2024-04-02 nightly release (c731673)

.ci/scripts/build_llama_android.sh

@@ -0,0 +1,51 @@
+#!/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 -exu
+
+# shellcheck source=/dev/null
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+install_executorch_and_backend_lib() {
+  echo "Installing executorch and xnnpack backend"
+  rm -rf cmake-android-out && mkdir cmake-android-out
+  ANDROID_NDK=/opt/ndk
+  BUCK2=buck2
+  ANDROID_ABI=arm64-v8a
+  cmake -DBUCK2="${BUCK2}" \
+    -DCMAKE_TOOLCHAIN_FILE="${ANDROID_NDK}/build/cmake/android.toolchain.cmake" \
+    -DANDROID_ABI="${ANDROID_ABI}" \
+    -DANDROID_PLATFORM=android-23 \
+    -DCMAKE_INSTALL_PREFIX=cmake-android-out \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
+    -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
+    -DEXECUTORCH_BUILD_XNNPACK=ON \
+    -DEXECUTORCH_BUILD_OPTIMIZED=ON \
+    -DXNNPACK_ENABLE_ARM_BF16=OFF \
+    -Bcmake-android-out .
+
+  cmake --build cmake-android-out -j4 --target install --config Release
+}
+
+build_llama_runner() {
+    echo "Building llama runner for Android..."
+    ANDROID_ABI=arm64-v8a
+    cmake -DBUCK2="${BUCK2}" \
+    -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK"/build/cmake/android.toolchain.cmake  \
+    -DANDROID_ABI="${ANDROID_ABI}" \
+    -DANDROID_PLATFORM=android-23 \
+    -DCMAKE_INSTALL_PREFIX=cmake-android-out \
+    -DCMAKE_BUILD_TYPE=Release -DPYTHON_EXECUTABLE=python \
+    -DEXECUTORCH_BUILD_OPTIMIZED=ON \
+    -Bcmake-android-out/examples/models/llama2 examples/models/llama2
+
+    cmake --build cmake-android-out/examples/models/llama2 -j4 --config Release
+}
+install_flatc_from_source
+install_executorch_and_backend_lib
+build_llama_runner

.ci/scripts/gather_test_models.py

@@ -13,7 +13,6 @@
 from examples.models import MODEL_NAME_TO_MODEL
 from examples.xnnpack import MODEL_NAME_TO_OPTIONS
 
-
 DEFAULT_RUNNERS = {
     "linux": "linux.2xlarge",
     "macos": "macos-m1-stable",
@@ -24,6 +23,7 @@
         "w2l": "linux.12xlarge",
         "ic4": "linux.12xlarge",
         "resnet50": "linux.12xlarge",
+        "llava_encoder": "linux.4xlarge",
         # This one causes timeout on smaller runner, the root cause is unclear (T161064121)
         "dl3": "linux.12xlarge",
         "emformer_join": "linux.12xlarge",
@@ -83,7 +83,17 @@ def model_should_run_on_event(model: str, event: str) -> bool:
     We put higher priority and fast models to pull request and rest to push.
     """
     if event == "pull_request":
-        return model in ["add", "ic3", "mv2", "mv3", "resnet18", "vit"]
+        return model in ["add", "ic3", "mv2", "mv3", "resnet18", "vit", "llava_encoder"]
+    return True
+
+
+def model_should_run_on_target_os(model: str, target_os: str) -> bool:
+    """
+    A helper function to decide whether a model should be tested on a target os (linux/macos).
+    For example, a big model can be disabled in macos due to the limited macos resources.
+    """
+    if target_os == "macos":
+        return model not in ["llava_encoder"]
     return True
 
 
@@ -119,6 +129,9 @@ def export_models_for_ci() -> dict[str, dict]:
         if not model_should_run_on_event(name, event):
             continue
 
+        if not model_should_run_on_target_os(name, target_os):
+            continue
+
         if backend == "xnnpack":
             if name not in MODEL_NAME_TO_OPTIONS:
                 continue

.ci/scripts/test.sh

@@ -67,6 +67,10 @@ test_model() {
     run_portable_executor_runner
     rm "./${MODEL_NAME}.pte"
   fi
+  if [[ "${MODEL_NAME}" == "llava_encoder" ]]; then
+    # Install requirements for llava
+    bash examples/models/llava_encoder/install_requirements.sh
+  fi
   # python3 -m examples.portable.scripts.export --model_name="llama2" should works too
   "${PYTHON_EXECUTABLE}" -m examples.portable.scripts.export --model_name="${MODEL_NAME}"
   run_portable_executor_runner

