Merge branch 'main' into wasm-bindings-js

Author: Conarnar

.ci/scripts/test_model.sh

@@ -232,21 +232,24 @@ test_model_with_qnn() {
 # @param should_test If true, build and test the model using the coreml_executor_runner.
 test_model_with_coreml() {
   local should_test="$1"
+  local test_with_pybindings="$2"
+  local dtype="$3"
 
   if [[ "${BUILD_TOOL}" != "cmake" ]]; then
     echo "coreml only supports cmake."
     exit 1
   fi
 
-  DTYPE=float16
+  RUN_WITH_PYBINDINGS=""
+  if [[ "${test_with_pybindings}" == true ]]; then
+    echo \"Running with pybindings\"
+    export RUN_WITH_PYBINDINGS="--run_with_pybindings"
+  fi
 
-  "${PYTHON_EXECUTABLE}" -m examples.apple.coreml.scripts.export --model_name="${MODEL_NAME}" --compute_precision "${DTYPE}" --use_partitioner
+  "${PYTHON_EXECUTABLE}" -m examples.apple.coreml.scripts.export --model_name="${MODEL_NAME}" --compute_precision ${dtype} --use_partitioner ${RUN_WITH_PYBINDINGS}
   EXPORTED_MODEL=$(find "." -type f -name "${MODEL_NAME}*.pte" -print -quit)
 
   if [ -n "$EXPORTED_MODEL" ]; then
-    EXPORTED_MODEL_WITH_DTYPE="${EXPORTED_MODEL%.pte}_${DTYPE}.pte"
-    mv "$EXPORTED_MODEL" "$EXPORTED_MODEL_WITH_DTYPE"
-    EXPORTED_MODEL="$EXPORTED_MODEL_WITH_DTYPE"
     echo "OK exported model: $EXPORTED_MODEL"
   else
     echo "[error] failed to export model: no .pte file found"
@@ -303,7 +306,15 @@ elif [[ "${BACKEND}" == *"coreml"* ]]; then
   if [[ "${BACKEND}" == *"test"* ]]; then
     should_test_coreml=true
   fi
-  test_model_with_coreml "${should_test_coreml}"
+  test_with_pybindings=false
+  if [[ "${BACKEND}" == *"pybind"* ]]; then
+    test_with_pybindings=true
+  fi
+  dtype=float16
+  if [[ "${BACKEND}" == *"float32"* ]]; then
+    dtype=float32
+  fi
+  test_model_with_coreml "${should_test_coreml}" "${test_with_pybindings}" "${dtype}"
   if [[ $? -eq 0 ]]; then
     prepare_artifacts_upload
   fi

.ci/scripts/test_qnn_static_llama.sh

@@ -41,6 +41,10 @@ exit_code1=$?
 $PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_llama_stories_110m --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir . --llama_artifacts . --enable_x86_64
 exit_code2=$?
 
+# Check BC
+bash backends/qualcomm/bc/test_qnn_static_llama_bc.sh
+exit_code3=$?
+
 # Check the exit codes and print messages
 if [ $exit_code1 -ne 0 ]; then
     echo "Static Llama compile only with weight sharing test failed. $exit_code1."
@@ -50,8 +54,12 @@ if [ $exit_code2 -ne 0 ]; then
     echo "Static Llama accuracy test failed. $exit_code2."
 fi
 
+if [ $exit_code3 -ne 0 ]; then
+    echo "Static Llama BACKWARD COMPATIBILITY test failed. $exit_code3."
+fi
+
 # Return failure if either program failed
-if [ $exit_code1 -ne 0 ] || [ $exit_code2 -ne 0 ]; then
+if [ $exit_code1 -ne 0 ] || [ $exit_code2 -ne 0 ] || [ $exit_code3 -ne 0 ]; then
     exit 1
 else
     exit 0

.github/workflows/trunk.yml

@@ -18,8 +18,8 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  test-models-macos:
-    name: test-models-macos
+  test-models-macos-cpu:
+    name: test-models-macos-cpu
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     strategy:
       matrix:
@@ -568,10 +568,12 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_model.sh ${{ matrix.model }} "cmake" "qnn"
 
-  test-apple-model:
-    name: test-apple-model
+  test-models-macos-coreml:
+    name: test-models-macos-coreml
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     strategy:
+      matrix:
+        model: [dl3, edsr, efficient_sam, emformer_join, emformer_transcribe, ic3, ic4, mobilebert, mv2, mv3, resnet50, vit, w2l]
       fail-fast: false
     with:
       runner: macos-m1-stable
@@ -580,7 +582,23 @@ jobs:
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 90
       script: |
+        MODEL_NAME=${{ matrix.model }}
         BUILD_TOOL=cmake
+        BACKEND="coreml-pybind"
+
+
+        # Set model specific overrides
+        if [[ "${MODEL_NAME}" == "mobilebert" ]]; then
+          # See https://github.com/pytorch/executorch/issues/12907
+          # mobilebert has nan output on FP16, and high MSE on fp32, so we disable runtime test now
+          BACKEND="coreml"
+        fi
+
+        if [[ "${MODEL_NAME}" == "efficient_sam" ]]; then
+          # See https://github.com/pytorch/executorch/issues/12906
+          # efficient_sam fails to run on CoreML
+          BACKEND="coreml"
+        fi
 
         bash .ci/scripts/setup-conda.sh
 
@@ -589,13 +607,28 @@ jobs:
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash backends/apple/coreml/scripts/install_requirements.sh
         echo "Finishing installing coreml."
 
