Merge branch 'main' into fix-android-perf-periodic-default-spec

Author: huydhn

.ci/scripts/build-qnn-sdk.sh

@@ -0,0 +1,19 @@
+#!/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
+
+build_qnn_backend() {
+  echo "Start building qnn backend."
+  export ANDROID_NDK_ROOT=/opt/ndk
+  export QNN_SDK_ROOT=/tmp/qnn/2.23.0.240531
+  export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
+
+  bash backends/qualcomm/scripts/build.sh --skip_aarch64 --job_number 2 --release
+}
+
+build_qnn_backend

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

@@ -0,0 +1,29 @@
+#!/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
+
+install_qnn() {
+  echo "Start installing qnn."
+  QNN_INSTALLATION_DIR=/tmp/qnn
+  mkdir -p "${QNN_INSTALLATION_DIR}"
+
+  curl -Lo /tmp/v2.23.0.24.06.24.zip "https://softwarecenter.qualcomm.com/api/download/software/qualcomm_neural_processing_sdk/v2.23.0.24.06.24.zip"
+  echo "Finishing downloading qnn sdk."
+  unzip -qo /tmp/v2.23.0.24.06.24.zip -d /tmp
+  echo "Finishing unzip qnn sdk."
+
+
+  # Print the content for manual verification
+  ls -lah "/tmp/qairt"
+  mv "/tmp/qairt"/* "${QNN_INSTALLATION_DIR}"
+  echo "Finishing installing qnn '${QNN_INSTALLATION_DIR}' ."
+
+  ls -lah "${QNN_INSTALLATION_DIR}"
+}
+
+install_qnn

.ci/scripts/test_llama.sh

@@ -72,6 +72,25 @@ fi
 
 echo "COREML option ${COREML}"
 
+if [[ "${MODE}" =~ .*qnn.* ]]; then
+  QNN=ON
+  export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
+  export QNN_SDK_ROOT=/tmp/qnn/2.23.0.240531
+  export LD_LIBRARY_PATH="${QNN_SDK_ROOT}/lib/x86_64-linux-clang"
+  export PYTHONPATH=".."
+  cp schema/program.fbs exir/_serialize/program.fbs
+  cp schema/scalar_type.fbs exir/_serialize/scalar_type.fbs
+  cp -f build-x86/backends/qualcomm/PyQnnManagerAdaptor.cpython-310-x86_64-linux-gnu.so backends/qualcomm/python
+  cp -f build-x86/backends/qualcomm/PyQnnWrapperAdaptor.cpython-310-x86_64-linux-gnu.so backends/qualcomm/python
+
+else
+  QNN=OFF
+  QNN_SDK_ROOT=""
+fi
+
+echo "QNN option ${QNN}"
+echo "QNN_SDK_ROOT: ${QNN_SDK_ROOT}"
+
 if [[ -z "${BUCK:-}" ]]; then
   BUCK=buck2
 fi
@@ -96,6 +115,8 @@ cmake_install_executorch_libraries() {
         -DEXECUTORCH_BUILD_XNNPACK="$XNNPACK" \
         -DEXECUTORCH_BUILD_MPS="$MPS" \
         -DEXECUTORCH_BUILD_COREML="$COREML" \
+        -DEXECUTORCH_BUILD_QNN="$QNN" \
+        -DQNN_SDK_ROOT="$QNN_SDK_ROOT" \
         -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
         -Bcmake-out .
     cmake --build cmake-out -j9 --target install --config Debug
@@ -176,6 +197,9 @@ fi
 if [[ "${COREML}" == "ON" ]]; then
   EXPORT_ARGS="${EXPORT_ARGS} -kv -v --coreml --disable_dynamic_shape"
 fi
+if [[ "${QNN}" == "ON" ]]; then
+  EXPORT_ARGS="${EXPORT_ARGS} -kv -v --qnn --disable_dynamic_shape"
+fi
 # Add dynamically linked library location
 $PYTHON_EXECUTABLE -m examples.models.llama2.export_llama ${EXPORT_ARGS}
 

.github/workflows/pull.yml

@@ -205,27 +205,13 @@ jobs:
 
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "cmake"
 
+        # install pybind
+        bash install_requirements.sh --pybind xnnpack
+
         # install Llava requirements
         bash examples/models/llama2/install_requirements.sh
         bash examples/models/llava/install_requirements.sh
 