.github/workflows/android.yml

@@ -1,4 +1,4 @@
-name: Build ExecuTorch Android demo apps
+name: Android
 
 on:
   push:

.github/workflows/apple.yml

@@ -68,6 +68,7 @@ jobs:
         FRAMEWORKS=(
           "executorch"
           "coreml_backend"
+          "custom_backend"
           "mps_backend"
           "optimized_backend"
           "portable_backend"
@@ -91,20 +92,12 @@ jobs:
 
         # Build Release iOS Frameworks
         PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
-        build/build_apple_frameworks.sh --coreml --mps --optimized --portable --quantized --xnnpack
+        build/build_apple_frameworks.sh --coreml --custom --mps --optimized --portable --quantized --xnnpack
 
         # Bundle Release iOS Frameworks
         for FRAMEWORK in "${FRAMEWORKS[@]}"; do (
-          cd cmake-out && zip -r "${RUNNER_TEMP}/artifacts/${FRAMEWORK}-${VERSION}.zip" "${FRAMEWORK}.xcframework"
-        ) done
-
-        # Build Debug iOS Frameworks
-        PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
-        build/build_apple_frameworks.sh --coreml --mps --optimized --portable --quantized --xnnpack --Debug
-
-        # Bundle Debug iOS Frameworks
-        for FRAMEWORK in "${FRAMEWORKS[@]}"; do (
-          cd cmake-out && zip -r "${RUNNER_TEMP}/artifacts/${FRAMEWORK}_debug-${VERSION}.zip" "${FRAMEWORK}_debug.xcframework"
+          cd cmake-out && \
+          zip -r "${RUNNER_TEMP}/artifacts/${FRAMEWORK}-${VERSION}.zip" "${FRAMEWORK}.xcframework"
         ) done
 
         popd

.github/workflows/pull.yml

@@ -114,6 +114,29 @@ jobs:
         # Test llama2
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh stories110M.pt "${BUILD_TOOL}" "${DTYPE}" "${MODE}"
 
+  test-llama-runner-linux-android:
+    name: test-llama-runner-linux-android
+    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    strategy:
+      matrix:
+        include:
+          - build-tool: cmake
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-clang12-android
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        BUILD_TOOL=${{ matrix.build-tool }}
+        PYTHON_EXECUTABLE=python \
+        bash .ci/scripts/build_llama_android.sh  "${BUILD_TOOL}"
+
   test-custom-ops-linux:
     name: test-custom-ops-linux
     uses: pytorch/test-infra/.github/workflows/linux_job.yml@main

.github/workflows/trunk.yml

@@ -159,6 +159,10 @@ jobs:
         git config --global user.name "Github Executorch"
         bash examples/arm/setup.sh --i-agree-to-the-contained-eula
 
+        # Increase number of files user can monitor to bypass buck failures.
+        # Hopefully this is high enough for this setup.
+        sudo sysctl fs.inotify.max_user_watches=1048576 # 1024 * 1024
+
         # Test ethos-u delegate examples with run.sh
         source examples/arm/ethos-u-scratch/setup_path.sh
         PYTHON_EXECUTABLE=python bash examples/arm/run.sh examples/arm/ethos-u-scratch/ buck2

.github/workflows/update-viablestrict.yml

@@ -20,6 +20,6 @@ jobs:
         with:
           repository: pytorch/executorch
           stable-branch: viable/strict
-          requires: '[\"pull\", \"lint\", \"trunk\", \"Build documentation\"]'
+          requires: '[\"pull\", \"lint\", \"trunk\", \"Build documentation\", \"Android\", \"Apple\"]'
           secret-bot-token: ${{ secrets.UPDATEBOT_TOKEN }}
           rockset-api-key: ${{ secrets.ROCKSET_API_KEY }}

.gitignore

@@ -5,12 +5,8 @@ cmake-android-out/
 cmake-ios-out/
 ethos-u-scratch/
 executorch.egg-info
+pip-out/
 __pycache__/
-build/lib/
-exir/_serialize/scalar_type.fbs
-exir/_serialize/program.fbs
-sdk/bundled_program/serialize/bundled_program_schema.fbs
-sdk/bundled_program/serialize/scalar_type.fbs
 
