2025-05-10 nightly release (b173722)

Author: pytorchbot

.ci/scripts/test_model.sh

@@ -87,10 +87,6 @@ test_model() {
     bash examples/models/llava/install_requirements.sh
     STRICT="--no-strict"
   fi
-  if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" || "$MODEL_NAME" == "llama3_2_text_decoder" ]]; then
-    # Install requirements for llama vision.
-    bash examples/models/llama3_2_vision/install_requirements.sh
-  fi
   if [[ "${MODEL_NAME}" == "qwen2_5" ]]; then
       # Install requirements for export_llama
       bash examples/models/llama/install_requirements.sh

.ci/scripts/unittest-linux.sh

@@ -24,9 +24,6 @@ if [[ "$BUILD_TOOL" == "cmake" ]]; then
     CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
     .ci/scripts/setup-linux.sh "$@"
 
-    # Install llama3_2_vision dependencies.
-    PYTHON_EXECUTABLE=python ./examples/models/llama3_2_vision/install_requirements.sh
-
     .ci/scripts/unittest-linux-cmake.sh
 elif [[ "$BUILD_TOOL" == "buck2" ]]; then
     # Removing this breaks sccache in the Buck build, apparently

.ci/scripts/unittest-macos.sh

@@ -29,7 +29,6 @@ if [[ "$BUILD_TOOL" == "cmake" ]]; then
     # Install llama3_2_vision dependencies.
     PYTHON_EXECUTABLE=python \
     ${CONDA_RUN} --no-capture-output \
-    ./examples/models/llama3_2_vision/install_requirements.sh
 
     .ci/scripts/unittest-macos-cmake.sh
 elif [[ "$BUILD_TOOL" == "buck2" ]]; then

.lintrunner.toml

@@ -220,6 +220,13 @@ exclude_patterns = [
     'extension/**',
     'kernels/optimized/**',
     # Justified <functional> include.
+    'kernels/portable/cpu/op_bitwise*.cpp',
+    'kernels/portable/cpu/op_eq.cpp',
+    'kernels/portable/cpu/op_ge.cpp',
+    'kernels/portable/cpu/op_gt.cpp',
+    'kernels/portable/cpu/op_le.cpp',
+    'kernels/portable/cpu/op_lt.cpp',
+    'kernels/portable/cpu/op_ne.cpp',
     'runtime/kernel/thread_parallel_interface.h',
     'scripts/**',
     'third-party/**',

CMakeLists.txt

@@ -59,6 +59,7 @@ print_configured_options()
 
 include(tools/cmake/Utils.cmake)
 include(CMakeDependentOption)
+include(ExternalProject)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
@@ -242,6 +243,8 @@ option(EXECUTORCH_USE_DL "Use libdl library" ON)
 
 option(EXECUTORCH_BUILD_CADENCE "Build the Cadence DSP backend" OFF)
 
+option(EXECUTORCH_BUILD_CORTEX_M "Build the Cortex-M backend" OFF)
+
 #
 # pthreadpool: build pthreadpool library. Disable on unsupported platforms
 #
@@ -258,6 +261,8 @@ cmake_dependent_option(
   "NOT EXECUTORCH_BUILD_ARM_BAREMETAL" OFF
 )
 
+add_subdirectory(third-party)
+
 if(EXECUTORCH_BUILD_EXTENSION_TRAINING)
   set(EXECUTORCH_BUILD_EXTENSION_DATA_LOADER ON)
   set(EXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR ON)
@@ -452,81 +457,6 @@ if(EXECUTORCH_USE_CPP_CODE_COVERAGE)
   endif()
 endif()
 
