaDemo APKs (`app-debug-LlamaDemo.apk`, `app-release-unsigned-LlamaDemo.apk`)
             - DeepLabV3Demo APKs (`app-debug-DeepLabV3Demo.apk`, `app-release-unsigned-DeepLabV3Demo.apk`)
+            - MV3Demo APKs (`app-debug-MV3Demo.apk`, `app-release-unsigned-MV3Demo.apk`)
 
             **Build Date:** ${{ steps.tag.outputs.build_date }}
             **Commit:** ${{ github.sha }}

.github/workflows/llm-android.yml

@@ -32,6 +32,10 @@ on:
         description: 'Custom URL for tokenizer file (only used when model_preset is custom)'
         required: false
         type: string
+      local_aar:
+        description: 'URL to download a local AAR file. When set, the workflow will download the AAR and use it instead of the Maven dependency.'
+        required: false
+        type: string
 
 permissions:
   contents: read
@@ -74,6 +78,12 @@ jobs:
       - name: Setup Gradle
         uses: gradle/actions/setup-gradle@v4
 
+      - name: Download local AAR
+        if: ${{ inputs.local_aar }}
+        run: |
+          mkdir -p llm/android/LlamaDemo/app/libs
+          curl -fL -o llm/android/LlamaDemo/app/libs/executorch.aar "${{ inputs.local_aar }}"
+
       - name: AVD cache
         uses: actions/cache@v4
         id: avd-cache
@@ -145,6 +155,7 @@ jobs:
         uses: reactivecircus/android-emulator-runner@v2
         env:
           MODEL_PRESET: ${{ inputs.model_preset || 'stories' }}
+          USE_LOCAL_AAR: ${{ inputs.local_aar != '' }}
         with:
           api-level: ${{ env.API_LEVEL }}
           arch: ${{ env.ARCH }}
@@ -154,7 +165,7 @@ jobs:
           emulator-options: -no-window -gpu swiftshader_indirect -noaudio -no-boot-anim -camera-back none -no-snapshot-save -memory 16384
           disable-animations: true
           working-directory: llm/android/LlamaDemo
-          script: bash ./scripts/run-ci-tests.sh "$MODEL_PRESET" "$MODEL_FILE" "$TOKENIZER_FILE"
+          script: bash ./scripts/run-ci-tests.sh "$MODEL_PRESET" "$MODEL_FILE" "$TOKENIZER_FILE" "$USE_LOCAL_AAR"
 
       - name: Add model response to summary
         if: always()

llm/android/LlamaDemo/app/src/androidTest/java/com/example/executorchllamademo/PresetSanityTest.kt

@@ -25,6 +25,7 @@ import androidx.test.ext.junit.runners.AndroidJUnit4
 import androidx.test.filters.LargeTest
 import org.junit.Assert.assertTrue
 import org.junit.Before
+import org.junit.Ignore
 import org.junit.Rule
 import org.junit.Test
 import org.junit.runner.RunWith
@@ -165,6 +166,7 @@ class PresetSanityTest {
      * 5. Type "Once upon a time" and send
      * 6. Verify response is generated
      */
+    @Ignore("Temporarily disabled")
     @Test
     fun testPresetModelDownloadAndChat() {
         composeTestRule.waitForIdle()

llm/android/LlamaDemo/app/src/main/java/com/example/executorchllamademo/PromptFormat.kt

@@ -85,12 +85,12 @@ object PromptFormat {
     @JvmStatic
     fun getLlavaPresetPrompt(): String {
         return "A chat between a curious human and an artificial intelligence assistant. The assistant" +
-                " gives helpful, detailed, and polite answers to the human's questions. USER: "
+                " gives helpful, detailed, and polite answers to the human's questions."
     }
 
     @JvmStatic
-    fun getLlavaFirstTurnUserPrompt(): String {
-        return "$USER_PLACEHOLDER ASSISTANT:"
+    fun getLlavaMultimodalUserPrompt(): String {
+        return "USER: $USER_PLACEHOLDER ASSISTANT:"
     }
 
     @JvmStatic

llm/android/LlamaDemo/app/src/main/java/com/example/executorchllamademo/ui/viewmodel/ChatViewModel.kt