-        # run export_llava.sh
-        python examples/models/llava/export_llava.py --use-sdpa-with-kv-cache --pte-name llava_custom_sdpa.pte
-
-        # verify file exists
-        if [ ! -f "llava_custom_sdpa.pte" ]; then
-            echo "llava_custom_sdpa.pte not found!"
-            exit 1
-        fi
-
-        python examples/models/llava/export_llava.py --no-use-sdpa-with-kv-cache --pte-name llava.pte
-
-        # verify file exists
-        if [ ! -f "llava.pte" ]; then
-            echo "llava.pte not found!"
-            exit 1
-        fi
-
         # run python unittest
         python -m unittest examples.models.llava.test.test_llava
 
@@ -383,3 +369,38 @@ jobs:
 
         # Run pytest with coverage
         pytest -c /dev/null -v -n auto --cov=./ --cov-report=xml backends/arm/test
+
+
+  test-llama-runner-qnn-linux:
+    name: test-llama-runner-qnn-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    strategy:
+      matrix:
+        dtype: [fp32]
+        build-tool: [cmake]
+        mode: [qnn]
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-clang12-android
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 900
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        DTYPE=${{ matrix.dtype }}
+        BUILD_TOOL=${{ matrix.build-tool }}
+        MODE=${{ matrix.mode }}
+
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
+        PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
+
+        # Setup executorch
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh buck2
+        # Install requirements for export_llama
+        PYTHON_EXECUTABLE=python bash examples/models/llama2/install_requirements.sh
+        # Test llama2
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh stories110M.pt "${BUILD_TOOL}" "${DTYPE}" "${MODE}"

.gitmodules

@@ -28,9 +28,6 @@
 [submodule "backends/xnnpack/third-party/pthreadpool"]
 	path = backends/xnnpack/third-party/pthreadpool
 	url = https://github.com/Maratyszcza/pthreadpool.git
-[submodule "examples/third-party/LLaVA"]
-	path = examples/third-party/LLaVA
-	url = https://github.com/haotian-liu/LLaVA.git
 [submodule "examples/third-party/fbjni"]
 	path = examples/third-party/fbjni
 	url = https://github.com/facebookincubator/fbjni.git

CMakeLists.txt

@@ -1,4 +1,5 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
+# Copyright 2024 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
@@ -179,6 +180,8 @@ option(EXECUTORCH_BUILD_GTESTS "Build googletest based test binaries" OFF)
 
 option(EXECUTORCH_BUILD_MPS "Build the MPS backend" OFF)
 
+option(EXECUTORCH_BUILD_NEURON "Build the backends/mediatek directory" OFF)
+
 option(EXECUTORCH_BUILD_PYBIND "Build the Python Bindings" OFF)
 
 option(EXECUTORCH_BUILD_QNN "Build the Qualcomm backend" OFF)
@@ -586,6 +589,10 @@ if(EXECUTORCH_BUILD_EXECUTOR_RUNNER)
     list(APPEND _executor_runner_libs portable_ops_lib)
   endif()
 
+  if(EXECUTORCH_ENABLE_EVENT_TRACER)
+    list(APPEND _executor_runner_libs etdump flatccrt)
+  endif()
+
   # Generate lib to register quantized ops
   if(EXECUTORCH_BUILD_KERNELS_QUANTIZED)
     list(APPEND _executor_runner_libs quantized_ops_lib)
@@ -624,6 +631,10 @@ if(EXECUTORCH_BUILD_EXTENSION_MODULE)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/module)
 endif()
 
+if(EXECUTORCH_BUILD_NEURON)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/mediatek)
+endif()
+
 if(EXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/runner_util)
 endif()

CONTRIBUTING.md

@@ -23,6 +23,8 @@ We actively welcome your pull requests (PRs).
    See the [testing section](#testing) for more information.
 1. If you've changed APIs or added a new tool or feature, [update the
    documentation](#updating-documentation).
+1. If you added an experimental API or deprecated an existing API, follow the
+   [API Life Cycle and Deprecation Policy](/docs/source/api-life-cycle.md).
 1. Make sure your code follows the [style guides](#coding-style) and passes the
    [lint checks](#lintrunner).
 1. If you haven't already, complete the [Contributor License Agreement ("CLA")](#contributor-license-agreement-cla).

