Update on "Serialize NamedDataStoreOutput into PTD."

Update PTD serialization to account for blobs from the NamedDataStoreOutput.

Something we can do in the future is to consolidate tensors (that go through the emitter) and blobs (that come from the NamedDataStore).

Differential Revision: [D70939807](https://our.internmc.facebook.com/intern/diff/D70939807/)

[ghstack-poisoned]

Author: lucylq

.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/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-08434df1f2f88c9770e59246caa2ff9c6f613270
+295f2ed4d103017f7e19a7b8263ece606cd629db

.ci/docker/common/install_android.sh

@@ -70,6 +70,7 @@ install_sdk() {
   # These are the tools needed to build Android apps
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "platforms;android-34"
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "build-tools;33.0.1"
+  yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "build-tools;35.0.0"
   # And some more tools for future emulator tests
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "platform-tools"
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "tools"

.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/docker/conda-env-ci.txt

@@ -1,4 +1,4 @@
-cmake=3.22.1
+cmake=3.26.4
 ninja=1.10.2
 libuv
 llvm-openmp

.ci/scripts/build_android_instrumentation.sh

@@ -0,0 +1,41 @@
+#!/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() {
+  mkdir -p extension/android/executorch_android/src/androidTest/resources
+  cp extension/module/test/resources/add.pte extension/android/executorch_android/src/androidTest/resources
+  pushd extension/android
+  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:testDebugUnitTest
+  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android: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 "${JAVA_LIBRARY_TEST_DIR}"
+  cp extension/android/executorch_android/build/outputs/apk/androidTest/debug/*.apk "${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

@@ -0,0 +1,85 @@
+#!/bin/bash
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# 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 -exu
+
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+export EXECUTORCH_ROOT="$(dirname "${BASH_SOURCE[0]}")/../.."
+
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+  PYTHON_EXECUTABLE=python3
+fi
+
+which "${PYTHON_EXECUTABLE}"
+
+# Update tokenizers submodule
+pushd $EXECUTORCH_ROOT/extension/llm/tokenizers
+echo "Update tokenizers submodule"
+git submodule update --init
+popd
+
+# Install ET with CMake
+cmake -DPYTHON_EXECUTABLE=python \
+    -DCMAKE_INSTALL_PREFIX=cmake-out \
+    -DEXECUTORCH_ENABLE_LOGGING=1 \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
+    -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
+    -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
+    -DEXECUTORCH_BUILD_XNNPACK=OFF \
+    -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
+    -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
+    -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
+    -Bcmake-out .
+cmake --build cmake-out -j16 --target install --config Release
+
+# Install llama runner with torchao
+cmake -DPYTHON_EXECUTABLE=python \
+    -DCMAKE_PREFIX_PATH=$(python -c 'from distutils.sysconfig import get_python_lib; print(get_python_lib())') \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
+    -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
+    -DEXECUTORCH_BUILD_XNNPACK=OFF \
+    -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
+    -DEXECUTORCH_BUILD_TORCHAO=ON \
+    -Bcmake-out/examples/models/llama \
+    examples/models/llama
+cmake --build cmake-out/examples/models/llama -j16 --config Release
+
+# Download stories llama110m artifacts
+download_stories_model_artifacts
+
+echo "Creating tokenizer.bin"
+$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+
+# Export model
+LLAMA_CHECKPOINT=stories110M.pt
+LLAMA_PARAMS=params.json
+MODEL_OUT=model.pte
+TOKENIZER=tokenizer.bin
+
+# Set low-bit quantization parameters
+QLINEAR_BITWIDTH=3 # Can be 1-8
+QLINEAR_GROUP_SIZE=128 # Must be multiple of 16
+QEMBEDDING_BITWIDTH=4 # Can be 1-8
+QEMBEDDING_GROUP_SIZE=32 # Must be multiple of 16
+
+${PYTHON_EXECUTABLE} -m examples.models.llama.export_llama \
+    --checkpoint "${LLAMA_CHECKPOINT:?}" \
+    --params "${LLAMA_PARAMS:?}" \
+    -kv \
+    --use_sdpa_with_kv_cache \
+    --output_name=${MODEL_OUT} \
+    -qmode "torchao:8da${QLINEAR_BITWIDTH}w" \
+    --group_size ${QLINEAR_GROUP_SIZE} \
+    -E "torchao:${QEMBEDDING_BITWIDTH},${QEMBEDDING_GROUP_SIZE}" \
+    --disable_dynamic_shape \
+    -d fp32
+
+# Test run
+./cmake-out/examples/models/llama/llama_main --model_path=$MODEL_OUT --tokenizer_path=$TOKENIZER --prompt="Once upon a time,"

.ci/scripts/test_llama_torchao_lowbit.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

@@ -8,17 +8,22 @@ 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/... + \
 //kernels/optimized/... + //kernels/portable/... + //kernels/quantized/... + \
 //kernels/test/... + //runtime/... + //schema/... + //test/... + //util/..."
 
+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 build //runtime/backend/... //runtime/core/... //runtime/executor: //runtime/kernel/... //runtime/platform/...
-buck2 test //runtime/backend/... //runtime/core/... //runtime/executor: //runtime/kernel/... //runtime/platform/...
+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

@@ -1,6 +1,9 @@
-#!/usr/bin/env python3
+#!/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

build/__init__.py renamed to .ci/scripts/unittest-macos-buck2.sh

@@ -19,14 +19,14 @@ export TMP_DIR=$(mktemp -d)
 export PATH="${TMP_DIR}:$PATH"
 trap 'rm -rfv ${TMP_DIR}' EXIT
 
-if [[ "$BUILD_TOOL" == "cmake" ]]; then
-    # Setup MacOS dependencies as there is no Docker support on MacOS atm
-    PYTHON_EXECUTABLE=python \
-    EXECUTORCH_BUILD_PYBIND=ON \
-    CMAKE_ARGS="-DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
-    ${CONDA_RUN} --no-capture-output \
-    .ci/scripts/setup-macos.sh "$@"
+# Setup MacOS dependencies as there is no Docker support on MacOS atm
+PYTHON_EXECUTABLE=python \
+EXECUTORCH_BUILD_PYBIND=ON \
+CMAKE_ARGS="-DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+${CONDA_RUN} --no-capture-output \
+.ci/scripts/setup-macos.sh "$@"
 
+if [[ "$BUILD_TOOL" == "cmake" ]]; then
     # Install llama3_2_vision dependencies.
     PYTHON_EXECUTABLE=python \
     ${CONDA_RUN} --no-capture-output \
@@ -35,6 +35,7 @@ if [[ "$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"