@@ -110,23 +110,26 @@ class ChatViewModel(application: Application) : AndroidViewModel(application), L
         val isLoadModel = updatedSettingsFields.isLoadModel
         if (isUpdated) {
             checkForClearChatHistory(updatedSettingsFields)
-            // Update media capabilities after settings are updated
-            setBackendMode(updatedSettingsFields.backendType)
 
             if (isLoadModel) {
+                // Update local copy BEFORE checking media capabilities
+                val settingsWithLoadFlagCleared = updatedSettingsFields.copy(isLoadModel = false)
+                currentSettingsFields = settingsWithLoadFlagCleared
+                demoSharedPreferences.saveModuleSettings(settingsWithLoadFlagCleared)
+
+                // Update media capabilities after settings are updated
+                setBackendMode(updatedSettingsFields.backendType)
+
                 loadLocalModelAndParameters(
                     updatedSettingsFields.modelFilePath,
                     updatedSettingsFields.tokenizerFilePath,
                     updatedSettingsFields.dataPath,
                     updatedSettingsFields.temperature.toFloat()
                 )
-                // Save with isLoadModel = false and update local copy to match,
-                // preventing duplicate "To get started..." messages on subsequent calls
-                val settingsWithLoadFlagCleared = updatedSettingsFields.copy(isLoadModel = false)
-                demoSharedPreferences.saveModuleSettings(settingsWithLoadFlagCleared)
-                currentSettingsFields = settingsWithLoadFlagCleared
             } else {
                 currentSettingsFields = updatedSettingsFields.copy()
+                // Update media capabilities after settings are updated
+                setBackendMode(updatedSettingsFields.backendType)
                 if (module == null) {
                     addSystemMessage(systemPromptMessage)
                 }
@@ -243,8 +246,9 @@ class ChatViewModel(application: Application) : AndroidViewModel(application), L
             modelInfo = "Successfully loaded model. $pteName and tokenizer $tokenizerName in ${loadDuration.toFloat() / 1000} sec. $capabilityText"
 
             if (currentSettingsFields.modelType == ModelType.LLAVA_1_5) {
-                ETLogging.getInstance().log("Llava start prefill prompt")
-                module?.prefillPrompt(PromptFormat.getLlavaPresetPrompt())
+                val llavaPresetPrompt = PromptFormat.getLlavaPresetPrompt()
+                ETLogging.getInstance().log("Llava start prefill prompt: $llavaPresetPrompt")
+                module?.prefillPrompt(llavaPresetPrompt)
                 ETLogging.getInstance().log("Llava completes prefill prompt")
             }
             loadSuccess = true
@@ -318,13 +322,13 @@ class ChatViewModel(application: Application) : AndroidViewModel(application), L
             val processedImageList = getProcessedImagesForModel(_selectedImages)
             if (processedImageList.isNotEmpty()) {
                 _messages.add(
-                    Message("Llava/Gemma - Starting image Prefill.", false, MessageType.SYSTEM, 0)
+                    Message("Starting image prefill.", false, MessageType.SYSTEM, 0)
                 )
                 executor.execute {
                     android.os.Process.setThreadPriority(android.os.Process.THREAD_PRIORITY_MORE_FAVORABLE)
                     ETLogging.getInstance().log("Starting runnable prefill image")
                     val img = processedImageList[0]
-                    ETLogging.getInstance().log("Start prefill image")
+                    ETLogging.getInstance().log("Starting prefill image")
                     if (currentSettingsFields.modelType == ModelType.LLAVA_1_5) {
                         module?.prefillImages(
                             img.getInts(),
@@ -333,7 +337,9 @@ class ChatViewModel(application: Application) : AndroidViewModel(application), L
                             ModelUtils.VISION_MODEL_IMAGE_CHANNELS
                         )
                     } else if (currentSettingsFields.modelType == ModelType.GEMMA_3) {
-                        module?.prefillPrompt(PromptFormat.getGemmaPreImagePrompt())
+                        val gemmaPreImagePrompt = PromptFormat.getGemmaPreImagePrompt()
+                        ETLogging.getInstance().log("Gemma prefill pre-image prompt: $gemmaPreImagePrompt")
+                        module?.prefillPrompt(gemmaPreImagePrompt)
                         module?.prefillImages(
                             img.getFloats(),
                             img.width,
@@ -375,8 +381,11 @@ class ChatViewModel(application: Application) : AndroidViewModel(application), L
         val rawPrompt = inputText
         val finalPrompt: String
 
