Merge branch 'main' into int16-tables

Author: zingo

.ci/docker/build.sh

@@ -29,6 +29,10 @@ case "${IMAGE_NAME}" in
     LINTRUNNER=""
     CLANG_VERSION=12
     ;;
+  executorch-ubuntu-22.04-gcc11-aarch64)
+    LINTRUNNER=""
+    GCC_VERSION=11
+    ;;
   executorch-ubuntu-22.04-linter)
     LINTRUNNER=yes
     CLANG_VERSION=12
@@ -44,6 +48,7 @@ case "${IMAGE_NAME}" in
   executorch-ubuntu-22.04-mediatek-sdk)
     MEDIATEK_SDK=yes
     CLANG_VERSION=12
+    ANDROID_NDK_VERSION=r27b
     ;;
   executorch-ubuntu-22.04-clang12-android)
     LINTRUNNER=""

.ci/docker/common/install_conda.sh

@@ -13,6 +13,9 @@ source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 install_miniconda() {
   BASE_URL="https://repo.anaconda.com/miniconda"
   CONDA_FILE="Miniconda3-py${PYTHON_VERSION//./}_${MINICONDA_VERSION}-Linux-x86_64.sh"
+  if [[ $(uname -m) == "aarch64" ]]; then 
+    CONDA_FILE="Miniconda3-py${PYTHON_VERSION//./}_${MINICONDA_VERSION}-Linux-aarch64.sh"
+  fi
 
   mkdir -p /opt/conda
   chown ci-user:ci-user /opt/conda
@@ -36,7 +39,7 @@ install_python() {
 
   # From https://github.com/pytorch/pytorch/blob/main/.ci/docker/common/install_conda.sh
   if [[ $(uname -m) == "aarch64" ]]; then
-    conda_install "openblas==0.3.28=*openmp*"
+    conda_install "openblas==0.3.29=*openmp*" -c conda-forge
   else
     conda_install mkl=2022.1.0 mkl-include=2022.1.0
   fi

.ci/scripts/build_android_instrumentation.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+# 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.
+
+set -ex
+
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+  PYTHON_EXECUTABLE=python3
+fi
+which "${PYTHON_EXECUTABLE}"
+
+build_android_test() {
+  pushd extension/android_test
+  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew testDebugUnitTest
+  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew build assembleAndroidTest
+  popd
+}
+
+collect_artifacts_to_be_uploaded() {
+  ARTIFACTS_DIR_NAME="$1"
+  # Collect Java library test
+  JAVA_LIBRARY_TEST_DIR="${ARTIFACTS_DIR_NAME}/library_test_dir"
+  mkdir -p "${JAVA_LIBRARY_TEST_DIR}"
+  cp extension/android_test/build/outputs/apk/debug/*.apk "${JAVA_LIBRARY_TEST_DIR}"
+  cp extension/android_test/build/outputs/apk/androidTest/debug/*.apk "${JAVA_LIBRARY_TEST_DIR}"
+}
+
+main() {
+  build_android_test
+  if [ -n "$ARTIFACTS_DIR_NAME" ]; then
+    collect_artifacts_to_be_uploaded ${ARTIFACTS_DIR_NAME}
+  fi
+}
+
+if [[ "${BASH_SOURCE[0]}" == "${0}" ]]; then
+  main "$@"
+fi

.ci/scripts/build_llama_android.sh

@@ -60,6 +60,5 @@ build_llama_runner() {
 
     cmake --build cmake-android-out/examples/models/llama -j4 --config Release
 }
-install_flatc_from_source
 install_executorch_and_backend_lib
 build_llama_runner

.ci/scripts/gather_benchmark_configs.py

@@ -263,7 +263,8 @@ def is_valid_huggingface_model_id(model_name: str) -> bool:
 def get_benchmark_configs() -> Dict[str, Dict]:  # noqa: C901
     """
     Gather benchmark configurations for a given set of models on the target operating system and devices.
-
+    CHANGE IF this function's return changed:
+        extract_model_info() in executorch/.github/scripts/extract_benchmark_results.py IF YOU CHANGE THE RESULT OF THIS FUNCTION.
     Args:
         None
 

