Merge remote-tracking branch 'origin/main' into start-pos-api-llava-7-9

Author: kirklandsign

.ci/docker/build.sh

@@ -54,13 +54,13 @@ case "${IMAGE_NAME}" in
   executorch-ubuntu-22.04-mediatek-sdk)
     MEDIATEK_SDK=yes
     CLANG_VERSION=12
-    ANDROID_NDK_VERSION=r27b
+    ANDROID_NDK_VERSION=r28c
     ;;
   executorch-ubuntu-22.04-clang12-android)
     LINTRUNNER=""
     CLANG_VERSION=12
     # From https://developer.android.com/ndk/downloads
-    ANDROID_NDK_VERSION=r27b
+    ANDROID_NDK_VERSION=r28c
     ;;
   *)
     echo "Invalid image name ${IMAGE_NAME}"

.ci/scripts/setup-samsung-linux-deps.sh

@@ -52,7 +52,7 @@ download_ai_lite_core() {
 install_enn_backend() {
   NDK_INSTALLATION_DIR=/opt/ndk
   rm -rf "${NDK_INSTALLATION_DIR}" && sudo mkdir -p "${NDK_INSTALLATION_DIR}"
-  ANDROID_NDK_VERSION=r27b
+  ANDROID_NDK_VERSION=r28c
 
   # build Exynos backend
   export ANDROID_NDK_ROOT=${ANDROID_NDK_ROOT:-/opt/ndk}

.ci/scripts/test-cuda-build.sh

@@ -9,14 +9,14 @@ set -exu
 
 CUDA_VERSION=${1:-"12.6"}
 
-echo "=== Testing ExecutorTorch CUDA ${CUDA_VERSION} Build ==="
+echo "=== Testing ExecuTorch CUDA ${CUDA_VERSION} Build ==="
 
-# Function to build and test ExecutorTorch with CUDA support
+# Function to build and test ExecuTorch with CUDA support
 test_executorch_cuda_build() {
     local cuda_version=$1
 
-    echo "Building ExecutorTorch with CUDA ${cuda_version} support..."
-    echo "ExecutorTorch will automatically detect CUDA and install appropriate PyTorch wheel"
+    echo "Building ExecuTorch with CUDA ${cuda_version} support..."
+    echo "ExecuTorch will automatically detect CUDA and install appropriate PyTorch wheel"
 
     # Check available resources before starting
     echo "=== System Information ==="
@@ -27,11 +27,11 @@ test_executorch_cuda_build() {
     nvcc --version || echo "nvcc not found"
     nvidia-smi || echo "nvidia-smi not found"
 
-    # Set CMAKE_ARGS to enable CUDA build - ExecutorTorch will handle PyTorch installation automatically
+    # Set CMAKE_ARGS to enable CUDA build - ExecuTorch will handle PyTorch installation automatically
     export CMAKE_ARGS="-DEXECUTORCH_BUILD_CUDA=ON"
 
-    echo "=== Starting ExecutorTorch Installation ==="
-    # Install ExecutorTorch with CUDA support with timeout and error handling
+    echo "=== Starting ExecuTorch Installation ==="
+    # Install ExecuTorch with CUDA support with timeout and error handling
     timeout 5400 ./install_executorch.sh || {
         local exit_code=$?
         echo "ERROR: install_executorch.sh failed with exit code: $exit_code"
@@ -41,15 +41,15 @@ test_executorch_cuda_build() {
         exit $exit_code
     }
 
-    echo "SUCCESS: ExecutorTorch CUDA build completed"
+    echo "SUCCESS: ExecuTorch CUDA build completed"
 
     # Verify the installation
-    echo "=== Verifying ExecutorTorch CUDA Installation ==="
+    echo "=== Verifying ExecuTorch CUDA Installation ==="
 