-        if (currentSettingsFields.modelType == ModelType.LLAVA_1_5 && shouldAddSystemPrompt) {
-            finalPrompt = PromptFormat.getLlavaFirstTurnUserPrompt()
+        if (currentSettingsFields.modelType == ModelType.LLAVA_1_5 && _selectedImages.isNotEmpty()) {
+            finalPrompt = PromptFormat.getLlavaMultimodalUserPrompt()
+                .replace(PromptFormat.USER_PLACEHOLDER, rawPrompt)
+        } else if (currentSettingsFields.modelType == ModelType.GEMMA_3 && _selectedImages.isNotEmpty()) {
+            finalPrompt = PromptFormat.getGemmaMultimodalUserPrompt()
                 .replace(PromptFormat.USER_PLACEHOLDER, rawPrompt)
         } else if (currentSettingsFields.modelType == ModelType.GEMMA_3 && _selectedImages.isNotEmpty()) {
             finalPrompt = PromptFormat.getGemmaMultimodalUserPrompt()
@@ -415,8 +424,10 @@ class ChatViewModel(application: Application) : AndroidViewModel(application), L
                 if (currentSettingsFields.modelType == ModelType.VOXTRAL && audioFileToPrefill != null) {
                     prefillVoxtralAudio(audioFileToPrefill!!, finalPrompt)
                     audioFileToPrefill = null
+                    ETLogging.getInstance().log("Running vision model inference.. prompt=(empty after audio prefill)")
                     module?.generate("", ModelUtils.VISION_MODEL_SEQ_LEN, this, false)
                 } else {
+                    ETLogging.getInstance().log("Running vision model inference.. prompt=$finalPrompt")
                     module?.generate(finalPrompt, ModelUtils.VISION_MODEL_SEQ_LEN, this, false)
                 }
             } else if (currentSettingsFields.modelType == ModelType.LLAMA_GUARD_3) {
@@ -459,9 +470,13 @@ class ChatViewModel(application: Application) : AndroidViewModel(application), L
             val bins = 128
             val frames = 3000
             val batchSize = floatCount / (bins * frames)
-            module?.prefillPrompt("<s>[INST][BEGIN_AUDIO]")
+            val preAudioPrompt = "<s>[INST][BEGIN_AUDIO]"
+            val postAudioPrompt = "$textPrompt[/INST]"
+            ETLogging.getInstance().log("Voxtral prefill pre-audio prompt: $preAudioPrompt")
+            module?.prefillPrompt(preAudioPrompt)
             module?.prefillAudio(floats, batchSize, bins, frames)
-            module?.prefillPrompt("$textPrompt[/INST]")
+            ETLogging.getInstance().log("Voxtral prefill post-audio prompt: $postAudioPrompt")
+            module?.prefillPrompt(postAudioPrompt)
         } catch (e: IOException) {
             Log.e("AudioPrefill", "Audio file error")
         }