.ci/scripts/gather_test_models.py

@@ -5,6 +5,9 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# WARNING: The CI runner logic should directly be in the corresponding yml files
+# This file will be deleted once the reference in periodic.yml is deleted.
+
 import itertools
 import json
 import os

build/test_ios_ci.sh renamed to .ci/scripts/test_ios_ci.sh

@@ -209,9 +209,14 @@ test_model_with_qnn() {
   EXPORTED_MODEL=$(find "./${EXPORT_SCRIPT}" -type f -name "${MODEL_NAME}*.pte" -print -quit)
 }
 
+# Run CoreML tests.
+#
+# @param should_test If true, build and test the model using the coreml_executor_runner.
 test_model_with_coreml() {
-  if [[ "${BUILD_TOOL}" == "buck2" ]]; then
-    echo "coreml doesn't support buck2."
+  local should_test="$1"
+
+  if [[ "${BUILD_TOOL}" != "cmake" ]]; then
+    echo "coreml only supports cmake."
     exit 1
   fi
 
@@ -229,6 +234,14 @@ test_model_with_coreml() {
     echo "No .pte file found"
     exit 1
   fi
+
+  # Run the model
+  if [ "${should_test}" = true ]; then
+    echo "Testing exported model with coreml_executor_runner..."
+    local out_dir=$(mktemp -d)
+    COREML_EXECUTOR_RUNNER_OUT_DIR="${out_dir}" examples/apple/coreml/scripts/build_executor_runner.sh
+    "${out_dir}/coreml_executor_runner" --model_path "${EXPORTED_MODEL}"
+  fi
 }
 
 test_model_with_mps() {
@@ -247,7 +260,11 @@ elif [[ "${BACKEND}" == *"qnn"* ]]; then
   fi
 elif [[ "${BACKEND}" == *"coreml"* ]]; then
   echo "Testing ${MODEL_NAME} with coreml..."
-  test_model_with_coreml
+  should_test_coreml=false
+  if [[ "${BACKEND}" == *"test"* ]]; then
+    should_test_coreml=true
+  fi
+  test_model_with_coreml "${should_test_coreml}"
   if [[ $? -eq 0 ]]; then
     prepare_artifacts_upload
   fi

.ci/scripts/test_model.sh

@@ -8,9 +8,7 @@ set -eux
 
 # TODO: expand this to //...
 # TODO: can't query cadence & vulkan backends
-# TODO: can't query //kernels/prim_ops because of a cpp_unittest and
-# broken code in shim to read oss.folly_cxx_tests. Sending fix but it
-# needs to propagate and we need a submodule update.
+# TODO: can't query //kernels/prim_ops because of non-buckified stuff in OSS.
 buck2 query "//backends/apple/... + //backends/example/... + \
 //backends/mediatek/... + //backends/test/... + //backends/transforms/... + \
 //backends/xnnpack/... + //configurations/... + //kernels/aten/... + \
@@ -19,8 +17,13 @@ buck2 query "//backends/apple/... + //backends/example/... + \
 
 UNBUILDABLE_OPTIMIZED_OPS_REGEX="gelu|fft_r2c|log_softmax"
 BUILDABLE_OPTIMIZED_OPS=$(buck2 query //kernels/optimized/cpu/... | grep -E -v $UNBUILDABLE_OPTIMIZED_OPS_REGEX)
+
+# TODO: build prim_ops_test_cpp again once supported_features works in
+# OSS buck.
+BUILDABLE_KERNELS_PRIM_OPS_TARGETS=$(buck2 query //kernels/prim_ops/... | grep -v prim_ops_test)
 # TODO: expand the covered scope of Buck targets.
 # //runtime/kernel/... is failing because //third-party:torchgen_files's shell script can't find python on PATH.
 # //runtime/test/... requires Python torch, which we don't have in our OSS buck setup.
-buck2 test $BUILDABLE_OPTIMIZED_OPS //kernels/portable/... //runtime/backend/... //runtime/core/... \
+buck2 test $BUILDABLE_OPTIMIZED_OPS //kernels/portable/... \
+      $BUILDABLE_KERNELS_PRIM_OPS_TARGETS //runtime/backend/... //runtime/core/... \
       //runtime/executor: //runtime/kernel/... //runtime/platform/...

