Merge branch 'main' into vela-4.4.1

Author: digantdesai

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

@@ -13,7 +13,7 @@ download_ai_lite_core() {
   API_BASE="https://soc-developer.semiconductor.samsung.com/api/v1/resource/ai-litecore/download"
   API_KEY=$SAMSUNG_AI_LITECORE_KEY
 
-  VERSION="0.5"
+  VERSION="0.7"
   OS_NAME="Ubuntu 22.04"
   OUT_FILE="/tmp/exynos-ai-litecore-v${VERSION}.tar.gz"
   TARGET_PATH="/tmp/exynos_ai_lite_core"
@@ -62,7 +62,7 @@ install_enn_backend() {
   export PYTHONPATH=${PYTHONPATH:-}:${EXECUTORCH_ROOT}/..
 }
 
-AI_LITE_CORE_VERSION=0.5.0
+AI_LITE_CORE_VERSION=0.7.0
 
 download_ai_lite_core ${AI_LITE_CORE_VERSION}
 install_enn_backend

.ci/scripts/test_ios_ci.sh

@@ -36,6 +36,7 @@ say() {
 
 say "Cloning the Demo App"
 
+git config --global http.postBuffer 524288000
 git clone --depth 1 https://github.com/meta-pytorch/executorch-examples.git
 
 say "Installing CoreML Backend Requirements"

.github/workflows/cuda.yml

@@ -136,6 +136,7 @@ jobs:
               -DEXECUTORCH_BUILD_CUDA=ON \
               -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
               -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
+              -DEXECUTORCH_BUILD_EXTENSION_NAMED_DATA_MAP=ON \
               -DEXECUTORCH_BUILD_TESTS=ON \
               -Bcmake-out .
         cmake --build cmake-out -j$(( $(nproc) - 1 )) --target voxtral_runner

.github/workflows/pull.yml

@@ -935,6 +935,12 @@ jobs:
           python -m executorch.examples.samsung.aot_compiler --model_name=$model -c E9955
         done
 
+        # Test quant models
+        model_scripts="deeplab_v3 edsr inception_v3 inception_v4 mobilenet_v2 mobilenet_v3 resnet18 resnet50 vit wav2letter"
+        for m_script in $model_scripts; do
+          python -m executorch.examples.samsung.scripts.${m_script} -c e9955 -p A8W8
+        done
+
         # Test ops
         python -m unittest discover -s backends/samsung/test/ops -p "test_*.py"
 

.github/workflows/trunk.yml

@@ -346,7 +346,7 @@ jobs:
         elif [[ ${{ matrix.os}} == "zephyr-preset" ]]; then
           setup_script_args="--target-toolchain zephyr"
           toolchain_prefix=arm-zephyr-eabi-
-          threshold="135168" # 132 KiB
+          threshold="135240" # 132 KiB
           toolchain_cmake=examples/zephyr/x86_64-linux-arm-zephyr-eabi-gcc.cmake
         else
           echo "Fail unsupport OS selection ${{ matrix.os }}"

CMakeLists.txt

@@ -266,6 +266,18 @@ if(EXECUTORCH_BUILD_PTHREADPOOL)
   executorch_move_interface_include_directories_to_build_time_only(
     pthreadpool_interface
   )
+
+  if(APPLE)
+    # Use hidden visibility for pthreadpool on Apple platforms to avoid issues
+    # with pthreadpool symbols from libtorch_cpu taking precedence over the ones
+    # from the pthreadpool library statically linked in _portable_lib. The
+    # pthreadpool public APIs are marked as weak by default on some Apple
+    # platforms, so setting to hidden visibility works around this by not
+    # putting the symbol in the indirection table. See
+    # https://github.com/pytorch/executorch/issues/14321 for more details.
+    target_compile_options(pthreadpool PRIVATE -fvisibility=hidden)
+  endif()
+
   install(
     TARGETS pthreadpool pthreadpool_interface fxdiv
     EXPORT ExecuTorchTargets

backends/aoti/CMakeLists.txt

@@ -40,13 +40,8 @@ target_compile_options(aoti_common PUBLIC -fexceptions -frtti -fPIC)
 # Ensure symbols are exported properly
 target_link_options(aoti_common PUBLIC -Wl,--export-dynamic)
 
-# Link against PyTorch libraries and standard libraries
-target_link_libraries(
-  aoti_common
-  PUBLIC extension_tensor ${CMAKE_DL_LIBS}
-         # Link PyTorch libraries for AOTI functions
-         ${TORCH_LIBRARIES}
-)
+# Link against ExecuTorch libraries and standard libraries
+target_link_libraries(aoti_common PUBLIC extension_tensor ${CMAKE_DL_LIBS})
 executorch_target_link_options_shared_lib(aoti_common)
 
 install(

backends/aoti/aoti_model_container.h

@@ -77,6 +77,8 @@ struct AOTIDelegateHandle {
   void* so_handle;
   std::string so_path;
   AOTInductorModelContainerHandle container_handle;
+  void* cuda_stream; // cudaStream_t stored as void* to avoid CUDA header
+                     // dependency
 };
 
 } // namespace aoti

backends/aoti/common_shims.cpp

@@ -127,11 +127,18 @@ int32_t aoti_torch_layout_strided() {
 }
 
 // Dtype constants - these return the PyTorch dtype codes
-// Currently only float32 is supported, but using robust enum-based approach
 int32_t aoti_torch_dtype_float32() {
   return 6; // PyTorch's float32 dtype code
 }
 
+int32_t aoti_torch_dtype_bfloat16() {
+  return 15; // PyTorch's bfloat16 dtype code
+}
+
+int32_t aoti_torch_dtype_int64() {
+  return 4; // PyTorch's int64 dtype code
+}
+
 // Cleanup functions
 void cleanup_tensor_metadata() {
   internal::tensor_to_sizes.clear();

backends/aoti/common_shims.h

@@ -58,6 +58,8 @@ AOTITorchError aoti_torch_get_dim(Tensor* tensor, int64_t* ret_dim);
 int32_t aoti_torch_device_type_cpu();
 int32_t aoti_torch_layout_strided();
 int32_t aoti_torch_dtype_float32();
+int32_t aoti_torch_dtype_bfloat16();
+int32_t aoti_torch_dtype_int64();
 
 // Autograd mode functions
 int32_t aoti_torch_grad_mode_is_enabled();