llm/android/LlamaDemo/docs/delegates/qualcomm_README.md

@@ -106,12 +106,12 @@ We support PTQ by default. The entire export may take ~20 minutes (Llama 3.1 8B)
 Examples:
 ```
 # 4 bits weight only quantize
-python -m extension.llm.export.export_llm base.checkpoint="${MODEL_DIR}/consolidated.00.pth" base.params="${MODEL_DIR}/params.json" model.use_kv_cache=True model.enable_dynamic_shape=False backend.qnn.enabled=True backend.qnn.quantization="qnn_16a4w" model.dtype_override="fp32" base.metadata='"{\"get_bos_id\":128000, \"get_eos_ids\":[128009, 128001]}"' export.output_name="test.pte"
+python -m extension.llm.export.export_llm base.checkpoint="${MODEL_DIR}/consolidated.00.pth" base.params="${MODEL_DIR}/params.json" model.use_kv_cache=True model.enable_dynamic_shape=False backend.qnn.enabled=True quantization.pt2e_quantize="qnn_16a4w" model.dtype_override="fp32" base.metadata='"{\"get_bos_id\":128000, \"get_eos_ids\":[128009, 128001]}"' export.output_name="test.pte"
 ```
 If the model is really big, it may require model sharding because the Qualcomm DSP is a 32bit system and has a 4GB size limit . For example for Llama 3 8B models, we need to shard the model into 4, but ExecuTorch still packages it into one PTE file. Here is an example:
 ```
 # 8 bits quantization with 4 shards
-python -m extension.llm.export.export_llm base.checkpoint="${MODEL_DIR}/consolidated.00.pth" base.params="${MODEL_DIR}/params.json" model.use_kv_cache=True model.enable_dynamic_shape=False backend.qnn.enabled=True backend.qnn.quantization="qnn_8a8w" model.dtype_override="fp32" backend.qnn.num_sharding=4 base.metadata='"{\"get_bos_id\":128000, \"get_eos_ids\":[128009, 128001]}"' export.output_name="test.pte"
+python -m extension.llm.export.export_llm base.checkpoint="${MODEL_DIR}/consolidated.00.pth" base.params="${MODEL_DIR}/params.json" model.use_kv_cache=True model.enable_dynamic_shape=False backend.qnn.enabled=True quantization.pt2e_quantize="qnn_8a8w" model.dtype_override="fp32" backend.qnn.num_sharding=4 base.metadata='"{\"get_bos_id\":128000, \"get_eos_ids\":[128009, 128001]}"' export.output_name="test.pte"
 ```
 Note: if you encountered issues below
 ```
@@ -163,7 +163,7 @@ To export Llama 3 8B instruct with the Qualcomm AI Engine Direct Backend, ensure
 * 8B models might need 16GB RAM on the device to run.
 ```
 # Please note that calibration_data must include the prompt template for special tokens.
-python -m extension.llm.export.export_llm base.tokenizer=<path_to_tokenizer.model> base.params=<path_to_params.json> base.checkpoint=<path_to_checkpoint_for_Meta-Llama-3-8B-Instruct> model.use_kv_cache=True backend.qnn.enabled=True backend.qnn.quantization="qnn_16a4w" model.enable_dynamic_shape=False backend.qnn.num_sharding=8 backend.qnn.calibration_tasks="wikitext" backend.qnn.calibration_limit=1 backend.qnn.calibration_seq_length=128 backend.qnn.optimized_rotation_path=<path_to_optimized_matrix> backend.qnn.calibration_data="<|start_header_id|>system<|end_header_id|>\n\nYou are a funny chatbot.<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nCould you tell me about Facebook?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+python -m extension.llm.export.export_llm base.tokenizer=<path_to_tokenizer.model> base.params=<path_to_params.json> base.checkpoint=<path_to_checkpoint_for_Meta-Llama-3-8B-Instruct> model.use_kv_cache=True backend.qnn.enabled=True quantization.pt2e_quantize="qnn_16a4w" model.enable_dynamic_shape=False backend.qnn.num_sharding=8 backend.qnn.calibration_tasks="wikitext" backend.qnn.calibration_limit=1 backend.qnn.calibration_seq_length=128 backend.qnn.optimized_rotation_path=<path_to_optimized_matrix> backend.qnn.calibration_data="<|start_header_id|>system<|end_header_id|>\n\nYou are a funny chatbot.<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nCould you tell me about Facebook?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
 ```
 
 ## Pushing Model and Tokenizer

llm/android/LlamaDemo/scripts/run-ci-tests.sh

@@ -6,7 +6,7 @@