-#
-# flatc: Flatbuffer commandline tool to generate .h files from .fbs files
-#
-cmake_dependent_option(
-  EXECUTORCH_BUILD_FLATC "Build the flatc executable." ON
-  "NOT FLATC_EXECUTABLE" OFF
-)
-
-set(FLATBUFFERS_BUILD_FLATC OFF CACHE BOOL "")
-set(FLATBUFFERS_BUILD_FLATHASH OFF CACHE BOOL "")
-set(FLATBUFFERS_BUILD_FLATLIB OFF CACHE BOOL "")
-set(FLATBUFFERS_BUILD_TESTS OFF CACHE BOOL "")
-set(FLATBUFFERS_INSTALL OFF CACHE BOOL "")
-# exir lets users set the alignment of tensor data embedded in the flatbuffer,
-# and some users need an alignment larger than the default, which is typically
-# 32.
-set(FLATBUFFERS_MAX_ALIGNMENT 1024)
-
-if(EXECUTORCH_BUILD_FLATC)
-  if(FLATC_EXECUTABLE)
-    # We could ignore this, but it could lead to confusion about which `flatc`
-    # is actually being used.
-    message(
-      FATAL_ERROR "May not set both EXECUTORCH_BUILD_FLATC and FLATC_EXECUTABLE"
-    )
-  endif()
-
-  # Build flatc for the *host* to generate files as part of the build step.
-  include(ExternalProject)
-  ExternalProject_Add(
-    flatbuffers
-    PREFIX ${CMAKE_CURRENT_BINARY_DIR}/third-party/flatbuffers
-    BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR}/third-party/flatbuffers
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/third-party/flatbuffers
-    CMAKE_ARGS -DFLATBUFFERS_BUILD_FLATC=ON
-               -DFLATBUFFERS_BUILD_FLATHASH=${FLATBUFFERS_BUILD_FLATHASH}
-               -DFLATBUFFERS_BUILD_FLATLIB=${FLATBUFFERS_BUILD_FLATLIB}
-               -DFLATBUFFERS_BUILD_TESTS=${FLATBUFFERS_BUILD_TESTS}
-               -DFLATBUFFERS_INSTALL=${FLATBUFFERS_INSTALL}
-               -DCMAKE_CXX_FLAGS="-DFLATBUFFERS_MAX_ALIGNMENT=${FLATBUFFERS_MAX_ALIGNMENT}"
-               # If building for iOS, "unset" these variables to rely on the host (macOS) defaults.
-               $<$<AND:$<BOOL:${CMAKE_TOOLCHAIN_IOS}>,$<BOOL:$<FILTER:${PLATFORM},EXCLUDE,^MAC>>>:-DCMAKE_OSX_SYSROOT=>
-    INSTALL_COMMAND ""
-    BUILD_BYPRODUCTS <BINARY_DIR>/flatc
-  )
-  ExternalProject_Get_Property(flatbuffers BINARY_DIR)
-  if(WIN32)
-    # flatbuffers does not use CMAKE_BUILD_TYPE. Internally, the build forces Release
-    # config, but from CMake's perspective the build type is always Debug.
-    set(FLATC_EXECUTABLE ${BINARY_DIR}/$<CONFIG>/flatc.exe)
-  elseif(CMAKE_GENERATOR STREQUAL "Xcode")
-    set(FLATC_EXECUTABLE ${BINARY_DIR}/$<CONFIG>/flatc)
-  else()
-    set(FLATC_EXECUTABLE ${BINARY_DIR}/flatc)
-  endif()
-  set(FLATC_EXECUTABLE_BUILT_FROM_SOURCE YES)
-endif()
-
-if(NOT FLATC_EXECUTABLE)
-  message(
-    WARNING "FLATC_EXECUTABLE not specified, looking for flatc"
-  )
-  find_program(FLATC_EXECUTABLE flatc)
-
-  if(NOT FLATC_EXECUTABLE)
-    message(FATAL_ERROR "FLATC_EXECUTABLE must be set when EXECUTORCH_BUILD_FLATC is disabled.")
-  endif()
-endif()
-
-add_executable(flatc IMPORTED GLOBAL)
-set_target_properties(flatc PROPERTIES IMPORTED_LOCATION ${FLATC_EXECUTABLE})
-
-if(FLATC_EXECUTABLE_BUILT_FROM_SOURCE)
-  add_dependencies(flatc flatbuffers)
-endif()
 
 #
 # program_schema: Generated .h files from schema/*.fbs inputs
@@ -715,6 +645,10 @@ if(EXECUTORCH_BUILD_XNNPACK)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/xnnpack)
 endif()
 
+if(EXECUTORCH_BUILD_CORTEX_M)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/cortex_m)
+endif()
+
 if(EXECUTORCH_BUILD_DEVTOOLS)
   if(NOT EXECUTORCH_BUILD_ARM_BAREMETAL)
     set(EXECUTORCH_BUILD_EXTENSION_DATA_LOADER

backends/apple/mps/CMakeLists.txt

@@ -42,7 +42,7 @@ endforeach()
 add_custom_command(
   OUTPUT ${_mps_schema__outputs}
   COMMAND
-    ${FLATC_EXECUTABLE} --cpp --cpp-std c++11 --scoped-enums -o
+    flatc --cpp --cpp-std c++11 --scoped-enums -o
     "${_mps_schema__include_dir}/executorch/backends/apple/mps"
     ${_mps_schema__srcs}
   WORKING_DIRECTORY ${EXECUTORCH_ROOT}

backends/arm/scripts/build_executorch.sh

@@ -129,6 +129,7 @@ cmake                                                 \
     -DEXECUTORCH_BUILD_ARM_BAREMETAL=ON               \
     -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON           \
     -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON       \
+    -DEXECUTORCH_BUILD_CORTEX_M=ON                    \
     -DEXECUTORCH_ENABLE_LOGGING=ON                    \
     ${build_devtools_flags}                           \
     ${build_with_etdump_flags}                        \

backends/arm/scripts/pre-push

@@ -8,11 +8,13 @@
 # non-interactive mode. "$#" gives the number of positional arguments.
 [ "$#" -eq 0 ] && is_script_interactive=1 || is_script_interactive=0
 