.ci/scripts/unittest-buck2.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+# 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.
+set -eux
+
+buck2 test //extension/apple:ExecuTorch

.ci/scripts/unittest-macos-buck2.sh

@@ -35,6 +35,7 @@ if [[ "$BUILD_TOOL" == "cmake" ]]; then
     .ci/scripts/unittest-macos-cmake.sh
 elif [[ "$BUILD_TOOL" == "buck2" ]]; then
     .ci/scripts/unittest-buck2.sh
+    # .ci/scripts/unittest-macos-buck2.sh
 else
     echo "Unknown build tool $BUILD_TOOL"
     exit 1

.ci/scripts/unittest-macos.sh

@@ -20,6 +20,14 @@ clean_executorch_install_folders() {
   ./install_executorch.sh --clean
 }
 
+update_tokenizers_git_submodule() {
+  echo "Updating tokenizers git submodule..."
+  git submodule update --init
+  pushd extension/llm/tokenizers
+  git submodule update --init
+  popd
+}
+
 install_executorch() {
   which pip
   # Install executorch, this assumes that Executorch is checked out in the
@@ -72,25 +80,6 @@ install_pytorch_and_domains() {
   sccache --show-stats || true
 }
 
-install_flatc_from_source() {
-  # NB: This function could be used to install flatbuffer from source
-  pushd third-party/flatbuffers || return
-
-  cmake -G "Unix Makefiles" -DCMAKE_BUILD_TYPE=Release
-  if [ "$(uname)" == "Darwin" ]; then
-    CMAKE_JOBS=$(( $(sysctl -n hw.ncpu) - 1 ))
-  else
-    CMAKE_JOBS=$(( $(nproc) - 1 ))
-  fi
-  cmake --build . -j "${CMAKE_JOBS}"
-
-  # Copy the flatc binary to conda path
-  EXEC_PATH=$(dirname "$(which python)")
-  cp flatc "${EXEC_PATH}"
-
-  popd || return
-}
-
 build_executorch_runner_buck2() {
   # Build executorch runtime with retry as this step is flaky on macos CI
   retry buck2 build //examples/portable/executor_runner:executor_runner
@@ -103,9 +92,14 @@ build_executorch_runner_cmake() {
   mkdir "${CMAKE_OUTPUT_DIR}"
 
   pushd "${CMAKE_OUTPUT_DIR}" || return
+  if [[ $1 == "Debug" ]]; then
+      CXXFLAGS="-fsanitize=address,undefined"
+  else
+      CXXFLAGS=""
+  fi
   # This command uses buck2 to gather source files and buck2 could crash flakily
   # on MacOS
-  retry cmake -DPYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" -DCMAKE_BUILD_TYPE="${1:-Release}" ..
+  CXXFLAGS="$CXXFLAGS" retry cmake -DPYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" -DCMAKE_BUILD_TYPE="${1:-Release}" ..
   popd || return
 
   if [ "$(uname)" == "Darwin" ]; then

.ci/scripts/utils.sh

@@ -0,0 +1,11 @@
+# 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.
+
+# This file is sourced into the environment before building a pip wheel. It
+# should typically only contain shell variable assignments. Be sure to export
+# any variables so that subprocesses will see them.
+
+source "${GITHUB_WORKSPACE}/${REPOSITORY}/.ci/scripts/wheel/envvar_base.sh"

.ci/scripts/wheel/__init__.py

@@ -8,16 +8,7 @@
 # should typically only contain shell variable assignments. Be sure to export
 # any variables so that subprocesses will see them.
 