 # LICENSE file in the root directory of this source tree.
 
 # CI test script for running instrumentation tests with pre-downloaded models
-# Usage: ./run-ci-tests.sh <model_preset> <model_file> <tokenizer_file>
+# Usage: ./run-ci-tests.sh <model_preset> <model_file> <tokenizer_file> [use_local_aar]
 #
 # This script is designed for CI environments where models are pre-downloaded
 # to /tmp/llama_models/ before the emulator starts.
@@ -16,11 +16,13 @@ set -ex
 MODEL_PRESET="$1"
 MODEL_FILE="$2"
 TOKENIZER_FILE="$3"
+USE_LOCAL_AAR="${4:-false}"
 
 echo "=== Test Configuration ==="
 echo "MODEL_PRESET: $MODEL_PRESET"
 echo "MODEL_FILE: $MODEL_FILE"
 echo "TOKENIZER_FILE: $TOKENIZER_FILE"
+echo "USE_LOCAL_AAR: $USE_LOCAL_AAR"
 
 echo "=== Emulator Memory Info ==="
 adb shell cat /proc/meminfo | head -5
@@ -90,9 +92,11 @@ adb logcat > /tmp/logcat.txt &
 LOGCAT_PID=$!
 
 echo "=== Starting Gradle ==="
-./gradlew connectedCheck \
-  -PskipModelDownload=true \
-  -PmodelPreset="$MODEL_PRESET"
+GRADLE_ARGS="-PskipModelDownload=true -PmodelPreset=\"$MODEL_PRESET\""
+if [ "$USE_LOCAL_AAR" = "true" ]; then
+  GRADLE_ARGS="$GRADLE_ARGS -PuseLocalAar=true"
+fi
+eval ./gradlew connectedCheck "$GRADLE_ARGS"
 TEST_EXIT_CODE=$?
 
 echo "=== Model directory after Gradle ==="

llm/apple/etLLM.xcodeproj/project.pbxproj

@@ -462,7 +462,7 @@
 			isa = XCRemoteSwiftPackageReference;
 			repositoryURL = "https://github.com/pytorch/executorch";
 			requirement = {
-				branch = "swiftpm-1.0.0";
+				branch = "swiftpm-1.1.0";
 				kind = branch;
 			};
 		};

mv3/android/MV3Demo/README.md

@@ -0,0 +1,91 @@
+# MV3 Android Demo
+
+This is a sample Android application demonstrating MobileNet v3 (MV3) image classification using PyTorch ExecuTorch.
+
+## Features
+
+- **MobileNet v3 Inference**: Runs the MV3 model on Android.
+- **Live Camera Feed**: Real-time classification on camera frames.
+- **Image Selection**: Pick images from the gallery for classification.
+- **Material Design 3**: Modern UI with a bottom app bar and intuitive controls.
+
+## Prerequisites
+
+- Android SDK (API 34+)
+- JDK 17
+- ExecuTorch libraries (configured via Gradle)
+
+## Setup
+
+1.  **Clone the repository**:
+    ```bash
+    git clone https://github.com/meta-pytorch/executorch-examples.git
+    cd executorch-examples/mv3/android/MV3Demo
+    ```
+
+2.  **Build the project**:
+    ```bash
+    ./gradlew :app:assembleDebug
+    ```
+
+3.  **Install on device**:
+    ```bash
+    ./gradlew :app:installDebug
+    ```
+
+## Usage
+
+- **Local Model**: The app attempts to load `mv3.pte` from the app's internal storage. If missing, it offers a download button (currently configured with a placeholder URL).
+- **Live Camera**: Grant camera permissions to use the real-time classification feature.
+- **Pick Image**: Select an image from your device to classify it.
+
+## Architecture
+
+- **UI**: Jetpack Compose
+- **Camera**: CameraX (Preview + ImageAnalysis)
+- **Inference**: ExecuTorch Android API
+- **Image Processing**: `TensorImageUtils` for bitmap-to-tensor conversion
+
+## Testing
+
+The app includes an instrumentation test that validates the complete image classification workflow.
+
+### What the test does
+
+1. Launches the app
+2. Downloads the MV3 model if not already present
+3. Downloads a cat image from HuggingFace
+4. Runs inference on the image
+5. Validates that the model correctly classifies it as a cat
+
+### Running the test
+
+1. **Connect a device or start an emulator**
+
+2. **Build and install the test APKs**:
+    ```bash
+    ./gradlew installDebug installDebugAndroidTest
+    ```
+
+3. **Run the test**:
+    ```bash
+    adb shell am instrument -w -r \
+      -e class 'org.pytorch.executorchexamples.mv3.UIWorkflowTest#testCatImageClassification' \
+      org.pytorch.executorchexamples.mv3.test/androidx.test.runner.AndroidJUnitRunner
+    ```
+
+    Or run all tests via Gradle:
+    ```bash
+    ./gradlew connectedDebugAndroidTest
+    ```
+
+### Test output
+
+The test logs classification results to logcat with the tag `MV3_RESULT`:
+```bash
+adb logcat -s MV3_RESULT
+```
+
+## License
+
+BSD-3-Clause