-RESET='\e[0m'
-RED='\e[31m'
-GREEN='\e[32m'
-YELLOW='\e[33m'
-BLUE='\e[34m'
+if [ $is_script_interactive -eq 1 ]; then
+    RESET='\e[0m'
+    RED='\e[31m'
+    GREEN='\e[32m'
+    YELLOW='\e[33m'
+    BLUE='\e[34m'
+fi
 
 INFO="${BLUE}[INFO]${RESET}"
 WARNING="${YELLOW}[WARNING]${RESET}"

backends/arm/test/test_arm_baremetal.sh

@@ -83,8 +83,6 @@ test_pytest_ops() { # Test ops and other things
 test_pytest_models() { # Test ops and other things
     echo "${TEST_SUITE_NAME}: Run pytest"
 
-    examples/models/llama3_2_vision/install_requirements.sh
-
     # Prepare for pytest
     backends/arm/scripts/build_executorch.sh
 
@@ -117,8 +115,6 @@ test_pytest_ops_ethosu_fvp() { # Same as test_pytest but also sometime verify us
 test_pytest_models_ethosu_fvp() { # Same as test_pytest but also sometime verify using Corstone FVP
     echo "${TEST_SUITE_NAME}: Run pytest with fvp"
 
-    examples/models/llama3_2_vision/install_requirements.sh
-
     # Prepare Corstone-3x0 FVP for pytest
     backends/arm/scripts/build_executorch.sh
     backends/arm/scripts/build_portable_kernels.sh
@@ -154,6 +150,13 @@ test_run_ethosu_fvp() { # End to End model tests using run.sh
     echo "${TEST_SUITE_NAME}: Test ethos-u target Ethos-U85"
     examples/arm/run.sh --et_build_root=arm_test/test_run --target=ethos-u85-128 --model_name=add
     examples/arm/run.sh --et_build_root=arm_test/test_run --target=ethos-u85-128 --model_name=mul
+
+    # Cortex-M op tests
+    examples/arm/run.sh --et_build_root=arm_test/test_run --target=ethos-u55-128 --model_name=qadd --bundleio
+    examples/arm/run.sh --et_build_root=arm_test/test_run --target=ethos-u55-128 --model_name=qops --bundleio
+    examples/arm/run.sh --et_build_root=arm_test/test_run --target=ethos-u55-128 --model_name=qops --bundleio --no_delegate --portable_kernels="aten::sub.out,aten::add.out,aten::mul.out"
+    examples/arm/run.sh --et_build_root=arm_test/test_run --target=ethos-u85-128 --model_name=qops --bundleio
+
     echo "${TEST_SUITE_NAME}: PASS"
     }
 

backends/cortex_m/CMakeLists.txt

@@ -0,0 +1,61 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Kernel library for Cortex-M operators. Please keep this file formatted by running:
+# ~~~
+# cmake-format -i CMakeLists.txt
+# ~~~
+cmake_minimum_required(VERSION 3.19)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+if(NOT CMAKE_CXX_STANDARD)
+  set(CMAKE_CXX_STANDARD 17)
+endif()
+
+# Source root directory for executorch.
+if(NOT EXECUTORCH_ROOT)
+  set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../..)
+endif()
+
+include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
+include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
+
+if(NOT PYTHON_EXECUTABLE)
+  resolve_python_executable()
+endif()
+
+# Cortex-M ops kernel sources
+set(_cortex_m_kernels__srcs
+    ${CMAKE_CURRENT_SOURCE_DIR}/ops/op_quantize_per_tensor.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/ops/op_dequantize_per_tensor.cpp
+)
+
+# Generate C++ bindings to register kernels into Executorch (for runtime).
+# Here select all ops in operators.yaml
+set(_yaml_file ${CMAKE_CURRENT_LIST_DIR}/ops/operators.yaml)
+gen_selected_ops(LIB_NAME "cortex_m_ops_lib" OPS_SCHEMA_YAML "${_yaml_file}")
+
+# Generate bindings for the kernels
+generate_bindings_for_kernels(
+  LIB_NAME "cortex_m_ops_lib" CUSTOM_OPS_YAML "${_yaml_file}"
+)
+message("Generated files ${gen_command_sources}")
+
+# Build a library for _cortex_m_kernels_srcs
+add_library(cortex_m_kernels ${_cortex_m_kernels__srcs})
+target_link_libraries(cortex_m_kernels PRIVATE executorch)
+target_compile_options(cortex_m_kernels PUBLIC ${_common_compile_options})
+
+# cortex_m_ops_lib: Register Cortex-M ops kernels into Executorch runtime
+gen_operators_lib(
+  LIB_NAME "cortex_m_ops_lib" KERNEL_LIBS cortex_m_kernels DEPS executorch
+)
+
+install(
+  TARGETS cortex_m_kernels cortex_m_ops_lib
+  DESTINATION lib
+  PUBLIC_HEADER DESTINATION include/executorch/backends/cortex_m/ops/
+)