-    # Test that ExecutorTorch was built successfully
+    # Test that ExecuTorch was built successfully
     python -c "
 import executorch
-print('SUCCESS: ExecutorTorch imported successfully')
+print('SUCCESS: ExecuTorch imported successfully')
 "
 
     # Test CUDA availability and show details
@@ -60,7 +60,7 @@ try:
     print('INFO: CUDA available:', torch.cuda.is_available())
 
     if torch.cuda.is_available():
-        print('SUCCESS: CUDA is available for ExecutorTorch')
+        print('SUCCESS: CUDA is available for ExecuTorch')
         print('INFO: CUDA version:', torch.version.cuda)
         print('INFO: GPU device count:', torch.cuda.device_count())
         print('INFO: Current GPU device:', torch.cuda.current_device())
@@ -74,16 +74,16 @@ try:
         print('SUCCESS: CUDA tensor operation completed on device:', z.device)
         print('INFO: Result tensor shape:', z.shape)
 
-        print('SUCCESS: ExecutorTorch CUDA integration verified')
+        print('SUCCESS: ExecuTorch CUDA integration verified')
     else:
-        print('WARNING: CUDA not detected, but ExecutorTorch built successfully')
+        print('WARNING: CUDA not detected, but ExecuTorch built successfully')
         exit(1)
 except Exception as e:
-    print('ERROR: ExecutorTorch CUDA test failed:', e)
+    print('ERROR: ExecuTorch CUDA test failed:', e)
     exit(1)
 "
 
-    echo "SUCCESS: ExecutorTorch CUDA ${cuda_version} build and verification completed successfully"
+    echo "SUCCESS: ExecuTorch CUDA ${cuda_version} build and verification completed successfully"
 }
 
 # Main execution

.github/workflows/_android.yml

@@ -54,7 +54,7 @@ jobs:
     # NB: Use metal install for KVM support to run the emulator faster
     runs-on: linux.24xl.spr-metal
     env:
-      ANDROID_NDK_VERSION: r27b
+      ANDROID_NDK_VERSION: r28c
       API_LEVEL: 34
     steps:
       - name: Setup SSH (Click me for login details)

CMakePresets.json

@@ -63,7 +63,8 @@
         "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/third-party/ios-cmake/ios.toolchain.cmake",
         "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/ios.cmake",
         "PLATFORM": "OS64",
-        "DEPLOYMENT_TARGET": "17.0"
+        "DEPLOYMENT_TARGET": "17.0",
+        "CMAKE_OSX_DEPLOYMENT_TARGET": "12.0"
       },
       "condition": {
         "lhs": "${hostSystemName}",
@@ -80,7 +81,8 @@
         "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/third-party/ios-cmake/ios.toolchain.cmake",
         "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/ios.cmake",
         "PLATFORM": "SIMULATORARM64",
-        "DEPLOYMENT_TARGET": "17.0"
+        "DEPLOYMENT_TARGET": "17.0",
+        "CMAKE_OSX_DEPLOYMENT_TARGET": "12.0"
       },
       "condition": {
         "lhs": "${hostSystemName}",

backends/cortex_m/CMakeLists.txt

@@ -36,7 +36,7 @@ set(CMSIS_NN_LOCAL_PATH
 # library is downloaded via FetchContent in the default/regular case.
 if(CMSIS_NN_LOCAL_PATH AND EXISTS "${CMSIS_NN_LOCAL_PATH}")
   message(STATUS "Using CMSIS-NN from specified path: ${CMSIS_NN_LOCAL_PATH}")
-  add_subdirectory(${CMSIS_NN_LOCAL_PATH} cmsis_nn_build)
+  add_subdirectory(${CMSIS_NN_LOCAL_PATH} _deps/cmsis_nn-build)
 else()
   # Use FetchContent with automatic fallback
   message(STATUS "Using CMSIS-NN via FetchContent")
@@ -48,23 +48,7 @@ else()
     GIT_SHALLOW TRUE
   )
 
