Merge branch 'main' into export-D78762534

Author: ethansfng

.ci/scripts/setup-qnn-deps.sh

@@ -7,47 +7,7 @@
 
 set -ex
 
-verify_pkg_installed() {
-  echo $(dpkg-query -W --showformat='${Status}\n' $1|grep "install ok installed")
-}
+source "$(dirname "${BASH_SOURCE[0]}")/../../backends/qualcomm/scripts/install_qnn_sdk.sh"
 
-install_qnn() {
-  echo "Start installing qnn."
-  QNN_INSTALLATION_DIR=/tmp/qnn
-  mkdir -p "${QNN_INSTALLATION_DIR}"
-
-  curl -Lo /tmp/v2.28.0.24.10.29.zip "https://softwarecenter.qualcomm.com/api/download/software/qualcomm_neural_processing_sdk/v2.28.0.241029.zip"
-  echo "Finishing downloading qnn sdk."
-  unzip -qo /tmp/v2.28.0.24.10.29.zip -d /tmp
-  echo "Finishing unzip qnn sdk."
-
-
-  # Print the content for manual verification
-  ls -lah "/tmp/qairt"
-  mv "/tmp/qairt"/* "${QNN_INSTALLATION_DIR}"
-  echo "Finishing installing qnn '${QNN_INSTALLATION_DIR}' ."
-
-  ls -lah "${QNN_INSTALLATION_DIR}"
-}
-
-setup_libc++() {
-  clang_version=$1
-  sudo apt-get update
-  pkgs_to_check=("libc++-${clang_version}-dev")
-  j=0
-  while [ $j -lt ${#pkgs_to_check[*]} ]; do
-    install_status=$(verify_pkg_installed ${pkgs_to_check[$j]})
-    if [ "$install_status" == "" ]; then
-      sudo apt-get install -y ${pkgs_to_check[$j]}
-      if [[ $? -ne 0 ]]; then
-        echo "ERROR: Failed to install required packages for libc++"
-        exit 1
-      fi
-    fi
-    j=$(( $j +1));
-  done
-}
-
-# This needs to match with the clang version from the Docker image
-setup_libc++ 12
+setup_libcpp 12
 install_qnn

.ci/scripts/unittest-linux.sh

@@ -22,7 +22,7 @@ if [[ "$BUILD_TOOL" == "cmake" ]]; then
 
     # We need the runner to test the built library.
     PYTHON_EXECUTABLE=python \
-    CMAKE_ARGS="-DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON -DEXECUTORCH_BUILD_TESTS=ON" \
+    CMAKE_ARGS="-DEXECUTORCH_BUILD_EXTENSION_EVALUE_UTIL=ON -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON -DEXECUTORCH_BUILD_TESTS=ON" \
     .ci/scripts/setup-linux.sh "$@"
 
     .ci/scripts/unittest-linux-cmake.sh

.ci/scripts/unittest-macos.sh

@@ -22,7 +22,7 @@ trap 'rm -rfv ${TMP_DIR}' EXIT
 # Setup MacOS dependencies as there is no Docker support on MacOS atm
 # We need the runner to test the built library.
 PYTHON_EXECUTABLE=python \
-CMAKE_ARGS="-DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON -DEXECUTORCH_BUILD_TESTS=ON" \
+CMAKE_ARGS="-DEXECUTORCH_BUILD_EXTENSION_EVALUE_UTIL=ON -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON -DEXECUTORCH_BUILD_TESTS=ON" \
 ${CONDA_RUN} --no-capture-output \
 .ci/scripts/setup-macos.sh "$@"
 

.github/workflows/trunk.yml

@@ -387,7 +387,7 @@ jobs:
         eval "$(conda shell.bash hook)"
 
         # Install requirements
-        ${CONDA_RUN} python install_executorch.py
+        ${CONDA_RUN} EXECUTORCH_BUILD_TORCHAO=1 python install_executorch.py
         ${CONDA_RUN} sh examples/models/llama/install_requirements.sh
 
         # Run test

CMakeLists.txt

@@ -416,7 +416,7 @@ target_link_libraries(executorch PRIVATE executorch_core)
 target_include_directories(executorch PUBLIC ${_common_include_directories})
 target_compile_definitions(executorch PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
 target_compile_options(executorch PUBLIC ${_common_compile_options})
-target_link_options_shared_lib(executorch)
+executorch_target_link_options_shared_lib(executorch)
 