-# Enable pybindings so that users can execute ExecuTorch programs from python.
-export EXECUTORCH_BUILD_PYBIND=1
-
-# Ensure that CMAKE_ARGS is defined before referencing it. Defaults to empty
-# if not defined.
-export CMAKE_ARGS="${CMAKE_ARGS:-}"
-
-# Link the XNNPACK backend into the pybindings runtime so that users can execute
-# ExecuTorch programs that delegate to it.
-CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_XNNPACK=ON"
+source "${GITHUB_WORKSPACE}/${REPOSITORY}/.ci/scripts/wheel/envvar_base.sh"
 
 # When building for macOS, link additional backends into the pybindings runtime.
 CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_COREML=ON"

build/packaging/env_var_script_linux.sh renamed to .ci/scripts/wheel/envvar_base.sh

@@ -13,14 +13,5 @@ set -euxo pipefail
 # not install them. TODO(dbort): Switch to using `python -m build --wheel`,
 # which does install them. Though we'd need to disable build isolation to be
 # able to see the installed torch package.
-readonly BUILD_DEPS=(
-  # This list must match the build-system.requires list from pyproject.toml.
-  "cmake"
-  "pip>=23"
-  "pyyaml"
-  "setuptools>=63"
-  "tomli"
-  "wheel"
-  "zstd"
-)
-pip install --progress-bar off "${BUILD_DEPS[@]}"
+
+"${GITHUB_WORKSPACE}/${REPOSITORY}/install_requirements.sh"

.ci/scripts/wheel/envvar_linux.sh

@@ -0,0 +1,66 @@
+# 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.
+
+import os
+import subprocess
+import sys
+from dataclasses import dataclass
+from functools import cache
+from typing import List
+
+
+@cache
+def _unsafe_get_env(key: str) -> str:
+    value = os.getenv(key)
+    if value is None:
+        raise RuntimeError(f"environment variable '{key}' is not set")
+    return value
+
+
+@cache
+def _repository_root_dir() -> str:
+    return os.path.join(
+        _unsafe_get_env("GITHUB_WORKSPACE"),
+        _unsafe_get_env("REPOSITORY"),
+    )
+
+
+# For some reason, we are unable to see the entire repo in the python path.
+# So manually add it.
+sys.path.append(_repository_root_dir())
+from examples.models import Backend, Model
+
+
+@dataclass
+class ModelTest:
+    model: Model
+    backend: Backend
+
+
+def run_tests(model_tests: List[ModelTest]) -> None:
+    # Why are we doing this envvar shenanigans? Since we build the testers, which
+    # uses buck, we cannot run as root. This is a sneaky of getting around that
+    # test.
+    #
+    # This can be reverted if either:
+    #   - We remove usage of buck in our builds
+    #   - We stop running the Docker image as root: https://github.com/pytorch/test-infra/issues/5091
+    envvars = os.environ.copy()
+    envvars.pop("HOME")
+
+    for model_test in model_tests:
+        subprocess.run(
+            [
+                os.path.join(_repository_root_dir(), ".ci/scripts/test_model.sh"),
+                str(model_test.model),
+                # What to build `executor_runner` with for testing.
+                "cmake",
+                str(model_test.backend),
+            ],
+            env=envvars,
+            check=True,
+            cwd=_repository_root_dir(),
+        )

build/packaging/env_var_script_m1.sh renamed to .ci/scripts/wheel/envvar_macos.sh

@@ -0,0 +1,19 @@
+#!/usr/bin/env python
+# 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.
+
+import test_base
+from examples.models import Backend, Model
+
+if __name__ == "__main__":
+    test_base.run_tests(
+        model_tests=[
+            test_base.ModelTest(
+                model=Model.Mv3,
+                backend=Backend.XnnpackQuantizationDelegation,
+            )
+        ]
+    )