 # Any exported models and profiling outputs
 *.pte

.gitmodules

@@ -59,3 +59,9 @@
 	path = third-party/lm-evaluation-harness
 	url = https://github.com/EleutherAI/lm-evaluation-harness
 	branch = v0.4.1
+[submodule "kernels/optimized/third-party/eigen"]
+	path = kernels/optimized/third-party/eigen
+	url = https://gitlab.com/libeigen/eigen.git
+[submodule "examples/third-party/LLaVA"]
+	path = examples/third-party/LLaVA
+	url = https://github.com/haotian-liu/LLaVA.git

CMakeLists.txt

@@ -13,6 +13,16 @@
 # cloning or pulling the upstream repo. Once this is done, you don't need to do
 # it again until you pull from the upstream repo again.
 #
+# NOTE: Build options can be configured by passing arguments to cmake. For
+# example, to enable the EXECUTORCH_BUILD_XNNPACK option, change the cmake
+# command to 'cmake -DEXECUTORCH_BUILD_XNNPACK=ON ..'.
+#[[
+  (rm -rf cmake-out \
+    && mkdir cmake-out \
+    && cd cmake-out \
+    && cmake ..)
+]]
+#
 # ### Build ###
 #
 # NOTE: The `-j` argument specifies how many jobs/processes to use when
@@ -45,6 +55,7 @@ endif()
 if(NOT CMAKE_BUILD_TYPE)
   set(CMAKE_BUILD_TYPE Debug)
 endif()
+
 # ------------------------------ OPTIONS -------------------------------------
 # WARNING: Please don't add example specific options in this CMakeLists.txt.
 # Instead please use `find_package(executorch REQUIRED)` in the example
@@ -131,6 +142,8 @@ option(EXECUTORCH_BUILD_ARM_BAREMETAL
 
 option(EXECUTORCH_BUILD_COREML "Build the Core ML backend" OFF)
 
+option(EXECUTORCH_BUILD_CUSTOM "Build the custom kernels" OFF)
+
 option(EXECUTORCH_BUILD_EXTENSION_DATA_LOADER "Build the Data Loader extension"
        OFF)
 
@@ -159,6 +172,61 @@ option(EXECUTORCH_BUILD_XNNPACK "Build the XNNPACK backend" OFF)
 
 option(EXECUTORCH_BUILD_VULKAN "Build the Vulkan backend" OFF)
 
+#
+# pthreadpool: build pthreadpool library. Disable on unsupported platforms
+#
+cmake_dependent_option(EXECUTORCH_BUILD_PTHREADPOOL "Build pthreadpool library."
+                       ON "NOT EXECUTORCH_BUILD_ARM_BAREMETAL" OFF)
+
+#
+# cpuinfo: build cpuinfo library. Disable on unsupported platforms
+#
+cmake_dependent_option(EXECUTORCH_BUILD_CPUINFO "Build cpuinfo library." ON
+                       "NOT EXECUTORCH_BUILD_ARM_BAREMETAL" OFF)
+
+if(EXECUTORCH_BUILD_CPUINFO)
+  # --- cpuinfo
+  set(CPUINFO_SOURCE_DIR "backends/xnnpack/third-party/cpuinfo")
+  set(CPUINFO_BUILD_TOOLS
+      OFF
+      CACHE BOOL "")
+  set(CPUINFO_BUILD_UNIT_TESTS
+      OFF
+      CACHE BOOL "")
+  set(CPUINFO_BUILD_MOCK_TESTS
+      OFF
+      CACHE BOOL "")
+  set(CPUINFO_BUILD_BENCHMARKS
+      OFF
+      CACHE BOOL "")
+  set(CPUINFO_LIBRARY_TYPE
+      "static"
+      CACHE STRING "")
+  set(CPUINFO_LOG_LEVEL
+      "error"
+      CACHE STRING "")
+  set(CLOG_SOURCE_DIR "${CPUINFO_SOURCE_DIR}/deps/clog")
+  add_subdirectory("${CPUINFO_SOURCE_DIR}")
+endif()
+
+if(EXECUTORCH_BUILD_PTHREADPOOL)
+  # --- pthreadpool
+  set(PTHREADPOOL_SOURCE_DIR "backends/xnnpack/third-party/pthreadpool")
+  set(PTHREADPOOL_BUILD_TESTS
+      OFF
+      CACHE BOOL "")
+  set(PTHREADPOOL_BUILD_BENCHMARKS
+      OFF
+      CACHE BOOL "")
+  set(PTHREADPOOL_LIBRARY_TYPE
+      "static"
+      CACHE STRING "")
+  set(PTHREADPOOL_ALLOW_DEPRECATED_API
+      ON
+      CACHE BOOL "")
+  add_subdirectory("${PTHREADPOOL_SOURCE_DIR}")
+endif()
+
 if(NOT PYTHON_EXECUTABLE)
   resolve_python_executable()
 endif()
@@ -305,6 +373,12 @@ endif()
 #
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/kernels/portable)
 