-        # Build and test coreml model
-        MODELS=(mv3 ic4 resnet50 edsr mobilebert w2l)
-        for MODEL_NAME in "${MODELS[@]}"; do
-          echo "::group::Exporting coreml model: $MODEL_NAME"
-          PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_model.sh "${MODEL_NAME}" "${BUILD_TOOL}" "coreml"
-          echo "::endgroup::"
+        PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_model.sh "${MODEL_NAME}" "${BUILD_TOOL}" "${BACKEND}"
+
+  test-models-macos-mps:
+    name: test-models-macos-mps
+    uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
+    strategy:
+      fail-fast: false
+    with:
+      runner: macos-m1-stable
+      python-version: '3.11'
+      submodules: 'recursive'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        BUILD_TOOL=cmake
+        bash .ci/scripts/setup-conda.sh
+
+        # Setup MacOS dependencies as there is no Docker support on MacOS atm
+        PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh --build-tool "${BUILD_TOOL}"
 
+        # Build and test mps model
+        for MODEL_NAME in mv3 ic4 resnet50 edsr mobilebert w2l; do
           echo "::group::Exporting mps model: $MODEL_NAME"
           PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_model.sh "${MODEL_NAME}" "${BUILD_TOOL}" "mps"
           echo "::endgroup::"

CMakeLists.txt

@@ -821,6 +821,10 @@ endif()
 if(EXECUTORCH_BUILD_VULKAN)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/vulkan)
 endif()
+if(EXECUTORCH_BUILD_VGF)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/arm)
+endif()
+
 
 if(EXECUTORCH_BUILD_ANDROID_JNI)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/android)

backends/apple/coreml/scripts/generate_test_models.sh

@@ -22,7 +22,7 @@ cd "$EXECUTORCH_ROOT_PATH"
 MODELS=("add" "add_mul" "mul" "mv3")
 for MODEL in "${MODELS[@]}"
 do
-  echo "Executorch: Generating $MODEL model" 
+  echo "Executorch: Generating $MODEL model"
   # TODO: Don't use the script in examples directory.
   python3 -m examples.apple.coreml.scripts.export --model_name "$MODEL" --save_processed_bytes
   mv -f "$MODEL""_coreml_all.pte" "$COREML_DIR_PATH/runtime/test/models"
@@ -36,7 +36,7 @@ COMPILE_MODELS=("add_mul")
 echo "Executorch: Generating compiled model"
 for MODEL in "${COMPILE_MODELS[@]}"
 do
-  echo "Executorch: Generating compiled $MODEL model" 
+  echo "Executorch: Generating compiled $MODEL model"
   python3 -m examples.apple.coreml.scripts.export --model_name "$MODEL" --compile
   mv -f "$MODEL""_compiled_coreml_all.pte" "$COREML_DIR_PATH/runtime/test/models"
 done

backends/arm/CMakeLists.txt

@@ -12,13 +12,17 @@ if(NOT EXECUTORCH_ROOT)
   set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../..)
 endif()
 
-add_compile_options("-Wall" "-Werror")
-
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 
 set(_common_include_directories ${EXECUTORCH_ROOT}/.. ${EXECUTORCH_ROOT}/runtime/core/portable_type/c10)
 add_compile_definitions(C10_USING_CUSTOM_GENERATED_MACROS)
 
+
+# bare metal backend builds
+if(EXECUTORCH_BUILD_ARM_BAREMETAL)
+
+add_compile_options("-Wall" "-Werror")
+
 # Third-party folder and Ethos-U driver inclued
 set(THIRD_PARTY_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/third-party")
 set(DRIVER_ETHOSU_INCLUDE_DIR "${THIRD_PARTY_ROOT}/ethos-u-core-driver/include")
@@ -36,3 +40,47 @@ target_include_directories(
 target_include_directories(
   executorch_delegate_ethos_u PUBLIC ${DRIVER_ETHOSU_INCLUDE_DIR}
 )
+
+# end config for bare metal builds
+endif()
+
+
+# VGF backend builds 
+if(EXECUTORCH_BUILD_VGF)
+
+# include libvgf
+set(LIBVGF_PATH "${EXECUTORCH_ROOT}/examples/arm/ethos-u-scratch/ml-sdk-for-vulkan-manifest/sw/vgf-lib/")
+
+set(VULKAN_THIRD_PARTY_PATH ${EXECUTORCH_ROOT}/backends/vulkan/third-party)
+set(VULKAN_HEADERS_PATH ${VULKAN_THIRD_PARTY_PATH}/Vulkan-Headers/include)
+set(VOLK_HEADERS_PATH ${VULKAN_THIRD_PARTY_PATH}/volk)
+
+set(LIBVGF_STATIC "${LIBVGF_PATH}/build/src/libvgf.a")
+set(LIBVGF_INCLUDE "${LIBVGF_PATH}/include/")
+
+add_library(vgf STATIC IMPORTED)
+set_property( TARGET vgf PROPERTY IMPORTED_LOCATION "${LIBVGF_STATIC}" )
+target_include_directories(vgf INTERFACE "${LIBVGF_INCLUDE}")
+
+# Add backend delegate for VGF
+set(_vgf_backend_sources backends/arm/runtime/VGFBackend.cpp
+			 backends/arm/runtime/VGFSetup.cpp )
+
+# vgf backend
+list(TRANSFORM  _vgf_backend_sources PREPEND "${EXECUTORCH_ROOT}/")
+add_library(vgf_backend ${_vgf_backend_sources})
+target_include_directories(
+  vgf_backend PUBLIC
+  ${_common_include_directories}
+  ${VULKAN_HEADERS_PATH}
+  ${VOLK_HEADERS_PATH}
+)
+target_compile_options(vgf_backend PRIVATE -DUSE_VULKAN_WRAPPER -DUSE_VULKAN_VOLK)
+
+
+target_link_libraries(vgf_backend PRIVATE executorch_core)
+target_link_libraries(vgf_backend PRIVATE vgf)
+executorch_target_link_options_shared_lib(vgf_backend)
+
+# end config for VGF builds
+endif()