-  FetchContent_GetProperties(cmsis_nn)
-  if(NOT cmsis_nn_POPULATED)
-    FetchContent_Populate(cmsis_nn)
-    add_subdirectory(${cmsis_nn_SOURCE_DIR} ${cmsis_nn_BINARY_DIR})
-  endif()
-endif()
-
-# Add MVEI define to cmsis-nn target
-if(TARGET cmsis-nn)
-  target_compile_definitions(cmsis-nn PUBLIC ARM_MATH_MVEI=1)
-  get_target_property(CMSIS_NN_INCLUDES cmsis-nn INTERFACE_INCLUDE_DIRECTORIES)
-  message(STATUS "CMSIS-NN include dirs: ${CMSIS_NN_INCLUDES}")
-else()
-  message(
-    FATAL_ERROR
-      "CMSIS-NN target not found. Check your CMSIS_NN_LOCAL_PATH or network connection."
-  )
+  FetchContent_MakeAvailable(cmsis_nn)
 endif()
 
 # Cortex-M ops kernel sources

backends/cortex_m/test/test_quantize_op_fusion_pass.py

@@ -313,7 +313,7 @@ def forward(self, x, y):
         # Apply passes
         transformed_program = self._apply_passes(edge_program)
 
-        # Generate ExecutorTorch program
+        # Generate ExecuTorch program
         executorch_program = transformed_program.to_executorch()
 
         # Verify the program contains the expected fused operator

backends/nxp/backend/edge_program_converter.py

@@ -134,6 +134,7 @@ def _process_nodes(self, nodes: list[Node], conversion_context: ConversionContex
 
         qdq_related_functions = [
             exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default,
+            exir_ops.edge.quantized_decomposed.dequantize_per_channel.default,
             exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
         ]
 
@@ -203,7 +204,8 @@ def _convert_qdq_cluster_q_dq_nodes(
         :param conversion_context: ConversionContext instance.
         """
         qdq_q_ops_converters = {
-            exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default: QDQDequantizeConverter,  # noqa F405
+            exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default: QDQPerTensorDequantizeConverter,  # noqa F405
+            exir_ops.edge.quantized_decomposed.dequantize_per_channel.default: QDQPerChannelDequantizeConverter,  # noqa F405
             exir_ops.edge.quantized_decomposed.quantize_per_tensor.default: QDQQuantizeConverter,  # noqa F405
         }
 

backends/nxp/backend/ir/converter/node_converters/ops_converters/__init__.py

@@ -41,7 +41,8 @@
     PermuteCopyConverter,
 )
 from executorch.backends.nxp.backend.ir.converter.node_converters.ops_converters.qdq_dequantize_converter import (
-    QDQDequantizeConverter,
+    QDQPerChannelDequantizeConverter,
+    QDQPerTensorDequantizeConverter,
 )
 from executorch.backends.nxp.backend.ir.converter.node_converters.ops_converters.qdq_quantize_converter import (
     QDQQuantizeConverter,
@@ -70,7 +71,8 @@
     "PermuteCopyConverter",
     "SoftmaxConverter",
     "ViewCopyConverter",
-    "QDQDequantizeConverter",
+    "QDQPerTensorDequantizeConverter",
+    "QDQPerChannelDequantizeConverter",
     "QDQQuantizeConverter",
     "ConstantPadNDConverter",
     "ReLUConverter",

backends/nxp/backend/ir/converter/node_converters/ops_converters/convolution_converter.py

@@ -321,6 +321,10 @@ def _convert_2d_conv(
                 t_op.tmp_inputs[1] = self.builder.create_transposed_tensor(
                     weight_tensor, perm
                 )
+
+                if t_op.tmp_inputs[1].quantization is not None:
+                    # Model is quantized
+                    t_op.tmp_inputs[1].quantization.quantized_dimension = 3
             else:
                 raise NotImplementedError("Dynamic Depthwise Conv weights.")