LICENSE

@@ -6,6 +6,7 @@ Copyright (c) Meta Platforms, Inc. and affiliates.
 Copyright 2023 Arm Limited and/or its affiliates.
 Copyright (c) Qualcomm Innovation Center, Inc.
 Copyright (c) 2023 Apple Inc.
+Copyright (c) 2024 MediaTek Inc.
 
 Redistribution and use in source and binary forms, with or without modification,
 are permitted provided that the following conditions are met:

backends/apple/coreml/CMakeLists.txt

@@ -13,6 +13,11 @@ if(NOT EXECUTORCH_ROOT)
   set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
 endif()
 
+if(EXECUTORCH_BUILD_SDK)
+# protobuf requires frtti
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -frtti" )
+endif()
+  
 option(COREML_BUILD_EXECUTOR_RUNNER "Build CoreML executor runner." OFF)
 
 # inmemoryfs sources
@@ -59,6 +64,7 @@ set(SDK_SOURCES
     runtime/sdk/ETCoreMLModelAnalyzer.mm
     runtime/sdk/ETCoreMLModelStructurePath.mm
     runtime/sdk/ETCoreMLOperationProfilingInfo.mm
+    runtime/sdk/ETCoreMLModelDebugInfo.mm
     runtime/sdk/ETCoreMLModelDebugger.mm
     runtime/sdk/ETCoreMLModelProfiler.mm
     runtime/sdk/ETCoreMLPair.mm
@@ -141,6 +147,8 @@ if(EXECUTORCH_BUILD_SDK)
   add_subdirectory(
     ${CMAKE_CURRENT_SOURCE_DIR}/third-party/coremltools/deps/protobuf/cmake
   )
+
+  target_link_options_shared_lib(libprotobuf-lite)
   target_link_libraries(coremldelegate PRIVATE libprotobuf-lite)
 endif()
 

backends/apple/coreml/runtime/delegate/ETCoreMLModelManager.mm

@@ -5,36 +5,37 @@
 //
 // Please refer to the license found in the LICENSE file in the root directory of the source tree.
 
-#import <ETCoreMLAsset.h>
-#import <ETCoreMLAssetManager.h>
-#import <ETCoreMLDefaultModelExecutor.h>
-#import <ETCoreMLLogging.h>
-#import <ETCoreMLModel.h>
-#import <ETCoreMLModelCompiler.h>
-#import <ETCoreMLModelExecutor.h>
-#import <ETCoreMLModelLoader.h>
-#import <ETCoreMLModelManager.h>
-#import <ETCoreMLStrings.h>
-#import <MLModel_Prewarm.h>
-#import <MLMultiArray_Copy.h>
+#import "ETCoreMLAsset.h"
+#import "ETCoreMLAssetManager.h"
+#import "ETCoreMLDefaultModelExecutor.h"
+#import "ETCoreMLLogging.h"
+#import "ETCoreMLModel.h"
+#import "ETCoreMLModelCompiler.h"
+#import "ETCoreMLModelExecutor.h"
+#import "ETCoreMLModelLoader.h"
+#import "ETCoreMLModelManager.h"
+#import "ETCoreMLStrings.h"
+#import "MLModel_Prewarm.h"
+#import "MLMultiArray_Copy.h"
 #import <filesystem>
-#import <inmemory_filesystem_utils.hpp>
+#import "inmemory_filesystem_utils.hpp"
 #import <iostream>
 #import <memory>
-#import <model_metadata.h>
-#import <multiarray.h>
-#import <objc_array_util.h>
+#import "model_metadata.h"
+#import "multiarray.h"
+#import "objc_array_util.h"
 #import <optional>
 #import <os/lock.h>
-#import <serde_json.h>
+#import "serde_json.h"
 #import <string>
 #import <system_error>
 #import <vector>
 
 #if ET_EVENT_TRACER_ENABLED
-#import <ETCoreMLModelAnalyzer.h>
-#import <ETCoreMLModelStructurePath.h>
-#import <objc_safe_cast.h>
+#import "ETCoreMLModelAnalyzer.h"
+#import "ETCoreMLModelDebugInfo.h"
+#import "ETCoreMLModelStructurePath.h"
+#import "objc_safe_cast.h"
 #endif
 
 namespace {
@@ -317,31 +318,14 @@ void add_compute_unit(std::string& identifier, MLComputeUnits compute_units) {
     return [[ETCoreMLAsset alloc] initWithBackingAsset:std::move(backingAsset.value())];
 }
 