 #
 # portable_ops_lib: A library to register core ATen ops using portable kernels,
@@ -548,6 +548,16 @@ if(EXECUTORCH_BUILD_EXTENSION_DATA_LOADER)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/data_loader)
 endif()
 
+if(EXECUTORCH_BUILD_EXTENSION_EVALUE_UTIL)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/evalue_util)
+  install(
+    DIRECTORY extension/evalue_util/
+    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/executorch/extension/evalue_util
+    FILES_MATCHING
+    PATTERN "*.h"
+  )
+endif()
+
 if(EXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/flat_tensor)
 endif()
@@ -576,6 +586,12 @@ endif()
 
 if(EXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/runner_util)
+  install(
+    DIRECTORY extension/runner_util/
+    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/executorch/extension/runner_util
+    FILES_MATCHING
+    PATTERN "*.h"
+  )
 endif()
 
 if(EXECUTORCH_BUILD_EXTENSION_TENSOR)
@@ -651,8 +667,7 @@ if(EXECUTORCH_BUILD_PYBIND)
 
   # util lib
   add_library(
-    util ${CMAKE_CURRENT_SOURCE_DIR}/extension/evalue_util/print_evalue.cpp
-         ${CMAKE_CURRENT_SOURCE_DIR}/extension/aten_util/aten_bridge.cpp
+    util ${CMAKE_CURRENT_SOURCE_DIR}/extension/aten_util/aten_bridge.cpp
   )
   target_include_directories(
     util PUBLIC ${_common_include_directories} ${TORCH_INCLUDE_DIRS}
@@ -690,12 +705,14 @@ endif()
 
 if(EXECUTORCH_BUILD_KERNELS_QUANTIZED)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantized)
-  target_link_options_shared_lib(quantized_ops_lib)
+  executorch_target_link_options_shared_lib(quantized_ops_lib)
 endif()
 
 if(EXECUTORCH_BUILD_EXECUTOR_RUNNER)
   # Baseline libraries that executor_runner will link against.
-  set(_executor_runner_libs executorch gflags)
+  set(_executor_runner_libs executorch extension_evalue_util
+                            extension_runner_util gflags
+  )
 
   if(EXECUTORCH_BUILD_KERNELS_OPTIMIZED)
     list(APPEND _executor_runner_libs optimized_native_cpu_ops_lib)

backends/apple/coreml/CMakeLists.txt

@@ -199,7 +199,7 @@ if(APPLE)
       ${CMAKE_CURRENT_SOURCE_DIR}/third-party/coremltools/deps/protobuf/cmake
     )
 
-    target_link_options_shared_lib(libprotobuf-lite)
+    executorch_target_link_options_shared_lib(libprotobuf-lite)
     target_link_libraries(coremldelegate PRIVATE libprotobuf-lite)
   endif()
 
@@ -210,7 +210,7 @@ if(APPLE)
             ${FOUNDATION_FRAMEWORK} ${SQLITE_LIBRARY}
   )
 
-  target_link_options_shared_lib(coremldelegate)
+  executorch_target_link_options_shared_lib(coremldelegate)
 
   if(EXECUTORCH_COREML_BUILD_EXECUTOR_RUNNER)
     target_link_libraries(

backends/apple/coreml/test/tester.py

@@ -4,23 +4,73 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-from typing import Any, List, Optional, Tuple
+import functools
+from typing import Any, List, Optional, Sequence, Tuple
 
+import coremltools as ct
 import executorch
 import executorch.backends.test.harness.stages as BaseStages
-
 import torch
+
+from executorch.backends.apple.coreml.compiler import CoreMLBackend
 from executorch.backends.apple.coreml.partition import CoreMLPartitioner
+from executorch.backends.apple.coreml.quantizer import CoreMLQuantizer
 from executorch.backends.test.harness import Tester as TesterBase
 from executorch.backends.test.harness.stages import StageType
 from executorch.exir import EdgeCompileConfig
 from executorch.exir.backend.partitioner import Partitioner
 