backends/arm/_passes/match_arg_ranks_pass.py

@@ -54,6 +54,9 @@ def __init__(self, exported_program):
         exir_ops.edge.aten.le.Tensor,
         exir_ops.edge.aten.pow.Tensor_Tensor,
         exir_ops.edge.aten.where.self,
+        exir_ops.edge.aten.bitwise_and.Tensor,
+        exir_ops.edge.aten.bitwise_xor.Tensor,
+        exir_ops.edge.aten.bitwise_or.Tensor,
     ]
 
     def _match_op_rank(self, graph_module, node, arg, max_rank):

backends/arm/_passes/replace_scalar_with_tensor_pass.py

@@ -34,6 +34,9 @@
     exir_ops.edge.aten.lt.Scalar: exir_ops.edge.aten.lt.Tensor,
     exir_ops.edge.aten.le.Scalar: exir_ops.edge.aten.le.Tensor,
     exir_ops.edge.aten.ne.Scalar: exir_ops.edge.aten.ne.Tensor,
+    exir_ops.edge.aten.bitwise_and.Scalar: exir_ops.edge.aten.bitwise_and.Tensor,
+    exir_ops.edge.aten.bitwise_or.Scalar: exir_ops.edge.aten.bitwise_or.Tensor,
+    exir_ops.edge.aten.bitwise_xor.Scalar: exir_ops.edge.aten.bitwise_xor.Tensor,
     torch.ops.aten.add.Scalar: torch.ops.aten.add.Tensor,
     torch.ops.aten.sub.Scalar: torch.ops.aten.sub.Tensor,
     torch.ops.aten.mul.Scalar: torch.ops.aten.mul.Tensor,
@@ -46,6 +49,9 @@
     torch.ops.aten.lt.Scalar: torch.ops.aten.lt.Tensor,
     torch.ops.aten.le.Scalar: torch.ops.aten.le.Tensor,
     torch.ops.aten.ne.Scalar: torch.ops.aten.ne.Tensor,
+    torch.ops.aten.bitwise_and.Scalar: torch.ops.aten.bitwise_and.Tensor,
+    torch.ops.aten.bitwise_or.Scalar: torch.ops.aten.bitwise_or.Tensor,
+    torch.ops.aten.bitwise_xor.Scalar: torch.ops.aten.bitwise_xor.Tensor,
 }
 
 

backends/arm/operator_support/ethos_u55_support.py

@@ -125,6 +125,9 @@ class EthosU55NotSupported(OperatorSupportBase):
         exir_ops.edge.aten.bitwise_and.Tensor,
         exir_ops.edge.aten.bitwise_or.Tensor,
         exir_ops.edge.aten.bitwise_xor.Tensor,
+        exir_ops.edge.aten.bitwise_and.Scalar,
+        exir_ops.edge.aten.bitwise_or.Scalar,
+        exir_ops.edge.aten.bitwise_xor.Scalar,
         exir_ops.edge.aten.bitwise_not,
         exir_ops.edge.aten.logical_and.default,
         exir_ops.edge.aten.logical_or.default,

backends/arm/operator_support/tosa_supported_operators.py

@@ -164,6 +164,9 @@ def is_node_supported(
             exir_ops.edge.aten.bitwise_and.Tensor,
             exir_ops.edge.aten.bitwise_or.Tensor,
             exir_ops.edge.aten.bitwise_xor.Tensor,
+            exir_ops.edge.aten.bitwise_and.Scalar,
+            exir_ops.edge.aten.bitwise_or.Scalar,
+            exir_ops.edge.aten.bitwise_xor.Scalar,
             exir_ops.edge.aten.expand_copy.default,
             exir_ops.edge.aten.cat.default,
             exir_ops.edge.aten.ceil.default,