-NSDictionary<ETCoreMLModelStructurePath *, NSString *> * _Nullable get_operation_path_to_symbol_name_map(const inmemoryfs::InMemoryFileSystem *inMemoryFS,
-                                                                                                         NSError * __autoreleasing *error) {
+ETCoreMLModelDebugInfo * _Nullable get_model_debug_info(const inmemoryfs::InMemoryFileSystem *inMemoryFS,
+                                                        NSError * __autoreleasing *error) {
     NSData *file_data = get_file_data(inMemoryFS, ETCoreMLStrings.debugInfoFileRelativePath);
     if (!file_data) {
         return nil;
     }
-    
-    id object = [NSJSONSerialization JSONObjectWithData:file_data options:(NSJSONReadingOptions)0 error:error];
-    if (!object) {
-        return nil;
-    }
-    
-    NSDictionary<NSString *, id> *json_dict = SAFE_CAST(object, NSDictionary);
-    NSMutableDictionary<ETCoreMLModelStructurePath *, NSString *> *result = [NSMutableDictionary dictionaryWithCapacity:json_dict.count];
-    NSDictionary<NSString *, NSArray<id> *> *debug_symbol_to_operation_path_map = SAFE_CAST(json_dict[ETCoreMLStrings.debugSymbolToOperationPathKeyName], NSDictionary);
-    for (NSString *symbol_name in debug_symbol_to_operation_path_map) {
-        NSArray<NSDictionary<NSString *, id> *> *components = SAFE_CAST(debug_symbol_to_operation_path_map[symbol_name], NSArray);
-        if (components.count == 0) {
-            continue;
-        }
-        ETCoreMLModelStructurePath *path = [[ETCoreMLModelStructurePath alloc] initWithComponents:components];
-        result[path] = symbol_name;
-    }
-    
-    return result;
+
+    return [ETCoreMLModelDebugInfo modelDebugInfoFromData:file_data error:error];
 }
 
 #endif
@@ -490,16 +474,16 @@ - (nullable NSURL *)compiledModelURLWithIdentifier:(NSString *)identifier
     }
 
     NSError *localError = nil;
-    NSDictionary<ETCoreMLModelStructurePath *, NSString *> *operation_path_to_symbol_name_map = get_operation_path_to_symbol_name_map(inMemoryFS,
-                                                                                                                                      &localError);
+    ETCoreMLModelDebugInfo *debug_info = get_model_debug_info(inMemoryFS, &localError);
     if (localError) {
         ETCoreMLLogError(localError, "Failed to parse debug info file");
     }
 
+
     return [[ETCoreMLModelAnalyzer alloc] initWithCompiledModelAsset:compiledModelAsset
                                                           modelAsset:modelAsset
+                                                      modelDebugInfo:debug_info
                                                             metadata:metadata
-                                       operationPathToDebugSymbolMap:operation_path_to_symbol_name_map
                                                        configuration:configuration
                                                         assetManager:self.assetManager
                                                                error:error];

backends/apple/coreml/runtime/delegate/ETCoreMLStrings.h

@@ -66,6 +66,8 @@ NS_ASSUME_NONNULL_BEGIN
 @property (class, copy, readonly, nonatomic, nullable) NSString* debugInfoFileRelativePath;
 /// The debug symbol to operation path key name.
 @property (class, copy, readonly, nonatomic, nullable) NSString* debugSymbolToOperationPathKeyName;
+/// The debug symbol to handles key name.
+@property (class, copy, readonly, nonatomic, nullable) NSString* debugSymbolToHandlesKeyName;
 
 @end
 

backends/apple/coreml/runtime/delegate/ETCoreMLStrings.mm

@@ -95,6 +95,11 @@ + (NSString *)debugSymbolToOperationPathKeyName {
     return ETCoreMLDebugSymbolToOperationPathKeyName;
 }
 
++ (NSString *)debugSymbolToHandlesKeyName {
+    static NSString * const ETCoreMLDebugSymbolToHandlesKeyName = @"debugSymbolToHandles";
+    return ETCoreMLDebugSymbolToHandlesKeyName;
+}
+
 + (nullable NSString *)assetsDirectoryPath {
     static dispatch_once_t onceToken;
     static NSString *result = nil;