+if(EXECUTORCH_BUILD_CUSTOM)
+  # TODO: move all custom kernels to ${CMAKE_CURRENT_SOURCE_DIR}/kernels/custom
+  add_subdirectory(
+    ${CMAKE_CURRENT_SOURCE_DIR}/examples/models/llama2/custom_ops)
+endif()
+
 if(EXECUTORCH_BUILD_OPTIMIZED)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/kernels/optimized)
 endif()
@@ -398,11 +472,6 @@ if(EXECUTORCH_BUILD_VULKAN)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/vulkan)
 endif()
 
-if(EXECUTORCH_BUILD_ANDROID_JNI)
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/examples/models/llama2/runner)
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/android)
-endif()
-
 if(EXECUTORCH_BUILD_QNN)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/qualcomm)
 endif()
@@ -445,7 +514,7 @@ if(EXECUTORCH_BUILD_PYBIND)
   # find pytorch lib, to allow pybind to take at::Tensor as input/output
   find_package(Torch CONFIG REQUIRED)
   find_library(TORCH_PYTHON_LIBRARY torch_python
-              PATHS "${TORCH_INSTALL_PREFIX}/lib")
+               PATHS "${TORCH_INSTALL_PREFIX}/lib")
 
   # compile options for pybind
 

Package.swift

@@ -11,13 +11,14 @@ import PackageDescription
 
 let url = "https://ossci-ios.s3.amazonaws.com/executorch"
 let version = "0.1.0"
-let coreml_sha256 = "e8c5000a389bdc98274aa0b359350a47e6d0cccb8af5efc46f814feac6afaf86"
-let executorch_sha256 = "e6c5d798b614a03ab8a4891caeaa8a7adf8d58ba29e767079321691ec9f1ffb4"
-let mps_sha256 = "3e54e3166b5e739cb3f76b2bc6f7b1982a0401821ab785a93120bacfde4bc1ee"
-let optimized_sha256 = "4d353f44badd321cf29fe548db9d66b493b93c6233a7e023988e256f0eefeaa1"
-let portable_sha256 = "c501f9b644a3e8a7bab62600b7802e4a9752fb789ba4fd02f46bec47858cec07"
-let quantized_sha256 = "4fb5f7216abc0ee16ece91a4bce822b06d67b52ca985c9eecbf9d3f8bd1ea1ba"
-let xnnpack_sha256 = "e610904cfd6e96f8f738c25a7bb4f6d7b86995b2cfeb72fc1f30523630dbb285"
+let coreml_sha256 = "78d853d87be478696e56e658aa4ff17d47ae185a9a6a36316c821fa8b2d3aacd"
+let custom_sha256 = "f059f6716298403dff89a952a70e323c54911be140d05f2467bd5cc61aaefae3"
+let executorch_sha256 = "ba9a0c2b061afaedbc3c5454040a598b1371170bd9d9a30b7163c20e23339841"
+let mps_sha256 = "39542a8671cca1aa627102aa47785d0f6e2dfe9a40e2c22288a755057b00fbfa"
+let optimized_sha256 = "1d84fa16197bb6f0dec01aaa29d2a140c0e14d8e5e92630a7b4dd6f48012506d"
+let portable_sha256 = "4993904f89ecb4476677ff3c072ed1a314a608170f10d364cfd23947851ccbf3"
+let quantized_sha256 = "8d35ee0e7ca77c19782eaea07a1888f576cda679f8a4a5edb03d80ebe858047e"
+let xnnpack_sha256 = "380e5185c4c48ede7cc0d0f0657ffb26df83cd9f55813d78593aea8a93942caf"
 
 struct Framework {
   let name: String
@@ -57,6 +58,10 @@ let frameworks = [
       "sqlite3",
     ]
   ),
+  Framework(
+    name: "custom_backend",
+    checksum: custom_sha256
+  ),
   Framework(
     name: "executorch",
     checksum: executorch_sha256