 
+def _create_default_partitioner(
+    minimum_deployment_target: Any = ct.target.iOS15,
+) -> CoreMLPartitioner:
+    return CoreMLPartitioner(
+        compile_specs=CoreMLBackend.generate_compile_specs(
+            minimum_deployment_target=minimum_deployment_target
+        )
+    )
+
+
+def _get_static_int8_linear_qconfig():
+    return ct.optimize.torch.quantization.LinearQuantizerConfig(
+        global_config=ct.optimize.torch.quantization.ModuleLinearQuantizerConfig(
+            quantization_scheme="symmetric",
+            activation_dtype=torch.quint8,
+            weight_dtype=torch.qint8,
+            weight_per_channel=True,
+        )
+    )
+
+
+class Quantize(BaseStages.Quantize):
+    def __init__(
+        self,
+        quantizer: Optional[CoreMLQuantizer] = None,
+        quantization_config: Optional[Any] = None,
+        calibrate: bool = True,
+        calibration_samples: Optional[Sequence[Any]] = None,
+        is_qat: Optional[bool] = False,
+    ):
+        super().__init__(
+            quantizer=quantizer
+            or CoreMLQuantizer(
+                quantization_config or _get_static_int8_linear_qconfig()
+            ),
+            calibrate=calibrate,
+            calibration_samples=calibration_samples,
+            is_qat=is_qat,
+        )
+
+
 class Partition(BaseStages.Partition):
-    def __init__(self, partitioner: Optional[Partitioner] = None):
+    def __init__(
+        self,
+        partitioner: Optional[Partitioner] = None,
+        minimum_deployment_target: Optional[Any] = ct.target.iOS15,
+    ):
         super().__init__(
-            partitioner=partitioner or CoreMLPartitioner,
+            partitioner=partitioner
+            or _create_default_partitioner(minimum_deployment_target),
         )
 
 
@@ -29,9 +79,12 @@ def __init__(
         self,
         partitioners: Optional[List[Partitioner]] = None,
         edge_compile_config: Optional[EdgeCompileConfig] = None,
+        minimum_deployment_target: Optional[Any] = ct.target.iOS15,
     ):
         super().__init__(
-            default_partitioner_cls=CoreMLPartitioner,
+            default_partitioner_cls=lambda: _create_default_partitioner(
+                minimum_deployment_target
+            ),
             partitioners=partitioners,
             edge_compile_config=edge_compile_config,
         )
@@ -43,13 +96,20 @@ def __init__(
         module: torch.nn.Module,
         example_inputs: Tuple[torch.Tensor],
         dynamic_shapes: Optional[Tuple[Any]] = None,
+        minimum_deployment_target: Optional[Any] = ct.target.iOS15,
     ):
         # Specialize for XNNPACK
         stage_classes = (
             executorch.backends.test.harness.Tester.default_stage_classes()
             | {
-                StageType.PARTITION: Partition,
-                StageType.TO_EDGE_TRANSFORM_AND_LOWER: ToEdgeTransformAndLower,
+                StageType.QUANTIZE: Quantize,
+                StageType.PARTITION: functools.partial(
+                    Partition, minimum_deployment_target=minimum_deployment_target
+                ),
+                StageType.TO_EDGE_TRANSFORM_AND_LOWER: functools.partial(
+                    ToEdgeTransformAndLower,
+                    minimum_deployment_target=minimum_deployment_target,
+                ),
             }
         )
 

backends/apple/mps/CMakeLists.txt

@@ -70,7 +70,7 @@ target_link_libraries(
           ${MPS_FRAMEWORK} ${MPS_GRAPH_FRAMEWORK}
 )
 
-target_link_options_shared_lib(mpsdelegate)
+executorch_target_link_options_shared_lib(mpsdelegate)
 target_compile_options(mpsdelegate PUBLIC ${_common_compile_options})
 target_compile_options(mpsdelegate PRIVATE "-fno-objc-arc")
 

backends/cadence/CMakeLists.txt

@@ -35,8 +35,8 @@ if(EXECUTORCH_CADENCE_CPU_RUNNER)
   # Find prebuilt libraries. executorch package should contain portable_ops_lib,
   # etdump, bundled_program.
   find_package(executorch CONFIG REQUIRED)
-  target_link_options_shared_lib(executorch)
-  target_link_options_shared_lib(portable_ops_lib)
+  executorch_target_link_options_shared_lib(executorch)
+  executorch_target_link_options_shared_lib(portable_ops_lib)
 
   target_include_directories(executorch INTERFACE ${_common_include_directories})
 

backends/cadence/aot/export_example.py

@@ -52,6 +52,13 @@ def export_model(
     # Prepare the model
     prepared_gm = prepare_pt2(model, example_inputs, quantizer)
 
+    # Calibrate the model
+    for samples in [example_inputs]:
+        prepared_gm(*samples)
+
+    # Prepare the model
+    prepared_gm = prepare_pt2(ep, quantizer)
+
     # Calibrate the model
     for samples in [example_inputs]:
         prepared_gm(*samples)